STATA

STRUCTURAL EQUATION MODELING

REFERENCE MANUAL
RELEASE 13

A Stata Press Publication

StataCorp LP
College Station, Texas

c 19852013 StataCorp LP
Copyright
All rights reserved
Version 13

Published by Stata Press, 4905 Lakeway Drive, College Station, Texas 77845
Typeset in TEX
ISBN-10: 1-59718-124-2
ISBN-13: 978-1-59718-124-2
This manual is protected by copyright. All rights are reserved. No part of this manual may be reproduced, stored
in a retrieval system, or transcribed, in any form or by any meanselectronic, mechanical, photocopy, recording, or
otherwisewithout the prior written permission of StataCorp LP unless permitted subject to the terms and conditions
of a license granted to you by StataCorp LP to use the software and documentation. No license, express or implied,
by estoppel or otherwise, to any intellectual property rights is granted by this document.
StataCorp provides this manual as is without warranty of any kind, either expressed or implied, including, but
not limited to, the implied warranties of merchantability and fitness for a particular purpose. StataCorp may make
improvements and/or changes in the product(s) and the program(s) described in this manual at any time and without
notice.
The software described in this manual is furnished under a license agreement or nondisclosure agreement. The software
may be copied only in accordance with the terms of the agreement. It is against the law to copy the software onto
DVD, CD, disk, diskette, tape, or any other medium for any purpose other than backup or archival purposes.
c 1979 by Consumers Union of U.S.,
The automobile dataset appearing on the accompanying media is Copyright
Inc., Yonkers, NY 10703-1057 and is reproduced by permission from CONSUMER REPORTS, April 1979.
Stata,

, Stata Press, Mata,

, and NetCourse are registered trademarks of StataCorp LP.

Stata and Stata Press are registered trademarks with the World Intellectual Property Organization of the United Nations.
NetCourseNow is a trademark of StataCorp LP.
Other brand and product names are registered trademarks or trademarks of their respective companies.
For copyright information about the software, type help copyright within Stata.

The suggested citation for this software is

StataCorp. 2013. Stata: Release 13 . Statistical Software. College Station, TX: StataCorp LP.

Contents
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
intro
intro
intro
intro
intro
intro
intro
intro
intro
intro
intro
intro

1
2
3
4
5
6
7
8
9
10
11
12

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Introduction
2
. . . . . . . . . . . . . . . . . . Learning the language: Path diagrams and command language
7
. . . . . . . . . . . . . . . . . . . Learning the language: Factor-variable notation (gsem only) 35
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Substantive concepts 42
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tour of models 61
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Comparing groups (sem only) 82
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Postestimation tests and predictions 89
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Robust and clustered standard errors 96
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Standard errors, the full story 98
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Fitting models with survey data (sem only) 102
. . . . . . . . . . . . . . . . . . . . . . . Fitting models with summary statistics data (sem only) 104
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Convergence problems and how to solve them 112

Builder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . SEM Builder 122

Builder, generalized . . . . . . . . . . . . . . . . . . . . . . . . . . . . . SEM Builder for generalized models 125
estat
estat
estat
estat
estat
estat
estat
estat
estat
estat
estat
estat
estat
estat

eform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Display exponentiated coefficients

eqgof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Equation-level goodness-of-fit statistics
eqtest . . . . . . . . . . . . . . . . . . . . . . . . . . . Equation-level test that all coefficients are zero
framework . . . . . . . . . . . . . . . . . . . . . Display estimation results in modeling framework
ggof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Group-level goodness-of-fit statistics
ginvariant . . . . . . . . . . . . . . . . . . . . . . Tests for invariance of parameters across groups
gof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Goodness-of-fit statistics
mindices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Modification indices
residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Display mean and covariance residuals
scoretests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Score tests
stable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Check stability of nonrecursive system
stdize . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Test standardized parameters
summarize . . . . . . . . . . . . . . . . . . . . . . Report summary statistics for estimation sample
teffects . . . . . . . . . . . . . . . . . . . . Decomposition of effects into total, direct, and indirect

example
example
example
example
example
example
example
example
example
example
example
example
example
example
example

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Single-factor measurement model

. . . . . . . . . . . . . . . . . . . . . . . . . . . . Creating a dataset from published covariances
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Two-factor measurement model
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Goodness-of-fit statistics
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Modification indices
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Linear regression
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nonrecursive structural model
. . . . . . . . . . . . . . . . . . . Testing that coefficients are equal, and constraining them
. . . . . . . . . . . . . . . . . . . . . . . . . . . Structural model with measurement component
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . MIMIC model
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . estat framework
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Seemingly unrelated regression
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Equation-level Wald test
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Predicted values
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Higher-order CFA
i

128
130
132
134
136
138
140
143
145
148
150
152
154
155
158
164
169
177
180
183
187
195
199
208
215
218
222
223
225

ii

Contents

example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
example
gsem
gsem
gsem
gsem
gsem
gsem
gsem

16
17
18
19
20
21
22
23
24
25
26
27g
28g
29g
30g
31g
32g
33g
34g
35g
36g
37g
38g
39g
40g
41g
42g
43g
44g
45g
46g

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Correlation
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Correlated uniqueness model
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Latent growth model
. . . . . . . . . . . . . . . . . . . . . . . . . . Creating multiple-group summary statistics data
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Two-factor measurement model by group
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Group-level goodness of fit
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Testing parameter equality across groups
. . . . . . . . . . . . . . . . . . . . . . . . . . Specifying parameter constraints across groups
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Reliability
. . . . . . . . . . . . . . . . . . . . . . . . . . Creating summary statistics data from raw data
. . . . . . . . . . . . . . . . . . . . . . . . . . . . Fitting a model with data missing at random
. . . . . . . . . . . . . . . . . . Single-factor measurement model (generalized response)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . One-parameter logistic IRT (Rasch) model
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Two-parameter logistic IRT model
. . . . . . . . . . . Two-level measurement model (multilevel, generalized response)
. . . . . . . . . . . . . . . . . . . Two-factor measurement model (generalized response)
. . . . . . . . . . . . . . . . . . . . Full structural equation model (generalized response)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Logistic regression
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . Combined models (generalized responses)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ordered probit and ordered logit
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . MIMIC model (generalized response)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Multinomial logistic regression
. . . . . . . . . . . . . . . . . . Random-intercept and random-slope models (multilevel)
. . . . . . . . . . . . . . . . . . . . . Three-level model (multilevel, generalized response)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Crossed models (multilevel)
. . . . . . . . . . . . . . . . . . . . Two-level multinomial logistic regression (multilevel)
. . . . . . . . . . . . . . . . . . . . . . . One- and two-level mediation models (multilevel)
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tobit regression
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Interval regression
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Heckman selection model
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Endogenous treatment-effects model

232
237
244
251
256
265
266
269
275
279
287
291
297
306
314
323
330
336
341
347
354
359
368
384
392
397
407
416
419
423
432

. . . . . . . . . . . . . . . . . . . . . . Generalized structural equation model estimation command

estimation options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Options affecting estimation
family-and-link options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Family-and-link options
model description options . . . . . . . . . . . . . . . . . . . . . . . . . . . . Model description options
path notation extensions . . . . . . . . . . . . . . . . . . . . Command syntax for path diagrams
postestimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Postestimation tools for gsem
reporting options . . . . . . . . . . . . . . . . . . . . . . . . . . Options affecting reporting of results

439
443
447
452
455
459
460

lincom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Linear combinations of parameters 462

lrtest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Likelihood-ratio test of linear hypothesis 463
methods and formulas for gsem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Methods and formulas 465
methods and formulas for sem . . . . . . . . . . . . . . . . . . . . . . . . . Methods and formulas for sem 478
nlcom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nonlinear combinations of parameters 490
predict after gsem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Generalized linear predictions, etc. 492
predict after sem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Factor scores, linear predictions, etc. 496
sem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Structural equation model estimation command 498
sem and gsem option constraints( ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . Specifying constraints 503
sem and gsem option covstructure( ) . . . . . . . . . . . . . . . . . Specifying covariance restrictions 505

Contents

sem
sem
sem
sem
sem
sem
sem
sem
sem
sem
sem
sem
sem
sem
ssd

and gsem option from( ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Specifying starting values

and gsem option reliability( ) . . . . . Fraction of variance not due to measurement error
and gsem path notation . . . . . . . . . . . . . . . . . . . . . . Command syntax for path diagrams
and gsem syntax options . . . . . . . . . . . . . . . . . Options affecting interpretation of syntax
estimation options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Options affecting estimation
group options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Fitting models on different groups
model description options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Model description options
option method( ) . . . . . . . . . . . . . . . . . . . . . Specifying method and calculation of VCE
option noxconditional . . . . . . Computing means, etc., of observed exogenous variables
option select( ) . . . . . . . . . . . . . . . . . . . . . . . . . . Using sem with summary statistics data
path notation extensions . . . . . . . . . . . . . . . . . . . . . Command syntax for path diagrams
postestimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Postestimation tools for sem
reporting options . . . . . . . . . . . . . . . . . . . . . . . . . . . Options affecting reporting of results
ssd options . . . . . . . . . . . . . . . . . . . . . . . . . Options for use with summary statistics data
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Making summary statistics data (sem only)

iii

508
511
514
520
521
523
525
527
529
532
534
538
540
542
544

test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wald test of linear hypotheses 548

testnl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Wald test of nonlinear hypotheses 550
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

552

Subject and author index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

565

Cross-referencing the documentation

When reading this manual, you will find references to other Stata manuals. For example,
[U] 26 Overview of Stata estimation commands
[XT] xtabond
[D] reshape

The first example is a reference to chapter 26, Overview of Stata estimation commands, in the Users
Guide; the second is a reference to the xtabond entry in the Longitudinal-Data/Panel-Data Reference
Manual; and the third is a reference to the reshape entry in the Data Management Reference Manual.
All the manuals in the Stata Documentation have a shorthand notation:
[GSM]
[GSU]
[GSW]
[U ]
[R]
[D ]
[G ]
[XT]
[ME]
[MI]
[MV]
[PSS]
[P ]
[SEM]
[SVY]
[ST]
[TS]
[TE]
[I]

Getting Started with Stata for Mac

Getting Started with Stata for Unix
Getting Started with Stata for Windows
Stata Users Guide
Stata Base Reference Manual
Stata Data Management Reference Manual
Stata Graphics Reference Manual
Stata Longitudinal-Data/Panel-Data Reference Manual
Stata Multilevel Mixed-Effects Reference Manual
Stata Multiple-Imputation Reference Manual
Stata Multivariate Statistics Reference Manual
Stata Power and Sample-Size Reference Manual
Stata Programming Reference Manual
Stata Structural Equation Modeling Reference Manual
Stata Survey Data Reference Manual
Stata Survival Analysis and Epidemiological Tables Reference Manual
Stata Time-Series Reference Manual
Stata Treatment-Effects Reference Manual:
Potential Outcomes/Counterfactual Outcomes
Stata Glossary and Index

[M ]

Mata Reference Manual

Acknowledgments
sem and gsem were developed by StataCorp.
Neither command would exist without the help of two people outside of StataCorp. We must thank
these two people profusely. They are
Jeroen Weesie, Department of Sociology at Utrecht University, The Netherlands
Sophia Rabe-Hesketh, University of California, Berkeley
Jeroen Weesie is responsible for the existence of the SEM project at StataCorp. While spending
his sabbatical with us, Jeroen expressedrepeatedlythe importance of SEM, and that enthusiasm
for SEM was disregardedrepeatedly. Not until after his sabbatical did StataCorp see the light. At
that point, we had him back, and back, and back, so that he could inspire us, guide us, tell us what
we had right, and, often, tell us what we had wrong.
Jeroen helped us with the math, the syntax, and system design, and, when we were too thick-headed,
he even wrote code. By the date of first shipment, all code had been rewritten by us, but design and
syntax for SEM still now and forever will show Jeroens influence.
Thank you, Jeroen Weesie, for teaching us SEM.
Sophia Rabe-Hesketh contributed a bit later, after the second project, GSEM, was well underway.
GSEM stands for generalized SEM. Sophia is the coauthor of gllamm and knows as much about
multilevel and structural equation modeling as anybody, and probably more. She helped us a lot
through her prolific published works; we did have her visit a few times, though, mainly because we
knew that features in GSEM would overlap with features in GLLAMM, and we wanted to straighten
out any difficulties that competing features might cause.
About the competing features, Sophia cared nothing. About the GSEM project, she was excited.
About syntax and computational methodswell, she straightened us out the first day, even on things
we thought we had settled. Today, enough of the underlying workings of GSEM are based on Sophias
and her coauthors publications that anyone who uses gsem should cite Rabe-Hesketh, Skrondal, and
Pickles (2004).
We are indebted to the works of Sophia Rabe-Hesketh, Anders Skrondal of the University of Oslo
and the Norwegian Institute of Public Health, and Andrew Pickles of the University of Manchester.

Reference
Rabe-Hesketh, S., A. Skrondal, and A. Pickles. 2004. Generalized multilevel structural equation modeling. Psychometrika
69: 167190.

Also see
[R] gllamm Generalized linear and latent mixed models

Title
intro 1 Introduction

Description

Remarks and examples

Also see

Description
SEM stands for structural equation model. Structural equation modeling is

1. A notation for specifying SEMs.

2. A way of thinking about SEMs.
3. Methods for estimating the parameters of SEMs.
Statas sem and gsem commands fit these models: sem fits standard linear SEMs, and gsem fits
generalized SEMs.
In sem, responses are continuous and models are linear regression.
In gsem, responses are continuous or binary, ordinal, count, or multinomial. Models are linear
regression, gamma regression, logit, probit, ordinal logit, ordinal probit, Poisson, negative binomial,
multinomial logit, and more.
sem fits models to single-level data.
gsem fits models to single-level or multilevel data. Latent variables can be included at any level.
gsem can fit models with mixed effects, including random effects such as unobserved effects within
patient, nested effects such as unobserved effects within patient within doctor, and crossed effects
such as unobserved effects within occupation and country.
Meanwhile, sem provides features not provided by gsem: standard errors adjusted for survey
sampling strategies and weights; easy testing for whether groups such as males and females differ;
estimation using observations with missing values under the assumption of joint normality; goodnessof-fit statistics, modification indices, tests of indirect effects, and more; and models fit using summarystatistic data.
There is obviously overlap between the capabilities of sem and gsem. In such cases, results will
be nearly equal. Results should be exactly equal because both commands are producing estimates of
the same mathematical model, but sem and gsem use different numerical machinery. sems machinery
requires less calculation and fewer approximations and so is faster and slightly more accurate.

Remarks and examples

Structural equation modeling encompasses a broad array of models from linear regression to
measurement models to simultaneous equations, including along the way confirmatory factor analysis
(CFA), correlated uniqueness models, latent growth models, multiple indicators and multiple causes
(MIMIC) models, and item-response theory (IRT) models.
Structural equation modeling is not just an estimation method for a particular model in the way
that Statas regress and probit commands are, or even in the way that stcox and mixed are.
Structural equation modeling is a way of thinking, a way of writing, and a way of estimating.
If you read the introductory manual pages in the front of this manual[SEM] intro 2, [SEM] intro 3,
and so onwe will do our best to familiarize you with SEM and our implementation of it.
2

intro 1 Introduction

Beginning with [SEM] intro 2, entitled Learning the language: Path diagrams and command
language, you will learn that
1. A particular SEM is usually described using a path diagram.
2. The sem and gsem commands allow you to use path diagrams to input models. In fact, the
sem and gsem share the same GUI, called the SEM Builder.
3. sem and gsem alternatively allow you to use a command language to input models. The
command language is similar to the path diagrams.
[SEM] intro 3, entitled Learning the language: Factor-variable notation (gsem only), amounts to a
continuation of [SEM] intro 2.
4. We teach you Statas factor-variable notation, a wonderfully convenient shorthand for including
categorical variables in models.
In [SEM] intro 4, entitled Substantive concepts, you will learn that
5. sem provides four different estimation methods; you need to specify the method appropriate
for the assumptions you are willing to make. For gsem, there are two estimation methods.
6. There are four types of variables in SEMs: A variable is observed or latent, and simultaneously
it is endogenous or exogenous. To this, sem and gsem add another type of variable, the error
variable. Error variables are latent exogenous variables with a fixed-unit path coefficient,
and they are associated with a single endogenous variable. Error variables are denoted with
an e. prefix, so if y1 is an endogenous variable, then e.y1 is the associated error variable.
7. It is easy to specify path constraints in SEMsyou just draw them, or omit drawing them,
on the diagram. It is similarly easy with the SEM Builder as well as with sems and gsems
command language.
8. Determining whether an SEM is identified can be difficult. We show you how to let the
software check for you.
9. Identification also includes normalization constraints. sem and gsem apply normalization
constraints automatically, but you can control that if you wish. Sometimes you might even
need to control it.
In [SEM] intro 5, entitled Tour of models,
10. We take you on a whirlwind tour of some of the models that sem and gsem can fit. This is
a fun and useful section because we give you an overview without getting lost in the details.
Then in [SEM] intro 6, entitled Comparing groups (sem only),
11. We show you a highlight of sem: its ability to take an SEM and data consisting of groups
sexes, age categories, and the likeand fit the model in an interacted way that makes it
easy for you to test whether and how the groups differ.
In [SEM] intro 7, entitled Postestimation tests and predictions,
12. We show you how to redisplay results (sem and gsem), how to obtain exponentiated
coefficients (gsem only), and how to obtain standardized results (sem only).
13. We show you how to obtain goodness-of-fit statistics (sem only).
14. We show you how to perform hypothesis tests, including tests for omitted paths, tests for
relaxing constraints, and tests for model simplification.
15. We show you how to display other results, statistics, and tests.

intro 1 Introduction

16. We show you how to obtain predictions of observed response variables and predictions of
latent variables. With gsem, you can obtain predicted means, probabilities, or counts that
take into account the predictions of the latent variables, or you can set the latent variables
to 0.
17. We show you how to access stored results.
In [SEM] intro 8, entitled Robust and clustered standard errors,
18. We mention that sem and gsem optionally provide robust standard errors and provide clustered
standard errors, which relaxes the assumption of independence of observations (or subjects)
to independence within clusters of observations (subjects).
In [SEM] intro 9, entitled Standard errors, the full story,
19. We provide lots of technical detail expanding on item 18.
In [SEM] intro 10, entitled Fitting models with survey data (sem only),
20. We explain how sem can be used with Statas svy: prefix to obtain results adjusted for
complex survey designs, including clustered sampling and stratification.
In [SEM] intro 11, entitled Fitting models with summary statistics data (sem only),
21. We show you how to use sem with summary statistics data such as the correlation or
covariance matrix rather than the raw data. Many sources, especially textbooks, publish data
in summary statistics form.
Finally, in [SEM] intro 12, entitled Convergence problems and how to solve them,
22. We regretfully inform you that some SEMs have difficulty converging. We figure 5% to 15%
of complicated models will cause difficulty. We show you what to do and it is not difficult.
In the meantime,
23. There are many examples that we have collected for you in [SEM] example 1, [SEM] example 2,
and so on. It is entertaining and informative simply to read the examples in order.
24. There is an alphabetical glossary in [SEM] Glossary, located at the end of the manual.
If you prefer, you can skip all this introductory material and go for the details. For the full
experience, go directly to [SEM] sem and [SEM] gsem. You will have no idea what we are talking
aboutwe promise.

intro 1 Introduction

The technical sections, in logical order, are

Estimation
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]
[SEM]

sem
gsem
sem and gsem path notation
sem path notation extensions
gsem path notation extensions
Builder
Builder, generalized
sem model description options
gsem model description options
sem group options
sem ssd options
sem estimation options
gsem estimation options
sem reporting options
gsem reporting options
sem and gsem syntax options
sem option noxconditional
sem option select( )
sem and gsem option covstructure( )
sem option method( )
sem and gsem option reliability( )
sem and gsem option from( )
sem and gsem option constraints( )
gsem family-and-link options
ssd (sem only)

Postestimation, summary of
[SEM] sem postestimation
[SEM] gsem postestimation
Reporting results
[R] estat
[SEM] estat eform (gsem only)
[SEM] estat teffects (sem only)
[SEM] estat residuals (sem only)
[SEM] estat framework (sem only)
Goodness-of-fit tests
[SEM] estat gof (sem only)
[SEM] estat eqgof (sem only)
[SEM] estat ggof (sem only)
[R] estat

intro 1 Introduction

Hypotheses tests
[SEM] estat mindices (sem only)
[SEM] estat eqtest (sem only)
[SEM] estat scoretests (sem only)
[SEM] estat ginvariant (sem only)
[SEM] estat stable (sem only)
[SEM] test
[SEM] lrtest
[SEM] testnl
[SEM] estat stdize (sem only)
Linear and nonlinear combinations of results
[SEM] lincom
[SEM] nlcom
Predicted values
[SEM] predict after sem
[SEM] predict after gsem
Methods and formulas
[SEM] methods and formulas for sem
[SEM] methods and formulas for gsem
Many of these sections are technical, but mostly in the computer sense of the word. We suggest that
when you read the technical sections, you skip to Remarks and examples. If you read the introductory
sections, you will already know how to use the commands, so there is little reason to confuse yourself
with syntax diagrams that are more precise than they are enlightening. However, the syntax diagrams
do serve as useful reminders.

Also see
[SEM] intro 2 Learning the language: Path diagrams and command language
[SEM] example 1 Single-factor measurement model

[SEM] Acknowledgments

Title
intro 2 Learning the language: Path diagrams and command language
Description

Remarks and examples

Reference

Also see

Description
Individual structural equation models are usually described using path diagrams. Path diagrams
are described here.
Path diagrams can be used in sems (gsems) GUI, known as the SEM Builder or simply the Builder,
as the input to describe the model to be fit. Path diagrams differ a little from author to author, and
sems and gsems path diagrams differ a little, too. For instance, we omit drawing the variances and
covariances between observed exogenous variables by default.
sem and gsem also provide a command-language interface. This interface is similar to path diagrams
but is typable.

Remarks and examples

Remarks are presented under the following headings:
Using path diagrams to specify standard linear SEMs
Specifying correlation
Using the command language to specify standard linear SEMs
Specifying generalized SEMs: Family and link
Specifying generalized SEMs: Family and link, multinomial logistic regression
Specifying generalized SEMs: Family and link, paths from response variables
Specifying generalized SEMs: Multilevel mixed effects (2 levels)
Specifying generalized SEMs: Multilevel mixed effects (3 levels)
Specifying generalized SEMs: Multilevel mixed effects (4+ levels)
Specifying generalized SEMs: Multilevel mixed effects with random intercepts
Specifying generalized SEMs: Multilevel mixed effects with random slopes

Using path diagrams to specify standard linear SEMs

In structural equation modeling, models are often illustrated in a path diagram such as this one:

x1

x2

x3

x4

e.x1

e.x2

e.x3
7

e.x4

intro 2 Learning the language: Path diagrams and command language

This diagram is composed of the following:

1. Boxes and circles with variable names written inside them.
a. Boxes contain variables that are observed in the data.
b. Circles contain variables that are unobserved, known as latent variables.
2. Arrows, called paths, that connect some of the boxes and circles.
a. When a path points from one variable to another, it means that the first variable
affects the second.
b. More precisely, if s d, it means to add k s to the linear equation for d. k is
called the path coefficient.
c. Sometimes small numbers are written along the arrow connecting two variables.
This means that k is constrained to be the value specified. (Some authors use the
term path coefficient to mean standardized path coefficient. We do not.)
d. When no number is written along the arrow, the corresponding coefficient is to be
estimated from the data. Sometimes symbols are written along the path arrow to
emphasize this and sometimes not.
e. The same path diagram used to describe the model can be used to display the
results of estimation. In that case, estimated coefficients appear along the paths.
3. There are other elements that may appear on the diagram to indicate variances and betweenvariable correlations. We will get to them later.
Thus the above figure corresponds to the equations

x1 = 1 + 1 X + e.x1
x2 = 2 + 2 X + e.x2
x3 = 3 + 3 X + e.x3
x4 = 4 + 4 X + e.x4
Theres a third way of writing this model, namely,
(x1<-X) (x2<-X) (x3<-X) (x4<-X)
This is the way we would write the model if we wanted to use sems or gsems command syntax
rather than drawing the model in the Builder. The full command we would type is
. sem (x1<-X) (x2<-X) (x3<-X) (x4<-X)
We will get to that later.
By the way, the above model is a linear single-level model. Linear single-level models can be fit
by sem or by gsem. sem is the preferred way to fit linear single-level models because it has added
features for these models that you might find useful later. Nonetheless, if you want to fit the model
with gsem, you would type
. gsem (x1<-X) (x2<-X) (x3<-X) (x4<-X)
Whether we use sem or gsem, we obtain the same results.

intro 2 Learning the language: Path diagrams and command language

However we write this model, what is it? It is a measurement model, a term loaded with meaning
for some researchers. X might be mathematical ability. x1 , x2 , x3 , and x4 might be scores from
tests designed to measure mathematical ability. x1 might be the score based on your answers to a
series of questions after reading this section.
The model we have just drawn, or written in mathematical notation, or written in Stata command
notation, can be interpreted in other ways, too. Look at this diagram:

y
1

x1

x2

x3

e.x1

e.x2

e.y

e.x3

Despite appearances, this diagram is identical to the previous diagram except that we have renamed
x4 to be y . The fact that we changed a name obviously does not matter substantively. The fact that
we have rearranged the boxes in the diagram is irrelevant, too; paths connect the same variables in
the same directions. The equations for the above diagrams are the same as the previous equations
with the substitution of y for x4 :

x1 = 1 + X1 + e.x1
x2 = 2 + X2 + e.x2
x3 = 3 + X3 + e.x3
y = 4 + X4 + e.y
The Stata command notation changes similarly:
(x1<-X) (x2<-X) (x3<-X) (y<-X)
Many people looking at the model written this way might decide that it is not a measurement
model but a measurement-error model. y depends on X but we do not observe X . We do observe
x1 , x2 , and x3 , each a measurement of X , but with error. Our interest is in knowing 4 , the effect
of true X on y .
A few others might disagree and instead see a model for interrater agreement. Obviously, we have
three or four raters who each make a judgment, and we want to know how well the judgment process
works and how well each of these raters performs.

10

intro 2 Learning the language: Path diagrams and command language

Specifying correlation
One of the key features of SEM is the ease with which you can allow correlation between latent
variables to adjust for the reality of the situation. In the measurement model in the previous section,
we drew the following path diagram:

x1

x2

x3

x4

e.x1

e.x2

e.x3

e.x4

which corresponded to the equations

x1 = 1 + X1 + e.x1
x2 = 2 + X2 + e.x2
x3 = 3 + X3 + e.x3
x4 = 4 + X4 + e.x4
(X, x1 , x2 , x3 , x4 , e.x1 , e.x2 , e.x3 , e.x4 ) i.i.d. with mean vector and covariance matrix
where i.i.d. means that observations are independent and identically distributed.
We must appreciate that and are estimated, just as are 1 , 1 , . . . , 4 , 4 . Some of the
elements of , however, are constrained to be 0; which elements are constrained is determined by
how we specify the model. In the above diagram, we drew the model in such a way that we assumed
that error variables were uncorrelated with each other. We could have drawn the diagram differently.

intro 2 Learning the language: Path diagrams and command language

11

If we wish to allow for a correlation between e.x2 and e.x3 , we add a curved path between the
variables:

x1

x2

x3

x4

e.x1

e.x2

e.x3

e.x4

The curved path states that there is a correlation to be estimated between the variables it connects. The
absence of a curved pathsay, between e.x1 and e.x4 means that the variables are constrained to be
uncorrelated. That is not to say that x1 and x4 are uncorrelated; obviously, they are correlated because
both are functions of the same X . Their corresponding error variables, however, are uncorrelated.
The equations for this model, in their full detail, are

x1 = 1 + X1 + e.x1
x2 = 2 + X2 + e.x2
x3 = 3 + X3 + e.x3
x4 = 4 + X4 + e.x4
(X, x1 , x2 , x3 , x4 , e.x1 , e.x2 , e.x3 , e.x4 ) i.i.d. with mean and variance

12

intro 2 Learning the language: Path diagrams and command language

is constrained such that

e.x1 ,e.x2 = e.x2 ,e.x1 = 0

e.x1 ,e.x3 = e.x3 ,e.x1 = 0
e.x1 ,e.x4 = e.x4 ,e.x1 = 0
e.x2 ,e.x4 = e.x4 ,e.x2 = 0
e.x3 ,e.x4 = e.x4 ,e.x3 = 0
X,e.x1 = e.x1 ,X = 0
X,e.x2 = e.x2 ,X = 0
X,e.x3 = e.x3 ,X = 0
X,e.x4 = e.x4 ,X = 0
is constrained such that

X = 0
e.x1 = 0
e.x2 = 0
e.x3 = 0
e.x4 = 0
The parameters to be estimated are

1 , 1 , . . . , 4 , 4 ,

Look carefully at the above list. You will not find a line reading

e.x2 ,e.x3 = e.x3 ,e.x2 = 0

although you will find lines constraining the other covariances between error terms to be 0. The line
is missing because we drew a curved path between e.x2 and e.x3 .
There are lots of other curved arrows we could have drawn. By not drawing them, we are asserting
that the corresponding covariance is 0.

intro 2 Learning the language: Path diagrams and command language

13

Some authors would draw the above model as

x1

x2

x3

x4

e.x1

e.x2

e.x3

e.x4

A curved path from a variable to itself indicates a variance. All covariances could be shown, including
those between latent variables and between observed exogenous variables.
When we draw diagrams, however, we will assume variance paths and omit drawing them, and we
will similarly assume but omit drawing covariances between observed exogenous variables (there are
no observed exogenous variables in this model). The Builder in sem mode has an option concerning
the latter. Like everyone else, we will not assume correlations between latent variables unless they
are shown.
In sems (gsems) command-language notation, curved paths between variables are indicated via
an option:
(x1<-X) (x2<-X) (x3<-X) (x4<-X), cov(e.x2*e.x3)

Using the command language to specify standard linear SEMs

You can describe your model to sem by using path diagrams with the Builder, or you can describe
your model by using sems command language. Here are the trade-offs:
1. If you use path diagrams, you can see the results of your estimation as path diagrams or as
standard computer output.
2. If you use the command language, only standard computer output is available.
3. Typing models in the command language is usually quicker than drawing them in the Builder.
4. You can type models in the command language and store them in do-files. By doing so,
you can more easily correct the errors you make.

14

intro 2 Learning the language: Path diagrams and command language

Translating from path diagrams to command language is easy.

1. Path diagrams have squares and circles to distinguish observed from latent variables.
In the command language, variables are assumed to be observed if they are typed in lowercase
and are assumed to be latent if the first letter is capitalized. Variable educ is observed, while
variable Knowledge or KNOWLEDGE is latent.
If the observed variables in your dataset have uppercase names, type rename all, lower
to covert them to lowercase; see [D] rename group.
2. When typing path diagrams in the command language, remember the /// continuation line
indicator. You may type
. sem (x1<-X) (x2<-X) (x3<-X) (y<-X)
or you may type
. sem (x1<-X) (x2<-X)
(x3<-X) (y<-X)

///

or you may type

. sem (x1<-X)
(x2<-X)
(x3<-X)
(y<-X)

///
///
///

3. In path diagrams, you draw arrows connecting variables to indicate paths. In the command
language, you type variable names and arrows. So long as your arrow points the correct
way, it does not matter which variable comes first. The following mean the same thing:
(x1 <- X)
(X -> x1)
4. In the command language, you may type multiple variables on either side of the arrow:
(X -> x1 x2 x3 x4)
The above means the same as
(X -> x1) (X -> x2) (X -> x3) (X -> x4)
which means the same as
(x1 <- X) (x2 <- X) (x3 <- X) (x4 <- X)
which means the same as
(x1 x2 x3 x4 <- X)
In a more complicated measurement model, we might have
(X Y -> x1 x2 x3) (X -> x4 x5) (Y -> x6 x7)
The above means the same as
(X -> x1 x2 x3 x4 x5)
(Y -> x1 x2 x3 x6 x7)

///

intro 2 Learning the language: Path diagrams and command language

15

which means
(X -> x1) (X -> x2) (X -> x3) (X -> x4) (X -> x5)
(Y -> x1) (Y -> x2) (Y -> x3) (Y -> x6) (Y -> x7)

///

5. In path diagrams, you are required to show the error variables. In the command language,
you may omit the error variables. sem knows that each endogenous variable needs an error
variable. You can type
(x1 <- X) (x2 <- X) (x3 <- X) (x4 <- X)
and that means the same thing as
(x1
(x2
(x3
(x4

<<<<-

X
X
X
X

e.x1)
e.x2)
e.x3)
e.x4)

///
///
///

except that we have lost the small numeric 1s we had next to the arrows from e.x1 to x1,
e.x2 to x2, and so on. To constrain the path coefficient, you type
(x1
(x2
(x3
(x4

<<<<-

X
X
X
X

e.x1@1)
e.x2@1)
e.x3@1)
e.x4@1)

///
///
///

It is easier to simply type

(x1 <- X) (x2 <- X) (x3 <- X) (x4 <- X)
but now you know that if you wanted to constrain the path coefficient x2<-X to be 2, you
could type
(x1 <- X) (x2 <- X@2) (x3 <- X) (x4 <- X)
If you wanted to constrain the path coefficients x2<-X and x3<-X to be equal, you could
type
(x1 <- X) (x2 <- X@b) (x3 <- X@b) (x4 <- X)
We have not covered symbolic constraints with path diagrams yet, but typing symbolic
names along the paths in either the Builder or the command language is how you constrain
coefficients to be equal.
6. Curved paths are specified with the cov() option after you have specified your model:
(x1 x2 x3 x4 <- X), cov(e.x2*e.x3)
If you wanted to allow for correlation of e.x2*e.x3 and e.x3*e.x4, you can specify that
in a single cov() option,
(x1 x2 x3 x4 <- X), cov(e.x2*e.x3 e.x3*e.x4)
or in separate cov() options:
(x1 x2 x3 x4 <- X), cov(e.x2*e.x3) cov(e.x3*e.x4)

16

intro 2 Learning the language: Path diagrams and command language

7. Nearly all the above applies equally to gsem. We have to say nearly because sometimes,
in some models, some concepts simply vanish. For instance, in a logistic model, there are no
error terms. For generalized responses with family Gaussian, link log, there are error terms,
but they cannot be correlated. Also, for responses with family Gaussian, link identity, and
censoring, there are error terms, but they cannot be correlated. gsem also takes observed
exogenous variables as given and so cannot estimate the covariances between them.

Specifying generalized SEMs: Family and link

We began this discussion by showing you a linear measurement model:

x1

x2

x3

x4

e.x1

e.x2

e.x3

e.x4

In this model, we observe continuous measurements x1, . . . , x4. What if x1, . . . , x4 were instead
binary outcomes such as success and failure or passed and failed? That is, what if x1, . . . , x4 took
on values of only 1 and 0?
In that case, we would want to fit a model appropriate to binary outcomes. Perhaps we want to
fit a logistic regression model or a probit model. To do either one, we will have to use gsem rather
than sem. We will use a probit model.
The path diagram for the measurement model with binary outcomes is

Bernoulli

Bernoulli

Bernoulli

Bernoulli

x1

x2

x3

x4

probit

probit

probit

probit

What were plain boxes for x1, . . . , x4 now say Bernoulli and probit at the top and bottom,
meaning that the variable is from the Bernoulli family and is using the probit link. In addition, e.x1,
. . . , e.x4 (the error terms for x1, . . . , x4) have disappeared. In the generalized linear framework,
there are error terms only for the Gaussian family.

intro 2 Learning the language: Path diagrams and command language

17

Perhaps some math will clear up the issue. The generalized linear model is

g{E(y | X)} = x
and in the case of probit, g{E(y | X)} = 1 {E(y | X)}, where () is the cumulative normal
distribution. Thus the equations are

1 {E(x1 | X)} = 1 + X1
1 {E(x2 | X)} = 2 + X2
1 {E(x3 | X)} = 3 + X3
1 {E(x4 | X)} = 4 + X4
Note that there are no error terms. Equivalently, the above can be written as
Pr(x1 = 1 | X) = (1 + X1 )
Pr(x2 = 1 | X) = (2 + X2 )
Pr(x3 = 1 | X) = (3 + X3 )
Pr(x4 = 1 | X) = (4 + X4 )
In gsems command language, we write this model as
(x1 x2 x3 x4<-X, family(bernoulli) link(probit))
or as
(x1<-X,
(x2<-X,
(x3<-X,
(x4<-X,

family(bernoulli)
family(bernoulli)
family(bernoulli)
family(bernoulli)

link(probit))
link(probit))
link(probit))
link(probit))

In the command language, you can simply type probit to mean family(bernoulli) link(probit),
so the model could also be typed as
(x1 x2 x3 x4<-X, probit)
or even as
(x1<-X, probit) (x2<-X, probit) (x3<-X, probit) (x4<-X, probit)
Whether you type family(bernoulli) link(probit) or type probit, when all the response
variables are probit, you can type
(x1 x2 x3 x4<-X), probit
or
(x1<-X) (x2<-X) (x3<-X) (x4<-X), probit

18

intro 2 Learning the language: Path diagrams and command language

The response variables do not have to be all from the same family and link. Perhaps x1, x2, and
x3 are pass/fail variables but x4 is a continuous variable. Then the model would be diagrammed as

Bernoulli

Bernoulli

Bernoulli

Gaussian

x1

x2

x3

x4

probit

probit

probit

identity

e.x4
The words Gaussian and identity now appear for variable x4 and e.x4 is back! Just as previously,
the generalized linear model is
g{E(y | X)} = x
and in the case of linear regression, g() = , so our fourth equation becomes

E(x4 | X) = X4
or

x4 = + X4 + e.x4
and the entire set of equations to be estimated is
Pr(x1 = 1 | X) = (1 + X1 )
Pr(x2 = 1 | X) = (2 + X2 )
Pr(x3 = 1 | X) = (3 + X3 )

x4 = 4 + X4 + e.x4
This can be written in the command language as
(x1 x2 x3<-X, family(bernoulli) link(probit))
(x4
<-X, family(gaussian) link(identity))
or as
(x1 x2 x3<-X, probit) (x4<-X, regress)
regress is a synonym for family(gaussian) link(identity). Because family(gaussian)
link(identity) is the default, we can omit the option altogether:
(x1 x2 x3<-X, probit) (x4<-X)

intro 2 Learning the language: Path diagrams and command language

19

We demonstrated generalized linear models above by using probit (family Bernoulli, link probit),
but we could just as well have used logit (family Bernoulli, link logit). Nothing changes except that in
the path diagrams, where you see probit, logit would now appear. Likewise, in the command, where
you see probit, logit would appear.
The same is true with almost all the other possible generalized linear models. What they all have
in common is
1. There are no e. error terms, except for family Gaussian.
2. Response variables appear the ordinary way except that the family is listed at the top of the
box and the link is listed at the bottom of the box, except for family multinomial, link logit
(also known as multinomial logistic regression or mlogit).
Concerning item 1, we just showed you a combined probit and linear regression with its e.x4
term. Linear regression is family Gaussian.
Concerning item 2, multinomial logistic regression is different enough that we need to show it to
you.

Specifying generalized SEMs: Family and link, multinomial logistic regression

Lets consider a multinomial logistic model in which y takes on one of four possible outcomes
and is determined by x1 and x2 . Such a model could be fit by Statas mlogit command:
. mlogit y x1 x2
The model could also be fit by gsem. The path diagram would be
multinomial

2.y
logit

x1
multinomial

3.y
logit

x2
multinomial

4.y
logit

Note that there are three boxes for y , boxes containing 2.y , 3.y , and 4.y . When specifying a multinomial
logistic regression model in which the dependent variables can take one of k different values, you
draw k 1 boxes. Names like 1.y , 2.y , . . . , mean y = 1, y = 2, and so on. Logically speaking, you
have to omit one of them. Which you omit is irrelevant and is known as the base outcome. Estimated
coefficients in the model relevant to the other outcomes will be measured as a difference from the
base outcome. We said which you choose is irrelevant, but it might be easier for you to interpret your
model if you chose a meaningful base outcome, such as the most frequent outcome. In fact, that is
just what Statas mlogit command does by default when you do not specify otherwise.

20

intro 2 Learning the language: Path diagrams and command language

In path diagrams, you may implicitly or explicitly specify the base outcome. We implicitly specified
the base by omitting 1.y from the diagram. We could have included it by drawing a box for 1.y
and labeling it 1b.y . Stata understands the b to mean base category. See [SEM] example 37g for an
example.
Once you have handled specification of the base category, you draw path arrows from the predictors
to the remaining outcome boxes. We drew paths from x1 and x2 to all of the outcome boxes, but if
we wanted to omit x1 to 3.y and 4.y , we could have omitted those paths.
Our example is simple in that y is the final outcome. If we had a more complex model where y s
outcome affected another response variable, arrows would connect all or some of 2.y , 3.y , and 4.y
to the other response variable.
The command syntax for our simple example is
(2.y 3.y 4.y<-x1 x2), mlogit
2.y, 3.y, and 4.y are examples of Statas factor-variable syntax. The factor-variable syntax has some
other features that can save typing in command syntax. i.y, for instance, means 1b.y, 2.y, 3.y,
and 4.y. It is especially useful because if we had more levels, say, 10, it would mean 1b.y, 2.y,
. . . , 10.y. To fit the model we diagrammed, we could type
(i.y<-x1 x2), mlogit
If we wanted to include some paths and not others, we need to use the more verbose syntax. For
instance, to omit paths from x1 to 3.y and 4.y , we would type
(2.y<-x1 x2) (3.y 4.y<-x2), mlogit
or
(i.y<-x2) (2.y<-x1), mlogit
For more information on specifying mlogit paths, see [SEM] intro 3, [SEM] example 37g, and
[SEM] example 41g.

Specifying generalized SEMs: Family and link, paths from response variables
When we draw a path from one response variable to another, we are stating that the first endogenous
variable is a predictor of the other endogenous variable. The diagram might look like this:

x2

x1

y1

y2

e.y1

e.y2

The command syntax might look like (y1<-x1) (y2<-y1 x2).

intro 2 Learning the language: Path diagrams and command language

21

The response variables in the model are linear, and note that there is a path from y1 to y2 . Could
we change y1 to be a member of the generalized linear family, such as probit, logit, and so on? It
turns out that we can:

x2

Bernoulli

x1

y1

y2

probit

e.y2
In the command syntax, we could write this model as (y1<-x1, probit) (y2<-y1 x2). (In this
case, the observed values and not the expectations are used to fit the y1->y2 coefficient. In general,
this is true for all generalized responses that are not family Gaussian, link identity.)
We can make the substitution from linear to generalized linear if the path from y1 appears in a
recursive part of the model. We will define recursive shortly, but trust us that the above model is
recursive all the way through. The substitution would not have been okay if the path had been in a
nonrecursive portion of the model. The following is a through-and-through nonrecursive model:

x1
y1

e.y1

y2

e.y2

x2

x3
It could be written in command syntax as (y1<-y1 x1 x2) (y2<-y2 x2 x3).
In this model, we could not change y1 to be family Bernoulli and link probit or any other
generalized linear response variable. If we tried to fit the model with such a change, we would get
an error message:
invalid path specification;
a loop among the paths between y1 and y2 is not allowed
r(198);

The software will spot the problem, but you can spot it for yourself.

22

intro 2 Learning the language: Path diagrams and command language

Nonrecursive models have loops. Do you see the loop in the above model? You will if you work
out the total effect of a change in y2 from a change in y1 . Assume a change in y1 . Then that change
directly affects y2 , the new value of y2 affects y1 , which in turn indirectly affects y2 again, which
affects y1 , and on and on.
Now follow the same logic with either the probit or the continuous recursive models above. The
change in y1 affects y2 and it stops right there.
We sympathize if you think that we have interchanged the terms recursive and nonrecursive.
Remember it this way: total effects in recursive models can be calculated nonrecursively because
the model itself is recursive, and total effects in nonrecursive models must be calculated recursively
because the model itself is nonrecursive.
Anyway, you may draw paths from generalized linear response variables to other response variables,
whether linear or generalized linear, as long as no loops are formed.
We gave special mention to multinomial logistic regression in the previous section because those
models look different from the other generalized linear models. Multinomial logistic regression has a
plethora of response variables. In the case of multinomial logistic regression, a recursive model with
a path from the multinomial logistic outcome y1 to (linear) y2 would look like this:

multinomial

x2

2.y1
logit
multinomial

x1

3.y1

y2

logit

multinomial

e.y2

4.y1
logit

In the command syntax, the model could be written as

(2.y1 3.y1 4.y1<-x1, mlogit) (y2<-2.y1 3.y1 4.y1 x2)
In multinomial logistic regression models, outcomes are numbered 1, 2, . . . , k . The outcomes
might correspond to walk, take public transportation, drive a car, and fly. In ordered probit and
logistic models, outcomes are also numbered 1, 2, . . . , k , and the outcomes are otherwise similar
to multinomial logistic models except that the outcomes are also naturally ordered. The outcomes
might be did poorly, did okay, did fairly well, and did well. After taking account of the ordering,
you would probably want to model your remaining response variables in the same way you did for
multinomial logistic regression. In that case, you would diagram the model like this:

intro 2 Learning the language: Path diagrams and command language

23

x2
2.y1

ordinal

x1

y1

3.y1

y2

logit

e.y2
4.y1
In the command syntax, the model could be written as
(y1<-x1, ologit) (y2<-2.y1 3.y1 4.y1 x2)
Unlike multinomial logistic regression, in which the k outcomes result in the estimation of k 1
equations, in ordered probit and logistic models, only one equation is estimated, and thus the response
variable is specified simply as y1 rather than 2.y1 , 3.y1 , and 4.y1 . Even so, you probably will want
the different outcomes to have separate coefficients in the y2 equation so that the effects of being in
groups 1, 2, 3, and 4 are 0 , 0 + 1 , 0 + 2 , and 0 + 3 , respectively. If you drew the diagram
with a single path from y1 to y2 , the effects of being in the groups would be 0 + 11 , 0 + 21 ,
0 + 31 , and 0 + 41 , respectively.

Specifying generalized SEMs: Multilevel mixed effects (2 levels)

The models above are all single-level models or, equivalently, observational-level models. Lets
reconsider our original measurement model:

x1

x2

x3

x4

e.x1

e.x2

e.x3

e.x4

24

intro 2 Learning the language: Path diagrams and command language

The data for this model would look something like this:
. list in 1/5

1.
2.
3.
4.
5.

x1

x2

x3

x4

76
106
99
101
104

76
110
106
85
82

63
114
104
77
108

490
778
757
637
654

Lets pretend that the observations are students and that x1, . . . , x4 are four test scores.
Now lets pretend that we have new data with students from different schools. A part of our data
might look like this:
. list in 1/10
school

x1

x2

x3

x4

1.
2.
3.
4.
5.

1
1
1
1
1

99
70
75
96
80

91
62
67
88
72

94
65
70
91
75

92
63
68
89
73

6.
7.
8.
9.
10.

2
2
2
2
2

113
112
97
84
89

119
118
103
90
95

118
117
102
89
94

119
118
103
90
95

We have four test scores from various students in various schools. Our corresponding model might
be

x1

e.x1

x2

e.x2

x3

e.x3

x4

e.x4

school1
The school1 inside double circles in the figure tells us, I am a latent variable at the school
levelmeaning that I am constant within school and vary across schoolsand I correspond to the
latent variable named M1. That is, double circles denote a latent variable named M#, where # is the
subscript of the variable name inside the double circles. Meanwhile, the variable name inside the
double circles specifies the level of the M# variable.

intro 2 Learning the language: Path diagrams and command language

25

The equivalent command-language form for this model is

(x1 x2 x3 x4<-X M1[school])
The M1 part of M1[school] is the latent variables name, while the [school] part means at the
school level. Thus M1[school] denotes latent variable M1, which varies at the school level or,
equivalently, which is constant within school.
By the way, we gave the latent variable the name M1 just to match what the Builder does by
default. In real life, we would probably give it a different name, such as S.
The mathematical way of writing this model is

x1 = 1 + 1 X + 1 M1,S + e.x1
x2 = 2 + 2 X + 2 M1,S + e.x2
x3 = 3 + 3 X + 3 M1,S + e.x3
x4 = 4 + 4 X + 4 M1,S + e.x4
where S = school number.
Thus we have three different ways of referring to the same thing: school1 inside double circles in
the path diagram corresponds to M1[school] in the command language, which in turn corresponds
to M1,S in the mathematical notation.
Rabe-Hesketh, Skrondal, and Pickles (2004) use boxes to identify different levels of the model.
Our path diagrams are similar to theirs, but they do not use double circles for multilevel latent
variables. We can put a box around the individual-level part of the model, producing something that
looks like this:

x1

e.x1

x2

e.x2

x3

e.x3

x4

e.x4

school1
You can do that in the Builder, but the box has no special meaning to gsem; however, adding the
box does make the diagram easier to understand in presentations.
However you diagram the model, this model is known as a two-level model. The first level is the
student or observational level, and the second level is the school level.

26

intro 2 Learning the language: Path diagrams and command language

Specifying generalized SEMs: Multilevel mixed effects (3 levels)

A three-level model adds another latent variable at yet another level. For instance, perhaps we
have data on four test scores for individual students, who are nested within school, which are nested
within counties. The beginning of our data might look like this:
. list in 1/10
county

school

x1

x2

x3

x4

1.
2.
3.
4.
5.

1
1
1
1
1

1
1
1
1
1

105
115
90
124
114

115
125
100
134
124

113
123
98
132
122

114
124
99
133
123

6.
7.
8.
9.
10.

1
1
1
1
1

2
2
2
2
2

115
112
116
102
109

128
125
129
115
122

125
122
126
112
119

127
124
128
114
121

Further down, the data might look like this:

. list in 991/l
county

school

x1

x2

x3

x4

991.
992.
993.
994.
995.

2
2
2
2
2

1
1
1
1
1

106
72
60
107
125

111
77
65
112
130

110
76
64
111
129

111
77
65
112
130

996.
997.
998.
999.
1000.

2
2
2
2
2

2
2
2
2
2

79
84
90
96
104

72
77
83
89
97

74
79
85
91
99

73
78
84
90
98

intro 2 Learning the language: Path diagrams and command language

27

We might diagram the student-within-school-within-county model as

county1

school2

x1

x2

x3

x4

e.x1

e.x2

e.x3

e.x4

Now we have two variables in double circles: county1 and school2.

county1 tells us, I am a latent variable at the county levelmeaning that I am constant within
county and vary across countiesand I correspond to the latent variable named M1.
school2 tells us, I am a latent variable at the school levelmeaning that I am constant within
school and vary across schoolsand I correspond to the latent variable named M2.
Do not read anything into the fact that the latent variables are named M1 and M2 rather than M2
and M1. When we diagrammed the figure in the Builder, we diagrammed county first and then we
added school. Had we diagrammed them the other way around, the latent variable names would
have been reversed. We can edit the names after the fact, but we seldom bother.
The equivalent command-language form for this three-level model is
(x1 x2 x3 x4<-X M1[county] M2[county>school])
or
(x1 x2 x3 x4<-X M1[county] M2[school<county])
The [school<county] part is usually spoken aloud as school within county, and
[county>school] is spoken aloud as county containing school. When we write [county>school]
or [school<county], we are saying that the counties contain schools or, equivalently, that school
is nested within county. Thus the model we are specifying is in fact a three-level nested model.
The order in which we specify the level effects is irrelevant; we could just as well type
(x1 x2 x3 x4<-X M2[county>school] M1[county])

28

intro 2 Learning the language: Path diagrams and command language

or
(x1 x2 x3 x4 <- X M2[school<county] M1[county])
The mathematical way of writing this model is

x1 = 1 + 1 X + 1 M1,C + 1 M2,S + e.x1

x2 = 2 + 2 X + 2 M1,C + 2 M2,S + e.x2
x3 = 3 + 3 X + 3 M1,C + 3 M2,S + e.x3
x4 = 4 + 4 X + 4 M1,C + 4 M2,S + e.x4
where C = county number and S = school number.
Three-level nested models are often double-boxed (or double-shaded) in presentation:

county1

school2

x1

x2

x3

x4

e.x1

e.x2

e.x3

e.x4

The darker box highlights the first level (student, in our case), and the lighter box highlights the
second level (school, in our case). As we previously mentioned, you can add these boxes using the
Builder, but they are purely aesthetic and have no meaning to gsem.

Specifying generalized SEMs: Multilevel mixed effects (4+ levels)

We have now diagrammed one-, two-, and three-level nested models. You can draw with the
Builder and fit with gsem higher-level models, but you will usually need lots of data to be successful
in fitting those models and, even so, you may run into other estimation problems.

intro 2 Learning the language: Path diagrams and command language

29

Specifying generalized SEMs: Multilevel mixed effects with random intercepts

Lets change gears and consider with the following simple model a linear regression of y on x1
and x2 :

x1

e.y

x2
In the command language, this model is specified as
(x1 x2->y)
or
(y<-x1 x2)
and the mathematical representation is

y = + x1 + x2 + e.y
Now assume that we have data not just on y , x1 , and x2 , but also on county of residence. If we
wanted to add a random intercepta random effectfor county, the diagram becomes

county1

x1

x2
The command-language equivalent is
(x1 x2 M1[county]->y)
or
(y<-x1 x2 M1[county])

e.y

30

intro 2 Learning the language: Path diagrams and command language

and the mathematical representation is

y = + x1 + x2 + M1,C + e.y
where C = county number. Actually, the model is

y = + x1 + x2 + M1,C + e.y
but is automatically constrained to be 1 by gsem. The software is not reading our mind; consider
a solution for M1,C with = 0.5. Then another equivalent solution is M1,C /2 with = 1, and
another is M1,C /4 with = 2, and on and on, because in all cases, M1,C will equal the same
value. The fact is that is unidentified. Whenever a latent variable is unidentified because of such
scaling considerations, gsem (and sem) automatically set the coefficient to 1 and thus set the scale.
This is a two-level model: the first level is the observational level and the second level is county.
Just as we demonstrated previously with the measurement model, we could have a three-level
nested model. We could imagine the observational level nested within the county level nested within
the state level. The path diagram would be

county1

x1

e.y

x2

state2
The command-language equivalent is
(y<-x1 x2 M1[county<state] M2[state])
and the mathematical representation is

y = + x1 + x2 + M1,C + M2,S + e.y

where C = county number and S = state number. Just as previously, the actual form of the equation
is
y = + x1 + x2 + M1,C + M2,S + e.y
but the constraints = = 1 are automatically applied by the software.

intro 2 Learning the language: Path diagrams and command language

31

You can specify higher-level models, but just as we mentioned when discussing the higher-level
measurement models, you will need lots of data to fit them successfully and you still may run into
other estimation problems.
You can also fit crossed models, such as county and occupation. Unlike county and state, where
a particular county appears only in one state, the same occupation will appear in more than one
county, which is to say, occupation and county are crossed, not nested. Except for a change in the
names of the variables, the path diagram for this model looks identical to the diagram for the state
and county-within-state model:

county1

x1

e.y

x2

occup2
When we enter the model into the Builder, however, we will specify that the two effects are crossed
rather than nested.
The command-language way of expressing this crossed model is
(y<-x1 x2 M1[county] M2[occupation])
and the mathematical representation is

y = + x1 + x2 + M1,C + M2,O + e.y

where C = county number and O = occupation number.
Higher-order crossed models are theoretically possible, and you will find gsem game to try to fit
them. However, you are unlikely to be successful unless you have lots of data so that all the effects
can be identified. In fact, you can specify crossed models anywhere you can specify nested models,
but the same comment applies.

32

intro 2 Learning the language: Path diagrams and command language

Specifying generalized SEMs: Multilevel mixed effects with random slopes

Perhaps we feel that county does not affect the intercept, but it affects the slope of x1. In that
case, we just shift the path from county1 to instead point to the path between y and x1 .

county1

x1

e.y

x2
The command-language equivalent for this model is
(y<-x1 c.x1#M1[county]) (y<-x2)
or
(y<-x1 c.x1#M1[county] x2)
To include a random slope (coefficient) on a variable in the command language, include the variable
on which you want the random slope just as you ordinarily would:
(y<-x1
Then, before closing the parenthesis, type
c.variable#latent variable[grouping variable]
In our case, the variable on which we want the random coefficient is x1, the latent variable we want
to create is named M1, and the grouping variable within which the latent variable will be constant
is county, so we type
(y<-x1 c.x1#M1[county]
Finally, finish off the command by typing the rest of the model:
(y<-x1 c.x1#M1[county] x2)
This is another example of Statas factor-variable notation.
The mathematical representation of the random-slope model is

y = + x1 + x2 + M1,C x1 + e.y
where C = county number and = 1.

intro 2 Learning the language: Path diagrams and command language

33

You can simultaneously specify both random slope and random intercept by putting together what
we have already done. The first time you see the combined path diagram, you may think there is a
mistake:

county2

county1

x1

e.y

x2
County appears twice in two different double circles, once with a subscript of 1 and the other
with a subscript of 2! We would gamble at favorable odds that you would have included county in
double circles once and then would have drawn two paths from it, one to the path between x1 and
y, and the other to y itself. That model, however, would make an extreme assumption.
Consider what you would be saying. There is one latent variable M1. The random intercepts would
be equal to M1. The random slopes would be equal to M1, a scalar replica of the random intercepts!
You would be constraining the random slopes and intercepts to be correlated 1.
What you usually want, however, is one latent variable for the random slopes and another for the
random intercepts. They might be correlated, but they are not related by a multiplicative constant.
Thus we also included a covariance between the two random effects.
The command-language equivalent of the correct path diagram (the one shown above) is
(y<-x1 x2 c.x1#M2[county] M1[county])
You will learn later that the command language assumes covariances exist between latent exogenous
variables unless an option is specified. Meanwhile, the Builder assumes those same covariances are
0 unless a curved covariance path is drawn.
The mathematical representation of the model is

y = + x1 + x2 + M2,C x1 + M1,C + e.y

where C = county number and = = 1.
And now you, too, are an expert on path diagrams.

Reference
Rabe-Hesketh, S., A. Skrondal, and A. Pickles. 2004. Generalized multilevel structural equation modeling. Psychometrika
69: 167190.

34

intro 2 Learning the language: Path diagrams and command language

Also see
[SEM] intro 1 Introduction
[SEM] intro 3 Learning the language: Factor-variable notation (gsem only)
[SEM] Builder SEM Builder
[SEM] Builder, generalized SEM Builder for generalized models
[SEM] sem and gsem path notation Command syntax for path diagrams

Title
intro 3 Learning the language: Factor-variable notation (gsem only)

Description

Remarks and examples

Also see

Description
This entry concerns gsem only; sem does not allow the use of factor variables.
gsem allows you to use Statas factor-variable notation in path diagrams (in the SEM Builder) and
in the command language. Use of the notation, though always optional, is sometimes useful:
1. Use of factor variables sometimes saves effort. For instance, rather than typing
. generate femXage = female * age

and including femXage in your model, you can directly include 1.female#c.age. You can
type 1.female#c.age with the gsem command or type it into an exogenous-variable box
in a path diagram.
2. Use of factor variables can save even more effort in command syntax. You can type things
like i.female i.skill i.female#i.skill to include main effects and interaction effects
of female interacted with indicators for all the different levels of skill.
3. Use of factor variables causes the postestimation commands margins, contrast, and
pwcompare to produce more useful output. For instance, they will show the effect of the
discrete change between female = 0 and female = 1 rather than the infinitesimal change
(slope) of female.
In the examples in the rest of this manual, we sometimes use factor-variable notation and, at other
times, ignore it. We might have variable smokes recording whether a person smokes. In one example,
you might see smokes directly included in the model and yet, in another example, we might have
the odd-looking 1.smokes, which is factor-variable notation for emphasizing that smokes is a 0/1
variable. We probably used 1.smokes for reason 3, although we might have done it just to emphasize
that variable smokes is indeed 0/1.
In other examples, we will use more complicated factor-variable notation, such as 1.female#c.age,
because it is more convenient.
You should follow the same rules. Use factor-variable notation when convenient or when you plan
subsequently to use postestimation commands margins, contrast, or pwcompare. If neither reason
applies, use factor-variable notation or not. There is no benefit from consistency in this case.

Remarks and examples

Remarks are presented under the following headings:
Specifying indicator variables
Specifying interactions with indicator variables
Specifying categorical variables
Specifying interactions with categorical variables
Specifying endogenous variables
Inconsistency between gsem and other estimation commands

35

36

intro 3 Learning the language: Factor-variable notation (gsem only)

Specifying indicator variables

We wish to fit the following model:

This model corresponds to the equation yi = xi + i . It can be fit using the Builder with the path
diagram above or using the command syntax by typing
. gsem (y<-x)

Say we now tell you that x is a 0/1 (binary) variable. Perhaps x indicates male/female. That
changes nothing mathematically. The way we drew the path diagram and specified the command
syntax are as valid when x is an indicator variable as they were when x was continuous.
Specifying 1.x is a way you can emphasize that x is 0/1:

1.x

That model will produce the same results as the first model. In command syntax, we can type
. gsem (y<-1.x)

The only real advantage of 1.x over x arises when we plan to use the postestimation command
margins, contrast, or pwcompare, as we mentioned above. If we fit the model using 1.x, those
commands will know x is 0/1 and will exploit that knowledge to produce more useful output.
In other cases, the #.varname notation can sometimes save us effort. #.varname means a variable
equal to 1 if varname = #. It is not required that varname itself be an indicator variable.
Lets say variable skill takes on the values 1, 2, and 3 meaning unskilled, skilled, and highly
skilled. Then 3.skill is an indicator variable for skill = 3. We could use 3.skill in path
diagrams or in the command language just as we previously used 1.x.

Specifying interactions with indicator variables

We can specify interactions with #. Say we wish to fit the model
yi = 0 + 1 femalei + 2 agei + 3 femalei agei + i

intro 3 Learning the language: Factor-variable notation (gsem only)

37

We could do that by specifying 1.female, age, and 1.female#c.age in a path diagram:

1.female

age

1.female#c.age

We can use 1.female, age, and 1.female#c.age in command syntax, too:

. gsem (y<-1.female age 1.female#c.age)

The c in c.age specifies that age is a continuous variable.

We can just as easily create interactions of indicator variables with other indicator variables. If
variable employed is a 0/1 variable, then 1.female#1.employed creates an indicator for employed
female.

Specifying categorical variables

We use the same #.varname notation to handle categorical variables as we did for indicator
variables. Consider the following model:

1.female

2.skill

3.skill

The same model would be specified in command syntax as

. gsem (y<-1.female 2.skill 3.skill)

The equation for the model is

yi = 0 + 1 1.femalei + 2 2.skilli + 3 3.skilli + i
However we express the model, we are including indicator variables for skill level 2 and for skill
level 3. We intentionally omit skill level 1.
Categorical variables require a base level, a level against which the other effects will be measured.
We omitted 1.skill to cause 1.skill to become the base level. This results in 2 measuring the
difference in y between skill levels 2 and 1, and 3 measuring the difference in y between skill
levels 3 and 1.

38

intro 3 Learning the language: Factor-variable notation (gsem only)

Omission is one way to specify the base level. The other way is to include all the skill levels and
indicate which one we want to be used as the base:

1.female

1b.skill
y

2.skill

3.skill

In command syntax, we type

. gsem (y<-1.female 1b.skill 2.skill 3.skill)

b specifies the base level. Specifying 1b.skill makes skill level 1 the base. Had we wanted skill
level 2 to be the base, we would have specified 2b.skill.
Specification by omission is usually easier for simple models. For more complicated models, which
might have paths from different skill levels going to different variables, it is usually better to specify
all the skill levels and all the relevant paths, mark one of the skill levels as the base, and let gsem
figure out the correct reduced-form model.
By the way, indicator variables are just a special case of categorical variables. Indicator variables
are categorical variables with two levels. Everything just said about categorical variable skill applies
equally to indicator variable female. We can specify 1.female by itself, and thus implicitly specify
that female 6= 1 is the base, or we can explicitly specify both 0b.female and 1.female.
The command language has a neat syntax for specifying all the indicator variables that can be
manufactured from a categorical variable. When you type i.skill, it is the same as typing 1b.skill
2.skill 3.skill. You can use the i. shorthand in the command language by typing
. gsem (y<-1.female i.skill)

or even by typing
. gsem (y<-i.female i.skill)

Most experienced Stata command-line users would type the model the second way, putting i. in
front of both skill and female. They would not bother to think whether variables are indicator or
categorical, nor would they concern themselves with remembering how the indicator and categorical
variables are coded.
You cannot use the i. notation in path diagrams, however. Path diagrams allow only one variable
per box. The i. notation produces at least two variables, and it usually produces a lot more.

intro 3 Learning the language: Factor-variable notation (gsem only)

39

Specifying interactions with categorical variables

We just discussed the model (y<-i.female i.skill). Obviously, the next step is to include an
interaction between female and skill level. To write such a model, we use the same # interaction
indicator that we have already discussed in the context of indicator variables: (y<-i.female i.skill
i.female#i.skill). We use the same notation because there is really no distinction between factor
variables and indicator variables. We can specify the model exactly as shown by using the command
syntax:
. gsem (y<-i.female i.skill i.female#i.skill)

In the Builder, we could diagram this model by specifying all the individual interactions but omitting
the base levels,

1.female

2.skill

3.skill

1.female#2.skill

1.female#3.skill

40

intro 3 Learning the language: Factor-variable notation (gsem only)

or by including them,

1.female

1b.skill

2.skill

3.skill

1.female#1b.skill

1.female#2.skill

1.female#3.skill

In the second diagram, we did not bother to include the base levels for the indicator variable female,
but we could have included them.
We can type the model in command syntax just as we have drawn it, either as
. gsem (y<-1.female
2.skill 3.skill
1.female#2.skill

1.female#3.skill)

. gsem (y<-1.female
1b.skill 2.skill
1.female#1b.skill

3.skill
1.female#2.skill

or as

1.female#3.skill)

But in command syntax, it is easiest to type

. gsem (y<-i.female i.skill i.female#i.skill)

Specifying endogenous variables

You may not use factor variables to specify endogenous variables except for the multinomial
logistic regression model. Multinomial logistic regression is also known as mlogit and as being family
multinomial, link logit.
In the mlogit case, you must specify a base level, which you can do implicitly by omission or
explicitly by including a b on one of the levels. See [SEM] example 37g and [SEM] example 41g for
examples.

intro 3 Learning the language: Factor-variable notation (gsem only)

41

Inconsistency between gsem and other estimation commands

gsem treats factor variables slightly differently than do all the other Stata estimation commands.
The differences are
1. When the #.varname notation is used, gsem will choose a base level from among the
unspecified levels. All other commands will pick one of the specified levels as a base. One
implication of this is that if all levels are specified using #.varname, gsem will not pick a
base whereas other commands will.
2. gsem will ignore unspecified elements of interactions that otherwise would be added when
explicit levels of factor variables are used.
There is an argument to be made that the other estimation commands ought to work like gsem.
The differences, however, do not affect the i.varname notation, which is how most users use factor
variables. The differences were introduced to deal with the Builders one-box-per-variable restriction,
which results in more frequent use of #.varname, which caused Statas developers to rethink defaults.
In the command language, you can ask gsem to use the same behavior as the other estimation
commands by specifying option fvstandard.

Also see
[SEM] intro 2 Learning the language: Path diagrams and command language
[SEM] intro 4 Substantive concepts
[SEM] Builder SEM Builder
[SEM] Builder, generalized SEM Builder for generalized models
[SEM] sem and gsem path notation Command syntax for path diagrams

Title
intro 4 Substantive concepts

Description

Remarks and examples

References

Also see

Description
The structural equation modeling way of describing models is deceptively simple. It is deceptive
because the machinery underlying structural equation modeling is sophisticated, complex, and sometimes temperamental, and it can be temperamental both in substantive statistical ways and in practical
computer ways.
Professional researchers need to understand these issues.

Remarks and examples

Remarks are presented under the following headings:
Differences in assumptions between sem and gsem
sem: Choice of estimation method
gsem: Choice of estimation method
Treatment of missing values
Variable types: Observed, latent, endogenous, exogenous, and error
Constraining parameters
Constraining path coefficients to specific values
Constraining intercepts to specific values (suppressing the intercept)
Constraining path coefficients or intercepts to be equal
Constraining covariances to be equal (or to specific values)
Constraining variances to specific values (or to be equal)
Identification 1: Substantive issues
Not all models are identified
How to count parameters
What happens when models are unidentified
How to diagnose and fix the problem
Identification 2: Normalization constraints (anchoring)
Why the problem arises
How the problem would manifest itself
How sem (gsem) solves the problem for you
Overriding sems (gsems) solution

42

intro 4 Substantive concepts

43

Differences in assumptions between sem and gsem

sem fits standard linear SEMs.
gsem fits generalized SEMs, by which we mean
1. gsem fits not just linear but generalized linear models,
2. gsem fits multilevel mixed models, and
3. gsem fits items 1 and 2 combined.
There is a difference in assumptions between standard linear SEMs and generalized SEMs.
Standard linear SEMs generally assume that the observed endogenous variables, the observed
exogenous variables, the latent endogenous variables, and the latent exogenous variables are jointly
distributed normally with mean and variance matrix . In this formulation, we include the e. error
variables among the latent exogenous variables.
Generalized SEMs drop the observed variables from the joint-normality assumption. Instead, generalized SEMs treat the observed exogenous variables as given and produce estimates conditional on
their values. This seems a minor difference, but for some researchers it has important implications.
Consider a researcher wishing to include a subjects age and age-squared among the observed
exogenous variables of a model. Variables age and age-squared simply cannot be jointly normally
distributed. Including age and age-squared violates the assumptions of standard linear SEMs, at least
officially. You may now suspect that such researchersperhaps including yourselfshould use gsem
and not sem. Because standard linear SEMs are a subset of generalized SEMs, that should not be much
of an inconvenience.
It is interesting to note, however, that sem and gsem produce the same numeric solutions for the
parameters and the standard errors when both can fit the same model.
All of which is to say that although it is typical to assume joint normality of all variables when
deriving the standard linear SEM, joint normality is not strictly necessary. The lesser assumption of
joint normality conditional on the observed exogenous variables is sufficient. Even the normality
assumption can be relaxed and replaced with i.i.d., and even the i.i.d. assumption can be relaxed.
We will discuss these issues below.
Even so, some structural equation modeling statistics are dependent on the full joint normality
assumption. Many goodness-of-fit tests fall into that category. Throughout this manual, we will be
explicit about when the full joint-normality assumption is required.
gsem never requires the full joint-normality assumption. Standard linear SEMs have a history that
includes the full joint-normality assumption; generalized SEMs have no such history. At the end of
the day, both can be considered M estimation.

sem: Choice of estimation method

sem provides four estimation methods: maximum likelihood (ML; the default), quasimaximum
likelihood (QML), asymptotic distribution free (ADF), and maximum likelihood with missing values
(MLMV).
Strictly speaking, the assumptions one must make to establish the consistency of the estimates and
their asymptotic normality is determined by the method used to estimate them. We want to give you
advice on when to use each.
1. ML is the method that sem uses by default. In sem, the function being maximized formally
assumes the full joint normality of all the variables, including the observed variables. But

44

intro 4 Substantive concepts

the full joint-normality assumption can be relaxed, and the substitute conditional-on-theobserved-exogenous-variables is sufficient to justify all reported estimates and statistics except
the log-likelihood value and the model-versus-saturated 2 test.
Relaxing the constraint that latent variables outside of the error variables are not normally
distributed is more questionable. In the measurement model (X->x1 x2 x3 x4), simulations
with the violently nonnormal X 2 (2) produced good results except for the standard
error of the estimated variance of X . Note that it was not the coefficient on X that was
estimated poorly, it was not the coefficients standard error, and it was not even the variance
of X that was estimated poorly. It was the standard error of the variance of X . Even so,
there are no guarantees.
sem uses method ML when you specify method(ml) or when you omit the method() option
altogether.
2. QML uses ML to fit the model parameters but relaxes the normality assumptions when
estimating the standard errors. QML handles nonnormality by adjusting standard errors.
Concerning the parameter estimates, everything just said about ML applies to QML because
those estimates are produced by ML.
Concerning standard errors, we theoretically expect consistent standard errors, and we
practically observe that in our simulations. In the measurement model with X 2 (2), we
even obtained good standard errors of the estimated variance of X . QML does not really fix
the problem of nonnormality of latent variables, but it does tend to do a better job.
sem uses method QML when you specify method(ml) vce(robust) or, because method(ml)
is the default, when you specify just vce(robust).
3. ADF makes no assumption of joint normality or even symmetry, whether for observed or latent
variables. Whereas QML handles nonnormality by adjusting standard errors and not point
estimates, ADF produces justifiable point estimates and standard errors under nonnormality.
For many researchers, this is most important for relaxing the assumption of normality of the
errors, and because of that, ADF is sometimes described that way. ADF in fact relaxes the
normality assumption for all latent variables.
Along the same lines, it is sometimes difficult to be certain exactly which normality
assumptions are being relaxed when reading other sources. It sometimes seems that ADF
uniquely relaxes the assumption of the normality of the observed variables, but that is not
true. Other methods, even ML, can handle that problem.
ADF is a form of weighted least squares (WLS). ADF is also a generalized method of moments
(GMM) estimator. In simulations of the measurement model with X 2 (2), ADF produces

excellent results, even for the standard error of the variance of X . Be aware, however, that
ADF is less efficient than ML when latent variables can be assumed to be normally distributed.
If latent variables (including errors) are not normally distributed, on the other hand, ADF
will produce more efficient estimates than ML or QML.

sem uses method ADF when you specify method(adf).

4. MLMV aims to retrieve as much information as possible from observations containing missing
values.
In this regard, sem methods ML, QML, and ADF do a poor job. They are known as listwise
deleters. If variable x1 appears someplace in the model and if x1 contains a missing value
in observation 10, then observation 10 simply will not be used. This is true whether x1 is
endogenous or exogenous and even if x1 appears in some equations but not in others.

intro 4 Substantive concepts

45

Method MLMV, on the other hand, is not a deleter at all. Observation 10 will be used in
making all calculations.
For method MLMV to perform what might seem like magic, joint normality of all variables
is assumed and missing values are assumed to be missing at random (MAR). MAR means
either that the missing values are scattered completely at random throughout the data or that
values more likely to be missing than others can be predicted by the variables in the model.
Method MLMV formally requires the assumption of joint normality of all variables, both
observed and latent. If your observed variables do not follow a joint normal distribution, you
may be better off using ML, QML, or ADF and simply omitting observations with missing
values.
sem uses method MLMV when you specify method(mlmv). See [SEM] example 26.

gsem: Choice of estimation method

gsem provides only two estimation methods: maximum likelihood (ML; the default) and quasimaximum likelihood (QML).
1. ML is the method that gsem uses by default. This is the same ML used by sem but applied
to a different likelihood function. The sem likelihood function assumes and includes joint
normality of all variables. The gsem likelihood function assumes only conditional normality.
Because likelihood functions are different, the likelihood values reported by sem and gsem
are not comparable except in the case where there are no observed exogenous variables.
ML formally assumes conditional normality, and thus latent variables are still assumed to

be normally distributed. What we said in the sem case about relaxing the assumption of
normality of the latent variables applies equally in the gsem case.
gsem uses method ML when you specify method(ml) or when you omit the method()
option altogether.
2. QML uses ML to fit the model but relaxes the conditional normality assumptions when
estimating the standard errors. QML handles nonnormality by adjusting standard errors.
Everything said about sems QML applies equally to gsems QML.
gsem uses method QML when you specify vce(robust).
Because the choice of method often affects convergence with sem, in the gsem case there is a
tendency to confuse choice of integration method with maximization method. However, there are
no issues related to assumptions about integration method; choice of integration method is purely a
mechanical issue. This is discussed in [SEM] intro 12.

Treatment of missing values

sem, sem with method MLMV, and gsem treat missing values differently.
1. sem by default is a listwise deleter.
If variable x1 appears in the model and if x1 contains missing in observation 10, then
observation 10 will not be used. This is true whether x1 is endogenous or exogenous and
even if x1 appears in some equations but not in others.

46

intro 4 Substantive concepts

2. sem with method MLMV is not a deleter at all; it uses all observations.
If variable x1 appears in the model and if x1 contains missing in observation 10, then
observation 10 will still be used. Doing this formally requires assuming the joint normality
of all observed variables and was discussed in item 4 of sem: Choice of estimation method.
3. gsem by default is an equationwise deleter.
The abridged meaning is that gsem will often be able to use more observations from the
data than sem will, assuming you do not use sem with method MLMV.
The full meaning requires some setup. Consider a model of at least five equations. Assume
that observed exogenous variable x1 appears in the first equation but not in equations 24;
that equation 1 predicts y1; that y1 appears as a predictor in equations 24; and that x1
contains missing in observation 10.
If endogenous variable y1 is latent or observed and of family Gaussian, link identity, but
without censoring, then
3.1 Observation 10 will be ignored in making calculations related to equation 1.
3.2 Observation 10 will also be ignored in making calculations related to equations
24 because y1, a function of x1, appears in them.
3.3 The calculations for the other equation(s) will include observation 10.
Alternatively, if y1 is observed and not family Gaussian, link identity, or has censoring, then
item 3.2 changes:
3.2 Observation 10 will be used in making calculations related to equations 24 even
though y1, a function of x1, appears in them.
As we said at the outset, the result of all of this is that gsem often uses more observations
than does sem (excluding method MLMV).
4. gsem has an option, listwise, that duplicates the sem rules. This is used in testing of
Stata. There is no reason you would want to specify the option.

Variable types: Observed, latent, endogenous, exogenous, and error

Structural equation models can contain four different types of variables:
1. observed exogenous
2. latent exogenous
3. observed endogenous
4. latent endogenous
As a software matter, it is useful to think as though there is a fifth type, too:
5. error
Errors are in fact a special case of latent exogenous variables, but there will be good reason to
consider them separately.
As a language matter, it is sometimes useful to think of there being yet another type of variable,
namely,
6. measure or measurement
Measurement variables are a special case of observed endogenous variables.

intro 4 Substantive concepts

47

Let us explain:
Observed.
A variable is observed if it is a variable in your dataset. In this documentation, we often refer
to observed variables with x1, x2, . . . , y1, y2, and so on, but in reality observed variables have
names such as mpg, weight, testscore, and so on.
Latent.
A variable is latent if it is not observed. A variable is latent if it is not in your dataset but you
wish it were. You wish you had a variable recording the propensity to commit violent crime, or
socioeconomic status, or happiness, or true ability, or even income. Sometimes, latent variables
are imagined variants of real variables, variables that are somehow better, such as being measured
without error. At the other end of the spectrum are latent variables that are not even conceptually
measurable.
In this documentation, latent variables usually have names such as L1, L2, F1, . . . , but in real
life the names are more descriptive such as VerbalAbility, SES, and so on. The sem and gsem
commands assume that variables are latent if the first letter of the name is capitalized, so we will
always capitalize our latent variable names.
Endogenous.
A variable is endogenous (determined within the system) if any path points to it.
Exogenous.
A variable is exogenous (determined outside the system) if paths only originate from it or,
equivalently, no path points to it.
Now that we have the above definitions, we can better understand the five types of variables:
1. Observed exogenous.
A variable in your dataset that is treated as exogenous in your model.
2. Latent exogenous.
An unobserved variable that is treated as exogenous in your model.
3. Observed endogenous.
A variable in your dataset that is treated as endogenous in your model.
4. Latent endogenous.
An unobserved variable that is treated as endogenous in your model.
5. Error.
Mathematically, error variables are just latent exogenous variables. In sem and gsem, however,
errors are different in that they have defaults different from the other latent exogenous variables.
Errors are named e. So, for example, the error variable associated with observed endogenous
variable y1 has the full name e.y1; the error variable associated with latent endogenous
variable L1 has the full name e.L1.
In sem, each endogenous variable has a corresponding e. variable.
In gsem, observed endogenous variables associated with family Gaussian have corresponding
e. variables, but other observed endogenous variables do not. All latent endogenous variables
have an associated e. variable, but that is not a special case because all latent endogenous
variables are assumed to be family Gaussian.
In the Builder, when you create an endogenous variable, the variables corresponding error
variable instantly springs into existence. The same happens in the command language, you
just do not see it. In addition, error variables automatically and inalterably have their path
coefficient constrained to be 1.

48

intro 4 Substantive concepts

The covariances between the different variable types are given in the table below. In the table,
1. 0 means 0 and there are no options or secret tricks to change the value from 0.
2. i means as implied by your model and beyond your control.
3. (#) means to see the numbered note below the table.
If an entry in the matrix below is 0 or i, then you may not draw curved paths between variables of
the specified types.

Observed exogenous
Latent exogenous
Observed endogenous
Latent endogenous
Error

Observed
exogenous

Latent
exogenous

Observed
endogenous

Latent
endogenous

Error

(1) & (2)

(3)
i
i
0

(4) & (5)

i
i
0

i
i
i

i
i

(6) & (7)

1. Variances of observed exogenous variables:

1.1 Builder, sem mode: Taken as given, assuming not using method MLMV. Can
be estimated or constrained; if so, you are assuming normality of the observed
exogenous variables unless using method ADF.
1.2 sem command: Same as item 1.1. Use var() option to estimate or constrain.
1.3 Builder, gsem mode: Taken as given. Cannot be estimated or constrained.
1.4 gsem command: Same as item 1.3.
2. Covariances between observed exogenous variables:
2.1 Builder, sem mode: Taken as given, assuming not using method MLMV. Can be
estimated or constrained by drawing curved paths between variables; if so, you are
assuming normality of the observed exogenous variables unless using method ADF.
2.2 sem command: Same as item 2.1. Use cov() option to estimate or constrain.
2.3 Builder, gsem mode: Taken as given. Cannot be estimated or constrained. Path may
not be drawn.
2.4 gsem command: Same as item 2.3.
3. Covariances between latent exogenous and observed exogenous variables:
3.1 Builder, sem mode: Constrained to be 0 unless a curved path is drawn between
variables, in which case it is estimated. Can be constrained.
3.2 sem command: Assumed to be nonzero and therefore estimated by default. Can be
constrained (even to 0) using cov() option.
3.3 Builder, gsem mode: Assumed to be nonzero. Cannot be estimated or constrained
because this covariance is not among the identified parameters of the generalized
SEM.
3.4 gsem command: Same as item 3.3.
4. Variances of latent exogenous variables:
4.1 Builder, sem mode: Variances are estimated and can be constrained.
4.2 sem command: Variances are estimated and can be constrained using var() option.

intro 4 Substantive concepts

49

4.3 Builder, gsem mode: Almost the same as item 4.1 except variances cannot be
constrained to 0.
4.4 gsem command: Almost the same as item 4.2 except variances cannot be constrained
to 0.
5. Covariances between latent exogenous variables:
5.1 Builder, sem mode: Assumed to be 0 unless a curved path is drawn between
variables. Path may include constraints.
5.2 sem command: Assumed to be nonzero and estimated, the same as if a curved
path without a constraint were drawn in the Builder. Can be constrained (even to
0) using cov() option.
5.3 Builder, gsem mode: Same as item 5.1.
5.4 gsem command: Same as item 5.2.
6. Variances of errors:
6.1 Builder, sem mode: Estimated. Can be constrained.
6.2 sem command: Estimated. Can be constrained using var() option.
6.3 Builder, gsem mode: Almost the same as item 6.1 except variances cannot be
constrained to 0.
6.4 gsem command: Almost the same as item 6.2 except variances cannot be constrained
to 0.
7. Covariances between errors:
7.1 Builder, sem mode: Assumed to be 0. Can be estimated by drawing curved paths
between variables. Can be constrained.
7.2 sem command: Assumed to be 0. Can be estimated or constrained using cov()
option.
7.3 Builder, gsem mode: Almost the same as item 7.1 except covariances between errors
cannot be estimated or constrained if one or both of the error terms correspond to a
generalized response with family Gaussian, link log, or link identity with censoring.
7.4 gsem command: Almost the same as item 7.2 except covariances between errors
cannot be estimated or constrained if one or both of the error terms correspond to a
generalized response with family Gaussian, link log, or link identity with censoring.
Finally, there is a sixth variable type that we sometimes find convenient to talk about:
Measure or measurement.
A measure variable is an observed endogenous variable with a path from a latent variable. We
introduce the word measure not as a computer term or even a formal modeling term but as a
convenience for communication. It is a lot easier to say that x1 is a measure of X than to say that
x1 is an observed endogenous variable with a path from latent variable X and so, in a real sense,
x1 is a measurement of X.

50

intro 4 Substantive concepts

In our measurement model,

x1

x2

x3

x4

e.x1

e.x2

e.x3

e.x4

the variables are

latent exogenous:
X
error:
e.x1, e.x2, e.x3, e.x4
observed endogenous: x1, x2, x3, x4
All the observed endogenous variables in this model are measures of X.

Constraining parameters
Constraining path coefficients to specific values

If you wish to constrain a path coefficient to a specific value, you just write the value next to the
path. In our measurement model without correlation of the residuals,

x1

x2

x3

x4

e.x1

e.x2

e.x3

e.x4

we indicate that the coefficients e.x1, . . . , e.x4 are constrained to be 1 by placing a small 1 along
the path. We can similarly constrain any path in the model.

intro 4 Substantive concepts

51

If we wanted to constrain 2 = 1 in the equation

x2 = 2 + X2 + e.x2
we would write a 1 along the path between X and x2 . If we were instead using sems or gsems
command language, we would write
(x1<-X) (x2<-X@1) (x3<-X) (x4<-X)
That is, you type an @ symbol immediately after the variable whose coefficient is being constrained,
and then you type the value.

Constraining intercepts to specific values (suppressing the intercept)

Constraining path coefficients is common. Constraining intercepts is less so, and usually when
the situation arises, you wish to constrain the intercept to 0, which is often called suppressing the
intercept.
Although it is unusual to draw the paths corresponding to intercepts in path diagrams, they are
assumed, and you could draw them if you wish. A more explicit version of our path diagram for the
measurement model is

x1

x2

x3

_cons

The path coefficient of cons to x1 corresponds to 1 in

x1 = 1 + X1 + e.x1

x4

52

intro 4 Substantive concepts

and the path coefficient of cons to x2 corresponds to 2 in

x2 = 2 + X2 + e.x2
and so on.
Obviously, if you wanted to constrain a particular intercept to a particular value, you would write
the value along the path. To constrain 2 = 0, you could draw

x1

x2

_cons

x3

x4

intro 4 Substantive concepts

53

Because intercepts are assumed, you could omit drawing the paths from cons to x1, cons to
x3, and cons to x4:

x1

x2

x3

x4

_cons
Just as with the Builder, the command language assumes paths from
variables, but you could type them if you wished:

cons to all endogenous

(x1<-X _cons) (x2<-X _cons) (x3<-X _cons) (x4<-X _cons)

If you wanted to constrain 2 = 0, you could type
(x1<-X _cons) (x2<-X _cons@0) (x3<-X _cons) (x4<-X _cons)
or you could type
(x1<-X) (x2<-X _cons@0) (x3<-X) (x4<-X)

54

intro 4 Substantive concepts

Constraining path coefficients or intercepts to be equal

If you wish to constrain two or more path coefficients to be equal, place a symbolic name along
the relevant paths:

X
myb

myb

x1

x2

x3

x4

In the diagram above, we constrain 2 = 3 because we stated that 2 = myb and 3 = myb.
You follow the same approach in the command language:
(x1<-X) (x2<-X@myb) (x3<-X@myb) (x4<-X)
This works the same way with intercepts. Intercepts are just paths from cons, so to constrain
intercepts to be equal, you add symbolic names to their paths. In the command language, you constrain
1 = 2 by typing
(x1<-X _cons@c) (x2<-X _cons@c) (x3<-X) (x4<-X)
See [SEM] example 8.

Constraining covariances to be equal (or to specific values)

If you wish to constrain covariances, usually you will want to constrain them to be equal instead of
to a specific value. If we wanted to fit our measurement model and allow correlation between e.x2
and e.x3 and between e.x3 and e.x4, and we wanted to constrain the covariances to be equal, we
could draw

intro 4 Substantive concepts

55

x1

x2

x3

x4

myc

myc

If you instead wanted to constrain the covariances to specific values, you would place the value
along the paths in place of the symbolic names.
In the command language, covariances (curved paths) are specified using the cov() option. To
allow covariances between e.x2 and e.x3 and between e.x3 and e.x4, you would type
(x1<-X) (x2<-X) (x3<-X) (x4<-X), cov(e.x2*e.x3) cov(e.x3*e.x4)
To constrain the covariances to be equal, you would type
(x1<-X) (x2<-X) (x3<-X) (x4<-X), cov(e.x2*e.x3@myc) cov(e.x3*e.x4@myc)

Constraining variances to specific values (or to be equal)

Variances are like covariances except that in path diagrams drawn by some authors, variances curve
back on themselves. In the Builder, variances appear inside or beside the box or circle. Regardless of
how they appear, variances may be constrained to normalize latent variables, although normalization
is handled by sem and gsem automatically (something we will explain in How sem (gsem) solves
the problem for you under Identification 2: Normalization constraints (anchoring) below).
In the Builder, you constrain variances by clicking on the variable and using the lock box to specify
the value, which can be a number or a symbol. In the command language, variances are specified
using the var() option as we will explain below.

56

intro 4 Substantive concepts

Lets assume that you want to normalize the latent variable X by constraining its variances to be 1.
You do that by drawing

x1

x2

x3

x4

e.x1

e.x2

e.x3

e.x4

In the command language, we specify this model as

(x1<-X) (x2<-X) (x3<-X) (x4<-X), var(X@1)
Constraining latent exogenous variables to have unit variance as an identifying restriction may be
desirable when you wish simultaneously to constrain their correlations with other latent exogenous
variables. sem and gsem allow you to constrain covariances, not correlations. Covariances are equal
to correlations when variances are 1.
In more complicated models, you may wish to constrain variances of latent exogenous variables
to be equal. You can do that by specifying a symbolic name.

Identification 1: Substantive issues

Not all models are identified

Just because you can draw the path diagram for a model, write its equations, or write it in
Statas command syntax, does not mean the model is identified. Identification refers to the conceptual
constraints on parameters of a model that are required for the models remaining parameters to
have a unique solution. A model is said to be unidentified if these constraints are not supplied.
These constraints are of two types: substantive constraints and normalization constraints. We will
begin by discussing substantive constraints because that is your responsibility; the software provides
normalization constraints automatically.
How to count parameters

If your model has K observed variables, then your data contain K(K + 1)/2 second-order moments,
and thus p, the number of parameters based on second-order moments that can be estimated, cannot
exceed K(K + 1)/2.
Every path in your model contributes 1 to p unless the parameter is constrained to a specific
value, and then it does not contribute at all. If two parameters are constrained to be equal, the two
parameters count as one. In counting p, you must remember to count the curved paths from latent
variables back to themselves, which is to say, the variances. Just counting the number of parameters
can be challenging. And even if p K(K + 1)/2, your model may not be identified. Identification
depends not only on the number of paths but also on their locations.

intro 4 Substantive concepts

57

Counting parameters can be even more difficult in the case of certain generalized linear (gsem)
models. For a discussion of this, see Skrondal and Rabe-Hesketh (2004, chap. 5).
Even in the non-gsem case, books have been written on this subject, and we will refer you to
them. A few are Bollen (1989), Brown (2006), Kline (2011), and Kenny (1979). We will refer you
to them, but do not be surprised if they refer you back to us. Brown (2006, 202) writes, Because
latent variable software programs are capable of evaluating whether a given model is identified, it is
often most practical to simply try to estimate the solution and let the computer determine the models
identification status. That is not bad advice.
What happens when models are unidentified

So what happens when you attempt to fit an unidentified model? In some cases, sem (gsem) will
tell you that your model is unidentified. If your model is unidentified for subtle substantive reasons,
however, you will see
initial values not feasible
r(1400);

or
Iteration
Iteration
Iteration
.
.
.
Iteration
.
.
.

50:
51:
52:

101:

log likelihood = -337504.44

log likelihood = -337504.44
log likelihood = -337504.44

log likelihood = -337504.44

(not concave)
(not concave)
(not concave)

(not concave)

In the latter case, sem (gsem) will iterate forever, reporting the same criterion value (such as log
likelihood) and saying not concave over and over again.
Observing periods of the not concave message is not concerning, so do not overreact at the first
occurrence. Become concerned when you see not concave and the criterion value is not changing,
and even then, stay calm for a short time because the value might be changing in digits you are not
seeing. If the iteration log continues to report the same value several times, however, press Break.
Your model is probably not identified.

How to diagnose and fix the problem

You must find and fix the problem.

If you get the initial values not feasible message, your goal is to find feasible initial values
and thereby either solve the problem or convert it into the infinitely long iteration log problem; see
[SEM] intro 12.
If you get the infinitely long log, rerun the model specifying sems (gsems) iterate(#) option,
where # is large enough to reach the not concave messages with constant criterion value. sem
(gsem) will iterate that many times and then report the results it has at that point. Look at the output
for missing standard errors. Those parameters are unidentified, and you need to think about either
changing your model so that they become identified or placing constraints on them.

58

intro 4 Substantive concepts

Identification 2: Normalization constraints (anchoring)

Normalization constraints (anchoring) are provided automatically by sem and gsem. It is rare
indeed that the automatically provided constraints will be the source of convergence problems.
Models with latent variables require normalization constraints because latent variables have no
natural scale. If constraints were not provided, the model would appear to the software the same as a
model with a substantive lack of identification; the estimation routine would iterate forever and never
arrive at a solution.
The sem and gsem commands automatically provide normalization constraints to prevent that from
happening.
Below we explain why the normalization constraints are required, which normalization constraints
sem and gsem automatically supply, how to override those automatic choices, and how to substitute
your own constraints should you desire.
Why the problem arises

Imagine a latent variable for propensity to be violent. Your imagination might supply a scale that
ranges from 0 to 1 or 1 to 100 or over other values, but regardless, the scale you imagine is arbitrary
in that one scale works as well as another.
Scales have two components: mean and variance. If you imagine a latent variable with mean 0
and your colleague imagines the same variable with mean 100, the difference can be accommodated
in the parameter estimates by an intercept changing by 100. If you imagine a standard deviation of
1 (variance 12 = 1) and your colleague imagines a standard deviation of 10 (variance 102 = 100),
the difference can be accommodated by a path coefficient differing by a multiplicative factor of 10.
You might measure an effect as being 1.1 and then your colleague would measure the same effect as
being 0.11, but either way you both will come to the same substantive conclusions.
How the problem would manifest itself

The problem is that different scales all work equally well, and the software will iterate forever,
jumping from one scale to another.
Another way of saying that the means and variances of latent variables are arbitrary is to say that
they are unidentified. Thats important because if you do not specify the scale you have in mind,
results of estimation will look just like substantive lack of identification.
sem (gsem) will iterate forever and never arrive at a solution.
How sem (gsem) solves the problem for you

You usually do not need to worry about this problem because sem (gsem) solves it for you. sem
(gsem) solves the unidentified scale problem by
1. Assuming that all latent exogenous variables have mean 0.
2. Assuming that all latent endogenous variables have intercept 0.
3. Setting the coefficients on paths from latent variables to the first observed endogenous
variable to be 1.
4. Setting the coefficients on paths from latent variables to the first latent endogenous variable
to be 1 if rule 3 does not applyif the latent variable is measured by other latent variables
only.

intro 4 Substantive concepts

59

Rules 3 and 4 are also known as the unit-loading rules. The variable to which the path coefficient
is set to 1 is said to be the anchor for the latent variable.
Applying those rules to our measurement model, when we type
(X->x1) (X->x2) (X->x3) (X->x4)
sem (gsem) acts as if we typed
(X@1->x1) (X->x2) (X->x3) (X->x4), means(X@0)
The above four rules are sufficient to provide a scale for latent variables for all models.

Overriding sems (gsems) solution

sem (gsem) automatically applies rules 1 through 4 to produce normalization constraints. There
are, however, other normalization constraints that would work as well. In what follows, we will
assume that you are well versed in deriving normalization constraints and just want to know how to
bend sem (gsem) to your will.
Before you do this, however, let us warn you that substituting your normalization rules for the
defaults can result in more iterations being required to fit your model. Yes, one set of normalization
constraints are as good as the next, but sems (gsem)s starting values are based on its default
normalization rules, which means that when you substitute your rules for the defaults, the required
number of iterations sometimes increases.
Lets return to the measurement model:
(X->x1) (X->x2) (X->x3) (X->x4)
As we said previously, type the above and sem (gsem) acts as if you typed
(X@1->x1) (X->x2) (X->x3) (X->x4), means(X@0)
If you wanted to assume instead that the mean of X is 100, you could type
(X->x1) (X->x2) (X->x3) (X->x4), means(X@100)
The means() option allows you to specify mean constraints, and you may do so for latent or observed
variables.
Lets leave the mean at 0 and specify that we instead want to constrain the second path coefficient
to be 1:
(X->x1) (X@1->x2) (X->x3) (X->x4)
We did not have to tell sem (gsem) not to constrain X->x1 to have coefficient 1. We just specified
that we wanted to constrain X->x2 to have coefficient 1. sem (gsem) takes all the constraints that
you specify and then adds whatever normalization constraints are needed to identify the model. If
what you have specified is sufficient, sem (gsem) does not add its constraints to yours.
Obviously, if we wanted to constrain the mean to be 100 and the second rather than the first path
coefficient to be 1, we would type
(X->x1) (X@1->x2) (X->x3) (X->x4), means(X@100)

60

intro 4 Substantive concepts

If we wanted to constrain the standard deviation of X to be 10 instead of constraining a path

coefficient to be 1, we would type
(X->x1) (X->x2) (X->x3) (X->x4), means(X@100) var(X@100)
Standard deviations are specified in variance units when you use the var() option.

References
Acock, A. C. 2013. Discovering Structural Equation Modeling Using Stata. Rev. ed. College Station, TX: Stata Press.
Bollen, K. A. 1989. Structural Equations with Latent Variables. New York: Wiley.
Brown, T. A. 2006. Confirmatory Factor Analysis for Applied Research. New York: Guilford Press.
Kenny, D. A. 1979. Correlation and Causality. New York: Wiley.
Kline, R. B. 2011. Principles and Practice of Structural Equation Modeling. 3rd ed. New York: Guilford Press.
Li, C. 2013. Littles test of missing completely at random. Stata Journal 13: 795809.
Skrondal, A., and S. Rabe-Hesketh. 2004. Generalized Latent Variable Modeling: Multilevel, Longitudinal, and
Structural Equation Models. Boca Raton, FL: Chapman & Hall/CRC.

Also see
[SEM] intro 3 Learning the language: Factor-variable notation (gsem only)
[SEM] intro 5 Tour of models
[SEM] sem and gsem path notation Command syntax for path diagrams
[SEM] sem and gsem option covstructure( ) Specifying covariance restrictions

Title
intro 5 Tour of models

Description

Remarks and examples

References

Also see

Description
Below is a sampling of SEMs that can be fit by sem or gsem.

Remarks and examples

If you have not read [SEM] intro 2, please do so. You need to speak the language. We also
recommend reading [SEM] intro 4, but that is not required.
Now that you speak the language, we can start all over again and take a look at some of the
classic models that sem and gsem can fit.
Remarks are presented under the following headings:
Single-factor measurement models
Itemresponse theory (IRT) models
Multiple-factor measurement models
Confirmatory factor analysis (CFA) models
Structural models 1: Linear regression
Structural models 2: Gamma regression
Structural models 3: Binary-outcome models
Structural models 4: Count models
Structural models 5: Ordinal models
Structural models 6: Multinomial logistic regression
Structural models 7: Dependencies between response variables
Structural models 8: Unobserved inputs, outputs, or both
Structural models 9: MIMIC models
Structural models 10: Seemingly unrelated regression (SUR)
Structural models 11: Multivariate regression
Structural models 12: Mediation models
Correlations
Higher-order CFA models
Correlated uniqueness model
Latent growth models
Models with reliability
Multilevel mixed-effects models

61

62

intro 5 Tour of models

Single-factor measurement models

A single-factor measurement model is

x1

x2

x3

x4

e.x1

e.x2

e.x3

e.x4

The model can be written in Stata command language as

(x1<-X) (x2<-X) (x3<-X) (x4<-X)
or as
(x1 x2 x3 x4<-X)
or as
(X->x1 x2 x3 x4)
or in other ways. All the equivalent ways really are equivalent; no subtle differences will subsequently
arise according to your choice.
The measurement model plays an important role in many other SEMs dealing with the observed
inputs and the observed outputs:

y1

y2

y3

...

Theoretical model stated

in terms of underlying
concepts (latent variables)

x1

x2

x3

...

intro 5 Tour of models

63

Because the measurement model is so often joined with other models, it is common to refer to the
coefficients on the paths from latent variables to observable endogenous variables as the measurement
coefficients and to refer to their intercepts as the measurement intercepts. The intercepts are usually not
shown in path diagrams. The other coefficients and intercepts are those not related to the measurement
issue.
The measurement coefficients are often referred to as loadings.
This model can be fit by sem or gsem. Use sem for standard linear models (standard means
single level); use gsem when you are fitting a multilevel model or when the response variables are
generalized linear such as probit, logit, multinomial logit, Poisson, and so on.
See the following examples:
1. [SEM] example 1. Single-factor measurement model.
2. [SEM] example 27g. Single-factor measurement model (generalized response).
3. [SEM] example 30g. Two-level measurement model (multilevel, generalized response).
4. [SEM] example 35g. Ordered probit and ordered logit.

Itemresponse theory (IRT) models

Itemresponse theory (IRT) models look like the following:

Bernoulli

Bernoulli

Bernoulli

Bernoulli

item1

item2

item3

item4

logit

logit

logit

logit

The items are the observed variables, and each has a 0/1 outcome measuring a latent variable. Often,
the latent variable represents ability. These days, it is traditional to fit IRT models using logistic
regression, but in the past, probit was used and they were called normal ogive models.
In one-parameter logistic models, also known as 1-PL models and Rasch models, constraints
are placed on the paths and perhaps the variance of the latent variable. Either path coefficients are
constrained to 1 or path coefficients are constrained to be equal and the variance of the latent variable
is constrained to be 1. Either way, this results in the negative of the intercepts of the fitted model
being a measure of difficulty.
1-PL and Rasch models can be fit treating the latent variableabilityas either fixed or random.
Abilities are treated as random with gsem.
In two-parameter logistic models (2-PL), no constraints are imposed beyond the one required to
identify the latent variable, which is usually done by constraining the variance to 1. This results
in path coefficients measuring discriminating ability of the items, and difficulty is measured by the
negative of the intercepts divided by the corresponding (slope) coefficient.

64

intro 5 Tour of models

IRT has been extended beyond 1-PL and 2-PL models, including extension to other types of
generalized responses.
IRT models, including the extensions, can be fit by gsem.

See the following examples:

1. [SEM] example 28g. One-parameter logistic IRT (Rasch) model.
2. [SEM] example 29g. Two-parameter logistic IRT model.

Multiple-factor measurement models

A two-factor measurement model is two one-factor measurement models with possible correlation
between the factors:

F1

F2

x1

x2

x3

x4

x5

x6

To obtain a correlation between F1 and F2, we drew a curved path.

The model can be written in Stata command language as
(F1->x1) (F1->x2) (F1->x3) (F2->x4) (F2->x5) (F2->x6)
In the command language, we do not have to include the cov(F1*F2) option because, by default,
sem assumes that exogenous latent variables are correlated with each other. This model can also be
written in any of the following ways:
(F1->x1 x2 x3) (F2->x4 x5 x6)
or
(x1 x2 x3<-F1) (x4 x5 x6<-F2)
or
(x1<-F1) (x2<-F1) (x3<-F1) (x4<-F2) (x5<-F2) (x6<-F2)
Two-factor measurement models can be fit by sem and gsem.
See the following examples:
1. [SEM] example 3. Two-factor measurement model.
2. [SEM] example 31g. Two-factor measurement model (generalized response).

intro 5 Tour of models

65

Confirmatory factor analysis (CFA) models

The measurement models just shown are also known as confirmatory factor analysis (CFA) models
because they can be analyzed using CFA.
In the single-factor model, after estimation, you might want to test that all the indicators have
significant loadings by using test; see [SEM] test. You might also want to test whether the correlations
between the errors should have been included in the model by using estat mindices; see [SEM] estat
mindices.
In the multiple-factor measurement model, you might want to test that any of the omitted paths
should in fact be included in the model. The omitted paths in the two-factor measurement model
above were F1 x4, F1 x5, F1 x6, and F2 x1, F2 x2, F2 x3. estat mindices
will perform these tests.
These tests are available after sem only. CFA is just a measurement model and can be fit by both
sem and gsem.
We show other types of CFA models below.
See the following example:
1. [SEM] example 5. Modification indices.

Structural models 1: Linear regression

Different authors define the meaning of structural models in different ways. Bollen (1989, 4)
defines a structural model as the parameters being not of a descriptive nature of association but instead
of a causal nature. By that definition, the measurement models above could be structural models, and
so could the linear regression below.
Others define structural models as models having paths reflecting causal dependencies between
endogenous variables, which would thus exclude the measurement model and linear regression. We
will show you a true structural model in the next example.
An example of a linear regression is

x1
x2

x3
This model can be written in Stata command language as
(y<-x1 x2 x3)
When you estimate a linear regression by using sem, you obtain the same point estimates as you
would with regress and the same standard errors up to a degree-of-freedom adjustment applied by
regress.

66

intro 5 Tour of models

Linear regression models can be fit by sem and gsem. gsem also has options for censoring.
See the following examples:
1. [SEM] example 6. Linear regression.
2. [SEM] example 38g. Random-intercept and random-slope models (multilevel).
3. [SEM] example 40g. Crossed models (multilevel).
4. [SEM] example 43g. Tobit regression.
5. [SEM] example 44g. Interval regression.

Structural models 2: Gamma regression

Gamma regression, also known as log-Gamma regression, is used when a continuous outcome
is nonnegative, when it ranges from zero to infinity, and often with positively skewed data. It is
appropriate when the error variance can be assumed to increase with the mean. Gamma regression
gives similar results to linear regression with a logged dependent variable.

x1
gamma

x2

y
log

x3

Gamma regression is fit by gsem; specify shorthand gamma or specify family(gamma) link(log).
You can fit exponential regressions using Gamma regression if you constrain the log of the scale
parameter to be 0; see [SEM] gsem family-and-link options.

Structural models 3: Binary-outcome models

Binary-outcome models have 0/1 response variables. These models include logistic regression (also
known as logit), probit, and complementary log-log (also known as cloglog) models.
A simple logistic regression model is

x1
Bernoulli

x2

y
logit

x3

intro 5 Tour of models

67

which in command syntax can be written as

(y<-x1 x2 x3, logit)
For the other binary-outcome models, all that changes in the diagram is the names of the family and
link; in the command language, the option name changes; see [SEM] gsem family-and-link options.
Usually, the observations in binary-outcome data record whether the event occurred, but the data
can instead record the number of events and the number of trials by changing the family from
Bernoulli to binomial; see [SEM] gsem family-and-link options.
Binary-outcome models can be fit by gsem.
See the following examples:
1. [SEM] example 33g. Logistic regression.
2. [SEM] example 27g. Single-factor measurement model (generalized response).
3. [SEM] example 34g. Combined models (generalized responses).

Structural models 4: Count models

Count models have response variables that are counts of things or events. These models include
Poisson and negative binomial models. A simple example of a Poisson model is

x1
Poisson

x2

y
log

x3

which in command syntax can be written as

(y<-x1 x2 x3, poisson)
For negative binomial models, in path diagrams, the family changes to nbreg and, in the command
syntax, the option changes to nbreg.
In Poisson models, both the mean and the variance are determined by a single parameter: the rate
at which the event occurs. One use of negative binomial models is to handle overdispersion, which
is when the variance is greater than the variance that would be predicted by a Poisson model.
The way we have diagrammed the model and written the command, we are assuming that each
observation was at risk for the same length of time. If observations varied in this, the command
would be
(y<-x1 x2 x3, poisson exposure(etime))
where variable etime records each observations exposure time. The diagram would not change, but
we would nonetheless enter the exposure time into the Builder. See [SEM] gsem family-and-link
options.

68

intro 5 Tour of models

Count models are fit by gsem.

See the following examples:
1. [SEM] example 34g. Combined models (generalized responses).
2. [SEM] example 39g. Three-level model (multilevel, generalized response).

Structural models 5: Ordinal models

Ordinal models have two or more possible outcomes that are ordered, such as responses of the
form a little, average, and a lot; or strongly disagree, disagree, . . . , strongly agree. The
outcomes are usually numbered 1, 2, . . . , k . These models include ordered probit, ordered logit, and
ordered complementary log-log (also known as ordered cloglog). A simple example of an ordered
probit model is

x1
ordinal

x2

y
probit

x3

which in command syntax can be written as

(y<-x1 x2 x3, oprobit)
All that changes for the other models is the link name that appears in the path diagram or the option
that appears in the command language.
Ordinal models are fit by gsem.
See the following examples:
1. [SEM] example 35g. Ordered probit and ordered logit.
2. [SEM] example 31g. Two-factor measurement model (generalized response).
3. [SEM] example 32g. Full structural equation model (generalized response).
4. [SEM] example 36g. MIMIC model (generalized response).

Structural models 6: Multinomial logistic regression

Multinomial logistic regression, also known as multinomial logit, is similar to ordinal models in
that it, too, deals with multiple responses; however, in multinomial logistic regression, the responses
cannot be ordered. Examples of multinomial logit response include method of transportation (car,
public transportation, etc.) or ice-cream flavor (vanilla, chocolate, etc.).
Just as with ordinal models, the outcome is usually recorded as a single variable containing 1, 2,
. . . , k , but in path diagrams, factor-variable notation is used to identify the outcomes:

intro 5 Tour of models

69

multinomial

1b.y
logit

x1
multinomial

2.y
logit

x2
multinomial

3.y
logit

In command syntax, this can be written as

(i.y<-x1 x2, mlogit)
In the example above, we have paths from all predictor variables to all outcomes. That is common
but not required.
The multinomial logistic regression model is fit by gsem.
See the following examples:
1. [SEM] example 37g. Multinomial logistic regression.
2. [SEM] example 41g. Two-level multinomial logistic regression (multilevel).

Structural models 7: Dependencies between response variables

An example of a structural model having paths between response (endogenous) variables is

x1

y1

y2

x2

x3

This model can be written in Stata command language as

(y1<-x1 x2 x3 x4) (y2<-y1 x2 x3)

x4

70

intro 5 Tour of models

In this example, all inputs and outputs are observed and the errors are assumed to be uncorrelated.
In these kinds of models, it is common to allow correlation between errors:

x1

y1

y2

x2

x3

x4

The model above can be written in Stata command language as

(y1<-x1 x2 x3 x4) (y2<-y1 x2 x3), cov(e.y1*e.y2)
This structural model is said to be overidentified. If we omitted y1 x4, the model would be
just-identified. If we also omitted y1 x1, the model would be unidentified.
When you fit the above model using sem, you obtain slightly different results from those you would
obtain with ivregress liml. This is because sem with default method(ml) produces full-information
maximum likelihood rather than limited-information maximum likelihood results.
Analysis of models for observed variables that include dependencies between endogenous variables
may also be referred to as path analysis. Acock (2013, chap. 2) discusses path analysis with sem in
more detail.
These models can be fit by sem or gsem. When using gsem to fit models with generalized response
variables, non-Gaussian responses and Gaussian responses with the log link, or censoring, can only
be included in recursive portions of models.
See the following examples:
1. [SEM] example 7. Nonrecursive structural model.
2. [SEM] example 34g. Combined models (generalized responses).
3. [SEM] example 42g. One- and two-level mediation models (multilevel).
4. [SEM] example 43g. Tobit regression.
5. [SEM] example 44g. Interval regression.
6. [SEM] example 45g. Heckman selection model.
7. [SEM] example 46g. Endogenous treatment-effects model.

intro 5 Tour of models

71

Structural models 8: Unobserved inputs, outputs, or both

Perhaps in a structural model such as

y1

y2

x1

x2

x3

the inputs x1, x2, and x3 are concepts and thus are not observed. Assume that we have measurements
for them. We can join this structural model example with a three-factor measurement model:

10

y1

y2

X1

X3

X2

z1

z2

z3

z4

z5

z6

z7

z8

Note the curved arrows denoting correlation between the pairs of X1, X2, and X3. In the previous
path diagram, we had no such arrows between the variables, yet we were still assuming that they
were there. In sems path diagrams, correlation between exogenous observed variables is assumed
and need not be explicitly shown. When we changed observed variables x1, x2, and x3 to be the
latent variables X1, X2, and X3, we needed to show explicitly the correlations we were allowing.
Correlation between latent variables is not assumed and must be shown.

72

intro 5 Tour of models

This model can be written in Stata command syntax as follows:

(y1<-X1
(X1->z1
(X2->z4
(X3->z6

X2) (y2<-y1 X2 X3)

z2 z3)
z5)
z7 z8),
cov(e.y1*e.y2)

///
///
///
///

We did not include the cov(X1*X2 X1*X3 X2*X3) option, although we could have. In the
command language, exogenous latent variables are assumed to be correlated with each other. If we
did not want X2 and X3 to be correlated, we would need to include the cov(X2*X3@0) option.
We changed x1, x2, and x3 to be X1, X2, and X3. In command syntax, variables beginning with a
capital letter are assumed to be latent. Alternatively, we could have left the names in lowercase and
specified the identities of the latent variables:
(y1<-x1
(x1->z1
(x2->z4
(x3->z6

x2) (y2<-y1 x2 x3)

z2 z3)
z5)
z7 z8),
cov(e.y1*e.y2)
latent(x1 x2 x3)

///
///
///
///
///

Just as we have joined an observed structural model to a measurement model to handle unobserved
inputs, we could join the above model to a measurement model to handle unobserved y1 and y2.
Models with unobserved inputs, outputs, or both can be fit by sem and gsem.
See the following examples:
1. [SEM] example 9. Structural model with measurement component.
2. [SEM] example 32g. Full structural equation model (generalized response).
4. [SEM] example 45g. Heckman selection model.
5. [SEM] example 46g. Endogenous treatment-effects model.

Structural models 9: MIMIC models

MIMIC stands for multiple indicators and multiple causes. An example of a MIMIC model is

c1
c2

c3

i1

i2

i3

In this model, the observed causes c1, c2, and c3 determine latent variable L, and L in turn determines
the observed indicators i1, i2, and i3.

intro 5 Tour of models

73

This model can be written in Stata command syntax as

(i1 i2 i3<-L) (L<-c1 c2 c3)
MIMIC models can be fit by sem and gsem.

See the following examples:

1. [SEM] example 10. MIMIC model.
2. [SEM] example 36g. MIMIC model (generalized response).

Structural models 10: Seemingly unrelated regression (SUR)

Seemingly unrelated regression (SUR) is like having two or more separate linear regressions but
allowing the errors to be correlated.
An example of an SUR model is

x1

y1

y2

x2

x3

x4

The model above can be written in Stata command syntax as

(y1<-x1 x2 x3) (y2<-x3 x4), cov(e.y1*e.y2)
In this example, the two regressions share a common exogenous variable, x3. They do not have
to share a common variable, or they could share more variables. If they shared all variables, results
would be the same as estimating multivariate regression, shown in the next example.
When you estimate an SUR with sem, you obtain the same point estimates as you would with
sureg if you specify suregs isure option, which causes sureg to iterate until it obtains the
maximum likelihood result. Standard errors will be different. If the model has exogenous variables
only on the right-hand side, then standard errors will be asymptotically identical and, although the
standard errors are different in finite samples, there is no reason to prefer one set over the other. If
the model being fit is recursive, standard errors produced by sem are better than those from sureg,
both asymptotically and in finite samples.
SUR models can be fit by sem and gsem. gsem will allow you to generalize the model to multilevel.
In cases other than family Gaussian, link identity, you can sometimes introduce latent variables to
create a similar effect.

See the following example:

1. [SEM] example 12. Seemingly unrelated regression.

74

intro 5 Tour of models

Structural models 11: Multivariate regression

A multivariate regression is just an SUR where the different dependent variables share the same
exogenous variables:

y1

y2

x1

x2

x3

The model above can be written in Stata command syntax as

(y1 y2<-x1 x2 x3), cov(e.y1*e.y2)
When you estimate a multivariate regression with sem, you obtain the same
p point estimates
as you would with mvreg and the same standard errors up to a multiplicative (N p 1)/N
degree-of-freedom adjustment applied by mvreg.
Multivariate regression can be fit by sem and gsem. gsem will allow you to generalize the model
to multilevel. In cases other than family Gaussian, link identity, you can sometimes introduce latent
variables to create a similar effect.
See the following example:
1. [SEM] example 12. Seemingly unrelated regression.

Structural models 12: Mediation models

Mediation models concern effects that pass through (are mediated by) other variables. Consider
response variable y affected by x mediated through m:

This can also be specified in command syntax as

(y<-m x) (m<-x)

intro 5 Tour of models

75

In this simple model, x has a direct effect on y and an indirect (mediated through m) effect. The
direct effect may be reasonable given the situation, or it may be included just so one can test whether
the direct effect is present. If both the direct and indirect effects are significant, the effect of x is
said to be partially mediated through m.
There are one-level mediation models and various two-level models, and lots of other variations,
too.
sem and gsem can both fit one-level linear models, but you will be better off using sem. gsem
can fit one-level generalized linear models and fit two-level (and higher) models, generalized linear
or not.
See the following example:
1. [SEM] example 42g. One- and two-level mediation models (multilevel).

Correlations
We are all familiar with correlation matrices of observed variables, such as
x1
1.0000
0.7700
0.0177

x1
x2
x3

x2

x3

1.0000
0.2229

1.0000

or covariances matrices, such as

x1
x2
x3

x1
662.172
62.5157
0.769312

x2

x3

9.95558
1.19118

2.86775

These results can be obtained from sem. The path diagram for the model is

x1

x2

x3

We could just as well leave off the curved paths because sem assumes them among observed exogenous
variables:

x1

x2

x3

Either way, this model can be written in Stata command syntax as

(<- x1 x2 x3)
That is, we simply omit specifying the target of the path, the endogenous variable.

76

intro 5 Tour of models

If we fit the model, we will obtain the covariance matrix by default. correlate with the
covariance option produces covariances that are divided by N 1 rather than by N . To match this
covariance exactly, you need to specify the nm1 option, which we can do in the command language
by typing
(<- x1 x2 x3), nm1
If we want correlations rather than covariances, we ask for them by specifying the standardized
option:
(<- x1 x2 x3), nm1 standardized
An advantage of obtaining correlation matrices from sem rather than from correlate is that you
can perform statistical tests on the results, such as that the correlation of x1 and x3 is equal to the
correlation of x2 and x3.
If you are willing to assume joint normality of the variables, you can obtain more efficient estimates
of the correlations in the presence of missing-at-random data by specifying the method(mlmv) option.
Correlations are fit using sem.
See the following example:
1. [SEM] example 16. Correlation.

Higher-order CFA models

Observed values sometimes measure traits or other aspects of latent variables, so we insert a
new layer of latent variables to reflect those traits or aspects. We have measurementssay, x1, . . . ,
x6all reflecting underlying factor F, but x1 and x2 measure one trait of F, x3 and x4 measure
another trait, and x5 and x6 measure yet another trait. This model can be drawn as

intro 5 Tour of models

77

x1

x2

x3

x4

x5

x6

The model can be written in command syntax as

(A->x1 x2) (B->x3 x4) (C->x5 x6) (A B C<-F)
Higher-order CFA models can be fit using sem or gsem.
See the following example:
1. [SEM] example 15. Higher-order CFA.

Correlated uniqueness model

Observed values sometimes are correlated just because of how the data are collected. Imagine
we have factor T1 representing a trait with measurements x1 and x4. Perhaps T1 is aggression, and
then x1 is self reported and x4 is reported by the spouse. Imagine we also have factor T2 with
measurements x2 and x5. Again, x2 is self reported and x5 is reported by the spouse. It would not
be unlikely that x1 and x2 are correlated and that x4 and x5 are correlated. That is exactly what the
correlated uniqueness model assumes:

78

intro 5 Tour of models

T1

T2

T3

x1

x2

x3

x4

x5

x6

x7

x8

x9

Data that exhibit this kind of pattern are known as multitraitmultimethod (MTMM) data. Researchers
historically looked at the correlations, but structural equation modeling allows us to fit a model that
incorporates the correlations.
The above model can be written in Stata command syntax as
(T1->x1 x4 x7)
(T2->x2 x5 x8)
(T3->x3 x6 x9),
cov(e.x1*e.x2 e.x1*e.x3 e.x2*e.x3)
cov(e.x4*e.x5 e.x4*e.x6 e.x5*e.x6)
cov(e.x7*e*x8 e.x7*e.x9 e.x8*e.x9)

///
///
///
///
///

An alternative way to type the above is to use the covstructure() option, which we can abbreviate
as covstruct():
(T1->x1 x4 x7)
(T2->x2 x5 x8)
(T3->x3 x6 x9),
covstruct(e.x1 e.x2 e.x3, unstructured)
covstruct(e.x4 e.x5 e.x6, unstructured)
covstruct(e.x7 e.x8 e.x9, unstructured)

///
///
///
///
///

Unstructured means that the listed variables have covariances. Specifying blocks of errors as unstructured would save typing if there were more variables in each block.
The correlated uniqueness model can be fit by sem or gsem, although we recommend use of sem in
this case. Gaussian responses with the identity link are allowed to have correlated uniqueness (error)
but only in the absence of censoring. gsem still provides the theoretical ability to fit these models in
multilevel contexts, but convergence may be difficult to achieve.
See the following example:
1. [SEM] example 17. Correlated uniqueness model.

intro 5 Tour of models

79

Latent growth models

A latent growth model is a variation on the measurement model. In our measurement model
examples, we have assumed four observed measurements of underlying factor X: x1, x2, x3, and x4.
In the command language, we can write this as
(X->x1) (X->x2) (X->x3) (X->x4)
Lets assume that the observed values are collected over time. x1 is observed at time 0, x2 at time
1, and so on. It thus may be more reasonable to assume that the observed values represent a base
value and a growth modeled with a linear trend. Thus we might write the model as
(B@1
(B@1
(B@1
(B@1

L@0->x1)
L@1->x2)
L@2->x3)
L@3->x4),
noconstant

///
///
///
///

The equations for this model are

x1 = B + 0L + e.x1
x2 = B + 1L + e.x2
x3 = B + 2L + e.x3
x4 = B + 3L + e.x4
and the path diagram is

1
1

L
0
1

3
2

x1

x2

x3

x4

In evaluating this model, it is useful to review the means of the latent exogenous variables. In most
models, latent exogenous variables have mean 0, and the means are thus uninteresting. sem usually
constrains latent exogenous variables to have mean 0 and does not report that fact.
In this case, however, we ourselves have placed constraints, and thus the means are identified and
in fact are an important point of the exercise. We must tell sem not to constrain the means of the
two latent exogenous variables B and L, which we do with the means() option:

80

intro 5 Tour of models

(B@1
(B@1
(B@1
(B@1

L@0->x1)
///
L@1->x2)
///
L@2->x3)
///
L@3->x4),
///
noconstant means(B L)

We must similarly specify the means() option when using the Builder.
Latent growth models can be fit with sem or gsem.
See the following example:
1. [SEM] example 18. Latent growth model.

Models with reliability

A typical solution for dealing with variables measured with error is to find multiple measurements
and use those measurements to develop a latent variable. See, for example, Single-factor measurement
models and Multiple-factor measurement models above.
When the reliability of the variables is knownreliability is measured as the fraction of variances
that is not due to measurement erroranother approach is available. This approach can be used
in place of or in addition to the use of multiple measurements. See [SEM] sem and gsem option
reliability( ).
Models with reliability can be fit with sem or gsem, although option reliability() is available
only when responses are Gaussian with the identity link and only in the absence of censoring.
See the following example:
1. [SEM] example 24. Reliability.

Multilevel mixed-effects models

Multilevel modeling concerns the inclusion of common, random effects across groups of the data,
which are known as levels. You have observations on students. Students attend schools. Or you have
observations on patients. Patients are served by hospitals. In either case, there may be an unmeasured
effect of the institution. Some schools are better than others while some schools are worse. The
same applies to hospitals. These unmeasured effects can be parameterized as a latent variable that is
constant within institution and varies across institution.
What we have just described is a two-level nested model. The first level, the lowest, is the
observational level. The second level is school or hospital.
In a three-level nested model, students are served by schools which are served by counties, or patients
are served by hospitals which are served by states. Whatever the example, there can be unmeasured
effects of each of those higher levels contributing to the effect observed at the observational level.
An alternative to nested models is crossed models. People with jobs work in an industry and in
a state. The same industries are found across states and the same states are found across industries,
yet both may have an effect on some aspect of the lives of their workers.

intro 5 Tour of models

81

Multilevel models are fit by gsem.

See the following examples:
1. [SEM] example 30g. Two-level measurement model (multilevel, generalized response).
2. [SEM] example 38g. Random-intercept and random-slope models (multilevel).
3. [SEM] example 39g. Three-level model (multilevel, generalized response).
4. [SEM] example 40g. Crossed models (multilevel).
5. [SEM] example 41g. Two-level multinomial logistic regression (multilevel).
6. [SEM] example 42g. One- and two-level mediation models (multilevel).

References
Acock, A. C. 2013. Discovering Structural Equation Modeling Using Stata. Rev. ed. College Station, TX: Stata Press.
Bauldry, S. 2014. miivfind: A command for identifying model-implied instrumental variables for structural equation
models in Stata. Stata Journal 14: 6075.
Bollen, K. A. 1989. Structural Equations with Latent Variables. New York: Wiley.

Also see
[SEM] intro 4 Substantive concepts
[SEM] intro 6 Comparing groups (sem only)
[SEM] example 1 Single-factor measurement model

Title
intro 6 Comparing groups (sem only)
Description

Remarks and examples

Reference

Also see

Description
sem has a unique feature not shared by gsem. You can easily compare groupscompare males
with females, compare age group 1 with age group 2 with age group 3, and so onwith respect to
any SEM. Said more technically, any model fit by sem can be simultaneously estimated for different
groups with some parameters constrained to be equal across groups and others allowed to vary, and
those estimates can be used to perform statistical tests for comparing the groups.

Remarks and examples

Remarks are presented under the following headings:
The generic SEM model
Fitting the model for different groups of the data
Which parameters vary by default, and which do not
Specifying which parameters are allowed to vary in broad, sweeping terms
Adding constraints for path coefficients across groups
Adding constraints for means, variances, or covariances across groups
Adding constraints for some groups but not others
Adding paths for some groups but not others
Relaxing constraints

The generic SEM model

In [SEM] intro 5, we noted that measurement models are often joined with other SEMs to produce

y1

y2

y3

...

Theoretical model stated

in terms of underlying
concepts (latent variables)

x1

x2

x3
82

...

intro 6 Comparing groups (sem only)

83

This can be written in the command syntax as

(Y1->...) (Y2->...)
(...)
(...)
(...)
(X1->...) (X2->...)

///
/// theoretical model stated in terms of
/// underlying concepts (latent variables)
///

where the middle part is the theoretical model stated in terms of underlying concepts Y1, Y2, X1, and
X2.
However we write the model, we are assuming the following:
1. The unobserved X1 and X2 are measured by the observed x1, x2, . . . .
2. The middle part is stated in terms of the underlying concepts X1, X2, Y1, and Y2.
3. The unobserved Y1 and Y2 are measured by the observed y1, y2, . . . .

Fitting the model for different groups of the data

We can fit this model for different groups (say, age groups) by specifying the group(varname)
option:
(Y1->...) (Y2->...)
(...)
(...)
(...)
(X1->...) (X2->...),
group(agegrp)

///
///
///
///
///

part 3
part 2
part 1

where agegrp is a variable in our dataset, perhaps taking on values 1, 2, 3, . . . . We can specify the
model by using the command language or by drawing the model in the Builder and then choosing
and filling in the group() option.
After estimation, you can use estat ginvariant (see [SEM] estat ginvariant) to obtain Wald
tests of whether constraints should be added and score tests of whether constraints should be relaxed.

Which parameters vary by default, and which do not

When we specify group(groupvar), the measurement parts of the model (parts 1 and 3) are
constrained by default to be the same across the groups, whereas the middle part (part 2) will have
separate parameters for each group. More specifically, parts 1 and 3 are constrained to be equal across
groups except that the variances of the errors will be estimated separately for each group.
If there is no measurement component to the modelif there are no latent variablesthen by
default all parameters are estimated separately for each group.

Specifying which parameters are allowed to vary in broad, sweeping terms

You can control which parameters are constrained to be equal across groups by specifying the
ginvariant() option:

84

intro 6 Comparing groups (sem only)

(Y1->...) (Y2->...)
///
part 3
(...)
///
(...)
///
part 2
(...)
///
(X1->...) (X2->...),
///
part 1
group(agegrp) ginvariant(classes)
The classes are as follows:
Class description

Class name

1. structural coefficients
2. structural intercepts

scoef
scons

3. measurement coefficients
4. measurement intercepts

mcoef
mcons

5. covariances of structural errors

6. covariances of measurement errors
7. covariances between structural and measurement errors

serrvar
merrvar
smerrcov

8. means of exogenous variables

9. covariances of exogenous variables

meanex
covex

(*)
(*)

10. all the above

all
(*)
11. none of the above
none
(*) Be aware that 8, 9, and 10 (meanex, covex, and all) exclude the
observed exogenous variablesthat is, they include only the latent
exogenous variablesunless you specify the noxconditional
option or the noxconditional option is otherwise implied; see
[SEM] sem option noxconditional. This is what you would desire
in most cases.
The default when ginvariant() is not specified is ginvariant(mcoef mcons):
(Y1->...) (Y2->...)
///
part 3, measurement
(...)
///
(...)
///
part 2, structural
(...)
///
(X1->...) (X2->...),
///
part 1, measurement
group(agegrp) ginvariant(mcoef mcons)
If you also wanted covariances of errors associated with measurement to be constrained across
groups, you could type
(Y1->...) (Y2->...)
///
part 3,
(...)
///
(...)
///
part 2,
(...)
///
(X1->...) (X2->...),
///
part 1,
group(agegrp) ginvariant(mcoef mcons

measurement
structural
measurement
merrvar)

intro 6 Comparing groups (sem only)

85

Adding constraints for path coefficients across groups

The ginvariant() option allows you to state in sweeping terms which parameters vary and
which are invariant across groups. You can also constrain individual parameters to be equal across
groups.
Pretend that in the substantive part of the generic model, we have Y1<-Y2. Assume that we fit
the model and allow the structural part to vary across groups:
(Y1->...) (Y2->...)
(...)
(Y1<-Y2)
(...)
(X1->...) (X2->...),
group(agegrp)

///
///
///
///
///

part 3, measurement
part 2, structural
part 1, measurement

In this model, the Y1<-Y2 path coefficient is allowed to vary across groups by default. We could
constrain the coefficient to be equal across groups by typing
(Y1->...) (Y2->...)
(...)
(Y1<-Y2@b)
(...)
(X1->...) (X2->...),
group(agegrp)

///
///
///
///
///

part 3, measurement
part 2, structural
part 1, measurement

Note the @b in (Y1<-Y2@b) where we previously typed Y1<-Y2.

Constraining a coefficient to equal a symbolic name such as b is how we usually constrain equality,
but in the usual case, the symbolic name appears at least twice in our model. For instance, we might
have (Y1<-Y2@b) and (Y1<-Y3@b) and thus constrain path coefficients to be equal.
In the case above, however, @b appears only once. Because we specified group(agegrp), results
are as if we specified this model separately for each age group, and in each group, we are specifying
@b. Thus we are constraining the path coefficient to be equal across all groups.

Adding constraints for means, variances, or covariances across groups

You use the same technique for adding constraints to means, variances, and covariances as you
would for adding constraints to path coefficients. Remember that means are specified by the means()
option, variances by the variance() option, and covariances by the covariance() option. The
variance() and covariance() options are abbreviated var() and cov(), respectively.
You can specify, for instance,
(Y1->...) (Y2->...)
(...)
(Y1<-Y2)
(...)
(X1->...) (X2->...),
group(agegrp)
means(X1@M)

///
///
///
///
///
///

part 3, measurement
part 2, structural
part 1, measurement

to constrain the mean of X1 to be the same across groups. The means would have been different
across groups by default.

86

intro 6 Comparing groups (sem only)

You can specify

(Y1->...) (Y2->...)
(...)
(Y1<-Y2)
(...)
(X1->...) (X2->...),
group(agegrp)
var(e.Y1@V)

///
///
///
///
///
///

part 3, measurement
part 2, structural
part 1, measurement

to constrain the variance of the error of Y1 to be the same across groups.

If we wanted to allow the errors of Y1 and Y2 to be correlated (by default, errors are uncorrelated),
we could add the cov() option:
(Y1->...) (Y2->...)
(...)
(...)
(...)
(X1->...) (X2->...),
group(agegrp)
var(e.Y1@V)
cov(e.Y1*e.Y2)

///
///
///
///
///
///
///

part 3, measurement
part 2, structural
part 1, measurement

If we then wanted to constrain the covariance to be the same across groups, we would type
(Y1->...) (Y2->...)
(...)
(...)
(...)
(X1->...) (X2->...),
group(agegrp)
var(e.Y1@V)
cov(e.Y1*e.Y2@C)

///
///
///
///
///
///
///

part 3, measurement
part 2, structural
part 1, measurement

Adding constraints for some groups but not others

Consider the following model:
... (Y1<-Y2) ..., group(agegrp)
Above we saw how to constrain the Y1<-Y2 path coefficient to be the same across groups:
... (Y1<-Y2@b) ..., group(agegrp)
To constrain the path coefficients Y1<-Y2 to be equal for groups 1 and 2 but leave the Y1<-Y2
path coefficients unconstrained for the remaining groups, we would type
... (Y1<-Y2) (1: Y1<-Y2@b) (2: Y1<-Y2@b) ..., group(agegrp)
Think of this as follows:
1. (Y1<-Y2): We set a path for all the groups.
2. (1: Y1<-Y2@b): We modify the path for agegrp = 1.
3. (2: Y1<-Y2@b): We modify the path for agegrp = 2.
4. We do not modify the path for any other agegrp value.

intro 6 Comparing groups (sem only)

87

The result is that we constrain age groups 1 and 2 to have the same value of the path, and we do
not constrain the path for the other age groups.
You can constrain variance and covariance estimates to be the same across some groups but not
others in the same way. You can specify, for instance,
..., group(agegrp) var(1: e.Y1@V) var(2: e.Y1@V)
or
..., group(agegrp) cov(e.Y1*e.Y2) cov(1: e.Y1*e.Y2@C) ///
cov(2: e.Y1*e.Y2@C)
Similarly, you can constrain means for some groups but not others, although this is rarely done:
..., group(agegrp) means(1: X1@b) means(2: X1@b)

Adding paths for some groups but not others

In the same way that you can constrain coefficients for some groups but not others, you can add
paths for some groups but not others. Consider the following model:
... (Y1<-Y2) ..., group(agegrp)
You can add the path Y1<-Y3 for groups 1 and 2 by typing
... (Y1<-Y2) (1: Y1<-Y3) (2: Y1<-Y3) ..., group(agegrp)
You can add covariances for some groups but not others in the same way. For instance, to allow
the errors of Y1 and Y2 to be correlated in groups 1 and 2 only, you can specify
..., group(agegrp) cov(1: e.Y1*e.Y2) cov(2: e.Y1*e.Y2)

Relaxing constraints
Just as you can specify
..., group(agegrp) ginvariant(classes)
and then add constraints, you can also specify
..., group(agegrp) ginvariant(classes)
and then relax constraints that the classes impose.
For instance, if we specified ginvariant(scoef), then we would be constraining (Y1<-Y2) to
be invariant across groups. We could then relax that constraint by typing
... (Y1<-Y2) (1: Y1<-Y2@b1) (2: Y1<-Y2@b2) ..., ///
group(agegrp) ginvariant(scoef)
The path coefficients would be free in groups 1 and 2 and constrained in the remaining groups, if
there are any. The path coefficient is free in group 1 because we specified symbolic name b1, and b1
appears nowhere else in the model. The path coefficient is free in group 2 because symbolic name b2
appears nowhere else in the model. If there are remaining groups and we want to relax the constraint
on them, too, we would need to add (3: Y1<-Y2@b3), and so on.

88

intro 6 Comparing groups (sem only)

The same technique can be used to relax constraints on means, variances, and covariances:
..., group(agegrp) ginvariant(... meanex ...) ///
means(1: X1@b1) means(2: X1@b2)
..., group(agegrp) ginvariant(... serrvar ...) ///
var(1: e.Y1@V1) var(2: e.Y1@V2)
..., group(agegrp) ginvariant(... serrvar ...) ///
cov(1: e.Y1*e.Y2@C1) cov(2: e.Y1*e.Y2@C2)

Reference
Acock, A. C. 2013. Discovering Structural Equation Modeling Using Stata. Rev. ed. College Station, TX: Stata Press.

Also see
[SEM] intro 5 Tour of models
[SEM] intro 7 Postestimation tests and predictions
[SEM] sem group options Fitting models on different groups
[SEM] sem and gsem option covstructure( ) Specifying covariance restrictions

Title
intro 7 Postestimation tests and predictions
Description

Remarks and examples

Also see

Description
After fitting a model with sem or gsem, you can perform statistical tests, obtain predicted values,
and more. Everything you can do is listed below.
sem and gsem vary in the tests and features available after estimation, and we mark whether each
test and feature is available after sem, gsem, or both.
Some tests and features after sem depend on the assumption of joint normality of the observed
variables and others do not. Whenever the joint-normality assumption is the required, we mention it
explicitly. If you do not believe your data meets the joint-normality assumption, you will want to
avoid those tests and features.
Some tests and features that are available after sem are not available after gsem because of the
joint-normality assumption. Others are not available for other reasons. If we do not mention the
joint-normality assumption, then that is not the cause. We usually leave unexplained the other reasons
that a test or feature might be inappropriate after gsem, because those reasons are not our focus here.

Remarks and examples

Remarks are presented under the following headings:
Replaying the model (sem and gsem)
Displaying odds ratios, incidence-rate ratios, etc. (gsem only)
Obtaining goodness-of-fit statistics (sem and gsem)
Performing tests for including omitted paths and relaxing constraints (sem only)
Performing tests of model simplification (sem and gsem)
Displaying other results, statistics, and tests (sem and gsem)
Obtaining predicted values (sem)
Obtaining predicted values (gsem)
Using contrast, pwcompare, and margins (sem and gsem)
Accessing stored results

Replaying the model (sem and gsem)

After estimation, you can type sem or gsem without arguments to display the estimation output:
. sem
(original output reappears)

If you wish to see results in the BentlerWeeks formulation, after sem estimation type
. estat framework
(output omitted )

See [SEM] example 11.

In many of the postestimation commands listed below for use after sem and gsem, you will need
to refer symbolically to particular coefficients. For instance, in the model
. sem ... (Y<-x1) ..., ... cov(e.Y1*e.Y2)
89

90

intro 7 Postestimation tests and predictions

the symbolic name of the coefficient corresponding to the path Y<-x1 is b[Y1:x1], and
the symbolic name of the coefficient corresponding to the covariance of e.Y1 and e.Y2 is
b[cov(e.Y1,e.Y2): cons].
Figuring out what the names are can be difficult, so instead, type
. sem, coeflegend
or
. gsem, coeflegend
With this command, sem (gsem) will produce a table looking very much like the estimation output
that lists the b[ ] notation for the estimated parameters in the model; see [SEM] example 8.

Displaying odds ratios, incidence-rate ratios, etc. (gsem only)

In some generalized linear response functions, exponentiated coefficients have a special meaning.
Those special meanings are
Common name

Family

Link

Meaning of exp(coef)

logit
ologit
mlogit
Poisson
nbreg

Bernoulli
ordinal
multinomial
Poisson
nbreg

logit
logit
logit
log
log

odds ratio
odds ratio
relative-risk ratio
incidence-rate ratio
incidence-rate ratio

gsem reports coefficients, not exponentiated coefficients. You can obtain exponentiated coefficients
and their standard errors by using estat eform after estimation. Using estat eform is no different
from redisplaying results. The syntax is
estat eform equationname
After gsem, equations are named after the dependent variable, so if you want to see the equation for
cases in exponentiated form, you can type estat eform cases.
See [SEM] example 33g, [SEM] example 34g, and [SEM] estat eform.

Obtaining goodness-of-fit statistics (sem and gsem)

One goodness-of-fit statistic and test is reported at the bottom of the sem output:
. sem
// redisplay results
Variables in structural equation model
(output omitted )
Structural equation model
(coefficient table omitted)
LR test of model vs. saturated: chi2(2)
=

1.78, Prob > chi2 = 0.4111

Be warned, this test is based on the assumption of joint normality of the observed variables, so
you may want to ignore it. The test is a goodness-of-fit test in badness-of-fit units; a significant
result implies that there may be missing paths in the models specification. More mathematically,
the null hypothesis of this test is that the fitted covariance matrix and mean vector of the observed
variables are equal to the matrix and vector observed in the population as measured by the sample.
Remember, however, the goal is not to maximize the goodness of fit. One must not add paths that
are not theoretically meaningful.

intro 7 Postestimation tests and predictions

91

In addition, other goodness-of-fit statistics are available:

1. (sem only.) Command estat gof reports a variety of goodness-of-fit statistics; see [SEM] example 4. Many but not all depend on the joint-normality assumption; see [SEM] estat gof.
2. (sem only.) Command estat eqgof reports R2 -like goodness-of-fit statistics for each
equation separately; see [SEM] example 3.
3. (sem only.) Command estat ggof reports goodness-of-fit statistics by group when you
have estimated using sems group() option; see [SEM] example 21.
4. (sem only.) Command estat residuals reports the element-by-element differences between
the observed and fitted covariance matrix, and the observed and fitted mean vector, optionally
in standardized or in normalized units; see [SEM] example 10.
5. (sem and gsem.) Command estat ic reports the Akaike and Bayesian information criteria;
see [R] estat ic.

Performing tests for including omitted paths and relaxing constraints (sem only)
1. (sem only.) Command estat mindices reports 2 modification indices and significance
values for each omitted path in the model, along with the expected parameter change; see
[SEM] example 5 and [SEM] example 9.
2. (sem only.) Command estat scoretests performs score tests on each of the linear
constraints placed on the paths and covariances; see [SEM] example 8.
3. (sem only.) Command estat ginvariant is for use when you have estimated using sems
group() option; see [SEM] intro 6. This command tests whether you can relax constraints
that parameters are equal across groups; see [SEM] example 22.

Performing tests of model simplification (sem and gsem)

1. (sem and gsem.) Command test reports Wald tests of single or multiple linear constraints.
See [SEM] example 8 and [SEM] test.
2. (sem and gsem.) Command lrtest reports likelihood-ratio tests of single or multiple linear
constraints. See [SEM] example 10, [SEM] lrtest, and [SEM] example 39g.
3. (sem only.) Command estat eqtest reports an overall Wald test for each equation in the
model, the test corresponding to all coefficients in the equation except the intercept being
simultaneously 0; see [SEM] example 13.
4. (sem only.) Command estat ginvariant is for use when you have estimated using sems
group() option; see [SEM] intro 6. This command tests whether parameters allowed to vary
across groups could be constrained; see [SEM] example 22.

Displaying other results, statistics, and tests (sem and gsem)

1. (sem only.) The estat stdize: command prefixused in front of test, testnl, lincom,
and nlcomallows you to perform tests on standardized coefficients. See [SEM] example 16
and [SEM] estat stdize.
2. (sem only.) Command estat teffects reports total effects of one variable on another
and decomposes the total effect into direct and indirect effects. Results can be reported in
standardized or unstandardized form. See [SEM] example 7 and [SEM] estat teffects.

92

intro 7 Postestimation tests and predictions

3. (sem only.) Command estat stable assesses the stability of nonrecursive structural equation
systems; see [SEM] example 7 and [SEM] estat stable.
4. (sem and gsem.) Command estat summarize reports summary statistics for the observed
variables used in the model; see [SEM] estat summarize.
5. (sem and gsem.) Command lincom reports the value, standard error, significance, and
confidence interval for linear combinations of estimated parameters; see [SEM] lincom.
6. (sem and gsem.) Command nlcom reports the value, standard error, significance, and
confidence interval for nonlinear (and linear) combinations of estimated parameters; see
[SEM] nlcom and [SEM] example 42g.
7. (sem and gsem.) Command estat vce reports the variancecovariance matrix of the estimated
parameters; see [R] estat vce.

Obtaining predicted values (sem)

You obtain predicted values with the predict command. Below we will write that predictions
are the expected values, but be aware that when there are latent variables in your model, predictions
are based on predicted scores; the scores can be inconsistent, and thus any prediction based on them
can be inconsistent.
Available are the following:
1. predict newvar, xb(odepvarname) creates new variable newvar containing the predicted
values for observed endogenous variable odepvarname.
predict stub*, xb creates new variables stub1, stub2, . . . containing the predicted values
for all the observed endogenous variables in the model.
These predicted values are the expected value of the xb given the values of the observed
exogenous variables.
2. predict newvar, latent(Lname) creates new variable newvar containing the predicted
values of the latent variable Lname, whether endogenous or exogenous.
predict stub*, latent creates new variables stub1, stub2, . . . containing the predicted
values for all the latent variables in the model.
Predicted values of latent variables, also known as predicted factor scores, are the expected
values of the variables given the values of the observed variables.
3. predict newvar, xblatent(Lname) creates new variable newvar containing the predicted
values for latent endogenous variable Lname.
predict stub*, xblatent creates new variables stub1, stub2, . . . containing the predicted
values for all the latent endogenous variables in the model.
predict with xblatent(Lname) differs from latent(Lname) in that the factor scores
predicted by latent() are then used with the linear equation for Lname to make the
prediction.
4. predict stub*, scores will create a slew of variables, one for each estimated parameter,
containing the observation-by-observation values of the first derivative, also known as scores.
This command is intended for use by programmers and may only be used after estimation
using method(ml) or method(mlmv).
See [SEM] example 14 and [SEM] predict after sem.

intro 7 Postestimation tests and predictions

93

Obtaining predicted values (gsem)

predict after gsem is more complicated than predict after sem because generalized SEMs are
more complicated. Two new issues arise: the prediction of generalized linear response variables as
opposed to linear response variables, and the optional multilevel nature of the model.
Lets start with the response functions of the observed endogenous variables. Corresponding to
predict, xb after sem are the following:
1. predict newvar, mu outcome(depvar) predicts mean responses in natural unitslinear,
probability, counts, and so on. Think g 1 (xi b) in the generalized linear model notation.
Calculations are made using the empirical Bayes predictions of latent variables, which can
be computationally intensive.
2. predict newvar, mu outcome(depvar) fixedonly also predicts mean responses in natural
units, and again it is appropriate to think g 1 (xi b). fixedonly includes only the fixed part
of the model and is equivalent to setting all latent variables to 0, their mean value.
3. Option pr is a synonym for mu when using family-and-link combinations that produce
probabilities, such as logit, probit, etc.
4. predict newvar, eta outcome(depvar) calculates the linear prediction xi b. Calculations
are made using the empirical Bayes predictions of latent variables, which can be numerically
intensive.
5. predict newvar, eta outcome(depvar) fixedonly also calculates the linear prediction
xi b, but includes only the fixed part of the model and is equivalent to setting all latent
variables to 0, their mean value.
6. Option outcome() varies to accommodate the multiple predictions produced by multinomial
logit, ordinal probit, ordinal logit, and ordinal clogclog.
a. outcome(depvar #) specifies that you want the prediction for depvar = #.
b. outcome(#.depvar) is another way of specifying that you want the prediction for
depvar = #.
c. outcome(depvar ##) specifies that you want the prediction for the #th category.
For multinomial logit, you use the above form of outcome() when requesting mu or eta
predictions.
For ordered outcomes, you use the above form of outcome() when requesting mu predictions.
For eta predictions, there is only one linear equation, and you use the usual outcome(depvar)
form.
7. Some GLM families, such as Poisson, allow specification of an offset or exposure. If you
use one of those families and you specify an offset or exposure with gsem when you fit the
model and now want predictions with offset = 0 (exposure = 1), you can specify option
nooffset.
8. To obtain predicted values for all observed endogenous variables, you can omit the outcome()
option and replace newvar with stub*that is, a piece of a new variable name followed by
*. This is not just convenient; it can also save computer time because some parts of the
calculation can be shared.

94

intro 7 Postestimation tests and predictions

For example, you can change

predict newvar, mu outcome(depvar)
to be
predict new*, mu
So much for the observed endogenous variables. Moving on to latent variables, whether endogenous
or exogenous:
9. predict newvar, latent(Lname) predicts the value of the specified latent variable by
using empirical Bayes means. You can specify that empirical Bayes modes be used by
specifying option mode; see item 12a below.
10. predict newvar, latent(Lname) se(another newvar) will predict both the latent variable
and its standard error.
11. You can use the same stub* trick mentioned for predicting observed endogenous variables
to obtain all the latent variables. Replace newvar with stub* and omit Lname to produce
predict stub*, latent
You can do this with the se() option, too:
predict stub*, latent se(another stub*)
12. There are four options having to do with the iterative procedure for obtaining the predicted
values:
a. mode specifies that empirical Bayes modes rather than means be predicted. Because
latent variables are assumed to be normally distributed, means and modes are usually
similar. They are, however, calculated differently. The predictions are the means
or modes of the empirical posterior distributions of the latent variables, which are
not necessarily normal. Therefore, the means can differ from the modes. Both are
computationally intensive.
The default empirical Bayes means calculation uses adaptive quadrature, which is
to say, numerical integration. All three of the options listed below apply to this
calculation.
The alternative empirical Bayes mode calculation uses an optimization method that
does not require adaptive quadrature. For small datasets, modes can usually be
calculated more quickly than means. For large datasets, however, predicting modes
can actually take longer. If you specify mode, of the three options listed below,
only tolerance() and iterate() are relevant.
b. intpoints(#) specifies the number of quadrature (numerical integration) points
to be used to compute the empirical Bayes means. The default is the value used in
estimation, which in turn defaults to 7.
c. tolerance(#) specifies the convergence tolerance in computing empirical Bayes
means and modes and defaults to the value used in estimation, which in turn defaults
to 1.0e8.
d. iterate(#) specifies the maximum number of iterations in computing empirical
Bayes means and modes and defaults to the value used in estimation, which in turn
defaults to 1,001.
That leaves us with only multilevel issues to discuss, and there is only one issue, and it is easy. If
you have a multilevel latent variable such as M1[school], you can predict it just as you would any
other latent variable:

intro 7 Postestimation tests and predictions

95

predict newvar, latent(M1[school])

If you predict all the latent variables with the stub* trick, any multilevel latent variables will be
automatically included among the new variables.
See [SEM] example 28g, [SEM] example 29g, and [SEM] predict after gsem.

Using contrast, pwcompare, and margins (sem and gsem)

constrast, pwcompare, and margins are postestimation commands of Stata. All three can be
used after gsem, and margins can be used after sem as well. Even so, these commands work best
when you specify your model with Statas factor-variable notation; see [SEM] intro 3. sem does not
allow that notation, and so obtaining desired results is a little more work.
1. Command contrast can be used after gsem to test linear hypotheses and make comparisons
involving factor variables and their interactions. contrast allows for performing ANOVAstyle tests of main effects, simple effects, or interaction effects. In addition, effects can be
decomposed into specific types of contrasts such as comparison against the grand mean,
orthogonal polynomial contrasts, and more. See [R] contrast.
2. Command pwcompare can be used after gsem to obtain all pairwise comparisons of marginal
linear predictions, including cell means, marginal means, etc. See [R] pwcompare.
3. Command margins is used after sem or gsem to obtain margins, meaning estimated marginal
means, predictive margins, adjusted predictions, and marginal effects. margins also allows
for computing contrasts and pairwise comparisons of both linear and nonlinear predictions.
See [R] margins, [R] margins, contrast, and [R] margins, pwcompare.

Accessing stored results

sem and gsem store all results in e(); see Stored results in [SEM] gsem. To get some idea of what
is stored in e() after sem or gsem estimation, type
. ereturn list
(output omitted )

You can save estimation results in files or temporarily in memory and do other useful things with
them; see [R] estimates.
Not stored by sem in e() are the BentlerWeeks matrices, but they can be obtained from the
r() stored results of estat framework. (The BentlerWeeks matrices are not relevant in the case
of gsem.)
See [SEM] sem and [SEM] estat framework.

Also see
[SEM] intro 6 Comparing groups (sem only)
[SEM] intro 8 Robust and clustered standard errors

Title
intro 8 Robust and clustered standard errors
Description

Options

Remarks and examples

Also see

Description
sem and gsem provide two options to modify how standard-error calculations are made:
vce(robust) and vce(cluster clustvar). These standard errors are less efficient than the default standard errors, but they are valid under less restrictive assumptions.
These options are allowed only when default estimation method method(ml) is used or when
option method(mlmv) is used. ml stands for maximum likelihood, and mlmv stands for maximum
likelihood with missing values; see [SEM] intro 4, [SEM] sem, and [SEM] gsem.
Also see [SEM] intro 9, entitled Standard errors, the full story.

Options
vce(vcetype) specifies how the VCE, and thus the standard errors, is calculated. VCE stands for
variancecovariance matrix of the estimators. The standard errors that sem and gsem report are
the square roots of the diagonal elements of the VCE.
vce(oim) is the default. oim stands for observed information matrix (OIM). The information matrix
is the matrix of second derivatives, usually of the log-likelihood function. The OIM estimator of
the VCE is based on asymptotic maximum-likelihood theory. The VCE obtained in this way is valid
if the errors are independent and identically distributed normal, although the estimated VCE is
known to be reasonably robust to violations of the normality assumption, at least as long as the
distribution is symmetric and normal-like.
vce(robust) specifies an alternative calculation for the VCE, called robust because the VCE
calculated in this way is valid under relaxed assumptions. The method is formally known as
the Huber/White/sandwich estimator. The VCE obtained in this way is valid if the errors are
independently distributed. It is not required that the errors follow a normal distribution, nor is it
required that they be identically distributed from one observation to the next. Thus the vce(robust)
VCE is robust to heteroskedasticity of the errors.
vce(cluster clustvar) is a generalization of the vce(robust) calculation that relaxes the
assumption of independence of the errors and replaces it with the assumption of independence
between clusters. Thus the errors are allowed to be correlated within clusters.

Remarks and examples

The vce() option is allowed by sem and gsem. In the rest of this entry, we will use sem in
illustrations, but everything we say applies equally to gsem.
The vce(robust) option,
. sem ..., ... vce(robust)
and the vce(cluster clustvar) option,
. sem ..., ... vce(cluster clustvar)
96

intro 8 Robust and clustered standard errors

97

relax assumptions that are sometimes unreasonable for a given dataset and thus produce more accurate
standard errors in those cases. Those assumptions are homoskedasticity of the variances of the errors
vce(robust)and independence of the observationsvce(cluster clustvar). vce(cluster clustvar) relaxes both assumptions.
Homoskedasticity means that the variances of the errors are the same from observation to observation.
Homoskedasticity can be unreasonable if, for instance, the error corresponds to a dependent variable
of income or socioeconomic status. It would not be unreasonable to instead assume that, in the data,
the variance of income or socioeconomic status increases as the mean increases. In such cases, rather
than typing
. sem (y<-...) (...) (...<-x1) (...<-x2)
you would type
. sem (y<-...) (...) (...<-x1) (...<-x2), vce(robust)
Independence implies that the observations are uncorrelated. If you have observations on people,
some of whom live in the same neighborhoods, it would not be unreasonable to assume instead that
the error of one person is correlated with those of others who live in the same neighborhood because
neighborhoods tend to be homogeneous. In such cases, if you knew the neighborhood, rather than
typing
. sem (y<-...) (...) (...<-x1) (...<-x2)
you would type
. sem (y<-...) (...) (...<-x1) (...<-x2), vce(cluster neighborhood)
Understand that if the assumptions of independent and identically distributed normal errors are
met, the vce(robust) and vce(cluster clustvar) standard errors are less efficient than the standard
vce(oim) standard errors. Less efficient means that for a given sample size, the standard errors jump
around more from sample to sample than would the vce(oim) standard errors. vce(oim) standard
errors are unambiguously best when the standard assumptions of homoskedasticity and independence
are met.

Also see
[SEM] intro 7 Postestimation tests and predictions
[SEM] intro 9 Standard errors, the full story
[SEM] sem option method( ) Specifying method and calculation of VCE
[SEM] gsem estimation options Options affecting estimation

Title
intro 9 Standard errors, the full story

Description

Options

Remarks and examples

Also see

Description
In [SEM] intro 8, we told you part of the story of the calculation of the VCE, the part we wanted
to emphasize. In this section, we tell you the full story.
We at Stata try to draw a clear distinction between method and technique. The method is the
process used to obtain the parameter estimates. The technique is the process used to obtain the
variancecovariance matrix of the parameter estimates, which is to say, the standard errors.
The literature does not always draw such clear distinctions.
sem and gsem provide the following methods and techniques:
Methods
ML
QML
MLMV
ADF

maximum likelihood
quasimaximum likelihood
maximum likelihood with missing values
asymptotic distribution free

Techniques
OIM
EIM
OPG

robust
clustered
bootstrap
jackknife

observed information matrix

expected information matrix
outer product of the gradients
Huber/White/sandwich estimator
generalized Huber/White/sandwich estimator
nonparametric bootstrap
delete-one jackknife

98

intro 9 Standard errors, the full story

They are allowed in the following combinations:

Command/
Method
sem
ML

MLMV

ADF

Allowed techniques

Comment

OIM
EIM
OPG

default

robust
clustered
bootstrap
jackknife

a.k.a. QML

OIM
EIM
OPG

default

robust
clustered
bootstrap
jackknife

a.k.a. QML

OIM
EIM

default; robust-like

bootstrap
jackknife
gsem
ML

OIM
OPG

default

robust
clustered
bootstrap
jackknife

a.k.a. QML

99

100

intro 9 Standard errors, the full story

Options
The corresponding options for sem and gsem to obtain each allowed method-and-technique
combination are

method()
sem
method(ml)

method(mlmv)

method(adf)

gsem
method(ml)

vce()

Comment

vce(oim)
vce(eim)
vce(opg)
vce(robust)
vce(cluster clustvar)
vce(bootstrap)
vce(jackknife)

default

vce(oim)
vce(eim)
vce(opg)
vce(robust)
vce(cluster clustvar)
vce(bootstrap)
vce(jackknife)

default

vce(oim)
vce(eim)
vce(bootstrap)
vce(jackknife)

default; vce(robust)-like

vce(oim)
vce(opg)
vce(robust)
vce(cluster clustvar)
bootstrap
jackknife

default

a.k.a. QML

a.k.a. QML

a.k.a. QML
no option; use bootstrap: prefix
no option; use jackknife: prefix

method(emethod) specifies the estimation method sem (gsem) is to use. If method() is not specified,
then method(ml) is assumed.
vce(vcetype) specifies the technique to be used to obtain the VCE. When vce() is not specified,
then vce(oim) is assumed.
In the case of gsem, vce(bootstrap) and vce(jackknife) are not allowed, although you can
obtain the bootstrap or jackknife results by prefixing the gsem command with the bootstrap: or
jackknife: prefix.
. bootstrap: gsem . . .
. jackknife: gsem . . .
See [R] bootstrap and [R] jackknife. If you are fitting a multilevel model, be sure to use bootstraps
and jackknifes cluster() and idcluster() options to obtain a correct resampling. If you
have a crossed model, you cannot resample in both dimensions; there is no solution to that problem,
and therefore you cannot use the bootstrap: or jackknife: prefix.

intro 9 Standard errors, the full story

101

Remarks and examples

In [SEM] intro 4, we gave reasons for why you might want to choose one method/technique over
another. In that section, we did not draw the clear distinction between method and technique that we
draw here.
The technique everyone is familiar with is OIM, although they might not know it by that name.
OIM is the inverse of the negative matrix of second derivatives. Note that it is also the default for
both sem and gsem.
The most common alternative technique chosen are robust and clusteroptions vce(robust) and
vce(cluster clustvar). That is because they relax assumptions; see [SEM] intro 8.
Aside: Robust and cluster are not allowed with method(adf). Nonetheless, the default OIM
technique is more robust when combined with method(adf) than OIM usually is. OIM can be
used even when errors are heteroskedastic. That is, the assumptions justifying the OIM calculation
when combined with method(adf) are they same as those justifying vce(robust)the errors
are merely assumed to be independent.
Technique EIMavailable with sem but not gsemhas similar properties to OIM. It is used in
performing score tests. The sem command secretly calculates the EIM when necessary so that you
can use postestimation score-test commands even if you estimate using a technique other than EIM.
EIM is available to you because the sem command needs EIM for its own hidden purposes.
Technique OPG is not used so much anymore, although historically it was popular because it took
less computer time to calculate than OIM.
For a discussion of bootstrap and jackknife variance estimation, see [R] bootstrap and [R] jackknife.

Also see
[SEM] intro 8 Robust and clustered standard errors
[SEM] intro 10 Fitting models with survey data (sem only)
[SEM] sem option method( ) Specifying method and calculation of VCE
[SEM] gsem estimation options Options affecting estimation

Title
intro 10 Fitting models with survey data (sem only)
Description

Remarks and examples

Also see

Description
Sometimes the data are not a simple random sample from the underlying population but instead
are based on a complex survey design that can include stages of clustered sampling and stratification.
Estimates produced by sem can be adjusted for these issues.
Adjustments for survey data are provided by sem but not gsem.

Remarks and examples

Data obtained from surveys, properly treated, produce different point estimates because some
observations represent a greater proportion of the underlying population than others. They also
produce different standard errors because the observation-to-observation (sample-to-sample) variation
is a function of the surveys design.
To obtain survey-corrected results, you first describe the characteristics of the survey with svyset:
. svyset county [pw=samplewgt], fpc(n_counties) strata(states) ||
school, fpc(n_schools)
||
student, fpc(n_students)

///
///

In the above, we are telling Stata that our data are from a three-stage sampling design. The first
stage samples without replacement counties within state; the second, schools within each sampled
county; and the third, students within schools.
Once we have done that, we can tell Stata to make the survey adjustment by prefixing statistical
commands with the svy: prefix:
. svy: regress test_result teachers_per_student sex ...

Point estimates and standard errors will be adjusted.

You can use the svy: prefix with sem:
. svy: sem (test_result<-...) ... (teachers_per_student->...) ...

See the Stata Survey Data Reference Manual for more information on this. From a survey perspective,
sem is not different from any other statistical command of Stata.
Once results are estimated, you do not include the svy: prefix in front of the postestimation
commands. You type, for instance,
. estat eqtest ...

You do not type svy: estat eqtest . . . .

Some postestimation procedures you might ordinarily perform can be inappropriate with survey
estimation results. This is because you no longer have a sample likelihood value. The postestimation
command lrtest is an example. If you attempt to use an inappropriate postestimation command,
you will be warned.
. lrtest ...
lrtest is not appropriate with survey estimation results
r(322);

102

intro 10 Fitting models with survey data (sem only)

Also see
[SEM] intro 9 Standard errors, the full story
[SEM] intro 11 Fitting models with summary statistics data (sem only)
[SVY] Stata Survey Data Reference Manual

103

Title
intro 11 Fitting models with summary statistics data (sem only)

Description

Remarks and examples

Reference

Also see

Description
In textbooks and research papers, the data used are often printed in summary statistic form. These
summary statistics include means, standard deviations or variances, and correlations or covariances.
These summary statistics can be used in place of the underlying raw data to fit models with sem.
Summary statistics data (SSD) are convenient for publication because of their terseness. By not
revealing individual responses, they do not violate participant confidentiality, which is sometimes
important.
Support for SSD is provided by sem but not by gsem.

Remarks and examples

Remarks are presented under the following headings:
Background
How to use sem with SSD
What you cannot do with SSD
Entering SSD
Entering SSD for multiple groups
What happens when you do not set all the summary statistics
Labeling SSD
Making summary statistics from data for use by others

Background
The structural equation modeling estimator is a function of the first and second momentsthe
means and covariancesof the data. Thus it is possible to obtain estimates of the parameters of an
SEM by using means and covariances. One does not need the original dataset.
In terms of sem, one can create a dataset containing these summary statistics and then use that
dataset to obtain fitted models. The sem command is used just as one would use it with the original,
raw data.

How to use sem with SSD

To use sem with SSD,
1. Enter the summary statistics by using the ssd command. How you do that is the topic of an
upcoming section.
2. Save the data just as you would any dataset, namely, with the save command.
3. Use the sem command just as you would ordinarily. You use the SSD if they are not already in
memory, and no special syntax or options are required by sem, except
104

intro 11 Fitting models with summary statistics data (sem only)

105

a. Do not use sems if exp or in range modifiers. You do not have the raw data in memory
and so you cannot select subsets of the data.
b. If you have entered summary statistics for groups of observations (for example, males and,
separately, females), use sems select() option if you want to fit the model with a subset
of the groups. That is, where you would ordinarily type
. sem ... if sex==1, ...
you instead type
. sem ..., ... select(1)
Where you would ordinarily type
. sem ... if region==1 | region==3, ...
you instead type
. sem ..., ... select(1 3)
See [SEM] example 3.

What you cannot do with SSD

With SSD in memory,
1. You cannot obtain robust standard errors, which you would normally do by specifying sem option
vce(robust).
2. You cannot obtain clustered standard errors, which you would normally do by specifying sem
option vce(cluster clustvar).
3. You cannot obtain survey-adjusted results, which you would normally do by specifying the
svy: prefix in front of the sem command.
4. You cannot obtain bootstrap or jackknife standard errors, which you would normally do by
specifying sem option vce(bootstrap) or vce(jackknife).
5. You cannot obtain VCE estimates from the observation-level outer product of the gradients, which
you would normally do by specifying vce(opg).
6. You cannot use weights, which you would normally do by specifying, for instance, [fw=varname].
7. You cannot restrict the estimation sample with if exp or in range.
8. You cannot fit the model by using maximum likelihood with missing values or by using the
asymptotic distribution free method, which you would normally do by specifying method(mlmv)
or method(adf).

Entering SSD
Entering SSD is easy. You need to see an example of how easy it is before continuing: see
[SEM] example 2.
What follows is an outline of the procedure. Let us begin with the data you need to have. You
have
1. The names of the variables. We will just call them x1, x2, and x3.
2. The number of observations, say, 74.

106

intro 11 Fitting models with summary statistics data (sem only)

3. The correlations, say,

1
-0.8072
0.3934

1
-0.5928

or you may have the covariances,

33.4722
-3.6294
1.0374

0.6043
-0.2120

0.2118

4. The variances: 33.4722, 0.6043, and 0.2118.

Or the standard deviations: 5.6855, 0.7774, and 0.4602.
Or neither.
If you have the covariances in step 3, you in fact have the variancesthey are just the diagonal
elements of the covariance matrixbut the software will not make you enter the values twice.
5. The means: 21.2973, 3.0195, and 0.2973.
Or not.
With that information at hand, do the following:
1. Start with no data in memory:
. clear all
2. Initialize the SSD by stating the names of the variables:
. ssd init x1 x2 x3
The remaining steps can be done in any order.
3. Set the number of observations:
. ssd set obs 74
4. Set the covariances:
. ssd set cov 33.4722 \ -3.6294 .6043 \ 1.0374 -.2120 .2118
Or the correlations:
. ssd set cor

1 \ -.8072 1 \ .3934 -.5928 1

5. If you set covariances in step 4, skip to step 6. Otherwise, if you have them, set the variances:
. ssd set var 33.4722 .6043 .2118
Or set the standard deviations:
. ssd set sd

5.6855 .7774 .4602

6. Set the means if you have them:

. ssd set means 21.2973 3.0195 .2973
7. If at any point you become confused as to what you have set and what remains to be set, type
. ssd status

intro 11 Fitting models with summary statistics data (sem only)

107

8. If you want to review what you have set, type

. ssd list
9. If you make a mistake, you can repeat any ssd set command by adding the replace option
to the end. For instance, you could reenter the means by typing
. ssd set means 21.2973 3.0195 .2973, replace
10. Save the dataset just as you would with any dataset:
. save mydata
You are now ready to use sem with the SSD. With the SSD in memory, you issue the sem command
just as you would if you had the raw data:
. sem ...

Entering SSD for multiple groups

You can enter summary statistics for groups of the data. Perhaps you have summary statistics not
for the data as a whole, but for males and for females, or for the young, for the middle-aged, and
for the old.
Lets pretend you have the following data:
The young:
observations:
correlations:

standard deviations:
means:
The middle-aged:
observations:
correlations:

standard deviations:
means:
The old:
observations:
correlations:

standard deviations:
means:

74
1
0.8072
0.3934
5.6855
21.2973

1
0.5928
0.7774
3.0195

1
0.4602
0.2973

141
1
0.5721
0.3843
4.9112
38.1512

1
0.4848
0.7010
5.2210

1
0.5420
0.2282

36
1
0.8222
0.3712
6.7827
58.7171

1
0.3113
0.7221
2.1511

1
0.4305
0.1623

The commands for entering these summary statistics are

.
.
.
.

ssd
ssd
ssd
ssd

init x1 x2 x3
set obs
74
set cor
1
\
set sd
5.6855

-.8072
.7774

1
\ .3934 -.5928 1
.4602

108

intro 11 Fitting models with summary statistics data (sem only)

.
.
.
.
.
.

ssd
ssd
ssd
ssd
ssd
ssd

set means 21.2973 3.0195

addgroup agecategory
set obs 141
set cor
1
\ -.5721
set sd
4.9112
.7010
set means 38.1512 5.2210

.
.
.
.
.
.

ssd addgroup
ssd set obs
36
ssd set cor
1
\
ssd set sd
6.7827
ssd set means 58.7171
save mygroupdata

-.8222
.7221
2.1511

.2973

1
\ .3843 -.4848 1
.5420
.2282

\ .3712 -.3113 1
.4305
.1623

The general procedure is as follows:

1. Enter the summary statistics for the first group just as outlined in the previous section.
2. Next add a group by typing
. ssd addgroup newgroupvar
In that one command, you are telling ssd that the summary statistics you entered in step 1 were
for a group you are now calling newgroupvar, and in particular they were for newgroupvar = 1.
You are also telling ssd that you now want to enter the summary statistics for the next group,
namely, newgroupvar = 2.
3. Enter the summary statistics for the second group in the same way you entered them for the first
group, just as outlined in the previous section.
4. If you have a third group, add it by typing
. ssd addgroup
In this case, you are telling ssd only one thing: that you now want to enter data for the next
group, namely, newgroupvar = 3.
5. Enter the summary statistics for the third group in the same way you entered them for the second
group, and just as outlined in the previous section.
6. If you want to add more groups, continue in the same way. Declare the next group of data by
typing
. ssd addgroup
and then enter the data by using the ssd set command.
7. If you mistakenly add a group and wish to rescind that, type
. ssd unaddgroup
8. If you wish to go back and modify the values entered for a previous group, put the group number
between ssd set and what is being setfor instance, type ssd set 2 observations . . . and
specify the replace option. For instance, to reenter the correlations for group 1, type
. ssd set 1 correlations values, replace

intro 11 Fitting models with summary statistics data (sem only)

109

What happens when you do not set all the summary statistics
You are required to set the number of observations and to set the covariances or the correlations.
Setting the variances (standard deviations) and setting the means are optional.
1. If you set correlations only, then
a. Means are assumed to be 0.
b. Standard deviations are assumed to be 1.
c. You will not be able to pool across groups if you have group data.
As a result of (a) and (b), the parameters sem estimates will be standardized even when you do
not specify sems standardized reporting option. Estimated means and intercepts will be 0.
Concerning (c), we need to explain. This concerns group data. If you type
. sem ...
then sem fits a model with all the data. sem does that whether you have raw data or SSD in
memory. If you have SSD with groupssay, males and females or age groups 1, 2, and 3sem
combines the summary statistics to obtain the summary statistics for the overall data. It is only
possible to do this when covariances and means are known for each group. If you set correlations
without variances or standard deviations and without means, the necessary statistics are not known
and the groups cannot be combined. Thus if you type
. sem ...
you will get an error message. You can still estimate using sem; you just have to specify on which
group you wish to run sem, and you do that with the select() option:
. sem ..., select(#)
2. If you set correlations and means,
a. Standard deviations are assumed to be 1.
b. You will not be able to pool across groups if you have group data.
This situation is nearly identical to situation 1. The only difference is that estimated means and
intercepts will be nonzero.
3. If you set correlations and standard deviations or variances, or if you set covariances only,
a. Means are assumed to be 0.
b. You will not be able to pool across groups if you have group data.
This situation is a little better than situation 1. Estimated intercepts will be 0, but the remaining
estimated coefficients will not be standardized unless you specify sems standardized reporting
option.

Labeling SSD
You may use the following commands on SSD, and you use them in the same way you would with
an ordinary dataset:
1. rename oldvarname newvarname
You may rename the variables; see [D] rename.
2. label data "dataset label"
You may label the dataset; see [D] label.

110

intro 11 Fitting models with summary statistics data (sem only)

3. label variable varname "variable label"

You may label variables; see [D] label.
4. label values groupvarname valuelabelname
You may place a value label on the group variable; see [D] label. The group variable always
takes on the values 1, 2, . . . .
5. note: my note
note varname: my note
You may places notes on the dataset or on its variables; see [D] notes.
Do not modify the SSD except by using the ssd command. Most importantly, do not drop variables
or observations.

Making summary statistics from data for use by others

If you have raw data and wish to make the summary statistics available for subsequent publication,
type
. ssd build varlist
where varlist lists the variables you wish to include in the dataset. The SSD will replace the raw data
you had in memory. The full syntax is
. ssd build varlist if exp in range
so you may specify if and in to restrict the observations that are included.
For instance, to build an SSD for variables occ prestige, income, and social status, type
. ssd build occ_prestige income social_status
If you wish to build the dataset to include separate groups for males and females, type
. ssd build occ_prestige income social_status, group(sex)
However the sex variable was coded in your original data, the two sexes will be now be coded
1 and 2 in the resulting SSD. Which sex is 1 and which is 2 will correspond to however sort
would have ordered sex in its original coding. For instance, if variable sex took on values male
and female, the resulting variable sex would take on values 1 corresponding to female and 2
corresponding to male.
Once you have built the SSD, you can describe it and list it:
. ssd describe
. ssd list
See [SEM] example 25.

Reference
Acock, A. C. 2013. Discovering Structural Equation Modeling Using Stata. Rev. ed. College Station, TX: Stata Press.

intro 11 Fitting models with summary statistics data (sem only)

Also see
[SEM] intro 10 Fitting models with survey data (sem only)
[SEM] intro 12 Convergence problems and how to solve them
[SEM] ssd Making summary statistics data (sem only)
[SEM] sem option select( ) Using sem with summary statistics data
[SEM] example 2 Creating a dataset from published covariances
[SEM] example 3 Two-factor measurement model
[SEM] example 19 Creating multiple-group summary statistics data
[SEM] example 25 Creating summary statistics data from raw data

111

Title
intro 12 Convergence problems and how to solve them
Description

Remarks and examples

Also see

Description
It can be devilishly difficult for software to obtain results for SEMs. Here is what can happen:
. sem ...
Variables in structural equation model
(output omitted )
Fitting target model:
initial values not feasible
r(1400);

or,
. gsem ...
Fitting fixed-effects model:
Iteration 0: log likelihood
Iteration 1: log likelihood
Iteration 2: log likelihood
(output omitted )
Refining starting values:
Grid node 0: log likelihood
Grid node 1: log likelihood
Grid node 2: log likelihood
Grid node 3: log likelihood
Fitting full model:
initial values not feasible
r(1400);

= -914.65237
= -661.32533
= -657.18568

=
=
=
=

.
.
.
.

or,
. sem ...
Endogenous variables
(output omitted )
Fitting target model:
Iteration 1: log likelihood
.
.
.
Iteration 50: log likelihood
Iteration 51: log likelihood
Iteration 52: log likelihood
.
.
.
Iteration 101: log likelihood
Break
r(1);

= ...

= -337504.44
= -337503.52
= -337502.13

(not concave)
(not concave)
(not concave)

= -337400.69

(not concave)

In the first two cases, sem and gsem gave up. The error message is perhaps informative if not
helpful. In the last case, sem (it could just as well have been gsem) iterated and iterated while
producing little improvement in the log-likelihood value. We eventually tired of watching a process
that was going nowhere slowly and pressed Break.
Now what?
112

intro 12 Convergence problems and how to solve them

113

Remarks and examples

Remarks are presented under the following headings:
Is your model identified?
Convergence solutions generically described
Temporarily eliminate option reliability()
Use default normalization constraints
Temporarily eliminate feedback loops
Temporarily simplify the model
Try other numerical integration methods (gsem only)
Get better starting values (sem and gsem)
Get better starting values (gsem)

Is your model identified?

You may not know whether your model is identified. Infinite iteration logs are an indication of
lack of identification or of poor starting values:
. sem ...
Endogenous variables
(output omitted )
Fitting target model:
Iteration
.
.
.
Iteration
Iteration
Iteration
.
.
.
Iteration
Break
r(1);

1:

50:
51:
52:

log likelihood = ...

log likelihood = -337504.44

log likelihood = -337503.52
log likelihood = -337502.13

(not concave)
(not concave)
(not concave)

101: log likelihood = -337400.69

(not concave)

If the problem is lack of identification, the criterion function being optimized (the log likelihood
in this case) will eventually stop improving at all and yet sem or gsem will continue iterating.
If the problem is poor starting values, the criterion function will continue to increase slowly.
So if your model might not be identified, do not press Break too soon.
There is another way to distinguish between identification problems and poor starting values. If
starting values are to blame, it is likely that a variance estimate will be heading toward 0. If the
problem is lack of identification, you are more likely to see an absurd path coefficient.
To distinguish between those two alternatives, you will need to rerun the model and specify the
iterate() option:
. sem ..., ... iterate(100)
(output omitted )

We omitted the output, but specifying iterate(100) allowed us to see the current parameter values
at the point. We chose to specify iterate(100) because we knew that the likelihood function was
changing slowly by that point.
If you are worried about model identification, you have a choice: Sit it out and do not press Break
too soon, or press Break and rerun.

114

intro 12 Convergence problems and how to solve them

If you discover that your model is not identified, see Identification 1: Substantive issues in
[SEM] intro 4.

Convergence solutions generically described

There are three generic solutions to convergence problems that we will use:
G1. The improved-starting-values procedure:
Obtain the current parameter values from a failed attempt, modify those lousy values to
make them better, use the improved values as starting values to try to fit the model again,
and repeat as necessary.
G2. The alternative-starting-values procedure:
Simplify the model to produce an easier-to-fit model, fit the simplified model, use the
simplified models solution as starting values to fit the original, more complicated model,
and repeat as necessary.
G3. The alternative-software-logic procedure:
Specify strange options that make the software behave differently in hopes that a different
approach will produce a solution. sem does not have any such strange options, but gsem
does. In following this approach, it does not matter whether we in fact understand what we
are doing because, once we find a solution, we can obtain the parameter values from the
successful model and use those values as starting values to fit the model without the strange
and confusing options.
Sometimes we will use all three of these procedures. We will talk about convergence and these
procedures substantively below, but before we do that, we want to show you how to use a tool that
you will need.
The generic solutions share a mechanical step in common, namely, obtaining parameter values
from one attempt at fitting the model to use as starting values in a subsequent attempt. Here is how
you do that:
. sem ..., ...
. matrix b = e(b)
. ...

// fit one model

// save parameter estimates in b
// optionally modify b

. sem ..., ... from(b)

// fit same model or different model

If the first sem or gsem command fails because you pressed Break or if the command issued
an error message, you must reissue the command adding option noestimate or iterate(#).
Specify noestimate if the failure came early before the iteration log started, and otherwise specify
iterate(#), making # the iteration number close to but before the failure occurred:
. sem ..., ... noestimate

or
. sem ..., ... iterate(50)

intro 12 Convergence problems and how to solve them

115

Once you have obtained the parameter values in b, you can list b,
. matrix b = e(b)
. matrix list b
b[1,10]
x1:
x1:
L
_cons
y1
1
96.284553
var(e.x1): var(e.x2):
_cons
_cons
y1
78.54289
113.62044

x2:
L
1.0971753
var(e.x3):
_cons
85.34721

x2:
_cons
97.284553
var(L):
_cons
120.45744

x3:
L
1.0814186

x3:
_cons
97.097561

x2:
L
1.0971753
var(e.x3):
_cons
85.34721

x2:
_cons
97.284553
var(L):
_cons
500

x3:
L
1.0814186

x3:
_cons
97.097561

and you can modify it:

. matrix b[1, 10] = 500
. matrix list b
b[1,10]
x1:
x1:
L
_cons
y1
1
96.284553
var(e.x1): var(e.x2):
_cons
_cons
y1
78.54289
113.62044

And, whether you modify it or not, you can use b as the starting values for another attempt of the
same model or for an attempt of a different model:
. sem ..., ... from(b)

Temporarily eliminate option reliability()

If you specified sems or gsems reliability() option, remove it and try fitting the model again.
If the model converges, then your estimate of the reliability is too low; see What can go wrong in
[SEM] sem and gsem option reliability( ).

Use default normalization constraints

Let sem and gsem provide their default normalization constraints. By default, sem and gsem
constrain all latent exogenous variables to have mean 0; constrain all latent endogenous variables
to have intercept 0; and constrain the paths from latent variables to the first observed endogenous
variable to have coefficient 1.
Replacing any of the above defaults can cause problems, but problems are most likely to arise if you
replace the last listed default. Do not constrain path coefficients merely to obtain model identification.
Let sem choose those constraints. It is possible that the default-coefficient-1 constraint is inappropriate
for how you want to interpret your model. Relax it anyway and reimpose it later.
If default constraints solve the problem, you are done unless you want to reimpose your original,
alternative constraints. That you do by typing
. sem ..., ...
. matrix b = e(b)
. sem ..., ... from(b)

// model with default constraints

// model with desired constraints

If you have multiple constraints that you want to reimpose, you may need to do them in sets.

116

intro 12 Convergence problems and how to solve them

Temporarily eliminate feedback loops

Check whether your model has any feedback loops, such as
. sem ... (y1<-y2 x2) (y2<-y1 x3) ...

In this example, variable y1 affects y2 affects y1. Models with such feedback loops are said to be
nonrecursive. Assume you had a solution to the above model. The results might be unstable in a
substantive sense; see nonrecursive (structural) model (system) in [SEM] Glossary. The problem is
that finding such truly unstable solutions is often difficult and the stability problem manifests itself
as a convergence problem.
If you have convergence problems and you have feedback loops, that is not proof that the underlying
values are unstable.
Regardless, temporarily remove the feedback loop,
. sem ... (y1<-y2 x2) (y2<-

x3) ...

and see whether the model converges. If it does, save the parameter estimates and refit the original
model with the feedback, but using the saved parameter estimates as starting values.
. matrix b = e(b)
. sem ... (y1<-y2 x2) (y2<-y1 x3) ..., ... from(b)

If the model converges, the feedback loop is probably stable. If you are using sem, you can check
for stability with estat stable. If the model does not converge, you now must find which other
variables need to have starting values modified.

Temporarily simplify the model

At this point, it is difficult to know whether you should temporarily simplify your model or simply
proceed with the subsequent steps and come back to simplification later should it be necessary. So
proceed in whatever order seems best to you.
The idea of temporarily simplifying the model is to simplify the model, get that model to converge,
and use the simplified models solution as starting values for the more complicated model.
The more orthogonal and independent you can make the pieces of the model, the better. Remove
covariances. If you have measurement, fit it separately. Basically, remove whatever you hold most
dear, because you are probably looking for subtle and correlated effects.
Once you find a simplified model that converges, do the following:
. sem ..., ...
. matrix b = e(b)
. sem ..., ... from(b)

// fit simplified model

// fit original model

Try other numerical integration methods (gsem only)

gsem provides four numerical integration methods just so you can try them. The intmethod()
option specifies the integration method.
1. intmethod(mvaghermite) is the default and performs mean-and-variance adaptive Gauss
Hermite quadrature. It is fast and accurate.
2. intmethod(mcaghermite) performs mode-and-curvature adaptive GaussHermite quadrature. It is accurate but not as fast. If you are fitting a multilevel model, there are cases where
method 1 will not work but this method will work.

intro 12 Convergence problems and how to solve them

117

3. intmethod(ghermite) performs nonadaptive GaussHermite quadrature. It is less accurate

but quicker, and the calculation it makes converges more readily than either of the above
methods.
4. intmethod(laplace) performs the Laplacian approximation instead of quadrature. It is
the least accurate, sometimes the fastest, and the calculation it makes has no convergence
issues whatsoever.
Try all of them. We view methods 1 and 2 as completely trustworthy. If your model will only
converge with method 3 or 4, we recommend using the result as starting values for method 1 or 2.
Thus you might type
. gsem ..., ... intmethod(ghermite)
. matrix b = e(b)
. gsem ..., ... from(b)

There is another option we should mention, namely, intpoints(#). Methods 1, 2, and 3 default
to using seven integration points. You can change that. A larger number of integration points produces
more accurate results but does not improve model convergence. You might reduce the number of
integration points to, say, 3. A lower number of integration points slightly improves convergence
of the model, and it certainly makes model fitting much quicker. Obviously, model results are less
accurate. We have been known to use fewer integration points, but mainly because of the speed issue.
We can experiment more quickly. At the end, we always return to the default number of integration
points.

Get better starting values (sem and gsem)

If you observe,
. sem ..., ...
Variables in structural equation model
(output omitted )
Fitting target model:
initial values not feasible
r(1400);

and you are using sem, we direct you back to Temporarily simplify the model. The problem that we
discuss here seldom reveals itself as initial values not feasible with sem. If you are using gsem,
we direct you to the next section, Get better starting values (gsem). It is not impossible that what we
discuss here is the solution to your problem, but it is unlikely.

118

intro 12 Convergence problems and how to solve them

We discuss here problems that usually reveal themselves by producing an infinite iteration log:
. sem ..., ...
Endogenous variables
(output omitted )
Fitting target model:
Iteration 1: log likelihood
.
.
.
Iteration 50: log likelihood
Iteration 51: log likelihood
Iteration 52: log likelihood
.
.
.
Iteration 101: log likelihood
Break
r(1);

= ...

= -337504.44
= -337503.52
= -337502.13

(not concave)
(not concave)
(not concave)

= -337400.69

(not concave)

The first thing to do is look at the parameter values. To do that, type

. sem ..., ... iterate(100)

We specified 100; you should specify an iteration value based on your log.
In most cases, you will discover that you have a variance of a latent exogenous variable going to
0, or you have a variance of an error (e.) variable going to 0.
Based on what you see, say that you suspect the problem is with the variance of the error of a
latent endogenous variable F going to 0, namely, e.F. You need to give that variance a larger starting
value, which you can do by typing
. sem ..., ... var(e.F, init(1))

or
. sem ..., ... var(e.F, init(2))

or
. sem ..., ... var(e.F, init(10))

We recommend choosing a value for the variance that is larger than you believe is necessary.
To obtain that value,
1. If the variable is observed, use summarize to obtain the summary statistics for the variable,
square the reported standard deviation, and then increase that by, say, 20%.
2. If the variable is latent, use summarize to obtain a summary of the latent variables
anchor variable and then follow the same rule: use summarize, square the reported standard
deviation, and then increase that by 20%. (The anchor variable is the variable whose path
is constrained to have coefficient 1.)
3. If the variable is latent and has paths only to other latent variables so that its anchor variable
is itself latent, follow the anchors paths to an observed variable and follow the same rule:
use summarize, square the reported standard deviation, and then increase that by 20%.
4. If you are using gsem to fit a multilevel model and the latent variable is at the observational
level, follow advice 2 above. If the latent variable is at a higher levelsay schooland
its anchor is x, a Gaussian response with the identity link, type

intro 12 Convergence problems and how to solve them

.
.
.
.

119

sort school
by school: egen avg = mean(x)
by school: gen touse = _n==1 if school<.
summarize avg if touse==1

Square the reported standard deviation and add 20%.

5. If you are using gsem and the anchor of the latent variable is not Gaussian with the identity
link, see the next section.
Do not dismiss the possibility that the bad starting values concern estimated parameters other than
variances of latent exogenous or error variables, although variances of those kinds of variables is
the usual case. Covariances are rarely the problem because covariances can take on any value and,
whether too small or too large, usually get themselves back on track as the iterative process proceeds.
If you need to specify an initial value for a covariance, the syntax is
. sem ..., ... cov(e.F*e.G, init(-25))

Substitute for 25 the value you consider reasonable.

The other possibility is that a path coefficient needs a better starting value, which is as unlikely as
a covariance being the problem, and for the same reasons. To set the initial value of a path coefficient,
add the init() option where the path is specified. Say the original sem command included y<-x1:
. sem ... (y<-x1 x2) ...

If you wanted to set the initial value of the path from x1 to 3, modify the command to read
. sem ... (y<-(x1, init(3)) x2) ...

Get better starting values (gsem)

What we say below applies regardless of how the convergence problem revealed itself. You might
have seen the error message initial values not feasible, or some other error message, or you might
have an infinite iteration log.
gsem provides two options to help you obtain better starting values: startvalues() and startgrid().
startvalues(svmethod) allows you to specify one of five starting-value-calculation methods:
zero, constantonly, fixedonly, ivloadings, and iv. By default, gsem uses ivloadings.
Evidently, that did not work for you. Try the others, starting with iv:
. gsem ..., ... startvalues(iv)

If that does not solve the problem, proceed through the others in the following order: fixedonly,
constantonly, and zero.
By the way, if you have starting values for some parameters but not othersperhaps you fit a
simplified model to get themyou can combine the options startvalues() and from():
. gsem ..., ...
. matrix b = e(b)
. gsem ..., ... from(b) startvalues(iv)

// simplified model
// full model

You can combine startvalues() with the init() option, too. We described init() in the previous
section.
The other special option gsem provides is startgrid(). startgrid() can be used with or without
startvalues(). startgrid() is a brute-force approach that tries various values for variances and
covariances and chooses the ones that work best.

120

intro 12 Convergence problems and how to solve them

1. You may already be using a default form of startgrid() without knowing it. If you see
gsem displaying Grid node 1, Grid node 2, . . . following Grid node 0 in the iteration log,
that is gsem doing a default search because the original starting values were not feasible.
The default form tries 0.1, 1, and 10 for all variances of all latent variables, by which
we mean the variances of latent exogenous variables and the variances of errors of latent
endogenous variables.
2. startgrid(numlist) specifies values to try for variances of latent variables.
3. startgrid(covspec) specifies the particular variances and covariances in which grid searches
are to be performed. Variances and covariances are specified in the usual way. startgrid(e.F e.F*e.L M1[school] G*H e.y e.y1*e.y2) specifies that 0.1, 1, and 10 be
tried for each member of the list.
4. startgrid(numlist covspec) allows you to combine the two syntaxes, and you can specify
multiple startgrid() options so that you can search the different ranges for different
variances and covariances.
Our advice to you is this:
1. If you got an iteration log and it did not contain Grid node 1, Grid node 2, . . . , then specify
startgrid(.1 1 10). Do that whether the iteration log was infinite or ended with some
other error. In this case, we know that gsem did not run startgrid() on its own because
it did not report Grid node 1, Grid node 2, etc. Your problem is poor starting values, not
infeasible ones.
A synonym for startgrid(.1 1 10) is just startgrid without parentheses.
Be careful, however, if you have a large number of latent variables. startgrid could run
a long time because it runs all possible combinations. If you have 10 latent variables, that
means 103 = 1,000 likelihood evaluations.
If you have a large number of latent variables, rerun your difficult gsem command including
option iterate(#) and look at the results. Identify the problematic variances and search
across them only. Do not just look for variances going to 0. Variances getting really big can
be a problem, too, and even reasonable values can be a problem. Use your knowledge and
intuition about the model.
Perhaps you will try to fit your model by specifying startgrid(.1 1 10 e.F L e.X).
Because values 0.1, 1, and 10 are the default, you could equivalently specify startgrid(e.F
L e.X).
Look at covariances as well as variances. If you expect a covariance to be negative and it is
positive, try negative starting values for the covariance by specifying startgrid(-.1 -1
-10 G*H).
Remember that you can have multiple startgrid() options, and thus you could specify
startgrid(e.F L e.X) startgrid(-.1 -1 -10 G*H).
2. If you got initial values not feasible, you know that gsem already tried the default
startgrid.
The default startgrid only tried the values 0.1, 1, and 10, and only tried them on the
variances of latent variables. You may need to try different values or try the same values on
covariances or variances of errors of observed endogenous variables.
We suggest you first rerun the model causing difficulty including the noestimate option.
We also direct you back to the idea of first simplifying the model; see Temporarily simplify
the model.

intro 12 Convergence problems and how to solve them

121

If, looking at the results, you have an idea of which variance or covariance is a problem,
or if you have few variances and covariances, we would recommend running startgrid()
first. On the other hand, if you have no idea as to which variance or covariance is the
problem and you have a large number of them, you will be better off if you first simplify
the model. If, after doing that, your simplified model does not include all the variances and
covariances, you can specify a combination of from() and startgrid().

Also see
[SEM] intro 11 Fitting models with summary statistics data (sem only)

Title
Builder SEM Builder
Description

Remarks and examples

Reference

Description
The SEM Builder lets you create path diagrams for SEMs, fit those models, and show results on the
path diagram. Here we discuss standard linear SEMs; see [SEM] Builder, generalized for information
on using the Builder to create models with generalized responses and multilevel structure.

Remarks and examples

Launch the SEM Builder by selecting the menu item Statistics > SEM (structural equation
modeling) > Model building and estimation. You can also type sembuilder in the Command
window.
Select the Add Observed Variable tool,
, and click within the diagram to add an observed
variable. Type a name in the Name control of the Contextual Toolbar to name the variable. If the
variable is not placed exactly where you want it, simply select the Select tool and drag the variable
to your preferred location.
122

Builder SEM Builder

123

After adding the variable, use the Variable control in the Contextual Toolbar to select a variable
from your dataset or type a variable name into the edit field.
Add latent variables to the model by using the

tool.

Clicking within the diagram with either the

or
tool places a variable of the default size
and shape. If you instead wish to create a variable of custom size and shape, click and hold the
mouse button and drag until the variable is your preferred size. Select the Properties button in the
Contextual Toolbar or double-click on a variable to launch a dialog box where you can change even
more properties of the variable. You can customize the size, shape, label font, and other appearance
characteristics of the variable from the Appearance tab of the dialog, but you will rarely do that.
More often, you will want to change the appearance of all variables or a class of variables
from the Settings menu. From the Settings menu, you can change any aspect of the appearance of
variables. The Settings > Variables menu lets you change the appearance of All variables, Latent
variables, Observed variables, Latent Exogenous variables, Latent Endogenous variables, Observed
Exogenous variables, Observed Endogenous variables, Error variables, Latent Error variables, and
Observed Error variables.
Draw paths between variables with the
tool. Simply click in the source variables box or oval
and drag to the target variables box or oval. The new path can connect variables either along the
line between their center points or at the edge nearest where you click and release. This connection
behavior is set in the Settings > Automation... dialog.
Whenever a path is connected to an exogenous variable, that variable becomes endogenous and
an error variable is created. If you do not like the direction of the error variable relative to the new
tool to select the endogenous variable and then use the
endogenous variable, first use the
buttons in the Contextual Toolbar to rotate the error variable.
Draw covariances between variables with the
tool. If you drag to the right, the curve of the
covariance will bend up. If you drag to the left, the curve will bend down. If you drag down, the
curve will bend right, and if you drag up, it will bend left. That does not matter greatly; if the curve
bends the opposite of what you want, simply click the Mirror button in the Contextual Toolbar.
As with variables, you can change more properties of a path or covariance (connection) by
double-clicking on it or by selecting Properties from the Contextual Toolbar. You can also change the
appearance of all connections from the Settings > Connections menu. From Settings > Connections,
you can change the appearance of All connections, all Paths, all Covariances, or all Error Paths.
Manage variables and connections on the diagram with the Select tool,
. As you select an
object, its Contextual Toolbar will appear. You can change the name and other properties of selected
variables or the properties of selected connections. You can also drag-select or hold the Shift key
and click to select multiple objects. Changing a property from the Contextual Toolbar or Properties
dialog will affect all the selected objects. You can also drag selected variables as a group to other
locations on the diagram.
You can place a measurement model (or measurement component of a larger model) on the
diagram with the
tool. Clicking on the diagram with this tool launches a dialog box where you
can name the latent variable, select or type the names of the measurement variables, and specify the
direction of the measurements relative to the latent variable. You can even set the spacing between
the measurement variables and the distance from the latent variable to the measurements. The
tool creates measurement components quickly and with even spacing. Similarly, neatly organized sets
of observed variables can be placed with the
tool; sets of latent variables, with the
tool; and
regression components, with the
tool.

124

Builder SEM Builder

Place annotations and other text by using the

tool.

We have ignored the locks, , in the Contextual Toolbars. These locks apply constraints to the
parameters of the SEM. You can constrain variances,
; means,
; intercepts,
; and path
coefficients and covariances,
. For example, select an exogenous variable or an error variable and
type a number in
to constrain that variance to a fixed value. Or, select three path variables and
type a name (a placeholder) in
to constrain all the path coefficients to be equal. You can type
numbers, names, or linear expressions in the
controls. The linear expressions can involve only
controls.
numbers and names that are used in other
Do not be afraid to try things. If you do not know what a tool, control, or dialog item does, try
it. If you do not like the result, click on the Undo button, , in the Standard Toolbar.
Click on
in the Standard Toolbar to fit the model. A dialog is launched that allows you to set
all the estimation options that are not defined by the path diagram. After estimation, some of the
estimation results are displayed on the path diagram. Use the Results tab of the Settings > Variables
and Settings > Connections dialogs to change what results are shown and how they appear (font
sizes, locations, etc.). Also, as you click on connections or variables, the Details pane displays all
the estimation results for the selected object.
If you wish to create another model derived from the current model, click
Toolbar.

in the Standard

Video example
SEM Builder in Stata
Visit the Stata YouTube channel for more videos.

Reference
Huber, C. 2012. Using Statas SEM features to model the Beck Depression Inventory. The Stata Blog: Not Elsewhere
Classified. http://blog.stata.com/2012/10/17/using-statas-sem-features-to-model-the-beck-depression-inventory/.

Title
Builder, generalized SEM Builder for generalized models

Description

Remarks and examples

Reference

Description
The SEM Builder allows you to create path diagrams of generalized SEMs, fit those models, and
show the results on the path diagram. Here we extend the discussion from [SEM] Builder. Read that
manual entry first.

Remarks and examples

Launch the SEM Builder by selecting the menu item Statistics > SEM (structural equation
modeling) > Model building and estimation. You can also type sembuilder in the Command
window. Switch the Builder into generalized SEM mode by clicking on the Change to Generalized
SEM button, . The generalized SEM mode allows you to add generalized responses and multilevel
latent variables to your path diagram. This mode also disallows constraints on the means or variances
of observed exogenous variables and on the variances of errors. In technical terms, the generalized
mode draws and fits gsem models rather than sem models; see [SEM] intro 1 for a discussion.
Aside from activating tools to add generalized responses and multilevel latent variables, everything
about the Builder works the same in generalized SEM mode as it did in standard SEM mode. You can
still add observed variables ( ) and latent variables ( ) and create connections using paths ( )
and covariances ( ).
You can now add to your diagram a special kind of observed variable called a generalized response
variable by using the
tool. A generalized response is a dependent variable that may be continuous,
count, binary, ordinal, or multinomial, and whose statistical properties are described by its family and
link; see [SEM] intro 2 for a discussion. Simply click on the
tool and then click in the diagram
to add a generalized response variable. After adding the generalized response, use the Contextual
Toolbar to do the following: 1) select its Family/Link and 2) select its variable from the Variable
control. You may also type a variable name into the Variable control.
You can also add multilevel latent variablesrandom effects assumed to have a Gaussian distribution
whose realizations are at the level of a grouping variable or variables. See [SEM] intro 2 for a discussion.
tool and then click on the diagram to add a multilevel latent variable. After adding the
Click on the
variable, click on the
button of the Name/Level control in the Contextual Toolbar. In the resulting
dialog box, choose the nesting depth of your latent variable, and then choose the variables that define
the nesting. For example, for a three-level model of doctors nesting patients nesting observations, you
would select 3 in the Nesting depth control, select doctor in the first variable control, and select
patient in the second variable control, leaving > Observations, which is fixed. You do not need
to change the default Base name, but you may type a different name of your choice. Do read the note
on the dialog carefully because it provides guidance on successfully creating multilevel models. Also
see [SEM] example 39g for an example that demonstrates creating a nested model in the Builder.
125

126

Builder, generalized SEM Builder for generalized models

As with standard observed and latent variables, you can also click and drag as you add generalized
responses and multilevel latent variables to control the size of the resulting variables. You can also
select the Properties button from the Contextual Toolbar or double-click on a variable to change
even more properties of the variable. As with standard variables, you can control their appearance
from the Appearance tab of the resulting dialog box or, more likely, control the appearance of all
variables of a given type from the Settings > Variables menu.
the

As with standard models, you create the structure of your model by connecting the variables with
tool. You may also specify covariances among variables with the
tool.

There is one other construct available in generalized mode that is not available in standard mode
paths to other paths. In addition to creating paths from a latent variable to an observed variable or
to another latent variable at the same nesting level, you can connect a standard or multilevel latent
variable to a path from one observed variable to another observed variable. This makes that path a
random path, or a random slope. As with other path connections, think of the path as adding the
source variable to the target. So we are adding a random variable to a path (or slope), making it a
random path. See [SEM] example 38g for an example of creating a random path in the Builder.
In generalized mode, the Add Measurement Component tool,
, allows the measurements to
be created as generalized responses. Simply click on the Make measurements generalized box on
the dialog box and select a Family/Link. It also allows the latent variable to be either standard
observation-level or multilevel with the radio control at the top of the dialog box.
The Add Observed Variables Set tool, , also allows you to create multiple generalized responses.
Click on the Make variables generalized responses box on the dialog box and select a Family/Link.
Creating multinomial logit responses can be tricky. You must create a generalized response for
each level of your outcome and use factor-variable notation to designate the levels of the outcome.
tool is the safest way to create a multinomial logit response. With your dataset in memory,
The
select the
tool and then click on the diagram. Click on the Make variables generalized responses
box on the dialog box, and then select Multinomial, Logit in the Family/Link control. Select your
multinomial response variable from the Variable control, and the levels for your multinomial response
will be automatically populated. The level designated with a b will be the base level. You can type
the b on another level to make it the base. When you click OK, multinomial generalized responses
will be created for each level of your variable. See [SEM] example 37g for an example of creating
multinomial logit responses in the Builder.
The Add Regression Component tool,
generalized response.

, also allows the dependent variable to be specified as a

Place shaded boxes on the diagram with the

tool. This tool can be used in both standard and
generalized mode. It is most useful in the generalized mode, where we can shade different nesting
levels of multilevel models.
As in standard mode, constraints are specified with lock controls
using the
button on the Standard Toolbar.

Video example
SEM Builder in Stata
Visit the Stata YouTube channel for more videos.

, and you fit the model by

Builder, generalized SEM Builder for generalized models

127

Reference
Huber, C. 2012. Using Statas SEM features to model the Beck Depression Inventory. The Stata Blog: Not Elsewhere
Classified. http://blog.stata.com/2012/10/17/using-statas-sem-features-to-model-the-beck-depression-inventory/.

Title
estat eform Display exponentiated coefficients
Syntax
Remarks and examples

Menu
Also see

Description

Options

Syntax
estat eform

eqnamelist

 

, level(#) display options

where eqnamelist is a list of equation names. In gsem, equation names correspond to the names of
the response variables. If no eqnamelist is specified, exponentiated results for the first equation
are shown.

Menu
Statistics

>

SEM (structural equation modeling)

>

Other

>

Display exponentiated coefficients

Description
estat eform is for use after gsem but not sem.
gsem reports coefficients. You can obtain exponentiated coefficients and their standard errors by
using estat eform after estimation to redisplay results.

Options
level(#); see [R] estimation options; default is level(95).
display options control the display of factor variables and more. Allowed display options are
noomitted, vsquish, noemptycells, baselevels, allbaselevels, nofvlabel, fvwrap(#),
fvwrapon(style), cformat(% fmt), pformat(% fmt), sformat(% fmt), and nolstretch. See
[R] estimation options.

Remarks and examples

In some generalized linear response functions, exponentiated coefficients have a special meaning.
Those special meanings are as follows:
Common name

Family

Link

Meaning of exp(coef)

logit
ologit
mlogit
Poisson
nbreg

Bernoulli
ordinal
multinomial
Poisson
nbreg

logit
logit
logit
log
log

odds ratio
odds ratio
relative-risk ratio
incidence-rate ratio
incidence-rate ratio

See [SEM] example 33g and [SEM] example 34g.

128

estat eform Display exponentiated coefficients

Also see
[SEM] gsem Generalized structural equation model estimation command
[SEM] gsem postestimation Postestimation tools for gsem
[SEM] intro 7 Postestimation tests and predictions
[SEM] example 33g Logistic regression
[SEM] example 34g Combined models (generalized responses)

129

Title
estat eqgof Equation-level goodness-of-fit statistics
Syntax
Remarks and examples

Menu
Stored results

Description
Reference

Option
Also see

Syntax
estat eqgof

, format(% fmt)

Menu
Statistics

>

SEM (structural equation modeling)

>

Goodness of fit

>

Equation-level goodness of fit

Description
estat eqgof is for use after sem but not gsem.
estat eqgof displays equation-by-equation goodness-of-fit statistics. Displayed are R2 and the
BentlerRaykov squared multiple-correlation coefficient (Bentler and Raykov 2000).
These two concepts of fit are equivalent for recursive SEMs and univariate linear regression. For
nonrecursive SEMs, these measures are distinct.
Equation-level variance decomposition is also reported, along with the overall model coefficient
of determination.

Option
format(% fmt) specifies the display format. The default is format(%9.0f).

Remarks and examples

See [SEM] example 3.
In rare circumstances, these equation-level goodness-of-fit measures in nonrecursive structural
equations have unexpected values. It is possible to obtain negative R2 and multiple-correlation values.
It is recommended to use the BentlerRaykov squared multiple correlations as a measure of explained
variance for nonrecursive systems that involve endogenous variables with reciprocal causations.

Stored results
estat eqgof stores the following in r():
Scalars
r(N groups)

r(CD # )
Matrices
r(nobs)

r(eqfit # )

number of groups
overall coefficient of determination (for group #)
sample size for each group
fit statistics (for group #)

130

estat eqgof Equation-level goodness-of-fit statistics

131

Reference
Bentler, P. M., and T. Raykov. 2000. On measures of explained variance in nonrecursive structural equation models.
Journal of Applied Psychology 85: 125131.

Also see
[SEM] example 3 Two-factor measurement model

[SEM] estat gof Goodness-of-fit statistics

[SEM] estat ggof Group-level goodness-of-fit statistics
[SEM] methods and formulas for sem Methods and formulas for sem
[SEM] sem postestimation Postestimation tools for sem

Title
estat eqtest Equation-level test that all coefficients are zero
Syntax
Remarks and examples

Menu
Stored results

Description
Also see

Option

Syntax
estat eqtest

, total nosvyadjust

Menu
Statistics

>

SEM (structural equation modeling)

>

Testing and CIs

>

Equation-level Wald tests

Description
estat eqtest is for use after sem but not gsem.
estat eqtest displays Wald tests that all coefficients excluding the intercept are 0 for each
equation in the model.

Option
total is for use when estimation was with sem, group(). It specifies that the tests be aggregated
across the groups.
nosvyadjust is for use with svy estimation commands. It specifies that the Wald test be carried out
without the default adjustment for the design degrees of freedom. That is to say the test is carried
out as W/k F (k, d) rather than as (d k + 1)W/(kd) F (k, d k + 1), where k is the
dimension of the test and d is the total number of sampled PSUs minus the total number of strata.

Remarks and examples

See [SEM] example 13.

Stored results
estat eqtest stores the following in r():
Scalars
r(N groups)
Matrices
r(nobs)
 
r(test # )
r(test total)

number of groups
sample size for each group
test statistics (for group #)
aggregated test statistics (total only)

132

estat eqtest Equation-level test that all coefficients are zero

Also see
[SEM] example 13 Equation-level Wald test

[SEM] test Wald test of linear hypotheses

[SEM] lrtest Likelihood-ratio test of linear hypothesis
[SEM] methods and formulas for sem Methods and formulas for sem
[SEM] sem postestimation Postestimation tools for sem

133

Title
estat framework Display estimation results in modeling framework
Syntax
Remarks and examples

Menu
Stored results

Description
Reference

Options
Also see

Syntax
estat framework

, options

options

Description

standardized
compact
fitted
format(% fmt)

report standardized results

display matrices in compact form
include fitted means, variances, and covariances
display format to use

Menu
Statistics

>

SEM (structural equation modeling)

>

Other

>

Report model framework

Description
estat framework is a postestimation command for use after sem but not gsem.
estat framework displays the estimation results as a series of matrices derived from the Bentler
Weeks form; see Bentler and Weeks (1980).

Options
standardized reports results in standardized form.
compact displays matrices in compact form. Zero matrices are displayed as a description. Diagonal
matrices are shown as a row vector.
fitted displays the fitted mean and covariance values.
format(% fmt) specifies the display format to be used. The default is format(%9.0g).

Remarks and examples

See [SEM] example 11.

134

estat framework Display estimation results in modeling framework

135

Technical note
If sems nm1 option was specified when the model was fit, all covariance matrices are calculated
using N 1 in the denominator instead of N .

Stored results
estat framework stores the following in r():
Scalars
r(N groups)
r(standardized)

number of groups
indicator of standardized results (+)

Matrices
r(nobs)

sample size for each group

# )

r(Beta

# )

coefficients of endogenous variables on exogenous variables (for group #)

intercepts (for group #) (*)

r(Gamma
r(alpha

# )

covariances of errors (for group #)

covariances of exogenous variables (for group #)

# )

r(Phi

# )

means of exogenous variables (for group #) (*)

fitted covariances (for group #)

r(kappa

r(Sigma
r(mu

# )

r(Psi

coefficients of endogenous variables on endogenous variables (for group #)

# )

# )

fitted means (for group #) (*)

(+) If r(standardized)=1, the returned matrices contain standardized values.

(*) If there are no estimated means or intercepts in the sem model, these matrices are not returned.

Reference
Bentler, P. M., and D. G. Weeks. 1980. Linear structural equations with latent variables. Psychometrika 45: 289308.

Also see
[SEM] example 11 estat framework

[SEM] intro 7 Postestimation tests and predictions (Replaying the model (sem and gsem))
[SEM] intro 7 Postestimation tests and predictions (Accessing stored results)
[SEM] methods and formulas for sem Methods and formulas for sem
[SEM] sem postestimation Postestimation tools for sem

Title
estat ggof Group-level goodness-of-fit statistics
Syntax
Remarks and examples

Menu
Stored results

Description
Also see

Option

Syntax
estat ggof

, format(% fmt)

Menu
Statistics

>

SEM (structural equation modeling)

>

Group statistics

>

Group-level goodness of fit

Description
estat ggof is for use after estimation with sem, group().
estat ggof displays, by group, the standardized root mean squared residual (SRMR), the coefficient
of determination (CD), and the model versus saturated 2 along with its associated degrees of freedom
and p-value.

Option
format(% fmt) specifies the display format. The default is format(%9.3f).

Remarks and examples

See [SEM] example 21.
estat ggof provides group-level goodness-of-fit statistics after estimation by sem, group(); see
[SEM] sem group options.

Stored results
estat ggof stores the following in r():
Scalars
r(N groups)
Matrices
r(gfit)

number of groups
fit statistics

136

estat ggof Group-level goodness-of-fit statistics

Also see
[SEM] example 21 Group-level goodness of fit

[SEM] sem group options Fitting models on different groups

[SEM] estat gof Goodness-of-fit statistics
[SEM] estat eqgof Equation-level goodness-of-fit statistics
[SEM] methods and formulas for sem Methods and formulas for sem
[SEM] sem postestimation Postestimation tools for sem

137

Title
estat ginvariant Tests for invariance of parameters across groups
Syntax
Remarks and examples

Menu
Stored results

Description
References

Options
Also see

Syntax
estat ginvariant

, options

options

Description

showpclass(classname)
class
legend

restrict output to parameters in the specified parameter class

include joint tests for parameter classes
include legend describing parameter classes

classname

Description

scoef
scons

structural coefficients
structural intercepts

mcoef
mcons

measurement coefficients
measurement intercepts

serrvar
merrvar
smerrcov

covariances of structural errors

covariances of measurement errors
covariances between structural and measurement errors

meanex
covex

means of exogenous variables

covariances of exogenous variables

all
none

all the above

none of the above

Menu
Statistics

>

SEM (structural equation modeling)

>

Group statistics

>

Test invariance of parameters across groups

Description
estat ginvariant is for use after estimation with sem, group(); see [SEM] sem group options.
estat ginvariant performs score tests (Lagrange multiplier tests) and Wald tests of whether
parameters constrained to be equal across groups should be relaxed and whether parameters allowed
to vary across groups could be constrained.
See Sorbom (1989) and Wooldridge (2010, 421428).
138

estat ginvariant Tests for invariance of parameters across groups

139

Options
showpclass(classname) displays tests for the classes specified. showpclass(all) is the default.
class displays a table with joint tests for group invariance for each of the nine parameter classes.
legend displays a legend describing the parameter classes. This option may only be used with the
class option.

Remarks and examples

See [SEM] example 22.

Stored results
estat ginvariant stores the following in r():
Scalars
r(N groups)
Matrices
r(nobs)
r(test)
r(test pclass)
r(test class)

number of groups
sample size for each group
Wald and score tests
parameter classes corresponding to r(test)
joint Wald and score tests for each class

References
Sorbom, D. 1989. Model modification. Psychometrika 54: 371384.
Wooldridge, J. M. 2010. Econometric Analysis of Cross Section and Panel Data. 2nd ed. Cambridge, MA: MIT Press.

Also see
[SEM] example 22 Testing parameter equality across groups

[SEM] estat mindices Modification indices

[SEM] estat scoretests Score tests
[SEM] methods and formulas for sem Methods and formulas for sem
[SEM] sem postestimation Postestimation tools for sem

Title
estat gof Goodness-of-fit statistics
Syntax
Remarks and examples

Menu
Stored results

Description
References

Options
Also see

Syntax
estat gof

, options

options

Description

stats(statlist)
nodescribe

statistics to be displayed
suppress descriptions of statistics

statlist

Description

chi2
rmsea
ic
indices
residuals
all

2 tests; the default

root mean squared error of approximation
information indices
indices for comparison against baseline
measures based on residuals
all the above

Note: The statistics reported by chi2, rmsea, and indices are dependent on the assumption of joint
normality of the observed variables.

Menu
Statistics

>

SEM (structural equation modeling)

>

Goodness of fit

>

Overall goodness of fit

Description
estat gof is for use after sem but not gsem.
estat gof displays a variety of overall goodness-of-fit statistics.

Options
stats(statlist) specifies the statistics to be displayed. The default is stats(chi2).
stats(chi2) reports the model versus saturated test and the baseline versus saturated test. The
saturated model is the model that fits the covariances perfectly.
The model versus saturated test is a repeat of the test reported at the bottom of the sem output.
In the baseline versus saturated test, the baseline model includes the means and variances of all
observed variables plus the covariances of all observed exogenous variables. For a covariance
model (a model with no endogenous variables), the baseline includes only the means and
variances of observed variables. Be aware that different authors define the baseline model
differently.
140

estat gof Goodness-of-fit statistics

141

stats(rmsea) reports the root mean squared error of approximation (RMSEA) and its 90%
confidence interval, and pclose, the p-value for a test of close fit, namely, RMSEA < 0.05. Most
interpreters of this test label the fit close if the lower bound of the 90% CI is below 0.05 and
label the fit poor if the upper bound is above 0.10. See Browne and Cudeck (1993).
stats(ic) reports the Akaike information criterion (AIC) and Bayesian (or Schwarz) information
criterion (BIC). These statistics are available only after estimation with sem method(ml) or
method(mlmv). These statistics are used not to judge fit in absolute terms but instead to
compare the fit of different models. Smaller values indicate a better fit. Be aware that there
are many variations (minor adjustments) to statistics labeled AIC and BIC. Reported here are
statistics that match estat ic; see [R] estat ic.
To compare models that use statistics based on likelihoods, such as AIC and BIC, models
should include the same variables; see [SEM] lrtest. See Akaike (1987), Schwarz (1978), and
Raftery (1993).
stats(indices) reports CFI and TLI, two indices such that a value close to 1 indicates a good
fit. CFI stands for comparative fit index. TLI stands for TuckerLewis index and is also known
as the nonnormed fit index. See Bentler (1990).
stats(residuals) reports the standardized root mean squared residual (SRMR) and the coefficient
of determination (CD).
A perfect fit corresponds to an SRMR of 0. A good fit is a small value, considered by some to
be limited to 0.08. SRMR is calculated using the first and second moments unless sem option
nomeans was specified or implied, in which case SRMR is calculated based on second moments
only. Some software packages ignore the first moments even when available. See Hancock and
Mueller (2006, 157).
Concerning CD, a perfect fit corresponds to a CD of 1. CD is like R2 for the whole model.
stats(all) reports all the statistics. You can also specify just the statistics you wish reported,
such as
. estat gof, stats(indices residuals)

nodescribe suppresses the descriptions of the goodness-of-fit measures.

Remarks and examples

See [SEM] example 4.

142

estat gof Goodness-of-fit statistics

Stored results
estat gof stores the following in r():
Scalars
r(chi2 ms)
r(df ms)
r(p ms)
r(chi2 bs)
r(df bs)
r(p bs)
r(rmsea)
r(lb90 rmsea)
r(ub90 rmsea)
r(pclose)
r(aic)
r(bic)
r(cfi)
r(tli)
r(cd)
r(srmr)
r(N groups)

test of target model against saturated model

degrees of freedom for r(chi2 ms)
p-value for r(chi2 ms)
test of baseline model against saturated model
degrees of freedom for r(chi2 bs)
p-value for r(chi2 bs)
root mean squared error of approximation
lower bound of 90% CI for RMSEA
upper bound of 90% CI for RMSEA
p-value for test of close fit: RMSEA < 0.05
Akaike information criterion
Bayesian information criterion
comparative fit index
TuckerLewis fit index
coefficient of determination
standardized root mean squared residual
number of groups

Matrices
r(nobs)

sample size for each group

References
Akaike, H. 1987. Factor analysis and AIC. Psychometrika 52: 317332.
Bentler, P. M. 1990. Comparative fit indexes in structural models. Psychological Bulletin 107: 238246.
Browne, M. W., and R. Cudeck. 1993. Alternative ways of assessing model fit. Reprinted in Testing Structural
Equation Models, ed. K. A. Bollen and J. S. Long, pp. 136162. Newbury Park, CA: Sage.
Hancock, G. R., and R. O. Mueller, ed. 2006. Structural Equation Modeling: A Second Course. Charlotte, NC:
Information Age Publishing.
Raftery, A. E. 1993. Bayesian model selection in structural equation models. Reprinted in Testing Structural Equation
Models, ed. K. A. Bollen and J. S. Long, pp. 163180. Newbury Park, CA: Sage.
Schwarz, G. 1978. Estimating the dimension of a model. Annals of Statistics 6: 461464.

Also see
[SEM] example 4 Goodness-of-fit statistics

[SEM] estat ggof Group-level goodness-of-fit statistics

[SEM] estat eqgof Equation-level goodness-of-fit statistics
[SEM] estat residuals Display mean and covariance residuals
[R] estat ic Display information criteria
[SEM] methods and formulas for sem Methods and formulas for sem
[SEM] sem postestimation Postestimation tools for sem

Title
estat mindices Modification indices
Syntax
Remarks and examples

Menu
Stored results

Description
References

Options
Also see

Syntax
estat mindices

, options

options

Description

showpclass(classname)

restrict output to parameters in the specified parameter classes

minchi2(#)

display only tests with modification index (MI) #

classname

Description

scoef
scons

structural coefficients
structural intercepts

mcoef
mcons

measurement coefficients
measurement intercepts

serrvar
merrvar
smerrcov

covariances of structural errors

covariances of measurement errors
covariances between structural and measurement errors

meanex
covex

means of exogenous variables

covariances of exogenous variables

all
none

all the above

none of the above

Menu
Statistics

>

SEM (structural equation modeling)

>

Testing and CIs

>

Modification indices

Description
estat mindices is for use after sem but not gsem.
estat mindices reports modification indices for omitted paths in the fitted model. Modification
indices are score tests (Lagrange multiplier tests) for the statistical significance of the omitted paths.
See Sorbom (1989) and Wooldridge (2010, 421428).

143

144

estat mindices Modification indices

Options
showpclass(classname) specifies that results be limited to parameters that belong to the specified
parameter classes. The default is showpclass(all).
minchi2(#) suppresses listing paths with modification indices (MIs) less than #. By default,
estat mindices lists values significant at the 0.05 level, corresponding to 2 (1) value
minchi2(3.8414588). Specify minchi2(0) if you wish to see all tests.

Remarks and examples

See [SEM] example 5.

Stored results
estat mindices stores the following in r():
Scalars
r(N groups)
Matrices
r(nobs)
r(mindices pclass)
r(mindices)

number of groups
sample size for each group
parameter class of modification indices
matrix containing the displayed table values

References
Sorbom, D. 1989. Model modification. Psychometrika 54: 371384.
Wooldridge, J. M. 2010. Econometric Analysis of Cross Section and Panel Data. 2nd ed. Cambridge, MA: MIT Press.

Also see
[SEM] example 5 Modification indices

[SEM] estat scoretests Score tests

[SEM] estat ginvariant Tests for invariance of parameters across groups
[SEM] methods and formulas for sem Methods and formulas for sem
[SEM] sem postestimation Postestimation tools for sem

Title
estat residuals Display mean and covariance residuals

Syntax
Remarks and examples

Menu
Stored results

Description
References

Options
Also see

Syntax
estat residuals

, options

options

Description

normalized
standardized
sample
nm1
zerotolerance(tol)
format(% fmt)

report normalized residuals

report standardized residuals
use sample covariances in residual variance calculations
use adjustment N 1 in residual variance calculations
apply tolerance to treat residuals as 0
display format

Menu
Statistics

>

SEM (structural equation modeling)

>

Goodness of fit

>

Matrices of residuals

Description
estat residuals is for use after sem but not gsem.
estat residuals displays the mean and covariance residuals. Normalized and standardized
residuals are available.
Both mean and covariance residuals are reported unless sems option nomeans was specified or
implied at the time the model was fit, in which case mean residuals are not reported.
estat residuals usually does not work following sem models fit with method(mlmv). It also
does not work if there are any missing values, which after all is the whole point of using method(mlmv).

Options
normalized and standardized are alternatives. If neither is specified, raw residuals are reported.
Normalized residuals and standardized residuals attempt to adjust the residuals in the same way,
but they go about it differently. The normalized residuals are always valid, but they do not follow
a standard normal distribution. The standardized residuals do follow a standard normal distribution
but only if they can be calculated; otherwise, they will equal missing values. When both can be
calculated (equivalent to both being appropriate), the normalized residuals will be a little smaller
than the standardized residuals. See Joreskog and Sorbom (1986).
145

146

estat residuals Display mean and covariance residuals

sample specifies that the sample variance and covariances be used in variance formulas to compute
normalized and standardized residuals. The default uses fitted variance and covariance values as
described by Bollen (1989).
nm1 specifies that the variances be computed using N 1 in the denominator rather than using sample
size N .
zerotolerance(tol) treats residuals within tol of 0 as if they were 0. tol must be a numeric
value less than 1. The default is zerotolerance(0), meaning that no tolerance is applied.
When standardized residuals cannot be calculated, it is because a variance calculated by the
Hausman (1978) theorem turns negative. Applying a tolerance to the residuals turns some residuals
into 0 and then division by the negative variance becomes irrelevant, and that may be enough to
solve the calculation problem.
format(% fmt) specifies the display format. The default is format(%9.3f).

Remarks and examples

See [SEM] example 10.

Stored results
estat residuals stores the following in r():
Scalars
r(N groups)

number of groups

Macros
r(sample)
r(nm1)

empty or sample, if sample was specified

empty or nm1, if nm1 was specified

Matrices
r(nobs)

sample size for each group

r(res mean

# )

raw mean residuals (for group #) (*)

raw covariance residuals (for group #)

# )

r(res cov

r(nres mean

r(nres cov

r(sres mean

r(sres cov

# )

normalized mean residuals (for group #) (*)

# )

normalized covariance residuals (for group #)

# )

standardized mean residuals (for group #) (*)

standardized covariance residuals (for group #)

# )

(*) If there are no estimated means or intercepts in the sem model, these matrices are not returned.

References
Bollen, K. A. 1989. Structural Equations with Latent Variables. New York: Wiley.
Hausman, J. A. 1978. Specification tests in econometrics. Econometrica 46: 12511271.
Joreskog, K. G., and D. Sorbom. 1986. Lisrel VI: Analysis of linear structural relationships by the method of
maximum likelihood. Mooresville, IN: Scientific Software.

estat residuals Display mean and covariance residuals

Also see
[SEM] example 10 MIMIC model

[SEM] estat gof Goodness-of-fit statistics

[SEM] estat ggof Group-level goodness-of-fit statistics
[SEM] estat eqgof Equation-level goodness-of-fit statistics
[SEM] methods and formulas for sem Methods and formulas for sem
[SEM] sem postestimation Postestimation tools for sem

147

Title
estat scoretests Score tests
Syntax
Remarks and examples

Menu
Stored results

Description
References

Option
Also see

Syntax
estat scoretests

, minchi2(#)

Menu
Statistics

>

SEM (structural equation modeling)

>

Testing and CIs

>

Score tests of linear constraints

Description
estat scoretests is for use after sem but not gsem.
estat scoretests displays score tests (Lagrange multiplier tests) for each of the user-specified
linear constraints imposed on the model when it was fit. See Sorbom (1989) and Wooldridge (2010,
421428).

Option
minchi2(#) suppresses output of tests with 2 (1) < #. By default, estat mindices lists values significant at the 0.05 level, corresponding to 2 (1) value minchi2(3.8414588). Specify
minchi2(0) if you wish to see all tests.

Remarks and examples

See [SEM] example 8.

Stored results
estat scoretests stores the following in r():
Scalars
r(N groups)

number of groups

Matrices
r(nobs)
r(Cns sctest)

sample size for each group

matrix containing the displayed table values

148

estat scoretests Score tests

149

References
Sorbom, D. 1989. Model modification. Psychometrika 54: 371384.
Wooldridge, J. M. 2010. Econometric Analysis of Cross Section and Panel Data. 2nd ed. Cambridge, MA: MIT Press.

Also see
[SEM] example 8 Testing that coefficients are equal, and constraining them

[SEM] estat mindices Modification indices

[SEM] estat ginvariant Tests for invariance of parameters across groups
[SEM] methods and formulas for sem Methods and formulas for sem
[SEM] sem postestimation Postestimation tools for sem

Title
estat stable Check stability of nonrecursive system
Syntax
Remarks and examples

Menu
Stored results

Description
Reference

Option
Also see

Syntax
estat stable

, detail

Menu
Statistics

>

SEM (structural equation modeling)

>

Other

>

Assess stability of nonrecursive systems

Description
estat stable is for use after sem but not gsem.
estat stable reports the eigenvalue stability index for nonrecursive models. The stability index
is computed as the maximum modulus of the eigenvalues for the matrix of coefficients on endogenous
variables predicting other endogenous variables. If the model was fit by sem with the group() option,
estat stable reports the index for each group separately.
There are two formulas commonly used to calculate the index. estat stable uses the formulation
of Bentler and Freeman (1983).

Option
detail displays the matrix of coefficients on endogenous variables predicting other endogenous
variables, also known as the matrix.

Remarks and examples

See nonrecursive (structural) model (system) in [SEM] Glossary. The issue of stability is described
there. Also see Remarks and examples of [SEM] estat teffects.

150

estat stable Check stability of nonrecursive system

151

Stored results
estat stable stores the following in r():
Scalars
r(N groups)

r(stindex

number of groups

# )

Matrices
r(nobs)

r(Beta

sample size for each group

# )

r(Im

coefficients of endogenous variables on endogenous variables (for group #)

# )

r(Re

stability index (for group #)

real parts of the eigenvalues of A (for group #)

imaginary parts of the eigenvalues of A (for group #)

# )

r(Modulus

# )

modulus of the eigenvalues of A (for group #)

Reference
Bentler, P. M., and E. H. Freeman. 1983. Tests for stability in linear structural equation systems. Psychometrika 48:
143145.

Also see
[SEM] estat teffects Decomposition of effects into total, direct, and indirect
[SEM] methods and formulas for sem Methods and formulas for sem
[SEM] sem postestimation Postestimation tools for sem

Title
estat stdize Test standardized parameters

Syntax

Menu

Description

Remarks and examples

Stored results

Also see

Syntax
estat stdize: test ...
estat stdize: lincom ...
estat stdize: testnl ...
estat stdize: nlcom ...

Menu
Statistics

>

SEM (structural equation modeling)

>

Testing and CIs

>

Testing standardized parameters

Description
estat stdize: is for use after sem but not gsem.
estat stdize: can be used to prefix test, lincom, testnl, and nlcom; see [SEM] test,
[SEM] lincom, [SEM] testnl, and [SEM] nlcom.
These commands without a prefix work in the underlying metric of SEM, which is to say path
coefficients, variances, and covariances. If the commands are prefixed with estat stdize:, they will
work in the metric of standardized coefficients and correlation coefficients. There is no counterpart
to variances in the standardized metric because variances are standardized to be 1.

Remarks and examples

See [SEM] example 16.
Exercise caution when using the estat stdize: prefix to perform tests on estimated second
moments (correlations). Do not test that correlations are 0. Instead, omit the estat stdize: prefix
and test that covariances are 0. Covariances are more likely to be normally distributed than are
correlations.

Stored results
Stored results are the results stored by the command being used with the estat stdize: prefix.

152

estat stdize Test standardized parameters

Also see
[SEM] example 16 Correlation

[SEM] test Wald test of linear hypotheses

[SEM] lincom Linear combinations of parameters
[SEM] testnl Wald test of nonlinear hypotheses
[SEM] nlcom Nonlinear combinations of parameters
[SEM] sem postestimation Postestimation tools for sem
[SEM] methods and formulas for sem Methods and formulas for sem

153

Title
estat summarize Report summary statistics for estimation sample

Syntax

Menu for estat

Description

Options

Stored results

Also see

Syntax
estat summarize

eqlist

 

, group estat summ options

Menu for estat

Statistics

>

Postestimation

>

Reports and statistics

Description
estat summarize is a standard postestimation command of Stata. This entry concerns use of
estat summarize after sem or gsem.
estat summarize reports the summary statistics in the estimation sample for the observed variables
in the model. estat summarize is mentioned here because
1. estat summarize cannot be used if sem was run on summary statistics data; see [SEM] intro 11.
2. estat summarize allows the additional option group after estimation by sem.
If you fit your model with gsem instead of sem, see [R] estat summarize.

Options
group may be specified if group(varname) was specified with sem at the time the model was fit.
It requests that summary statistics be reported by group.
estat summ options are the standard options allowed by estat summarize and are outlined in
Options of [R] estat summarize.

Stored results
See Stored results of [R] estat summarize.

Also see
[R] estat summarize Summarize estimation sample
[SEM] sem postestimation Postestimation tools for sem
[SEM] gsem postestimation Postestimation tools for gsem

154

Title
estat teffects Decomposition of effects into total, direct, and indirect
Syntax
Remarks and examples

Menu
Stored results

Description
References

Options
Also see

Syntax
estat teffects

, options

options

Description

compact
standardized
nolabel
nodirect
noindirect
nototal

do not display effects with no path

report standardized effects
display group values, not labels
do not display direct effects
do not display indirect effects
do not display total effects

display options

control column formats, row spacing, and display of omitted paths

Menu
Statistics

>

SEM (structural equation modeling)

>

Testing and CIs

>

Direct and indirect effects

Description
estat teffects is for use after sem but not gsem.
estat teffects reports direct, indirect, and total effects for each path (Sobel 1987), along with
standard errors obtained by the delta method.

Options
compact is a popular option. Consider the following model:
. sem (y1<-y2 x1) (y2<-x2)

x2 has no direct effect on y1 but does have an indirect effect. estat teffects formats all its
effects tables the same way by default, so there will be a row for the direct effect of x2 on y1
just because there is a row for the indirect effect of x2 on y1. The value reported for the direct
effect, of course, will be 0. compact says to omit these unnecessary rows.
standardized reports effects in standardized form, but standard errors of the standardized effects
are not reported.
nolabel is relevant only if estimation was with sems group() option and the group variable has a
value label. Groups are identified by group value rather than label.
nodirect, noindirect, and nototal suppress the display of the indicated effect. The default is to
display all effects.
155

156

estat teffects Decomposition of effects into total, direct, and indirect

display options: noomitted, vsquish, cformat(% fmt), pformat(% fmt), sformat(% fmt), and
nolstretch; see [R] estimation options. Although estat teffects is not an estimation command, it allows these options.

Remarks and examples

See [SEM] example 7.
Direct effects are the path coefficients in the model.
Indirect effects are all mediating effects. For instance, consider
. sem ... (y1<-y2) (y1<-x2) (y2<-x3) ..., ...

The direct effect of y2 on y1 is the path coefficient (y1<-y2).

In this example, changes in x3 affect y1, too. That is called the indirect effect and is the product
of the path coefficients (y2<-x3) and (y1<-y2). If there were other paths in the model such that
y1 changed when x3 changed, those effects would be added to the indirect effect as well. estat
teffects reports total indirect effects.
The total effect is the sum of the direct and indirect effects.
When feedback loops are present in the model, such as
. sem ... (y1<-y2) (y1<-x2) (y2<-x3 y1) ..., ...

care must be taken when interpreting indirect effects. The feedback loop is when a variable indirectly
affects itself, as y1 does in the example; y1 affects y2 and y2 affects y1. Thus in calculating the
indirect effect, the sum has an infinite number of terms although the term values get smaller and
smaller and thus usually converge to a finite result. It is important that you check nonrecursive models
for stability; see Bollen (1989, 397) and see [SEM] estat stable. Caution: if the model is unstable,
the calculation of the indirect effect can sometimes still converge to a finite result.

Stored results
estat teffects stores the following in r():
Scalars
r(N groups)
Matrices
r(nobs)
r(direct)
r(indirect)
r(total)
r(V direct)
r(V indirect)
r(V total)

number of groups
sample size for each
direct effects
indirect effects
total effects
covariance matrix of
covariance matrix of
covariance matrix of

group

the direct effects

the indirect effects
the total effects

estat teffects with the standardized option additionally stores the following in r():
Matrices
r(direct std)
r(indirect std)
r(total std)

standardized direct effects

standardized indirect effects
standardized total effects

estat teffects Decomposition of effects into total, direct, and indirect

157

References
Bollen, K. A. 1989. Structural Equations with Latent Variables. New York: Wiley.
Sobel, M. E. 1987. Direct and indirect effects in linear structural equation models. Sociological Methods and Research
16: 155176.

Also see
[SEM] estat stable Check stability of nonrecursive system
[SEM] methods and formulas for sem Methods and formulas for sem
[SEM] sem postestimation Postestimation tools for sem

Title
example 1 Single-factor measurement model
Description

Remarks and examples

Reference

Also see

Description
The single-factor measurement model is demonstrated using the following data:
. use http://www.stata-press.com/data/r13/sem_1fmm
(single-factor measurement model)
. summarize
Obs
Mean
Std. Dev.
Variable
x1
x2
x3
x4

123
123
123
123

96.28455
97.28455
97.09756
690.9837

14.16444
16.14764
15.10207
77.50737

Min

Max

54
64
62
481

131
135
138
885

. notes
_dta:
1. fictional data
2. Variables x1, x2, and x3 each contain a test score designed to measure X.
The test is scored to have mean 100.
3. Variable x4 is also designed to measure X, but designed to have mean 700.

See Single-factor measurement models in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Single-factor measurement model
Fitting the same model with gsem
Fitting the same model with the Builder
The measurement-error model interpretation

Single-factor measurement model

Below we fit the following model:

x1

x2

x3

x4

e.x1

e.x2

e.x3
158

e.x4

example 1 Single-factor measurement model

159

. sem (x1 x2 x3 x4 <- X)

Endogenous variables
Measurement:

x1 x2 x3 x4

Exogenous variables
Latent:

Fitting target model:

Iteration 0:
Iteration 1:
Iteration 2:

log likelihood = -2081.0258

log likelihood = -2080.986
log likelihood = -2080.9859

Structural equation model

Estimation method = ml
Log likelihood
= -2080.9859
( 1)

Number of obs

123

[x1]X = 1
OIM
Std. Err.

Coef.
Measurement
x1 <X
_cons

1
96.28455

(constrained)
1.271963
75.70

P>|z|

[95% Conf. Interval]

0.000

93.79155

98.77755

x2 <X
_cons

1.172364
97.28455

.1231777
1.450053

9.52
67.09

0.000
0.000

.9309398
94.4425

1.413788
100.1266

X
_cons

1.034523
97.09756

.1160558
1.356161

8.91
71.60

0.000
0.000

.8070579
94.43953

1.261988
99.75559

X
_cons

6.886044
690.9837

.6030898
6.960137

11.42
99.28

0.000
0.000

5.704009
677.3421

8.068078
704.6254

80.79361
96.15861
99.70874
353.4711
118.2068

11.66414
13.93945
14.33299
236.6847
23.82631

60.88206
72.37612
75.22708
95.14548
79.62878

107.2172
127.7559
132.1576
1313.166
175.4747

x3 <-

x4 <-

var(e.x1)
var(e.x2)
var(e.x3)
var(e.x4)
var(X)

LR test of model vs. saturated: chi2(2)

1.78, Prob > chi2 = 0.4111

The equations for this model are

x1 = 1 + X1 + e.x1
x2 = 2 + X2 + e.x2
x3 = 3 + X3 + e.x3
x4 = 4 + X4 + e.x4
Notes:
1. Variable X is latent exogenous and thus needs a normalizing constraint. The variable is anchored
to the first observed variable, x1, and thus the path coefficient is constrained to be 1. See
Identification 2: Normalization constraints (anchoring) in [SEM] intro 4.

160

example 1 Single-factor measurement model

2. The path coefficients for X->x1, X->x2, and X->x3 are 1 (constrained), 1.17, and 1.03. Meanwhile,
the path coefficient for X->x4 is 6.89. This is not unexpected; we at StataCorp generated this
data, and the true coefficients are 1, 1, 1, and 7.
3. A test for model versus saturated is reported at the bottom of the output; the 2 (2) statistic
is 1.78 and its significance level is 0.4111. We cannot reject the null hypothesis of this test.
This test is a goodness-of-fit test in badness-of-fit units; a significant result implies that there
may be missing paths in the models specification.
More mathematically, the null hypothesis of the test is that the fitted covariance matrix and mean
vector of the observed variables are equal to the matrix and vector observed in the population.

Fitting the same model with gsem

sem and gsem produce the same results for standard linear SEMs. We are going to demonstrate
that just this once.
. gsem (x1 x2 x3 x4 <- X)
Fitting fixed-effects model:
Iteration 0:
log likelihood = -2233.3283
Iteration 1:
log likelihood = -2233.3283
Refining starting values:
Grid node 0:
log likelihood =
Fitting full model:
Iteration 0:
log likelihood =
Iteration 1:
log likelihood =
Iteration 2:
log likelihood =
Generalized structural equation
Log likelihood = -2080.9859
( 1) [x1]X = 1
Coef.

-2081.0303
-2081.0303
-2080.9861
-2080.9859
model

Std. Err.

Number of obs

123

P>|z|

[95% Conf. Interval]

0.000

93.79155

98.77755

x1 <X
_cons

1
96.28455

(constrained)
1.271962
75.70

X
_cons

1.172365
97.28455

.1231778
1.450052

9.52
67.09

0.000
0.000

.9309411
94.4425

1.413789
100.1266

X
_cons

1.034524
97.09756

.1160559
1.35616

8.91
71.60

0.000
0.000

.8070585
94.43954

1.261989
99.75559

X
_cons

6.886053
690.9837

.6030902
6.96013

11.42
99.28

0.000
0.000

5.704018
677.3421

8.068088
704.6253

var(X)

118.2064

23.8262

79.62858

175.474

var(e.x1)
var(e.x2)
var(e.x3)
var(e.x4)

80.79381
96.15857
99.70883
353.4614

11.66416
13.93942
14.33298
236.6835

60.88222
72.37613
75.22718
95.14011

107.2175
127.7558
132.1577
1313.168

x2 <-

x3 <-

x4 <-

example 1 Single-factor measurement model

161

Notes:
1. Results are virtually the same. Coefficients differ in the last digit; for instance, x2<-X was 1.172364
and now it is 1.172365. The same is true of standard errors, etc. Meanwhile, variance estimates
are usually differing in the next-to-last digit; for instance, var(e.x2) was 96.15861 and is now
96.15857.
These are the kind of differences we would expect to see. gsem follows a different approach
for obtaining results that involves far more numeric machinery, which correspondingly results in
slightly less accuracy.
2. The log-likelihood values reported are the same. This model is one of the few models we could
have chosen where sem and gsem would produce the same log-likelihood values. In general, gsem
log likelihoods are on different metrics from those of sem. In the case where the model does not
include observed exogenous variables, however, they share the same metric.
3. There is no reason to use gsem over sem when both can fit the same model. sem is slightly more
accurate, is quicker, and has more postestimation features.

Fitting the same model with the Builder

Use the diagram above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/sem_1fmm

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Create the measurement component for X.
Select the Add Measurement Component tool, , and then click in the diagram about one-third
of the way down from the top and slightly left of the center.
In the resulting dialog box,
a. change the Latent variable name to X;
b. select x1, x2, x3, and x4 by using the Measurement variables control;
c. select Down in the Measurement direction control;
d. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
Notice that the constraints of 1 on the paths from the error terms to the observed measures are
implied, so we do not need to add these to our diagram.
4. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
SEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem sem_1fmm

162

example 1 Single-factor measurement model

The measurement-error model interpretation

As we pointed out in Using path diagrams to specify standard linear SEMs in [SEM] intro 2, if
we rename variable x4 to be y, we can reinterpret this measurement model as a measurement-error
model. In this interpretation, X is the unobserved true value. x1, x2, and x3 are each measurements
of X, but with error. Meanwhile, y (x4) is really something else entirely. Perhaps y is earnings, and
we believe
y = 4 + 4 X + e.y
We are interested in 4 , the effect of true X on y.
If we were to go back to the data and type regress y x1, we would obtain an estimate of 4 ,
but we would expect that estimate to be biased toward 0 because of the errors-in-variable problem.
The same applies for y on x2 and y on x3. If we do that, we obtain

4 based on regress y x1 4.09

4 based on regress y x2 3.71
4 based on regress y x3 3.70
In the sem output above, we have an estimate of 4 with the bias washed away:

4 based on sem (y<-X) 6.89

The number 6.89 is the value reported for (x4<-X) in the sem output.
That 4 might be 6.89 seems plausible because we expect that the estimate should be larger than
the estimates we obtain using the variables measured with error. In fact, we can tell you that the 6.89
estimate is quite good because we at StataCorp know that the true value of 4 is 7. Here is how we
manufactured this fictional dataset:
set
set
gen
gen
gen
gen
gen

seed 12347
obs 123
X = round(rnormal(0,10))
x1 = round(100 + X + rnormal(0, 10))
x2 = round(100 + X + rnormal(0, 10))
x3 = round(100 + X + rnormal(0, 10))
x4 = round(700 + 7*X + rnormal(0, 10))

The data recorded in sem 1fmm.dta were obviously generated using normality, the same assumption
that is most often used to justify the SEM maximum likelihood estimator. In [SEM] intro 4, we explained
that the normality assumption can be relaxed and conditional normality can usually be substituted in
its place.
So lets consider nonnormal data. Lets make X be 2 (2), a violently nonnormal distribution,
resulting in the data-manufacturing code
set
set
gen
gen
gen
gen
gen

seed 12347
obs 123
X = (rchi2(2)-2)*(10/2)
x1 = round(100 + X + rnormal(0, 10))
x2 = round(100 + X + rnormal(0, 10))
x3 = round(100 + X + rnormal(0, 10))
x4 = round(700 + 7*X + rnormal(0, 10))

example 1 Single-factor measurement model

163

All the rnormal() functions remaining in our code have to do with the assumed normality of
the errors. The multiplicative and additive constants in the generation of X simply rescale the 2 (2)
variable to have mean 100 and standard deviation 10, which would not be important except for the
subsequent round() functions, which themselves were unnecessary except that we wanted to produce
a pretty dataset when we created the original sem 1fmm.dta.
In any case, if we rerun the commands with these data, we obtain

4 based on regress y x1 3.93

4 based on regress y x2 4.44
4 based on regress y x3 3.77
4 based on sem (y<-X) 6.70
We will not burden you with the details of running simulations to assess coverage; we will just
tell you that coverage is excellent: reported test statistics and significance levels can be trusted.
By the way, errors in the variables is something that does not go away with progressively
larger sample sizes. Change the code above to produce a 100,000-observation dataset instead of a
123-observation one, and you will obtain

4 based on regress y x1 3.51

4 based on regress y x2 3.51
4 based on regress y x3 3.48
4 based on sem (y<-X) 7.00

Reference
Acock, A. C. 2013. Discovering Structural Equation Modeling Using Stata. Rev. ed. College Station, TX: Stata Press.

Also see
[SEM] sem Structural equation model estimation command
[SEM] gsem Generalized structural equation model estimation command
[SEM] intro 5 Tour of models
[SEM] example 3 Two-factor measurement model
[SEM] example 24 Reliability
[SEM] example 27g Single-factor measurement model (generalized response)

Title
example 2 Creating a dataset from published covariances

Description

Remarks and examples

Reference

Also see

Description
Williams, Eaves, and Cox (2002) publish covariances from their data. We will use those published
covariances in [SEM] example 3 to fit an SEM.
In this example, we show how to create the summary statistics dataset (SSD) that we will analyze
in that example.

Remarks and examples

Remarks are presented under the following headings:
Background
Creating the SSD
At this point, we could save the dataset and stop
Labeling the SSD
Listing the SSD

For more explanation, also see [SEM] intro 11.

Background
In Williams, Eaves, and Cox (2002), the authors report a covariance matrix in a table that looks
something like this:
Affective
1
Affective
1
2
.
.
5

2038.035

Miniscale
... 5

1631.766
1932.163

...
...

Cognitive
1
.
.

164

Cognitive
1
2

Miniscale
...

1721.830 659.795 779.519

1688.292 702.969 790.448

...
...

2061.875

775.118

871.211

...

630.518

500.128

...

example 2 Creating a dataset from published covariances

165

Creating the SSD

. clear all
. ssd init a1 a2 a3 a4 a5 c1 c2 c3 c4 c5
Summary statistics data initialized. Next use, in any order,
ssd set observations (required)
It is best to do this first.
ssd set means (optional)
Default setting is 0.
ssd set variances or ssd set sd (optional)
Use this only if you have set or will set correlations and, even
then, this is optional but highly recommended. Default setting is 1.
ssd set covariances or ssd set correlations (required)
. ssd set obs 216
(value set)
Status:
observations:
set
means: unset
variances or sd: unset
covariances or correlations: unset (required to be set)
. #delimit ;
delimiter now ;
. ssd set cov 2038.035
>
1631.766
>
1393.567
>
1657.299
>
1721.830
>
659.795
>
779.519
>
>
899.912
>
1087.409
>
>
775.235
>
677.865
>
>
821.170
>
686.391
(values set)
Status:

\
1932.163
1336.871
1647.164
1688.292
702.969
790.448
879.179

\
1313.809 \
1273.261 2034.216 \
1498.401 1677.767 2061.875 \
585.019 656.527 775.118 630.518 \
653.734 764.755 871.211 500.128 741.767 \
750.037

897.441 1008.287 648.935 679.970

\
739.157 659.867
855.272 \
785.419
622.830

751.860

895.616 550.476 603.950

669.951 802.825
728.674 ;

877.681 491.042 574.775

observations:
means:
variances or sd:
covariances or correlations:
. #delimit cr
delimiter now cr

set
unset
set
set

Notes:
1. We used #delimit to temporarily set the end-of-line character to semicolon. That was not
necessary, but it made it easier to enter the data in a way that would be subsequently more
readable. You can use #delimit only in do-files; see [P] #delimit.
2. We recommend entering SSD by using do-files. That way, you can edit the file and get it right.
3. We did not have to reset the delimiter. We could have entered the numbers on one (long) line.
That works well when there are only a few summary statistics.

166

example 2 Creating a dataset from published covariances

At this point, we could save the dataset and stop

We could save the dataset and stop right now if we wished:
. save sem_2fmm
file sem_2fmm.dta saved

Obviously, we can save the dataset anytime we wish. We know we could stop because ssd status
tells us whether there is anything more that we need to define:
. ssd status
Status:
observations:
means:
variances or sd:
covariances or correlations:

set
unset
set
set

Notes:
1. The means have not been set. The authors did not provide the means.
2. ssd status would mention if anything that was not set was required to be set.

Labeling the SSD

If we were to use ssd describe to describe these data, the output would look like this:
. ssd describe
Summary statistics data
obs:
216
vars:
10
variable name

variable label

a1
a2
a3
a4
a5
c1
c2
c3
c4
c5

We can add labels and notes to our dataset:

. label data "Affective and cognitive arousal"
. label var a1 "affective arousal 1"
. label var a2 "affective arousal 2"
. label var a3 "affective arousal 3"
. label var a4 "affective arousal 4"
. label var a5 "affective arousal 5"
. label var c1 "cognitive arousal 1"
. label var c2 "cognitive arousal 2"
. label var c3 "cognitive arousal 3"
. label var c4 "cognitive arousal 4"

example 2 Creating a dataset from published covariances

167

. label var c5 "cognitive arousal 5"

. #delimit ;
delimiter now ;
. notes: Summary statistics data containing published covariances
>
from Thomas O. Williams, Ronald C. Eaves, and Cynthia Cox,
>
2 Apr 2002, "Confirmatory factor analysis of an instrument
>
designed to measure affective and cognitive arousal",
>
_Educational and Psychological Measurement_,
>
vol. 62 no. 2, 264-283. ;
. notes: a1-a5 report scores from 5 miniscales designed to measure
>
affective arousal. ;
. notes: c1-c5 report scores from 5 miniscales designed to measure
>
cognitive arousal. ;
. notes: The series of tests, known as the VST II
>
(Visual Similes Test II) were administered to 216 children
>
ages 10 to 12. The miniscales are sums of scores of
>
5 to 6 items in VST II. ;
. #delimit cr
delimiter now cr
. ssd describe
Summary statistics data
obs:
216
Affective and cognitive arousal
vars:
10
(_dta has notes)
variable name

variable label

a1
a2
a3
a4
a5
c1
c2
c3
c4
c5

affective
affective
affective
affective
affective
cognitive
cognitive
cognitive
cognitive
cognitive

arousal
arousal
arousal
arousal
arousal
arousal
arousal
arousal
arousal
arousal

1
2
3
4
5
1
2
3
4
5

. save sem_2fmm, replace

file sem_2fmm.dta saved

Notes:
1. You can label the variables and the data, and you can add notes just as you would to any dataset.
2. You save and use SSD just as you save and use any dataset.

168

example 2 Creating a dataset from published covariances

Listing the SSD

. ssd list
Observations = 216
Means undefined; assumed to be 0
Variances implicitly defined; they are the diagonal of the covariance
matrix.
Covariances:
a1
a2
a3
a4
a5
c1
c2
2038.035
1631.766 1932.163
1393.567 1336.871 1313.809
1657.299 1647.164 1273.261 2034.216
1721.83 1688.292 1498.401 1677.767 2061.875
659.795
702.969
585.019
656.527
775.118
630.518
779.519
790.448
653.734
764.755
871.211
500.128
741.767
899.912
879.179
750.037
897.441 1008.287
648.935
679.97
775.235
739.157
659.867
751.86
895.616
550.476
603.95
821.17
785.419
669.951
802.825
877.681
491.042
574.775
c3
1087.409
677.865
686.391

c4

c5

855.272
622.83

728.674

Reference
Williams, T. O., Jr., R. C. Eaves, and C. Cox. 2002. Confirmatory factor analysis of an instrument designed to
measure affective and cognitive arousal. Educational and Psychological Measurement 62: 264283.

Also see
[SEM] example 3 Two-factor measurement model

[SEM] ssd Making summary statistics data (sem only)

Title
example 3 Two-factor measurement model

Description

Remarks and examples

References

Also see

Description
The multiple-factor measurement model is demonstrated using summary statistics dataset (SSD)
sem 2fmm.dta:
. use http://www.stata-press.com/data/r13/sem_2fmm
(Affective and cognitive arousal)
. ssd describe
Summary statistics data from
http://www.stata-press.com/data/r13/sem_2fmm.dta
obs:
216
Affective and cognitive arousal
vars:
10
25 May 2013 10:11
(_dta has notes)
variable name

variable label

a1
a2
a3
a4
a5
c1
c2
c3
c4
c5

affective
affective
affective
affective
affective
cognitive
cognitive
cognitive
cognitive
cognitive

arousal
arousal
arousal
arousal
arousal
arousal
arousal
arousal
arousal
arousal

1
2
3
4
5
1
2
3
4
5

. notes
_dta:
1. Summary statistics data containing published covariances from Thomas O.
Williams, Ronald C. Eaves, and Cynthia Cox, 2 Apr 2002, "Confirmatory
factor analysis of an instrument designed to measure affective and
cognitive arousal", _Educational and Psychological Measurement_, vol. 62
no. 2, 264-283.
2. a1-a5 report scores from 5 miniscales designed to measure affective
arousal.
3. c1-c5 report scores from 5 miniscales designed to measure cognitive
arousal.
4. The series of tests, known as the VST II (Visual Similes Test II) were
administered to 216 children ages 10 to 12. The miniscales are sums of
scores of 5 to 6 items in VST II.

See [SEM] example 2 to learn how we created this SSD.

169

170

example 3 Two-factor measurement model

Remarks and examples

Remarks are presented under the following headings:
Fitting multiple-factor measurement models
Displaying standardized results
Fitting the model with the Builder
Obtaining equation-level goodness of fit by using estat eqgof

See Multiple-factor measurement models in [SEM] intro 5 for background.

Fitting multiple-factor measurement models

Below we fit the model shown by Kline (2005, 7074, 184), namely,

Cognitive

Affective

a1

a2

a3

a4

a5

c1

c2

c3

c4

c5

10

. sem (Affective -> a1 a2 a3 a4 a5) (Cognitive -> c1 c2 c3 c4 c5)

Endogenous variables
Measurement: a1 a2 a3 a4 a5 c1 c2 c3 c4 c5
Exogenous variables
Latent:
Affective Cognitive
Fitting target model:
Iteration
Iteration
Iteration
Iteration

0:
1:
2:
3:

log
log
log
log

likelihood
likelihood
likelihood
likelihood

=
=
=
=

-9542.8803
-9539.5505
-9539.3856
-9539.3851

example 3 Two-factor measurement model

Structural equation model
Estimation method = ml
Log likelihood
= -9539.3851
( 1)
( 2)

Number of obs

216

[a1]Affective = 1
[c1]Cognitive = 1

Coef.

OIM
Std. Err.

P>|z|

[95% Conf. Interval]

Measurement
a1 <Affective

a2 <Affective

.9758098

.0460752

21.18

0.000

.885504

1.066116

a3 <Affective

.8372599

.0355086

23.58

0.000

.7676643

.9068556

a4 <Affective

.9640461

.0499203

19.31

0.000

.866204

1.061888

a5 <Affective

1.063701

.0435751

24.41

0.000

.9782951

1.149107

c1 <Cognitive

c2 <Cognitive

1.114702

.0655687

17.00

0.000

.9861901

1.243215

c3 <Cognitive

1.329882

.0791968

16.79

0.000

1.174659

1.485105

c4 <Cognitive

1.172792

.0711692

16.48

0.000

1.033303

1.312281

c5 <Cognitive

1.126356

.0644475

17.48

0.000

1.000041

1.252671

var(e.a1)
var(e.a2)
var(e.a3)
var(e.a4)
var(e.a5)
var(e.c1)
var(e.c2)
var(e.c3)
var(e.c4)
var(e.c5)
var(Affect~e)
var(Cognit~e)

384.1359
357.3524
154.9507
496.4594
191.6857
171.6638
171.8055
276.0144
224.1994
146.8655
1644.463
455.9349

43.79119
41.00499
20.09026
54.16323
28.07212
19.82327
20.53479
32.33535
25.93412
18.5756
193.1032
59.11245

307.2194
285.3805
120.1795
400.8838
143.8574
136.894
135.9247
219.3879
178.7197
114.6198
1306.383
353.6255

480.3095
447.4755
199.7822
614.8214
255.4154
215.2649
217.1579
347.2569
281.2527
188.1829
2070.034
587.8439

cov(Affec~e,
Cognitive)

702.0736

85.72272

534.0601

870.087

(constrained)

(constrained)

LR test of model vs. saturated: chi2(34)

8.19
=

0.000

88.88, Prob > chi2 = 0.0000

171

172

example 3 Two-factor measurement model

Notes:
1. In [SEM] example 1, we ran sem on raw data. In this example, we run sem on SSD. There are
no special sem options that we need to specify because of this.
2. The estimated coefficients reported above are unstandardized coefficients or, if you prefer, factor
loadings.
3. The coefficients listed at the bottom of the coefficient table that start with e. are the estimated
error variances. They represent the variance of the indicated measurement that is not measured
by the respective latent variables.
4. The above results do not match exactly (Kline 2005, 184). If we specified sem option nm1, results
are more likely to match to 3 or 4 digits. The nm1 option says to divide by N 1 rather than
by N in producing variances and covariances.

Displaying standardized results

The output will be easier to interpret if we display standardized values for paths rather than path
coefficients. A standardized value is in standard deviation units. It is the change in one variable given
a change in another, both measured in standard deviation units. We can obtain standardized values
by specifying sems standardized option, which we can do when we fit the model or when we
replay results:

example 3 Two-factor measurement model

. sem, standardized
Structural equation model
Estimation method = ml
Log likelihood
= -9539.3851
( 1) [a1]Affective = 1
( 2) [c1]Cognitive = 1

Number of obs

OIM
Std. Err.

Standardized

Coef.

Measurement
a1 <Affective

.9003553

.0143988

a2 <Affective

.9023249

a3 <Affective

216

P>|z|

[95% Conf. Interval]

62.53

0.000

.8721342

.9285765

.0141867

63.60

0.000

.8745195

.9301304

.9388883

.0097501

96.29

0.000

.9197784

.9579983

a4 <Affective

.8687982

.0181922

47.76

0.000

.8331421

.9044543

a5 <Affective

.9521559

.0083489

114.05

0.000

.9357923

.9685195

c1 <Cognitive

.8523351

.0212439

40.12

0.000

.8106978

.8939725

c2 <Cognitive

.8759601

.0184216

47.55

0.000

.8398544

.9120658

c3 <Cognitive

.863129

.0199624

43.24

0.000

.8240033

.9022547

c4 <Cognitive

.8582786

.0204477

41.97

0.000

.8182018

.8983554

c5 <Cognitive

.8930346

.0166261

53.71

0.000

.8604479

.9256212

var(e.a1)
var(e.a2)
var(e.a3)
var(e.a4)
var(e.a5)
var(e.c1)
var(e.c2)
var(e.c3)
var(e.c4)
var(e.c5)
var(Affect~e)
var(Cognit~e)

.1893602
.1858097
.1184887
.2451896
.0933991
.2735248
.2326939
.2550083
.2633578
.2024893
1
1

.0259281
.0256021
.0183086
.0316107
.015899
.0362139
.0322732
.0344603
.0350997
.0296954
.
.

.1447899
.1418353
.0875289
.1904417
.0669031
.2110086
.1773081
.1956717
.2028151
.1519049
.
.

.2476506
.2434179
.1603993
.3156764
.1303885
.354563
.3053806
.3323385
.3419733
.2699183
.
.

cov(Affec~e,
Cognitive)

.8108102

.0268853

.758116

.8635045

30.16

LR test of model vs. saturated: chi2(34)

0.000

88.88, Prob > chi2 = 0.0000

173

174

example 3 Two-factor measurement model

Notes:
1. In addition to obtaining standardized coefficients, the standardized option reports estimated
error variances as the fraction of the variance that is unexplained. Error variances were previously
unintelligible numbers such as 384.136 and 357.352. Now they are 0.189 and 0.186.
2. Also listed in the sem output are variances of latent variables. In the previous output, latent
variable Affective had variance 1,644.46 with standard error 193. In the standardized output, it
has variance 1 with standard error missing. The variances of the latent variables are standardized
to 1, and obviously, being a normalization, there is no corresponding standard error.
3. We can now see at the bottom of the coefficient table that affective and cognitive arousal are
correlated 0.81 because standardized covariances are correlation coefficients.
4. The standardized coefficients for this model can be interpreted as the correlation coefficients
between the indicator and the latent variable because each indicator measures only one factor.
For instance, the standardized path coefficient a1<-Affective is 0.90, meaning the correlation
between a1 and Affective is 0.90.

Fitting the model with the Builder

Use the diagram above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/sem_2fmm

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Change the size of the observed variables rectangles.
From the SEM Builder menu, select Settings > Variables > All Observed....
In the resulting dialog box, change the first size to .38 and click on OK.
4. Create the measurement component for affective arousal.
Select the Add Measurement Component tool, , and then click in the diagram about one-third
of the way down from the top and one-fourth of the way in from the left.
In the resulting dialog box,
a. change the Latent variable name to Affective;
b. select a1, a2, a3, a4, and a5 by using the Measurement variables control;
c. select Down in the Measurement direction control;
d. click on OK.
If you wish, move this component by clicking on any variable and dragging it.
5. Create the measurement component for cognitive arousal.
Repeat the process from item 4, but place the measurement component about one-third of the
way down from the top and three-fourths of the way in from the left. Label the latent variable
Cognitive, and select measurement variables c1, c2, c3, c4, and c5. Drag to reposition if
desired.

example 3 Two-factor measurement model

175

6. Correlate the latent factors.

a. Select the Add Covariance tool,

b. Click in the upper-right quadrant of the Affective oval (it will highlight when you hover
over it), and drag a covariance to the upper-left quadrant of the Cognitive oval (it will
highlight when you can release to connect the covariance).
7. Clean up.
If you do not like where a covariance has been connected to its variable, use the Select tool,
, to click on the covariance, and then simply click on where it connects to an oval and drag
the endpoint. You can also change the bow of the covariance by dragging the control point that
extends from one end of the selected covariance.
8. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
SEM estimation options dialog box.
9. Show standardized estimates.
From the SEM Builder menu, select View > Standardized Estimates.
You can open a completed diagram in the Builder by typing
. webgetsem sem_2fmm

Obtaining equation-level goodness of fit by using estat eqgof

That the correlation between a1 and Affective is 0.90 implies that the fraction of the variance
of a1 explained by Affective is 0.902 = 0.81, and left unexplained is 1 0.81 = 0.19. Instead
of manually calculating the proportion of variance explained by indicators, we can use the estat
eqgof command:
. estat eqgof
Equation-level goodness of fit

depvars

fitted

Variance
predicted

residual

R-squared

mc

2028.598
1923.217
1307.726
2024.798
2052.328
627.5987
738.3325
1082.374
851.311
725.3002

1644.463
1565.865
1152.775
1528.339
1860.643
455.9349
566.527
806.3598
627.1116
578.4346

384.1359
357.3524
154.9507
496.4594
191.6857
171.6638
171.8055
276.0144
224.1994
146.8655

.8106398
.8141903
.8815113
.7548104
.9066009
.7264752
.7673061
.7449917
.7366422
.7975107

.9003553
.9023249
.9388883
.8687982
.9521559
.8523351
.8759601
.863129
.8582786
.8930346

mc2

observed
a1
a2
a3
a4
a5
c1
c2
c3
c4
c5
overall

.8106398
.8141903
.8815113
.7548104
.9066009
.7264752
.7673061
.7449917
.7366422
.7975107

.9949997

mc = correlation between depvar and its prediction

mc2 = mc^2 is the Bentler-Raykov squared multiple correlation coefficient

176

example 3 Two-factor measurement model

Notes:
1. fitted reports the fitted variance of each of the endogenous variables, whether observed or
latent. In this case, we have observed endogenous variables.
2. predicted reports the variance of the predicted value of each endogenous variable.
3. residual reports the leftover residual variance.
4. R-squared reports R2 , the fraction of variance explained by each indicator. The fraction of the
variance of Affective explained by a1 is 0.81, just as we calculated by hand at the beginning
of this section. The overall R2 is also called the coefficient of determination.
5. mc stands for multiple correlation, and mc2 stands for multiple-correlation squared. R-squared,
mc, and mc2 all report the relatedness of the indicated dependent variable with the models linear
prediction. In recursive models, all three statistics are really the same number. mc is equal to the
square root of R-squared, and mc2 is equal to R-squared.
In nonrecursive models, these three statistics are different and each can have problems. R-squared
and mc can actually become negative! That does not mean the model has negative predictive
power or that it might not even have reasonable predictive power. mc2 = mc2 is recommended
by Bentler and Raykov (2000) to be used instead of R-squared for nonrecursive systems.
In [SEM] example 4, we examine the goodness-of-fit statistics for this model.
In [SEM] example 5, we examine modification indices for this model.

References
Acock, A. C. 2013. Discovering Structural Equation Modeling Using Stata. Rev. ed. College Station, TX: Stata Press.
Bentler, P. M., and T. Raykov. 2000. On measures of explained variance in nonrecursive structural equation models.
Journal of Applied Psychology 85: 125131.
Kline, R. B. 2005. Principles and Practice of Structural Equation Modeling. 2nd ed. New York: Guilford Press.

Also see
[SEM] example 1 Single-factor measurement model
[SEM] example 2 Creating a dataset from published covariances
[SEM] example 20 Two-factor measurement model by group
[SEM] example 26 Fitting a model with data missing at random
[SEM] example 31g Two-factor measurement model (generalized response)

[SEM] sem Structural equation model estimation command

[SEM] estat eqgof Equation-level goodness-of-fit statistics

Title
example 4 Goodness-of-fit statistics

Description

Remarks and examples

Reference

Also see

Description
Here we demonstrate estat gof. See [SEM] intro 7 and [SEM] estat gof.
This example picks up where [SEM] example 3 left off:
. use http://www.stata-press.com/data/r13/sem_2fmm
. sem (Affective -> a1 a2 a3 a4 a5) (Cognitive -> c1 c2 c3 c4 c5)

Remarks and examples

When we fit this model in [SEM] example 3, at the bottom of the output, we saw
. sem (Affective -> a1 a2 a3 a4 a5) (Cognitive -> c1 c2 c3 c4 c5)
(output omitted )
LR test of model vs. saturated: chi2(34)
=
88.88, Prob > chi2 = 0.0000

Most texts refer to this test against the saturated model as the model 2 test.
These results indicate poor goodness of fit; see [SEM] example 1. The default goodness-of-fit
statistic reported by sem, however, can be overly influenced by sample size, correlations, variance
unrelated to the model, and multivariate nonnormality (Kline 2011, 201).
Goodness of fit in cases of sem is a measure of how well you fit the observed moments, which in
this case are the covariances between all pairs of a1, . . . , a5, c1, . . . , c5. In a measurement model,
the assumed underlying causes are unobserved, and in this example, those unobserved causes are the
latent variables Affective and Cognitive. It may be reasonable to assume that the observed a1,
. . . , a5, c1, . . . , c5 can be filtered through imagined variables Affective and Cognitive, but that
can be reasonable only if not too much information contained in the original variables is lost. Thus
goodness-of-fit statistics are of great interest to those fitting measurement models. Goodness-of-fit
statistics are of far less interest when all variables in the model are observed.

177

178

example 4 Goodness-of-fit statistics

Other goodness-of-fit statistics are available.

. estat gof, stats(all)
Fit statistic

Value

Description

Likelihood ratio
chi2_ms(34)
p > chi2
chi2_bs(45)
p > chi2

88.879
0.000
2467.161
0.000

Population error
RMSEA
90% CI, lower bound
upper bound
pclose

0.086
0.065
0.109
0.004

Information criteria
AIC
BIC

19120.770
19191.651

Baseline comparison
CFI
TLI

0.977
0.970

Comparative fit index

Tucker-Lewis index

Size of residuals
SRMR
CD

0.022
0.995

Standardized root mean squared residual

Coefficient of determination

model vs. saturated

baseline vs. saturated

Root mean squared error of approximation

Probability RMSEA <= 0.05

Akaikes information criterion

Bayesian information criterion

Notes:
1. Desirable values vary from test to test.
2. We asked for all the goodness-of-fit tests. We could have obtained specific tests from the above
output by specifying the appropriate option; see [SEM] estat gof.
3. Under likelihood ratio, estat gof reports two tests. The first is a repeat of the model 2
test reported at the bottom of the sem output. The saturated model is the model that fits the
covariances perfectly. We can reject at the 5% level (or any other level) that the model fits as
well as the saturated model.
The second test is a baseline versus saturated comparison. The baseline model includes the mean
and variances of all observed variables plus the covariances of all observed exogenous variables.
Different authors define the baseline differently. We can reject at the 5% level (or any other level)
that the baseline model fits as well as the saturated model.
4. Under population error, the RMSEA value is reported along with the lower and upper bounds
of its 90% confidence interval. Most interpreters of this test check whether the lower bound is
below 0.05 or the upper bound is above 0.10. If the lower bound is below 0.05, then they would
not reject the hypothesis that the fit is close. If the upper bound is above 0.10, they would not
reject the hypothesis that the fit is poor. The logic is to perform one test on each end of the 90%
confidence interval and thus have 95% confidence in the result. This models fit is not close, and
its upper limit is just over the bounds of being considered poor.
Pclose, a commonly used word in reference to this test, is the probability that the RMSEA value is
less than 0.05, interpreted as the probability that the predicted moments are close to the moments
in the population. This models fit is not close.

example 4 Goodness-of-fit statistics

179

5. Under information criteria are reported AIC and BIC, which contain little information by themselves
but are often used to compare models. Smaller values are considered better.
6. Under baseline comparison are reported CFI and TLI, two indices such that a value close to 1
indicates a good fit. TLI is also known as the nonnormed fit index.
7. Under size of residuals is reported the standardized root mean squared residual (SRMR) and the
coefficient of determination (CD).
A perfect fit corresponds to an SRMR of 0, and a good fit corresponds to a small value,
considered by some to be limited at 0.08. The model fits well by this standard.
The CD is like an R2 for the whole model. A value close to 1 indicates a good fit.
estat gof provides multiple goodness-of-fit statistics because, across fields, different researchers
use different statistics. You should not print them all and look for the one reporting the result you
seek.

Reference
Kline, R. B. 2011. Principles and Practice of Structural Equation Modeling. 3rd ed. New York: Guilford Press.

Also see
[SEM] example 3 Two-factor measurement model
[SEM] example 21 Group-level goodness of fit

[SEM] estat gof Goodness-of-fit statistics

Title
example 5 Modification indices

Description

Remarks and examples

Reference

Also see

Description
Here we demonstrate the use of estat mindices; see [SEM] intro 7 and [SEM] estat mindices.
This example picks up where [SEM] example 3 left off:
. use http://www.stata-press.com/data/r13/sem_2fmm
. sem (Affective -> a1 a2 a3 a4 a5) (Cognitive -> c1 c2 c3 c4 c5)

Remarks and examples

When we fit this model in [SEM] example 4, we allowed the latent variables to be correlated. We
typed
. sem (Affective -> a1 a2 a3 a4 a5) (Cognitive -> c1 c2 c3 c4 c5)

and by default in the command language, latent exogenous variables are assumed to be correlated
unless we specify otherwise. Had we used the Builder, the latent exogenous variables would have
been assumed to be uncorrelated unless we had drawn the curved path between them.
The original authors who collected these data analyzed them assuming no covariance, which we
could obtain by typing
. sem (Affective -> a1 a2 a3 a4 a5) (Cognitive -> c1 c2 c3 c4 c5), ///
cov(Affective*Cognitive@0)

It was Kline (2005, 7074, 184) who allowed the covariance. Possibly he did that after looking at
the modification indices.
The modification indices report statistics on all omitted paths. Lets begin with the model without
the covariance:

180

example 5 Modification indices

181

. sem (Affective -> a1 a2 a3 a4 a5) (Cognitive -> c1 c2 c3 c4 c5),

>
cov(Affective*Cognitive@0)
(output omitted )
. estat mindices
Modification indices

MI

df

P>MI

EPC

Standard
EPC

Cognitive

8.059

0.00

.1604476

.075774

Affective

5.885

0.02

.0580897

.087733

5.767
7.597
14.300
4.071
21.183
25.232
4.209
11.326
8.984
12.668
4.483
128.482

1
1
1
1
1
1
1
1
1
1
1
1

0.02
0.01
0.00
0.04
0.00
0.00
0.04
0.00
0.00
0.00
0.03
0.00

84.81133
-81.82092
129.761
-45.44807
-116.8181
118.4674
39.07999
66.3965
-47.31483
-80.98353
38.6556
704.4469

.1972802
-.2938627
.3110565
-.1641344
-.4267012
.6681337
.184049
.3098331
-.2931597
-.333871
.2116015
.8094959

Measurement
a5 <-

c5 <-

cov(e.a1,e.a4)
cov(e.a1,e.a5)
cov(e.a2,e.a4)
cov(e.a2,e.c4)
cov(e.a3,e.a4)
cov(e.a3,e.a5)
cov(e.a5,e.c4)
cov(e.c1,e.c3)
cov(e.c1,e.c5)
cov(e.c3,e.c4)
cov(e.c4,e.c5)
cov(Affective,Cognitive)

EPC = expected parameter change

Notes:
1. Four columns of results are reported.
a. MI stands for modification index and is an approximation to the change in the models
goodness-of-fit 2 if the path were added.
b. df stands for degrees of freedom and is the number that would be added to d of the 2 (d).
c. P>MI is the value of the significance of 2 (df).
d. EPC stands for expected parameter change and is an approximation to the value of the
parameter if it were not constrained to 0. It is reported in unstandardized (column 3) and
standardized (column 4) units.
2. There are lots of significant omitted paths in the above output.
3. Paths are listed only if the modification index is significant at the 0.05 level, corresponding to
2 (1) value 3.8414588. You may specify the minchi2() option to use different 2 (1) values.
Specify minchi2(0) if you wish to see all tests.
4. The omitted path between Affective and Cognitive has the largest change in 2 observed.
Perhaps this is why Kline (2005, 7074, 184) allowed a covariance between the two latent
variables. The standardized EPC reports the relaxed-constraint correlation value, which is the
value reported for the unconstrained correlation path in [SEM] example 3.
Another way of dealing with this significant result would be to add a direct path between the
variables, but that perhaps would have invalidated the theory being proposed. The original authors
instead proposed a second-order model postulating that Affective and Cognitive are themselves
measurements of another latent variable that might be called Arousal.

182

example 5 Modification indices

Reference
Kline, R. B. 2005. Principles and Practice of Structural Equation Modeling. 2nd ed. New York: Guilford Press.

Also see
[SEM] example 3 Two-factor measurement model

[SEM] estat mindices Modification indices

Title
example 6 Linear regression

Description

Remarks and examples

Also see

Description
Linear regression is demonstrated using auto.dta:
. sysuse auto
(1978 Automobile Data)

See Structural models 1: Linear regression in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Fitting linear regression models
Displaying standardized results
Fitting the model with the Builder

Fitting linear regression models

The first two examples in [R] regress are
. regress mpg weight c.weight#c.weight foreign
. regress, beta

This model corresponds to

weight

weight2

mpg

foreign

183

184

example 6 Linear regression

To fit this model with sem, we type

. generate weight2 = weight^2
. sem (mpg <- weight weight2 foreign)
Endogenous variables
Observed: mpg
Exogenous variables
Observed: weight weight2 foreign
Fitting target model:
Iteration 0:
log likelihood = -1909.8206
Iteration 1:
log likelihood = -1909.8206
Structural equation model
Estimation method = ml
Log likelihood
= -1909.8206

Coef.
Structural
mpg <weight
weight2
foreign
_cons
var(e.mpg)

OIM
Std. Err.

-.0165729
1.59e-06
-2.2035
56.53884

.0038604
6.08e-07
1.03022
6.027559

10.19332

1.675772

LR test of model vs. saturated: chi2(1)

Number of obs

-4.29
2.62
-2.14
9.38

P>|z|

0.000
0.009
0.032
0.000

74

[95% Conf. Interval]

-.0241392
4.00e-07
-4.222695
44.72504

-.0090067
2.78e-06
-.1843056
68.35264

7.385485

14.06865

0.00, Prob > chi2 = 1.0000

Notes:
1. We wished to include variable weight2 in our model. Because sem does not allow Statas
factor-variable notation, we first had to generate new variable weight2.
2. Reported coefficients match those reported by regress.
3. Reported standard errors (SEs) differ slightly from those reported by regress. For instance, the
SE for foreign is reported here as 1.03, whereas regress reported 1.06. SEM is an asymptotic
estimator, and sem divides variances and covariances by N = 74, the number of observations.
regress provides
p unbiased finite-sample estimates and divides by N k 1 = 74 3 1 = 70.
Note that 1.03 74/70 = 1.06.
4. sem reports z statistics whereas regress reports t statistics.
5. Reported confidence intervals differ slightly between sem and regress because of the (N
k 1)/N issue.
6. sem reports
the point estimate of e.mpg as 10.19332. regress reports the root MSE as 3.2827.
p
And 10.19332 74/70 = 3.2827.

Displaying standardized results

To obtain standardized coefficients from regress, you specify the beta option. To obtain standardized coefficients from sem, you specify the standardized option.

example 6 Linear regression

. sem, standardized
Structural equation model
Estimation method = ml
Log likelihood
= -1909.8206
OIM
Std. Err.

Standardized

Coef.

Structural
mpg <weight
weight2
foreign
_cons

-2.226321
1.32654
-.17527
9.839209

.4950378
.498261
.0810378
.9686872

.308704

.0482719

var(e.mpg)

LR test of model vs. saturated: chi2(1)

Number of obs

-4.50
2.66
-2.16
10.16

P>|z|

0.000
0.008
0.031
0.000

185

74

[95% Conf. Interval]

-3.196577
.3499662
-.3341011
7.940617

-1.256064
2.303113
-.0164389
11.7378

.2272168

.4194152

0.00, Prob > chi2 = 1.0000

regress simply reports standardized coefficients in an extra column. All other results are reported
in unstandardized form. sem updates the entire output with the standardized values.

Fitting the model with the Builder

Use the diagram above for reference.
1. Open the dataset and create the additional variable weight2.
In the Command window, type
. use http://www.stata-press.com/data/r13/auto
. generate weight2 = weight^2

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Create the regression component for the mpg outcome.
Select the Add Regression Component tool,

, and then click in the center of the diagram.

In the resulting dialog box,

a. select mpg in the Dependent variable control;
b. select weight, weight2, and foreign by using the Independent variables control;
c. select Left in the Independent variables direction control;
d. click on OK.
If you wish, move this component by clicking on any variable and dragging it.
4. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
SEM estimation options dialog box.
5. Show standardized estimates.
From the SEM Builder menu, select View > Standardized Estimates.

186

example 6 Linear regression

You can open a completed diagram in the Builder by typing

. webgetsem sem_regress

Also see
[SEM] example 12 Seemingly unrelated regression
[SEM] example 38g Random-intercept and random-slope models (multilevel)
[SEM] example 43g Tobit regression
[SEM] example 44g Interval regression

[SEM] sem Structural equation model estimation command

Title
example 7 Nonrecursive structural model

Description

Remarks and examples

References

Also see

Description
To demonstrate a nonrecursive structural model with all variables observed, we use data from
Duncan, Haller, and Portes (1968):
. use http://www.stata-press.com/data/r13/sem_sm1
(Structural model with all observed values)
. ssd describe
Summary statistics data from
http://www.stata-press.com/data/r13/sem_sm1.dta
obs:
329
Structural model with all obse..
vars:
10
25 May 2013 10:13
(_dta has notes)
variable name

variable label

r_intel
r_parasp
r_ses
r_occasp
r_educasp
f_intel
f_parasp
f_ses
f_occasp
f_educasp

respondents intelligence
respondents parental aspiration
respondents family socioeconomic status
respondents occupational aspiration
respondents educational aspiration
friends intelligence
friends parental aspiration
friends family socioeconomic status
friends occupational aspiration
friends educational aspiration

. notes
_dta:
1. Summary statistics data from Duncan, O.D., Haller, A.O., and Portes, A.,
1968, "Peer Influences on Aspirations: A Reinterpretation", _American
Journal of Sociology_ 74, 119-137.
2. The data contain 329 boys with information on five variables and the same
information for each boys best friend.

If you typed ssd status, you would learn that this dataset contains the correlation matrix only.
Variances (standard deviations) and means are undefined. Thus we need to use this dataset cautiously.
It is always better if you enter the variances and means if you have them.
That these data are the correlations only will not matter for how we will use them.

187

188

example 7 Nonrecursive structural model

Remarks and examples

See Structural models 7: Dependencies between response variables in [SEM] intro 5 for background.
Remarks are presented under the following headings:
Fitting the model
Fitting the model with the Builder
Checking stability with estat stable
Reporting total, direct, and indirect effects with estat teffects

Fitting the model

In the referenced paper above, the authors fit the following model:

r_intel
r_occasp

f_occasp

r_ses

f_ses

f_intel

example 7 Nonrecursive structural model

189

. sem (r_occasp <- f_occasp r_intel r_ses f_ses)

>
(f_occasp <- r_occasp f_intel f_ses r_ses),
>
cov(e.r_occasp*e.f_occasp) standardized
Endogenous variables
Observed: r_occasp f_occasp
Exogenous variables
Observed: r_intel r_ses f_ses f_intel
Fitting target model:
Iteration 0:
log likelihood = -2617.0489
Iteration 1:
log likelihood = -2617.0489
Structural equation model
Estimation method = ml
Log likelihood
= -2617.0489

Number of obs

329

OIM
Std. Err.

P>|z|

[95% Conf. Interval]

.2773441
.2854766
.1570082
.0973327

.1281904
.05
.0520841
.060153

2.16
5.71
3.01
1.62

0.031
0.000
0.003
0.106

.0260956
.1874783
.0549252
-.020565

.5285926
.3834748
.2590912
.2152304

f_occ~p <r_occasp
r_ses
f_ses
f_intel

.2118102
.0794194
.1681772
.3693682

.156297
.0587732
.0537199
.0525924

1.36
1.35
3.13
7.02

0.175
0.177
0.002
0.000

-.0945264
-.0357739
.062888
.2662891

.5181467
.1946127
.2734663
.4724474

var(e.r_oc~p)
var(e.f_oc~p)

.6889244
.6378539

.0399973
.039965

.6148268
.5641425

.7719519
.7211964

cov(e.r_oc~p,
e.f_occasp)

-.2325666

.2180087

-.6598558

.1947227

Standardized

Coef.

Structural
r_occ~p <f_occasp
r_intel
r_ses
f_ses

LR test of model vs. saturated: chi2(0)

-1.07
=

0.286

0.00, Prob > chi2 =

Notes:
1. We specified the standardized option, but in this case that did not matter much because
these data are based on the correlation coefficients only, so standardized values are equal to
unstandardized values. The exception is the correlation between the latent endogenous variables,
as reflected in the correlation of their errors, and we wanted to show that results match those in
the original paper.
2. Nearly all results match those in the original paper. The authors normalized the errors to have a
variance of 1; sem normalizes the paths from the errors to have coefficient 1. While you can apply
most normalizing constraints any way you wish, sem restricts errors to have path coefficients of
1 and this cannot be modified. You could, however, prove to yourself that sem would produce
the same variances as the authors produced by typing
. sem, coeflegend
. display sqrt(_b[var(e.r_occasp):_cons])
. display sqrt(_b[var(e.f_occasp):_cons])

190

example 7 Nonrecursive structural model

because the coefficients would be the standard deviations of the errors estimated without the
variance-1 constraint. Thus all results match. We replayed results by using the coeflegend
option so that we would know what to type to refer to the two error variances, namely,
b[var(e.r occasp): cons] and b[var(e.f occasp): cons].

Fitting the model with the Builder

Use the diagram above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/sem_sm1

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Create the four independent variables.
, and then click in the diagram about one-third
Select the Add Observed Variables Set tool,
of the way in from the left and one-fourth of the way up from the bottom.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. use the Variables control to select the four variables in this order: r intel, r ses, f ses,
and f intel;
c. select Vertical in the Orientation control;
d. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
4. Create the two dependent variables.
Select the Add Observed Variables Set tool, , and then click in the diagram about two-thirds
of the way in from the left, vertically aligned with the top of the f intel rectangle.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. use the Variables control to select the two variable names r occasp and f occasp;
c. select Vertical in the Orientation control;
d. select the Distances tab;
e. select .5 (inch) from the Distance between variables control;
f. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
5. Create paths from the independent variables to the dependent variables.
a. Select the Add Path tool,

b. Click in the right side of the r intel rectangle (it will highlight when you hover over it),
and drag a path to the left side of the r occasp rectangle (it will highlight when you can
release to connect the path).

example 7 Nonrecursive structural model

191

c. Continuing with the

tool, create the following paths by first clicking in the right side of
the rectangle for the independent variable and dragging it to the left side of the rectangle
for the dependent variable.
r_ses -> r_occasp
r_ses -> f_occasp
f_ses -> r_occasp
f_ses -> f_occasp
f_intel -> f_occasp
6. Create paths for the feedback loop between endogenous variables.
.

a. Continue with the Add Path tool,

b. Click in the bottom of the r occasp rectangle, slightly to the left of the center, and drag
a path to the f occasp rectangle.
c. Click in the top of the f occasp rectangle, slightly to the right of the center, and drag a
path to the r occasp rectangle.
7. Correlate the error terms.
a. Select the Add Covariance tool,

b. Click in the error term for r occasp (the circle labeled 1 ), and drag a covariance to the
error term for f occasp (the circle labeled 2 ).
8. Clean up.
If you do not like where a path has been connected to its variable, use the Select tool,
,
to click on the path, and then simply click on where it connects to a rectangle and drag the
endpoint. Similarly, you can change where the covariance connects to the error terms by clicking
on the covariance and dragging the endpoint. You can also change the bow of the covariance
by clicking on the covariance and dragging the control point that extends from one end of the
selected covariance.
9. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
SEM estimation options dialog box.
10. Show standardized estimates.
From the SEM Builder menu, select View > Standardized Estimates.
Tips: When you draw paths that should be exactly horizontal or vertical, such as the two between
r occasp and f occasp, holding the Shift key as you drag the path will guarantee that the line
is perfectly vertical. Also, when drawing paths from the independent variables to the dependent
variables, you may find it more convenient to change the automation settings as described in the
tips of [SEM] example 9. However, this does not work for the feedback loop between the dependent
variables.
You can open a completed diagram in the Builder by typing
. webgetsem sem_sm1

192

example 7 Nonrecursive structural model

Checking stability with estat stable

. estat stable
Stability analysis of simultaneous equation systems
Eigenvalue stability condition
Eigenvalue
-.2423722
.2423722

Modulus
.242372
.242372

stability index = .2423722

All the eigenvalues lie inside the unit circle.
SEM satisfies stability condition.

Notes:
1. estat stable is for use on nonrecursive models. Recursive models are, by design, stable.
2. Stability concerns whether the parameters of the model are such that the model would blow up
if it were operated over and over again. If the results are found not to be stable, then that casts
questions about the validity of the model.
3. The stability is the maximum of the moduli, and the moduli are the absolute values of the
eigenvalues. Usually, the two eigenvalues are not identical, but it is a property of this model that
they are.
4. If the stability index is less than 1, then the reported estimates yield a stable model.
In the next section, we use estat teffects to estimate total effects. That is appropriate only if
the model is stable, as we find that it is.

example 7 Nonrecursive structural model

Reporting total, direct, and indirect effects with estat teffects

. estat teffects
Direct effects

Coef.

OIM
Std. Err.

P>|z|

2.15
5.47
2.98
1.61

0.031
0.000
0.003
0.107

.0249748
.1831662
.0536534
-.0209901

.5297134
.3877869
.260363
.2156555

1.35

0.176

-.09472

.5183404

1.35
3.09
6.62

0.178
0.002
0.000

-.0360411
.0615838
.2600142

.1948799
.2747705
.4787223

OIM
Std. Err.

P>|z|

Structural
r_occ~p <r_occasp
f_occasp
r_intel
r_ses
f_ses
f_intel

0
.2773441
.2854766
.1570082
.0973327
0

(no path)
.1287622
.0522001
.052733
.0603699
(no path)

f_occ~p <r_occasp
f_occasp
r_intel
r_ses
f_ses
f_intel

.2118102
0
0
.0794194
.1681772
.3693682

.1563958
(no path)
(no path)
.0589095
.0543854
.0557939

[95% Conf. Interval]

Indirect effects

Coef.

[95% Conf. Interval]

Structural
r_occ~p <r_occasp
f_occasp
r_intel
r_ses
f_ses
f_intel

.0624106
.0173092
.0178168
.0332001
.0556285
.1088356

.0460825
.0080361
.0159383
.0204531
.0292043
.052243

1.35
2.15
1.12
1.62
1.90
2.08

0.176
0.031
0.264
0.105
0.057
0.037

-.0279096
.0015587
-.0134217
-.0068872
-.0016109
.0064411

.1527307
.0330597
.0490552
.0732875
.112868
.21123

f_occ~p <r_occasp
f_occasp
r_intel
r_ses
f_ses
f_intel

.0132192
.0624106
.0642406
.0402881
.0323987
.0230525

.0097608
.0289753
.0490164
.0315496
.0262124
.0202112

1.35
2.15
1.31
1.28
1.24
1.14

0.176
0.031
0.190
0.202
0.216
0.254

-.0059115
.0056201
-.0318298
-.021548
-.0189765
-.0165607

.0323499
.1192011
.160311
.1021242
.083774
.0626657

193

194

example 7 Nonrecursive structural model

Total effects

Coef.

OIM
Std. Err.

P>|z|

[95% Conf. Interval]

Structural
r_occ~p <r_occasp
f_occasp
r_intel
r_ses
f_ses
f_intel

.0624106
.2946533
.3032933
.1902083
.1529612
.1088356

.0460825
.1367983
.0509684
.050319
.050844
.052243

1.35
2.15
5.95
3.78
3.01
2.08

0.176
0.031
0.000
0.000
0.003
0.037

-.0279096
.0265335
.2033971
.091585
.0533089
.0064411

.1527307
.5627731
.4031896
.2888317
.2526136
.21123

f_occ~p <r_occasp
f_occasp
r_intel
r_ses
f_ses
f_intel

.2250294
.0624106
.0642406
.1197074
.2005759
.3924207

.1661566
.0289753
.0490164
.0483919
.0488967
.0502422

1.35
2.15
1.31
2.47
4.10
7.81

0.176
0.031
0.190
0.013
0.000
0.000

-.1006315
.0056201
-.0318298
.0248611
.10474
.2939478

.5506903
.1192011
.160311
.2145537
.2964118
.4908936

Note:
1. In the path diagram we drew for this model, you can see that the intelligence of the respondent,
r intel, has both direct and indirect effects on the occupational aspiration of the respondent,
r occasp. The tables above reveal that

0.3033 = 0.2855 + 0.0178

where 0.2855 is the direct effect and 0.0178 is the indirect effect.

References
Acock, A. C. 2013. Discovering Structural Equation Modeling Using Stata. Rev. ed. College Station, TX: Stata Press.
Duncan, O. D., A. O. Haller, and A. Portes. 1968. Peer influences on aspirations: A reinterpretation. American Journal
of Sociology 74: 119137.

Also see
[SEM] example 8 Testing that coefficients are equal, and constraining them

[SEM] sem Structural equation model estimation command

[SEM] estat stable Check stability of nonrecursive system
[SEM] estat teffects Decomposition of effects into total, direct, and indirect

Title
example 8 Testing that coefficients are equal, and constraining them
Description

Remarks and examples

Also see

Description
This example continues where [SEM] example 7 left off, where we typed
.
.
.
.

use http://www.stata-press.com/data/r13/sem_sm1
ssd describe
notes
sem (r_occasp <- f_occasp r_intel r_ses f_ses) ///
(f_occasp <- r_occasp f_intel f_ses r_ses), ///
cov(e.r_occasp*e.f_occasp) standardized
. estat stable
. estat teffects

Remarks and examples

Remarks are presented under the following headings:
Using test to evaluate adding constraints
Refitting the model with added constraints
Using estat scoretests to test whether constraints can be relaxed

We want to show you how to evaluate potential constraints after estimation, how to fit a model
with constraints, and how to evaluate enforced constraints after estimation.
Obviously, in a real analysis, if you evaluated potential constraints after estimation, there would
be no reason to evaluate enforced constraints after estimation, and vice versa.

Using test to evaluate adding constraints

In this model of respondents and corresponding friends, it would be surprising if the coefficients
relating friends characteristics to respondents occupational aspirations and vice versa were not equal.
It would also be surprising if coefficients relating a respondents characteristics to his occupational
aspirations were not equal to those of his friends characteristics to his occupational aspirations. The
paths that we suspect should be equal are
r_intel
r_ses
f_ses
f_occasp

->
->
->
->

r_occasp
r_occasp
r_occasp
r_occasp

f_intel
f_ses
r_ses
r_occasp

->
->
->
->

f_occasp
f_occasp
f_occasp
f_occasp

You are about to learn that to test whether those paths have equal coefficients, you type
. test (_b[r_occasp:r_intel ]==_b[f_occasp:f_intel ])
(_b[r_occasp:r_ses
]==_b[f_occasp:f_ses
])
(_b[r_occasp:f_ses
]==_b[f_occasp:r_ses
])
(_b[r_occasp:f_occasp]==_b[f_occasp:r_occasp])

195

///
///
///

196

example 8 Testing that coefficients are equal, and constraining them

In Stata, b[ ] is how one accesses the estimated parameters. It is difficult to remember what the
names are. To determine the names of the parameters, replay the sem results with the coeflegend
option:
. sem, coeflegend
Structural equation model
Estimation method = ml
Log likelihood
= -2617.0489
Coef.

Number of obs

.2773441
.2854766
.1570082
.0973327

_b[r_occasp:f_occasp]
_b[r_occasp:r_intel]
_b[r_occasp:r_ses]
_b[r_occasp:f_ses]

f_occ~p <r_occasp
r_ses
f_ses
f_intel

.2118102
.0794194
.1681772
.3693682

_b[f_occasp:r_occasp]
_b[f_occasp:r_ses]
_b[f_occasp:f_ses]
_b[f_occasp:f_intel]

var(e.r_oc~p)
var(e.f_oc~p)

.6868304
.6359151

_b[var(e.r_occasp):_cons]
_b[var(e.f_occasp):_cons]

cov(e.r_oc~p,
e.f_occasp)

-.1536992

_b[cov(e.r_occasp,e.f_occasp):_cons]

LR test of model vs. saturated: chi2(0)

0.00, Prob > chi2 =

With the parameter names at hand, to perform the test, we can type
(_b[r_occasp:r_intel ]==_b[f_occasp:f_intel ])
(_b[r_occasp:r_ses
]==_b[f_occasp:f_ses
])
(_b[r_occasp:f_ses
]==_b[f_occasp:r_ses
])
(_b[r_occasp:f_occasp]==_b[f_occasp:r_occasp])
[r_occasp]r_intel - [f_occasp]f_intel = 0
[r_occasp]r_ses - [f_occasp]f_ses = 0
[r_occasp]f_ses - [f_occasp]r_ses = 0
[r_occasp]f_occasp - [f_occasp]r_occasp = 0
chi2( 4) =
Prob > chi2 =

329

Legend

Structural
r_occ~p <f_occasp
r_intel
r_ses
f_ses

. test
>
>
>
( 1)
( 2)
( 3)
( 4)

1.61
0.8062

We cannot reject the constraint, just as we expected.

example 8 Testing that coefficients are equal, and constraining them

Refitting the model with added constraints

We could refit the model with these constraints by typing
. sem (r_occasp <- f_occasp@b1 r_intel@b2 r_ses@b3 f_ses@b4)
>
(f_occasp <- r_occasp@b1 f_intel@b2 f_ses@b3 r_ses@b4),
>
cov(e.r_occasp*e.f_occasp)
Endogenous variables
Observed: r_occasp f_occasp
Exogenous variables
Observed: r_intel r_ses f_ses f_intel
Fitting target model:
Iteration 0:
log likelihood = -2617.8735
Iteration 1:
log likelihood = -2617.8705
Iteration 2:
log likelihood = -2617.8705
Structural equation model
Number of obs
Estimation method = ml
Log likelihood
= -2617.8705
( 1) [r_occasp]f_occasp - [f_occasp]r_occasp = 0
( 2) [r_occasp]r_intel - [f_occasp]f_intel = 0
( 3) [r_occasp]r_ses - [f_occasp]f_ses = 0
( 4) [r_occasp]f_ses - [f_occasp]r_ses = 0

Coef.

329

OIM
Std. Err.

P>|z|

[95% Conf. Interval]

Structural
r_occ~p <f_occasp
r_intel
r_ses
f_ses

.2471578
.3271847
.1635056
.088364

.1024504
.0407973
.0380582
.0427106

2.41
8.02
4.30
2.07

0.016
0.000
0.000
0.039

.0463588
.2472234
.0889129
.0046529

.4479568
.4071459
.2380984
.1720752

f_occ~p <r_occasp
r_ses
f_ses
f_intel

.2471578
.088364
.1635056
.3271847

.1024504
.0427106
.0380582
.0407973

2.41
2.07
4.30
8.02

0.016
0.039
0.000
0.000

.0463588
.0046529
.0889129
.2472234

.4479568
.1720752
.2380984
.4071459

var(e.r_oc~p)
var(e.f_oc~p)

.6884513
.6364713

.0538641
.0496867

.5905757
.5461715

.8025477
.7417005

cov(e.r_oc~p,
e.f_occasp)

-.1582175

.1410111

-.4345942

.1181592

LR test of model vs. saturated: chi2(4)

-1.12
=

0.262

1.64, Prob > chi2 = 0.8010

Using estat scoretests to test whether constraints can be relaxed

. estat scoretests
(no score tests to report; all chi2 values less than 3.841458820694123)

197

198

example 8 Testing that coefficients are equal, and constraining them

No tests were reported because no tests were individually significant at the 5% level. We can
obtain all the individual tests by adding the minchi2(0) option, which we can abbreviate to min(0):
. estat scoretests, min(0)
Score tests for linear constraints
( 1) [r_occasp]f_occasp - [f_occasp]r_occasp = 0
( 2) [r_occasp]r_intel - [f_occasp]f_intel = 0
( 3) [r_occasp]r_ses - [f_occasp]f_ses = 0
( 4) [r_occasp]f_ses - [f_occasp]r_ses = 0

(
(
(
(

1)
2)
3)
4)

chi2

df

P>chi2

0.014
1.225
0.055
0.136

1
1
1
1

0.91
0.27
0.81
0.71

Notes:
1. When we began this example, we used test to evaluate potential constraints that we were
considering. We obtained an overall 2 (4) statistic of 1.61 and thus could not reject the constraints
at any reasonable level.
2. We then refit the model with those constraints.
3. For pedantic reasons, now we use estat scoretests to evaluate relaxing constraints included
in the model. estat scoretests does not report a joint test. You cannot sum the 2 values
to obtain a joint test statistic. Thus we learn only that the individual constraints should not be
relaxed at reasonable confidence levels.
4. Thus when evaluating multiple constraints, it is better to fit the model without the constraints
and use test to evaluate them jointly.

Also see
[SEM] example 7 Nonrecursive structural model

[SEM] sem Structural equation model estimation command

[SEM] sem and gsem path notation Command syntax for path diagrams
[SEM] test Wald test of linear hypotheses
[SEM] estat scoretests Score tests

Title
example 9 Structural model with measurement component
Description

Remarks and examples

References

Also see

Description
To demonstrate a structural model with a measurement component, we use data from Wheaton
et al. (1977):
. use http://www.stata-press.com/data/r13/sem_sm2
(Structural model with measurement component)
. ssd describe
Summary statistics data from
http://www.stata-press.com/data/r13/sem_sm2.dta
obs:
932
Structural model with measurem..
vars:
13
25 May 2013 11:45
(_dta has notes)
variable name

variable label

educ66
occstat66
anomia66
pwless66
socdist66
occstat67
anomia67
pwless67
socdist67
occstat71
anomia71
pwless71
socdist71

Education, 1966
Occupational status, 1966
Anomia, 1966
Powerlessness, 1966
Latin American social distance, 1966
Occupational status, 1967
Anomia, 1967
Powerlessness, 1967
Latin American social distance, 1967
Occupational status, 1971
Anomia, 1971
Powerlessness, 1971
Latin American social distance, 1971

. notes
_dta:
1. Summary statistics data from Wheaton, B., Muthen B., Alwin, D., &
Summers, G., 1977, "Assessing reliability and stability in panel models",
in D. R. Heise (Ed.), _Sociological Methodology 1977_ (pp. 84-136), San
Francisco: Jossey-Bass, Inc.
2. Four indicators each measured in 1966, 1967, and 1981, plus another
indicator (educ66) measured only in 1966.
3. Intended use: Create structural model relating Alienation in 1971,
Alienation in 1967, and SES in 1966.

See Structural models 8: Unobserved inputs, outputs, or both in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Fitting the model
Fitting the model with the Builder
Evaluating omitted paths with estat mindices
Refitting the model

199

200

example 9 Structural model with measurement component

Fitting the model

Simplified versions of the model fit by the authors of the referenced paper appear in many SEM
software manuals. A simplified model is

anomia67

pwless67

anomia71

pwless71

Alien67

Alien71

SES

educ66

occstat66

. sem
///
>
(anomia67 pwless67 <- Alien67)
/// measurement
>
(anomia71 pwless71 <- Alien71)
/// measurement
>
(Alien67 <- SES)
/// structural
>
(Alien71 <- Alien67 SES)
/// structural
>
(
SES -> educ66 occstat66)
// measurement
Endogenous variables
Measurement: anomia67 pwless67 anomia71 pwless71 educ66
Latent:
Alien67 Alien71
Exogenous variables
Latent:
SES
Fitting target model:
Iteration 0:
log likelihood = -15249.988
Iteration 1:
log likelihood = -15246.584
Iteration 2:
log likelihood = -15246.469
Iteration 3:
log likelihood = -15246.469

piece
piece
piece
piece
piece
occstat66

example 9 Structural model with measurement component

Structural equation model
Estimation method = ml
Log likelihood
= -15246.469
( 1)
( 2)
( 3)

Number of obs

201

932

[anomia67]Alien67 = 1
[anomia71]Alien71 = 1
[educ66]SES = 1

Coef.

OIM
Std. Err.

P>|z|

[95% Conf. Interval]

Structural
Alien67 <SES

-.6140404

.0562407

-10.92

0.000

-.7242701

-.5038107

Alien71 <Alien67
SES

.7046342
-.1744153

.0533512
.0542489

13.21
-3.22

0.000
0.001

.6000678
-.2807413

.8092007
-.0680894

Measurement
anom~67 <Alien67
_cons

1
13.61

(constrained)
.1126205
120.85

0.000

13.38927

13.83073

pwle~67 <Alien67
_cons

.8884887
14.67

20.59
146.44

0.000
0.000

.8039034
14.47365

.9730739
14.86635

anom~71 <Alien71
_cons

1
14.13

(constrained)
.1158943
121.92

0.000

13.90285

14.35715

pwle~71 <Alien71
_cons

.8486022
14.9

20.44
144.03

0.000
0.000

.7672235
14.69723

.9299808
15.10277

educ66 <SES
_cons

1
10.9

(constrained)
.1014894
107.40

0.000

10.70108

11.09892

occs~66 <SES
_cons

5.331259
37.49

.4307503
.6947112

0.000
0.000

4.487004
36.12839

6.175514
38.85161

var(e.ano~67)
var(e.pwl~67)
var(e.ano~71)
var(e.pwl~71)
var(e.educ66)
var(e.occ~66)
var(e.Ali~67)
var(e.Ali~71)
var(SES)

4.009921
3.187468
3.695593
3.621531
2.943819
260.63
5.301416
3.737286
6.65587

.3582978
.283374
.3911512
.3037908
.5002527
18.24572
.483144
.3881546
.6409484

3.365724
2.677762
3.003245
3.072483
2.109908
227.2139
4.434225
3.048951
5.511067

4.777416
3.794197
4.54755
4.268693
4.107319
298.9605
6.338201
4.581019
8.038482

.0431565
.1001798

.0415205
.1034537

LR test of model vs. saturated: chi2(6)

12.38
53.96

71.62, Prob > chi2 = 0.0000

Notes:
1. Measurement component: In both 1967 and 1971, anomia and powerlessness are used to measure endogenous latent variables representing alienation for the same two years. Education and
occupational status are used to measure the exogenous latent variable SES.

202

example 9 Structural model with measurement component

2. Structural component: SES->Alien67 and SES->Alien71, and Alien67->Alien71.

3. The model versus saturated 2 test indicates that the model is a poor fit.

Fitting the model with the Builder

Use the diagram above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/sem_sm2

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Create the measurement component for alienation in 1967.
Select the Add Measurement Component tool, , and then click in the diagram about halfway
down from the top and one-third of the way in from the left.
In the resulting dialog box,
a. change the Latent variable name to Alien67;
b. select anomia67 and pwless67 by using the Measurement variables control;
c. select Up in the Measurement direction control;
d. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
4. Create the measurement component for alienation in 1971.
Repeat the process from item 3, but place the measurement component about halfway down from
the top and two-thirds of the way in from the left. Label the latent variable Alien71, and select
measurement variables anomia71 and pwless71. Again drag to reposition if desired.
5. Create the measurement component for socioeconomic status.
Select the Add Measurement Component tool, , and then click in the diagram about one-third
of the way up from the bottom and one-fourth of the way in from the left.
In the resulting dialog box,
a. change the Latent variable name to SES;
b. select educ66 and occstat66 by using the Measurement variables control;
c. select Down in the Measurement direction control;
d. click on OK.
6. Create a path from Alien67 to Alien71.
a. Select the Add Path tool,

b. Click in the right half of the Alien67 oval (it will highlight when you hover over it), and
drag a path to the left half of the Alien71 oval (it will highlight when you can release to
connect the path).

example 9 Structural model with measurement component

203

7. Create a path from SES to Alien67.

Continuing with the
tool, click in the upper portion of the SES oval, and drag a path to the
lower-left quadrant of the Alien67 oval.
8. Create a path from SES to Alien71.
Continuing with the
tool, click in the upper-right quadrant of the SES oval, and drag a path
to the lower-left quadrant of the Alien71 oval.
9. Clean up the direction of the errors.
The error on Alien67 may have been created under the path from Alien67 to Alien71. If
so, choose the Select tool,
, and then click in the Alien67 oval. Click on one of the Error
, in the Contextual Toolbar until the error is where you want it.
Rotation buttons,
10. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
SEM estimation options dialog box.
Tips: We took some care above to click in specific quadrants of the ovals when creating paths.
This is because, by default, the Builder places the ends of paths on the boundary of the ovals and
rectangles closest to where we click. When you create paths, it is often convenient to change the
automation settings (Settings > Automation...) so that Connection attachment points is Attach based
on position of variables rather than Attach nearest to drop point. With the Connection attachment
points setting, the paths connect nicely regardless of where we click or drop in the variables ovals
and rectangles.
Note that when you create covariances with the Covariances tool,
is sometimes preferred.
You can open a completed diagram in the Builder by typing
. webgetsem sem_sm2

, Attach nearest to drop point

204

example 9 Structural model with measurement component

Evaluating omitted paths with estat mindices

That the model is a poor fit leads us to look at the modification indices:
. estat mindices
Modification indices

MI

df

P>MI

EPC

Standard
EPC

anomia71
pwless71
educ66

51.977
32.517
5.627

1
1
1

0.00
0.00
0.02

.3906425
-.2969297
.0935048

.4019984
-.2727609
.0842631

anomia71
pwless71
educ66

41.618
23.622
6.441

1
1
1

0.00
0.00
0.01

-.3106995
.2249714
-.0889042

-.3594367
.2323233
-.0900664

anomia67
pwless67

58.768
38.142

1
1

0.00
0.00

.429437
-.3873066

.4173061
-.3347904

anomia67
pwless67

46.188
27.760

1
1

0.00
0.00

-.3308484
.2871709

-.3601641
.2780833

anomia67
pwless67

4.415
6.816

1
1

0.04
0.01

.1055965
-.1469371

.1171781
-.1450411

63.786
49.892
6.063
49.876
37.357
7.752

1
1
1
1
1
1

0.00
0.00
0.01
0.00
0.00
0.01

1.951578
-1.506704
.5527612
-1.534199
1.159123
-.5557802

.5069627
-.3953794
.1608845
-.4470094
.341162
-.1814365

Measurement
anomia67 <-

pwless67 <-

anomia71 <-

pwless71 <-

educ66 <-

cov(e.anomia67,e.anomia71)
cov(e.anomia67,e.pwless71)
cov(e.anomia67,e.educ66)
cov(e.pwless67,e.anomia71)
cov(e.pwless67,e.pwless71)
cov(e.pwless67,e.educ66)
EPC = expected parameter change

Notes:
1. There are lots of statistically significant paths we could add to the model.
2. Some of those statistically significant paths also make theoretical sense.
3. Two in particular that make theoretical sense are the covariances between e.anomia67 and
e.anomia71 and between e.pwless67 and e.pwless71.

example 9 Structural model with measurement component

Refitting the model

Lets refit the model and include those two previously excluded covariances:
. sem
>
(anomia67 pwless67 <- Alien67)
>
(anomia71 pwless71 <- Alien71)
>
(Alien67 <- SES)
>
(Alien71 <- Alien67 SES)
>
(
SES -> educ66 occstat66)
>
, cov(e.anomia67*e.anomia71)
>
cov(e.pwless67*e.pwless71)

///
///
///
///
///
///
///

measurement
measurement
structural
structural
measurement

Endogenous variables
Measurement: anomia67 pwless67 anomia71 pwless71 educ66 occstat66
Latent:
Alien67 Alien71
Exogenous variables
Latent:
SES
Fitting target
Iteration 0:
Iteration 1:
Iteration 2:
Iteration 3:
Iteration 4:

model:
log likelihood
log likelihood
log likelihood
log likelihood
log likelihood

=
=
=
=
=

-15249.988
-15217.95
-15213.126
-15213.046
-15213.046

piece
piece
piece
piece
piece

205

206

example 9 Structural model with measurement component

Structural equation model
Estimation method = ml
Log likelihood
= -15213.046
( 1)
( 2)
( 3)

Number of obs

932

[anomia67]Alien67 = 1
[anomia71]Alien71 = 1
[educ66]SES = 1

Coef.

OIM
Std. Err.

P>|z|

[95% Conf. Interval]

Structural
Alien67 <SES

-.5752228

.057961

-9.92

0.000

-.6888244

-.4616213

Alien71 <Alien67
SES

.606954
-.2270301

.0512305
.0530773

11.85
-4.28

0.000
0.000

.5065439
-.3310596

.707364
-.1230006

Measurement
anom~67 <Alien67
_cons

1
13.61

(constrained)
.1126143
120.85

0.000

13.38928

13.83072

pwle~67 <Alien67
_cons

.9785952
14.67

15.79
146.43

0.000
0.000

.8571117
14.47365

1.100079
14.86635

anom~71 <Alien71
_cons

1
14.13

(constrained)
.1159036
121.91

0.000

13.90283

14.35717

pwle~71 <Alien71
_cons

.9217508
14.9

15.43
144.03

0.000
0.000

.8046968
14.69724

1.038805
15.10276

educ66 <SES
_cons

1
10.9

(constrained)
.1014894
107.40

0.000

10.70108

11.09892

occs~66 <SES
_cons

5.22132
37.49

.425595
.6947112

0.000
0.000

4.387169
36.12839

6.055471
38.85161

var(e.ano~67)
var(e.pwl~67)
var(e.ano~71)
var(e.pwl~71)
var(e.educ66)
var(e.occ~66)
var(e.Ali~67)
var(e.Ali~71)
var(SES)

4.728874
2.563413
4.396081
3.072085
2.803674
264.5311
4.842059
4.084249
6.796014

.456299
.4060733
.5171156
.4360333
.5115854
18.22483
.4622537
.4038995
.6524866

3.914024
1.879225
3.490904
2.326049
1.960691
231.1177
4.015771
3.364613
5.630283

5.713365
3.4967
5.535966
4.057398
4.009091
302.7751
5.838364
4.957802
8.203105

1.622024

.3154267

5.14

0.000

1.003799

2.240249

.3399961

.2627541

1.29

0.196

-.1749925

.8549847

cov(e.ano~67,
e.anomia71)
cov(e.pwl~67,
e.pwless71)

.0619825
.1001814

.0597225
.1034517

LR test of model vs. saturated: chi2(4)

12.27
53.96

4.78, Prob > chi2 = 0.3111

example 9 Structural model with measurement component

207

Notes:
1. We find the covariance between e.anomia67 and e.anomia71 to be significant (Z = 5.14).
2. We find the covariance between e.pwless67 and e.pwless71 to be insignificant at the 5%
level (Z = 1.29).
3. The model versus saturated 2 test indicates that the model is a good fit.

References
Acock, A. C. 2013. Discovering Structural Equation Modeling Using Stata. Rev. ed. College Station, TX: Stata Press.
Wheaton, B., B. Muthen, D. F. Alwin, and G. F. Summers. 1977. Assessing reliability and stability in panel models.
In Sociological Methodology 1977, ed. D. R. Heise, 84136. San Francisco: Jossey-Bass.

Also see
[SEM] example 32g Full structural equation model (generalized response)

[SEM] estat mindices Modification indices

[SEM] test Wald test of linear hypotheses

Title
example 10 MIMIC model

Description

Remarks and examples

Reference

Also see

Description
To demonstrate a MIMIC model, we use the following summary statistics data:
. use http://www.stata-press.com/data/r13/sem_mimic1
(Multiple indicators and multiple causes)
. ssd describe
Summary statistics data from
http://www.stata-press.com/data/r13/sem_mimic1.dta
obs:
432
Multiple indicators and multip..
vars:
5
25 May 2013 10:13
(_dta has notes)
variable name

variable label

occpres
income
s_occpres
s_income
s_socstat

occupational prestige, two-digit Dunca..

total family income in units of $2000,..
subjective occupational prestige
subjective income
subjective overall social status

. notes
_dta:
1. Summary statistics data from Kluegel, J. R., R. Singleton, Jr., and C. E.
Starnes, 1977, "Subjective class identification: A multiple indicator
approach", _American Sociological Review_, 42: 599-611.
2. Data is also analyzed in Bollen, K. A. 1989, _Structural Equations with
Latent Variables_, New York: John Wiley & Sons, Inc.
3. The summary statistics represent 432 white adults included in the sample
for the 1969 Gary Area Project for the Institute of Social Research at
Indiana University.
4. The three subjective variables are measures of socioeconomic status based
on an individuals perception of their own income, occupational prestige,
and social status.
5. The income and occpres variables are objective measures of income and
occupational prestige, respectively.

See Structural models 9: MIMIC models in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Fitting the MIMIC model
Fitting the MIMIC model with the Builder
Evaluating the residuals with estat residuals
Performing likelihood-ratio tests with lrtest

208

example 10 MIMIC model

209

Fitting the MIMIC model

Based on the data referenced above, Bollen (1989, 397399) fits a MIMIC model, the path diagram
of which is
s_income

SubjSES

s_occpres

s_socstat

occpres

income

Bollen includes paths that he constrains and we do not show. Our model is nonetheless equivalent to
the one he shows. In his textbook, Bollen illustrates various ways the same model can be written.

210

example 10 MIMIC model

. sem (SubjSES -> s_income s_occpres s_socstat) (SubjSES <- income occpres)
Endogenous variables
Measurement: s_income s_occpres s_socstat
Latent:
SubjSES
Exogenous variables
Observed:
income occpres
Fitting target model:
Iteration 0:
log likelihood = -4252.1834 (not concave)
Iteration 1:
log likelihood = -4022.8854 (not concave)
Iteration 2:
log likelihood = -3994.3867
Iteration 3:
log likelihood = -3978.7751 (not concave)
Iteration 4:
log likelihood = -3974.6636
Iteration 5:
log likelihood = -3972.0269
Iteration 6:
log likelihood = -3971.9238
Iteration 7:
log likelihood = -3971.9236
Structural equation model
Number of obs
=
432
Estimation method = ml
Log likelihood
= -3971.9236
( 1) [s_income]SubjSES = 1

Coef.

OIM
Std. Err.

P>|z|

[95% Conf. Interval]

.0138498
.0012464

5.97
3.71

0.000
0.000

.0555869
.0021847

.1098772
.0070704

(constrained)
.0794151
12.10

0.000

.8055583

1.11686

Structural
SubjSES <income
occpres

.082732
.0046275

Measurement
s_inc~e <SubjSES
_cons

1
.9612091

s_occ~s <SubjSES
_cons

.7301352
1.114563

.0832915
.0656195

8.77
16.99

0.000
0.000

.5668869
.9859512

.8933835
1.243175

s_soc~t <SubjSES
_cons

.9405161
1.002113

.0934855
.0706576

10.06
14.18

0.000
0.000

.7572878
.863627

1.123744
1.1406

var(e.s_in~e)
var(e.s_oc~s)
var(e.s_so~t)
var(e.Subj~S)

.2087546
.2811852
.1807129
.1860097

.0254098
.0228914
.0218405
.0270476

.1644474
.2397153
.1425987
.1398822

.2649996
.3298291
.2290146
.2473481

LR test of model vs. saturated: chi2(4)

26.65, Prob > chi2 = 0.0000

Notes:
1. In this model, there are three observed variables that record the respondents ideas of their
perceived socioeconomic status (SES). One is the respondents general idea of his or her SES
(s socstat); another is based on the respondents income (s income); and the last is based
on the respondents occupational prestige (s occpres). Those three variables form the latent
variable SubjSES.
2. The other two observed variables are the respondents income (income) and occupation, the latter
measured by the two-digit Duncan SEI scores for occupations (occpres). These two variables
are treated as predictors of SubjSES.

example 10 MIMIC model

211

3. In the model, item 1 is viewed as subjective and item 2 is viewed as objective.

4. All variables are statistically significant at the 5% level, but the model versus saturated test
suggests that we are not modeling the covariances well.

Fitting the MIMIC model with the Builder

Use the diagram above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/sem_mimic1

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Create the measurement component for subject socioeconomic status.
Select the Add Measurement Component tool,

, and then click near the center of the diagram.

In the resulting dialog box,

a. change the Latent variable name to SubjSES;
b. select s income, s occpress, and s socstat by using the Measurement variables control;
c. select Right in the Measurement direction control;
d. click on OK.
If you wish, move this component by clicking on any variable and dragging it.
4. Create the two variables for the formative indicators of socioeconomic status.
a. Select the Add Observed Variable tool,
, and then click in the diagram to add the new
variable slightly above and to the left of SubjSES. After adding it, you can click inside the
rectangle and move the variable if you wish.
b. In the Contextual Toolbar, select occpress by using the Variable control.
c. Continuing with the
occpress variable.

tool, click in the diagram to add another new variable below the

d. In the Contextual Toolbar, select income by using the Variable control.

5. Create the paths for the formative indicators of socioeconomic status.
a. Select the Add Path tool,

b. Click in the lower-right quadrant of the occpress rectangle (it will highlight when you
hover over it), and drag a path to the upper-left quadrant of the SubjSES oval (it will
highlight when you can release to connect the path).
c. Continuing with the
tool, click in the upper-right quadrant of the income rectangle, and
drag a path to the lower-left quadrant of the SubjSES oval.

212

example 10 MIMIC model

6. Clean up the direction of the errors.

The error on SubjSES is likely to have been created atop one of the existing paths. Choose the
Select tool, , and then click in the SubjSES oval. Click on one of the Error Rotation buttons,
, in the Contextual Toolbar until the error is where you want it.
7. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
SEM estimation options dialog box.
Tip: See the tips of [SEM] example 9 to make creating paths somewhat easier than described
above.
You can open a completed diagram in the Builder by typing
. webgetsem sem_mimic1

Evaluating the residuals with estat residuals

Remember that SEM fits covariances and means. Residuals in the SEM sense thus refer to covariances
and means. If we are not fitting well, we can examine the residuals.
. estat residuals, normalized
Residuals of observed variables
Mean residuals
s_income

s_occpres

s_socstat

income

occpres

0.000
0.000

0.000
0.000

0.000
0.000

0.000
0.000

0.000
0.000

s_income

s_occpres

s_socstat

income

occpres

0.000
-0.009
0.000
0.101
-0.856

0.000
0.008
-0.079
1.482

0.000
-0.053
0.049

0.000
0.000

0.000

s_income

s_occpres

s_socstat

income

occpres

0.000
-0.425
0.008
1.362
-1.221

0.000
0.401
-1.137
2.234

0.000
-0.771
0.074

0.000
0.000

0.000

raw
normalized
Covariance residuals

s_income
s_occpres
s_socstat
income
occpres

Normalized covariance residuals

s_income
s_occpres
s_socstat
income
occpres

Notes:
1. The residuals can be partitioned into two subsets: mean residuals and covariance residuals.
2. The normalized option caused the normalized residuals to be displayed.
3. Concerning mean residuals, the raw residuals and the normalized residuals are shown on a
separate line of the first table.

example 10 MIMIC model

213

4. Concerning covariance residuals, the raw residuals and the normalized residuals are shown in
separate tables.
5. Distinguish between normalized residuals and standardized residuals. Both are available from
estat residuals; if we wanted standardized residuals, we would have specified the standardized option instead of or along with normalized.
6. Both normalized and standardized residuals attempt to adjust the residuals in the same way. The
normalized residuals are always valid, but they do not follow a standard normal distribution.
The standardized residuals do follow a standard normal distribution if they can be calculated;
otherwise, they will equal missing values. When both can be calculated (equivalent to both being
appropriate), the normalized residuals will be a little smaller than the standardized residuals.
7. The normalized covariance residuals between income and s income and between occpres and
s occpres are large.

Performing likelihood-ratio tests with lrtest

Thus Bollen (1989, 397399) suggests adding a direct path from the objective measures to the
corresponding subjective measures. We are about to fit the model
(SubjSES -> s_income s_occpres s_socstat)
(SubjSES <- income occpres)
(s_income <- income)
(s_occpres <- occpres)

///
///
///
//

<- new
<- new

For no other reason than we want to demonstrate the likelihood-ratio test, we will then use lrtest
rather than test to test the joint significance of the new paths. lrtest compares the likelihood
values of two fitted models. Thus we will use lrtest to compare this new model with the one above.
To do that, we must plan ahead and store in memory the currently fit model:
. estimates store mimic1

Alternatively, we could skip that and calculate the joint significance of the two new paths by using
a Wald test and the test command.
In any case, having stored the current estimates under the name mimic1, we can now fit our new
model:
. sem (SubjSES -> s_income s_occpres s_socstat)
>
(SubjSES <- income occpres)
>
(s_income <- income)
>
(s_occpres <- occpres)
Endogenous variables
Observed:
s_income s_occpres
Measurement: s_socstat
Latent:
SubjSES
Exogenous variables
Observed:
income occpres
Fitting target model:
Iteration 0:
log likelihood = -4267.0974 (not concave)
Iteration 1:
log likelihood = -4022.6745 (not concave)
Iteration 2:
log likelihood = -3977.0648
Iteration 3:
log likelihood = -3962.9229
Iteration 4:
log likelihood = -3962.1604
Iteration 5:
log likelihood = -3960.8404
Iteration 6:
log likelihood = -3960.7133
Iteration 7:
log likelihood = -3960.7111
Iteration 8:
log likelihood = -3960.7111

214

example 10 MIMIC model

Structural equation model
Estimation method = ml
Log likelihood
= -3960.7111
( 1) [s_income]SubjSES = 1

Coef.

Number of obs

OIM
Std. Err.

432

P>|z|

[95% Conf. Interval]

0.000
0.000

.025242
.7293239

.081243
1.035739

Structural
s_inc~e <SubjSES
income
_cons

1
.0532425
.8825314

s_occ~s <SubjSES
occpres
_cons

.783781
.0045201
1.06586

.1011457
.0013552
.0696058

7.75
3.34
15.31

0.000
0.001
0.000

.585539
.0018641
.9294353

.982023
.0071762
1.202285

SubjSES <income
occpres

.0538025
.0034324

.0129158
.0011217

4.17
3.06

0.000
0.002

.028488
.0012339

.0791171
.0056309

Measurement
s_soc~t <SubjSES
_cons

1.195539
1.07922

.1582735
.078323

7.55
13.78

0.000
0.000

.8853282
.9257097

1.505749
1.23273

var(e.s_in~e)
var(e.s_oc~s)
var(e.s_so~t)
var(e.Subj~S)

.2292697
.2773786
.1459009
.1480275

.0248905
.0223972
.028228
.0278381

.1853261
.2367783
.0998556
.1023918

.2836329
.3249407
.2131785
.2140029

(constrained)
.0142862
3.73
.0781685
11.29

LR test of model vs. saturated: chi2(2)

4.22, Prob > chi2 = 0.1211

Now we can perform the likelihood-ratio test:

. lrtest mimic1 .
Likelihood-ratio test
(Assumption: mimic1 nested in .)

LR chi2(2) =
Prob > chi2 =

22.42
0.0000

Notes:
1. The syntax of lrtest is lrtest modelname1 modelname2. We specified the first model name
as mimic1, the model we previously stored. We specified the second model name as period (.),
meaning the model most recently fit. The order in which we specify the names is irrelevant.
2. We find the two added paths to be extremely significant.

Reference
Bollen, K. A. 1989. Structural Equations with Latent Variables. New York: Wiley.

Also see
[SEM] example 36g MIMIC model (generalized response)

[SEM] sem Structural equation model estimation command

[SEM] estat residuals Display mean and covariance residuals
[SEM] lrtest Likelihood-ratio test of linear hypothesis

Title
example 11 estat framework

Description

Remarks and examples

Also see

Description
To demonstrate estat framework, which displays results in BentlerWeeks form, we continue
where [SEM] example 10 left off:
.
.
.
.
.
.
.

use http://www.stata-press.com/data/r13/sem_mimic1
ssd describe
notes
sem (SubjSES -> s_income s_occpres s_socstat)
///
(SubjSES <- income occpres)
estat residuals, normalized
estimates store mimic1
sem (SubjSES -> s_income s_occpres s_socstat)
///
(SubjSES <- income occpres)
///
(s_income <- income)
///
(s_occpres <- occpres)
lrtest mimic1 .

See Structural models 9: MIMIC models in [SEM] intro 5 for background.

Remarks and examples

If you prefer to see SEM results reported in BentlerWeeks form, type estat framework after
estimating with sem. Many people find BentlerWeeks form helpful in understanding how the model
is fit.
[SEM] example 10 ended by fitting
. sem (SubjSES
(SubjSES
(s_income
(s_occpres

->
<<<-

s_income s_occpres s_socstat)

income occpres)
income)
occpres)

///
///
///

In BentlerWeeks form, the output appears as

. estat framework, fitted
Endogenous variables on endogenous variables
observed
Beta
s_income s_occpres s_socstat

latent
SubjSES

observed
s_income
s_occpres
s_socstat

0
0
0

0
0
0

0
0
0

1
.783781
1.195539

latent
SubjSES

215

216

example 11 estat framework

Exogenous variables on endogenous variables
observed
Gamma
income
occpres
observed
s_income
s_occpres
s_socstat

.0532425
0
0

0
.0045201
0

latent
SubjSES

.0538025

.0034324

Covariances of error variables

observed
e.s_inc~e e.s_occ~s
Psi

e.s_soc~t

latent
e.SubjSES

observed
e.s_income
e.s_occpres
e.s_socstat

.2292697
0
0

.2773786
0

.1459009

latent
e.SubjSES

.1480275

Intercepts of endogenous variables

observed
alpha
s_income s_occpres

s_socstat

latent
SubjSES

1.07922

_cons

.8825314

1.06586

Covariances of exogenous variables

observed
Phi
income
occpres
observed
income
occpres

4.820021
13.62431

451.6628

Means of exogenous variables

observed
kappa
income

occpres

mean

36.698

5.04

example 11 estat framework

Fitted covariances of observed and latent variables
observed
Sigma
s_income s_occpres s_socstat

latent
SubjSES

217

observed
income

observed
s_income
s_occpres
s_socstat

.4478609
.1614446
.225515

.4086519
.1738222

.392219

latent
SubjSES

.1886304

.1453924

.2060311

.1723333

observed
income
occpres

.5627232
3.008694

.3014937
3.831184

.3659463
2.729776

.3060932
2.283302

4.820021
13.62431

Sigma

observed
occpres

observed
occpres

451.6628

Fitted means of observed and latent variables

mu

observed
s_income

s_occpres

s_socstat

latent
SubjSES

observed
income

mu

1.548

1.543

1.554

.3971264

5.04

mu

observed
occpres

mu

36.698

Notes:
1. BentlerWeeks form is a vector and matrix notation for the estimated parameters of the model.
The matrices are known as , , , , , and . Those Greek names are spelled out in the
labels, along with a header stating what each contains.
2. We specified estat framework option fitted. That caused estat framework to list one more
matrix and one more vector at the end: and . These two results are especially interesting to
those wishing to see the ingredients of the residuals reported by estat residuals.
3. One of the more useful results reported by estat framework, fitted is the matrix, which
reports all estimated covariances in a readable format and includes the model-implied covariances
that do not appear in sems ordinary output.
4. estat framework also allows the standardized option if you want standardized output.

Also see
[SEM] example 10 MIMIC model

[SEM] estat framework Display estimation results in modeling framework

Title
example 12 Seemingly unrelated regression
Description

Remarks and examples

Also see

Description
sem can be used to estimate seemingly unrelated regression. We will use auto.dta, which surely
needs no introduction:
. sysuse auto
(1978 Automobile Data)

See Structural models 10: Seemingly unrelated regression (SUR) in [SEM] intro 5.

Remarks and examples

Remarks are presented under the following headings:
Fitting the seemingly unrelated regression model
Fitting the model with the Builder

Fitting the seemingly unrelated regression model

We fit the following model:

mpg
price

weight

displacement

foreign

length
. sem (price <- foreign mpg displacement)
>
(weight <- foreign length),
>
cov(e.price*e.weight)
Endogenous variables
Observed: price weight
Exogenous variables
Observed: foreign mpg displacement length
Fitting target model:
Iteration 0:
log likelihood = -2150.9983
Iteration 1:
log likelihood = -2138.5739
Iteration 2:
log likelihood = -2133.3461
Iteration 3:
log likelihood = -2133.1979
Iteration 4:
log likelihood = -2133.1956
Iteration 5:
log likelihood = -2133.1956

218

example 12 Seemingly unrelated regression

Structural equation model
Estimation method = ml
Log likelihood
= -2133.1956

Coef.

OIM
Std. Err.

Number of obs

74

P>|z|

[95% Conf. Interval]

Structural
price <foreign
mpg
displace~t
_cons

2940.929
-105.0163
17.22083
4129.866

724.7311
57.93461
4.5941
1984.253

4.06
-1.81
3.75
2.08

0.000
0.070
0.000
0.037

1520.482
-218.566
8.216558
240.8022

4361.376
8.53347
26.2251
8018.931

weight <foreign
length
_cons

-153.2515
30.73507
-2711.096

76.21732
1.584743
312.6813

-2.01
19.39
-8.67

0.044
0.000
0.000

-302.6347
27.62903
-3323.94

-3.868275
33.84111
-2098.252

var(e.price)
var(e.weight)

4732491
60253.09

801783.1
9933.316

3395302
43616.45

6596312
83235.44

cov(e.price,
e.weight)

209268

73909.54

64407.92

354128

LR test of model vs. saturated: chi2(3)

2.83
=

0.005

219

38.86, Prob > chi2 = 0.0000

Notes:
1. Point estimates are the same as reported by
. sureg (price foreign mpg displ) (weight foreign length), isure

suregs isure option is required to make sureg iterate to the maximum likelihood estimate.
2. If you wish to compare the estimated variances and covariances after estimation by sureg, type
. matrix list e(Sigma)

sureg does not estimate standard errors on variances and covariances.

3. Standard errors will be different between sem and sureg. In this case, there is no reason to
prefer one set of standard errors over the other, and standard errors are asymptotically equivalent.
This is a case of exogenous variables only on the right-hand side. When the model being fit is
recursive, standard errors produced by sem are better than those from sureg, both asymptotically
and in finite samples.
4. One reason you might want to use sem is that sem will provide robust standard errors whereas
sureg does not.
5. Multivariate regression can be viewed as seemingly unrelated regression. You just need to specify
the same regressors for each equation. In that case, standard p
errors reported by sem will be the same
as those reported by mvreg if one applies the multiplicative (N p 1)/N degree-of-freedom
adjustment.

220

example 12 Seemingly unrelated regression

Fitting the model with the Builder

Use the diagram above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/auto

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Change the size of the observed variables rectangles.
From the SEM Builder menu, select Settings > Variables > All Observed....
In the resulting dialog box, change the first size to .85 and click on OK.
4. Create the four independent variables.
Select the Add Observed Variables Set tool, , and then click in the diagram about one-fourth
of the way in from the left and one-fourth of the way up from the bottom.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. use the Variables control to select the four variables in this order: mpg, displacement,
foreign, and length;
c. select Vertical in the Orientation control;
d. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
5. Create the two dependent variables.
, and then click about two-thirds of the way in
Select the Add Observed Variables Set tool,
from the left and vertically aligned with the top of the length rectangle.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. use the Variables control to select the variables price and weight;
c. select Vertical in the Orientation control;
d. select the Distances tab;
e. select .5 (inch) from the Distance between variables control;
f. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
6. Create paths from the independent variables to the dependent variables.
a. Select the Add Path tool,

b. Click in the right side of the mpg rectangle (it will highlight when you hover over it), and
drag a path to the left side of the price rectangle (it will highlight when you can release
to connect the path).

example 12 Seemingly unrelated regression

221

c. Continuing with the

tool, create the following paths by clicking first in the right side of
the rectangle for the independent variable and dragging it to the left side of the rectangle
for the dependent variable:
displacement -> price
foreign -> price
foreign -> weight
length -> weight
7. Correlate the error terms.
a. Select the Add Covariance tool,

b. Click in the 1 circle (it will highlight when you hover over it), and drag a covariance to
the 2 circle (it will highlight when you can release to connect the covariance).
8. Clean up.
,
If you do not like where a path has been connected to its variables, use the Select tool,
to click on the path, and then simply click on where it connects to a rectangle and drag the
endpoint. Similarly, you can change where the covariance connects to the error terms by clicking
on the covariance and dragging the endpoint. You can also change the bow of the covariance
by clicking on the covariance and dragging the control point that extends from one end of the
selected covariance.
9. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
SEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem sem_sureg

Also see
[SEM] example 13 Equation-level Wald test

[SEM] sem Structural equation model estimation command

Title
example 13 Equation-level Wald test

Description

Remarks and examples

Also see

Description
We demonstrate estat eqtest. See [SEM] intro 7 and [SEM] estat eqtest.
This example picks up where [SEM] example 12 left off:
. use http://www.stata-press.com/data/r13/auto
. sem (price <- foreign mpg displacement)
(weight <- foreign length),
cov(e.price*e.weight)

///
///

Remarks and examples

We have fit a two-equation model with equations for endogenous variables price and weight.
There happen to be two equations, the model happens to be a seemingly unrelated regression, and
the endogenous variables happen to be observed, but none of that is important right now.
estat eqtest displays equation-by-equation Wald tests that all coefficients excluding the intercepts
are 0.
. estat eqtest
Wald tests for equations

observed
price
weight

chi2

df

36.43
633.34

3
2

0.0000
0.0000

Note:
1. The null hypothesis for this test is that the coefficients other than the intercepts are 0. We can
reject that null hypothesis for each equation.

Also see
[SEM] example 12 Seemingly unrelated regression

[SEM] intro 7 Postestimation tests and predictions

[SEM] estat eqtest Equation-level test that all coefficients are zero

222

Title
example 14 Predicted values

Description

Remarks and examples

Also see

Description
We demonstrate the use of predict. See [SEM] intro 7 and [SEM] predict after sem.
This example picks up where the first part of [SEM] example 1 left off:
. use http://www.stata-press.com/data/r13/sem_1fmm
. sem (x1 x2 x3 x4 <- X)

Remarks and examples

predict can create new variables containing predicted values of 1) observed endogenous variables,
2) latent variables, whether endogenous or exogenous, and 3) latent endogenous variables. In the case
of latent variables, item 2 corresponds to the factor score and item 3 is the linear prediction.
Below we demonstrate 1 and 2:
. predict x1hat x2hat, xb(x1 x2)
. predict Xhat, latent(X)

You specify options on predict to specify what you want predicted and how. Because of the differing
options, the two commands could not have been combined into one command.
Our dataset now contains three new variables. Below we compare the three variables with the
original x1 and x2 by using first summarize and then correlate:
. summarize x1 x1hat x2 x2hat Xhat
Obs
Mean
Variable
x1
x1hat
x2
x2hat
Xhat

123
123
123
123
123

96.28455
96.28455
97.28455
97.28455
-1.66e-08

Std. Dev.
14.16444
10.65716
16.14764
12.49406
10.65716

Min

Max

54
68.42469
64
64.62267
-27.85986

131
122.9454
135
128.5408
26.66084

Notes:
1. Means of x1hat and x1 are identical; means of x2hat and x2 are identical.
2. The standard deviation of x1hat is less than that of x1; the standard deviation of x2hat is less
than that of x2. Some of the variation in x1 and x2 is not explained by the model.
3. Standard deviations of x1hat and Xhat are equal. This is because in

x 1 = b0 + b1 X + e 1
coefficient b1 was constrained to be equal to 1 because of the anchoring normalization constraint;
see Identification 2: Normalization constraints (anchoring) in [SEM] intro 4.

223

224

example 14 Predicted values

The mean of Xhat in the model above is 1.66e08 rather than 0. Had we typed
. predict double Xhat, latent(X)

the mean would have been 1.61e15.

. correlate x1 x1hat x2 x2hat Xhat
(obs=123)
x1
x1hat
x1
x1hat
x2
x2hat
Xhat

1.0000
0.7895
0.5826
0.7895
0.7895

1.0000
0.8119
1.0000
1.0000

x2

x2hat

Xhat

1.0000
0.8119
0.8119

1.0000
1.0000

1.0000

Notes:
1. Both x1hat and x2hat correlate 1 with Xhat. That is because both are linear functions of Xhat
alone.
2. That x1hat and x2hat correlate 1 is implied by item 1, directly above.
3. That Xhat, x1hat, and x2hat all have the same correlation with x1 and with x2 is also implied
by item 1, directly above.

Also see
[SEM] example 1 Single-factor measurement model

[SEM] intro 7 Postestimation tests and predictions

[SEM] predict after sem Factor scores, linear predictions, etc.

Title
example 15 Higher-order CFA

Description

Remarks and examples

Reference

Also see

Description
sem can be used to estimate higher-order confirmatory factor analysis models.
. use http://www.stata-press.com/data/r13/sem_hcfa1
(Higher-order CFA)
. ssd describe
Summary statistics data from
http://www.stata-press.com/data/r13/sem_hcfa1.dta
obs:
251
Higher-order CFA
vars:
16
25 May 2013 11:26
(_dta has notes)
variable name

variable label

phyab1
phyab2
phyab3
phyab4
appear1
appear2
appear3
appear4
peerrel1
peerrel2
peerrel3
peerrel4
parrel1
parrel2
parrel3
parrel4

Physical ability 1
Physical ability 2
Physical ability 3
Physical ability 4
Appearance 1
Appearance 2
Appearance 3
Appearance 4
Relationship w/ peers 1
Relationship w/ peers 2
Relationship w/ peers 3
Relationship w/ peers 4
Relationship w/ parent 1
Relationship w/ parent 2
Relationship w/ parent 3
Relationship w/ parent 4

. notes
_dta:
1. Summary statistics data from Marsh, H. W. and Hocevar, D., 1985,
"Application of confirmatory factor analysis to the study of
self-concept: First- and higher order factor models and their invariance
across groups", _Psychological Bulletin_, 97: 562-582.
2. Summary statistics based on 251 students from Sydney, Australia in Grade
5.
3. Data collected using the Self-Description Questionnaire and includes
sixteen subscales designed to measure nonacademic traits: four intended
to measure physical ability, four intended to measure physical
appearance, four intended to measure relations with peers, and four
intended to measure relations with parents.

See Higher-order CFA models in [SEM] intro 5 for background.

225

226

example 15 Higher-order CFA

Remarks and examples

Remarks are presented under the following headings:
Fitting the model
Fitting the model with the Builder

Fitting the model

We fit the following model:

phyab1

phyab2

phyab3

phyab4

appear1

appear2

appear3

10

appear4

Phys

Appear

12

peerrel1

13

peerrel2

14

peerrel3

15

peerrel4

17

parrel1

18

parrel2

19

parrel3

20

parrel4

Peer

11

Par

16

Nonacad

example 15 Higher-order CFA

. sem (Phys -> phyab1 phyab2 phyab3 phyab4)
>
(Appear -> appear1 appear2 appear3 appear4)
>
(Peer -> peerrel1 peerrel2 peerrel3 peerrel4)
>
(Par -> parrel1 parrel2 parrel3 parrel4)
>
(Nonacad -> Phys Appear Peer Par)
Endogenous variables
Measurement:

phyab1 phyab2 phyab3 phyab4 appear1 appear2 appear3 appear4

peerrel1 peerrel2 peerrel3 peerrel4 parrel1 parrel2 parrel3
parrel4
Latent:
Phys Appear Peer Par
Exogenous variables
Latent:
Nonacad
Fitting target model:
Iteration 0:
log likelihood = -7686.6699 (not concave)
Iteration 1:
log likelihood = -7643.7387 (not concave)
Iteration 2:
log likelihood = -7616.2966 (not concave)
Iteration 3:
log likelihood = -7597.6133
Iteration 4:
log likelihood = -7588.9515
Iteration 5:
log likelihood = -7585.3162
Iteration 6:
log likelihood = -7584.8125
Iteration 7:
log likelihood = -7584.7885
Iteration 8:
log likelihood = -7584.7881
Structural equation model
Number of obs
=
Estimation method = ml
Log likelihood
= -7584.7881
( 1) [phyab1]Phys = 1
( 2) [appear1]Appear = 1
( 3) [peerrel1]Peer = 1
( 4) [parrel1]Par = 1
( 5) [Phys]Nonacad = 1

Coef.

OIM
Std. Err.

251

P>|z|

[95% Conf. Interval]

Structural
Phys <Nonacad

Appear <Nonacad

2.202491

.3975476

5.54

0.000

1.423312

2.98167

Peer <Nonacad

1.448035

.2921383

4.96

0.000

.8754549

2.020616

Par <Nonacad

.569956

.1382741

4.12

0.000

.2989437

.8409683

Measurement
phyab1 <Phys
_cons

1
8.2

(constrained)
.1159065
70.75

0.000

7.972827

8.427173

phyab2 <Phys
_cons

.9332477
8.23

0.000
0.000

.68125
7.990479

1.185245
8.469521

(constrained)

.1285726
.122207

7.26
67.34

227

228

example 15 Higher-order CFA

phyab3 <Phys
_cons

1.529936
8.17

.1573845
.1303953

9.72
62.66

0.000
0.000

1.221468
7.91443

1.838404
8.42557

phyab4 <Phys
_cons

1.325641
8.56

.1338053
.1146471

9.91
74.66

0.000
0.000

1.063387
8.335296

1.587894
8.784704

appear1 <Appear
_cons

1
7.41

(constrained)
.1474041
50.27

0.000

7.121093

7.698907

appear2 <Appear
_cons

1.0719
7

.0821893
.1644123

13.04
42.58

0.000
0.000

.9108121
6.677758

1.232988
7.322242

appear3 <Appear
_cons

1.035198
7.17

.0893075
.1562231

11.59
45.90

0.000
0.000

.8601581
6.863808

1.210237
7.476192

appear4 <Appear
_cons

.9424492
7.4

.0860848
.1474041

10.95
50.20

0.000
0.000

.7737262
7.111093

1.111172
7.688907

peerr~1 <Peer
_cons

1
8.81

(constrained)
.1077186
81.79

0.000

8.598875

9.021125

peerr~2 <Peer
_cons

1.214379
7.94

.1556051
.1215769

7.80
65.31

0.000
0.000

.9093989
7.701714

1.51936
8.178286

peerr~3 <Peer
_cons

1.667829
7.52

.190761
.1373248

8.74
54.76

0.000
0.000

1.293944
7.250848

2.041714
7.789152

peerr~4 <Peer
_cons

1.363627
8.29

.159982
.1222066

8.52
67.84

0.000
0.000

1.050068
8.050479

1.677186
8.529521

parrel1 <Par
_cons

1
9.35

(constrained)
.0825215
113.30

0.000

9.188261

9.511739

parrel2 <Par
_cons

1.159754
9.13

.184581
.0988998

6.28
92.32

0.000
0.000

.7979822
8.93616

1.521527
9.32384

parrel3 <Par
_cons

2.035143
8.67

.2623826
.1114983

7.76
77.76

0.000
0.000

1.520882
8.451467

2.549403
8.888533

parrel4 <Par
_cons

1.651802
9

.2116151
.0926003

7.81
97.19

0.000
0.000

1.237044
8.818507

2.06656
9.181493

var(e.phyab1)
var(e.phyab2)
var(e.phyab3)
var(e.phyab4)
var(e.appe~1)

2.07466
2.618638
1.231013
1.019261
1.986955

.2075636
.252693
.2062531
.1600644
.2711164

1.705244
2.167386
.8864333
.7492262
1.520699

2.524103
3.163841
1.70954
1.386621
2.596169

example 15 Higher-order CFA

var(e.appe~2)
var(e.appe~3)
var(e.appe~4)
var(e.peer~1)
var(e.peer~2)
var(e.peer~3)
var(e.peer~4)
var(e.parr~1)
var(e.parr~2)
var(e.parr~3)
var(e.parr~4)
var(e.Phys)
var(e.Appear)
var(e.Peer)
var(e.Par)
var(Nonacad)

2.801673
2.41072
2.374508
1.866632
2.167766
1.824346
1.803918
1.214141
1.789125
1.069717
.8013735
.911538
1.59518
.2368108
.3697854
.3858166

.3526427
.300262
.2872554
.18965
.2288099
.2516762
.212599
.1195921
.1748043
.1767086
.121231
.1933432
.3704939
.1193956
.0915049
.1237638

2.189162
1.888545
1.873267
1.529595
1.762654
1.392131
1.431856
1.000982
1.477322
.7738511
.5957527
.6014913
1.011838
.0881539
.2276755
.2057449

LR test of model vs. saturated: chi2(100) =

229

3.585561
3.077276
3.009868
2.277933
2.665984
2.390749
2.272659
1.472692
2.166738
1.478702
1.077963
1.381403
2.514828
.6361528
.600597
.7234903

219.48, Prob > chi2 = 0.0000

Notes:
1. The idea behind this model is that physical ability, appearance, and relationships with peers and
parents may be determined by a latent variable containing nonacademic traits. This model was
suggested by Bollen (1989, 315).
2. sem automatically provided normalization constraints for the first-order factors Phys, Appear,
Peer, and Par. Their path coefficients were set to 1.
3. sem automatically provided a normalization constraint for the second-order factor Nonacad. Its
path coefficient was set to 1.

Fitting the model with the Builder

Use the diagram above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/sem_hcfa1

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Enlarge the size of the canvas to accommodate the length of the diagram.
Click on the Adjust Canvas Size button,
7 (inches), and then click on OK.

, in the Standard Toolbar, change the second size to

4. Change the size of the observed variables rectangles.

a. In the SEM Builder menu, select Settings > Variables > All Observed....
b. In the resulting dialog box, change the second size to .25 and click on OK.
5. Create the measurement component for physical ability.
Select the Add Measurement Component tool, . Then using the darker one-inch grid lines in
the background as a guide, click in the diagram about two inches in from the left and one inch
down from the top.

230

example 15 Higher-order CFA

In the resulting dialog box,

a. change the Latent variable name to Phys;
b. select phyab1, phyab2, phyab3, and phyab4 by using the Measurement variables control;
c. select Left in the Measurement direction control;
d. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
6. Create the remaining first-order measurement components.
a. Repeat the process from item 5, but place the measurement component on the grid line two
inches in from the left and about two and one-half inches down from the top. Label the
latent variable Appear, and select measurement variables appear1, appear2, appear3,
and appear4.
b. Repeat the process from item 5, but place the measurement component on the grid line two
inches in from the left and about four inches down from the top. Label the latent variable
Peer, and select measurement variables peerrel1, peerrel2, peerrel3, and peerrel4.
c. Repeat the process from item 5, but place the measurement component on the grid line two
inches in from the left and about five and one-half inches down from the top. Label the
latent variable Par, and select measurement variables parrel1, parrel2, parrel3, and
parrel4.
7. Create the second-order latent variable.
a. Select the Add Latent Variable tool, , and then click in the diagram about two inches in
from the right and vertically centered between the Appear and Peer latent variables.
b. In the Contextual Toolbar, type Nonacad in the Name control and press Enter.
8. Create paths from Nonacad to each of the first-order latent variables.
a. Select the Add Path tool,

b. Click in the upper-left quadrant of the Nonacad oval (it will highlight when you hover over
it), and drag a path to the lower-left quadrant of the Phys oval (it will highlight when you
can release to connect the path).
tool, create the following paths by clicking first on the left side of
c. Continuing with the
the Nonacad variable and dragging to the right side of the first-order latent variable.
Nonacad -> Appear
Nonacad -> Peer
Nonacad -> Par
9. Clean up the direction of the errors.
We want the errors for each of the latent variables to be below the latent variable. The errors
for Phys, Appear, and Peer are likely to have been created in other directions.
a. Choose the Select tool,

b. Click in the Phys oval.

c. Click on one of the Error Rotation buttons,
is below the latent variable.

, in the Contextual Toolbar until the error

Repeat this for all errors on latent variables that are not below the latent variable.

example 15 Higher-order CFA

231

10. Clean up the paths.

If you do not like where a path has been connected to its variable, use the Select tool,
, to
click on the path, and then simply click on where it connects to an oval and drag the endpoint.
11. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
SEM estimation options dialog box.
Tip: See the tips of [SEM] example 9 to make creating paths somewhat easier than described
above.
You can open a completed diagram in the Builder by typing
. webgetsem sem_hcfa1

Reference
Bollen, K. A. 1989. Structural Equations with Latent Variables. New York: Wiley.

Also see
[SEM] sem Structural equation model estimation command

Title
example 16 Correlation
Description

Remarks and examples

Also see

Description
sem can be used to produce correlations or covariances between exogenous variables. The advantages
of using sem over Statas correlate command are that you can perform statistical tests on the results
and that you can handle missing values in a more elegant way.
To demonstrate these features, we use
. use http://www.stata-press.com/data/r13/census13
(1980 Census data by state)
. describe
Contains data from http://www.stata-press.com/data/r13/census13.dta
obs:
50
1980 Census data by state
vars:
9
9 Apr 2013 10:09
size:
1,600

variable name
state
brate
pop
medage
division
region
mrgrate
dvcrate
medagesq

storage
type

display
format

value
label

long
long
long
float
int
int
float
float
float

%13.0g
%10.0g
%12.0gc
%9.2f
%8.0g
%-8.0g
%9.0g
%9.0g
%9.0g

state1

division
cenreg

variable label
State
Birth rate
Population
Median age
Census Division
Census region

Sorted by:

See Correlations in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Using sem to obtain correlation matrices
Fitting the model with the Builder
Testing correlations with estat stdize and test

Using sem to obtain correlation matrices

We fit the following model:

mrgrate

dvcrate

232

medage

example 16 Correlation

233

This model does nothing more than estimate the covariances (correlations), something we could
obtain from the correlate command by typing
. correlate mrgrate dvcrate medage
(obs=50)
mrgrate dvcrate
medage
mrgrate
dvcrate
medage

1.0000
0.7700
-0.0177

1.0000
-0.2229

1.0000

. correlate mrgrate dvcrate medage, covariance

(obs=50)
mrgrate dvcrate
medage
mrgrate
dvcrate
medage

.000662
.000063 1.0e-05
-.000769 -.001191

2.86775

As explained in Correlations in [SEM] intro 5, to see results presented as correlations rather than
as covariances, we specify sems standardized option:
. sem ( <- mrgrate dvcrate medage), standardized
Exogenous variables
Observed: mrgrate dvcrate medage
Fitting target model:
Iteration 0:
log likelihood = 258.58985
Iteration 1:
log likelihood = 258.58985
Structural equation model
Estimation method = ml
Log likelihood
= 258.58985

Standardized

Coef.

Number of obs

50

OIM
Std. Err.

P>|z|

[95% Conf. Interval]

4.60
8.75
9.97

0.000
0.000
0.000

.4210282
1.981634
14.15611

1.045474
3.125947
21.08556

.
.
.

.
.
.

mean(mrgrate)
mean(dvcrate)
mean(medage)

.7332509
2.553791
17.62083

.1593002
.291922
1.767749

var(mrgrate)
var(dvcrate)
var(medage)

1
1
1

.
.
.

.7699637

.0575805

13.37

0.000

.6571079

.8828195

-.0176541

.1413773

-0.12

0.901

-.2947485

.2594403

-.222932

.1343929

-1.66

0.097

-.4863373

.0404732

cov(mrgrate,
dvcrate)
cov(mrgrate,
medage)
cov(dvcrate,
medage)

LR test of model vs. saturated: chi2(0)

0.00, Prob > chi2 =

234

example 16 Correlation

Note:
1. The correlations reported are

mrgrate and dvcrate

mrgrate and medage
dvcrate and medage

sem
0.7699637
0.0176541
0.222932

correlate
0.7700
0.0177
0.2229

Fitting the model with the Builder

Use the diagram above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/census13

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Create the set of observed variables.
, and then click in the diagram about halfway
Select the Add Observed Variables Set tool,
down from the top and a quarter of the way in from the left.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. use the Variables control to select the variables in this order: mrgrate, dvcrate, and
medage;
c. select Horizontal in the Orientation control;
d. select the Distances tab;
e. select .5 (inch) in the Distance between variables control;
f. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
Be sure you select the observed variables in the order indicated above; otherwise, the instructions
below for creating covariances will not be correct.
4. Correlate each pair of variables.
a. Select the Add Covariance tool,

b. Click in the top of the mrgrate rectangle, slightly to the right of the center (it will highlight
when you hover over it), and drag a path to the top of the dvcrate rectangle, slightly to
the left of the center (it will highlight when you can release to connect the covariance).
c. Click in the top of the dvcrate rectangle, slightly to the right of the center, and drag a
path to the top of the medage rectangle, slightly to the left of the center.
d. Click in the top of the mrgrate rectangle, slightly to the left of the center, and drag a path
to the top of the medage rectangle, slightly to the right of the center.

example 16 Correlation

235

5. Clean up.
If you do not like where a covariance has been connected to its variable, use the Select tool,
, to click on the covariance, and then simply click on where it connects to an oval and drag
the endpoint. You can also change the bow of the covariance by dragging the control point that
extends from one end of the selected covariance.
6. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
SEM estimation options dialog box.
7. Show standardized estimates.
From the SEM Builder menu, select View > Standardized Estimates.
You can open a completed diagram in the Builder by typing
. webgetsem sem_corr

Testing correlations with estat stdize and test

We can test whether the correlations between median age and marriage and divorce rates are equal
with test by typing
. estat stdize:
///
test _b[cov(medage,mrgrate):_cons] = _b[cov(medage,dvcrate):_cons]

We must prefix test with estat stdize because otherwise we would be testing equality of
covariances; see Displaying other results, statistics, and tests (sem and gsem) in [SEM] intro 7 and
see [SEM] estat stdize.
That we refer to the two correlations (covariances) by typing b[cov(medage,mrgrate): cons]
and b[cov(medage,dvcrate): cons] is something nobody remembers and that we remind
ourselves of by redisplaying sem results with the coeflegend option:
. sem, coeflegend
Structural equation model
Estimation method = ml
Log likelihood
= 258.58985
Coef.

Number of obs

50

Legend

mean(mrgrate)
mean(dvcrate)
mean(medage)

.0186789
.0079769
29.54

_b[mean(mrgrate):_cons]
_b[mean(dvcrate):_cons]
_b[mean(medage):_cons]

var(mrgrate)
var(dvcrate)
var(medage)

.0006489
9.76e-06
2.8104

_b[var(mrgrate):_cons]
_b[var(dvcrate):_cons]
_b[var(medage):_cons]

.0000613

_b[cov(mrgrate,dvcrate):_cons]

-.0007539

_b[cov(mrgrate,medage):_cons]

-.0011674

_b[cov(dvcrate,medage):_cons]

cov(mrgrate,
dvcrate)
cov(mrgrate,
medage)
cov(dvcrate,
medage)

LR test of model vs. saturated: chi2(0)

0.00, Prob > chi2 =

236

example 16 Correlation

We can now obtain the test:

. estat stdize:
>
test _b[cov(medage,mrgrate):_cons] = _b[cov(medage,dvcrate):_cons]
( 1) [cov(mrgrate,medage)]_cons - [cov(dvcrate,medage)]_cons = 0
chi2( 1) =
4.78
Prob > chi2 =
0.0288

Note:
1. We can reject the test at the 5% level.

Also see
[SEM] test Wald test of linear hypotheses
[SEM] estat stdize Test standardized parameters
[R] correlate Correlations (covariances) of variables or coefficients

Title
example 17 Correlated uniqueness model

Description

Remarks and examples

Reference

Also see

Description
To demonstrate a correlated uniqueness model, we use the following summary statistics data:
. use http://www.stata-press.com/data/r13/sem_cu1
(Correlated uniqueness)
. ssd describe
Summary statistics data from
http://www.stata-press.com/data/r13/sem_cu1.dta
obs:
500
Correlated uniqueness
vars:
9
25 May 2013 10:12
(_dta has notes)
variable name

variable label

par_i
szt_i
szd_i
par_c
szt_c
szd_c
par_o
szt_o
szd_o

self-report inventory for paranoid

self-report inventory for schizotypal
self-report inventory for schizoid
clinical interview rating for paranoid
clinical interview rating for schizoty..
clinical interview rating for schizoid
observer rating for paranoid
observer rating for schizotypal
observer rating for schizoid

. notes
_dta:
1. Summary statistics data for Multitrait-Multimethod matrix (a specific
kind of correlation matrix) and standard deviations from Brown, Timothy
A., 2006, _Confirmatory Factor Analysis for Applied Research_, New York,
NY: The Guilford Press.
2. Summary statistics represent a sample of 500 patients who were evaluated
for three personality disorders using three different methods.
3. The personality disorders include paranoid, schizotypal, and schizoid.
4. The methods of evaluation include a self-report inventory, ratings from a
clinical interview, and observational ratings.

See Correlated uniqueness model in [SEM] intro 5 for background.

237

238

example 17 Correlated uniqueness model

Remarks and examples

Remarks are presented under the following headings:
Fitting the model
Fitting the model with the Builder

Fitting the model

We fit the following model:

Par

Szt

Szd

par_i

szt_i

szd_i

par_c

szt_c

szd_c

par_o

szt_o

szd_o

. sem (Par -> par_i par_c par_o)

>
(Szt -> szt_i szt_c szt_o)
>
(Szd -> szd_i szd_c szd_o),
>
covstr(e.par_i e.szt_i e.szd_i, unstructured)
>
covstr(e.par_c e.szt_c e.szd_c, unstructured)
>
covstr(e.par_o e.szt_o e.szd_o, unstructured)
>
standardized
Endogenous variables
Measurement: par_i par_c par_o szt_i szt_c szt_o szd_i szd_c szd_o
Exogenous variables
Latent:
Par Szt Szd
Fitting target
Iteration 0:
Iteration 1:
Iteration 2:
Iteration 3:
Iteration 4:
Iteration 5:
Iteration 6:
Iteration 7:

model:
log likelihood
log likelihood
log likelihood
log likelihood
log likelihood
log likelihood
log likelihood
log likelihood

=
=
=
=
=
=
=
=

-10210.31
-10040.188
-9971.4015
-9918.0037
-9883.6368
-9880.0242
-9879.9961
-9879.9961

(not concave)
(not concave)

example 17 Correlated uniqueness model

Structural equation model
Estimation method = ml
Log likelihood
= -9879.9961
( 1)
( 2)
( 3)

Number of obs

500

[par_i]Par = 1
[szt_i]Szt = 1
[szd_i]Szd = 1
OIM
Std. Err.

Standardized

Coef.

Measurement
par_i <Par

.7119709

.0261858

par_c <Par

.8410183

par_o <Par

P>|z|

[95% Conf. Interval]

27.19

0.000

.6606476

.7632941

.0242205

34.72

0.000

.7935469

.8884897

.7876062

.0237685

33.14

0.000

.7410209

.8341916

szt_i <Szt

.7880887

.0202704

38.88

0.000

.7483594

.8278179

szt_c <Szt

.7675732

.0244004

31.46

0.000

.7197493

.8153972

szt_o <Szt

.8431662

.0181632

46.42

0.000

.807567

.8787653

szd_i <Szd

.7692321

.0196626

39.12

0.000

.7306942

.80777

szd_c <Szd

.8604596

.0179455

47.95

0.000

.8252871

.8956321

szd_o <Szd

.8715597

.0155875

55.91

0.000

.8410086

.9021107

var(e.par_i)
var(e.par_c)
var(e.par_o)
var(e.szt_i)
var(e.szt_c)
var(e.szt_o)
var(e.szd_i)
var(e.szd_c)
var(e.szd_o)
var(Par)
var(Szt)
var(Szd)

.4930975
.2926882
.3796764
.3789163
.4108313
.2890708
.408282
.2596093
.2403837
1
1
1

.0372871
.0407398
.0374404
.0319498
.0374582
.0306291
.0302501
.0308827
.027171
.
.
.

.4251739
.2228049
.3129503
.3211966
.3436006
.2348623
.3530966
.2056187
.192616
.
.
.

.5718722
.3844905
.4606295
.4470082
.4912169
.3557912
.4720922
.3277766
.2999976
.
.
.

239

240

example 17 Correlated uniqueness model

cov(e.par_i,
e.szt_i)
cov(e.par_i,
e.szd_i)
cov(e.par_c,
e.szt_c)
cov(e.par_c,
e.szd_c)
cov(e.par_o,
e.szt_o)
cov(e.par_o,
e.szd_o)
cov(e.szt_i,
e.szd_i)
cov(e.szt_c,
e.szd_c)
cov(e.szt_o,
e.szd_o)
cov(Par,Szt)
cov(Par,Szd)
cov(Szt,Szd)

.2166732

.0535966

4.04

0.000

.1116258

.3217207

.4411039

.0451782

9.76

0.000

.3525563

.5296515

-.1074802

.0691107

-1.56

0.120

-.2429348

.0279743

-.2646125

.0836965

-3.16

0.002

-.4286546

-.1005705

.4132457

.0571588

7.23

0.000

.3012165

.5252749

.3684402

.0587572

6.27

0.000

.2532781

.4836022

.7456394

.0351079

21.24

0.000

.6768292

.8144496

-.3296552

.0720069

-4.58

0.000

-.4707861

-.1885244

.4781276
.3806759
.3590146
.3103837

.0588923
.045698
.0456235
.0466126

8.12
8.33
7.87
6.66

0.000
0.000
0.000
0.000

.3627009
.2911095
.2695941
.2190246

.5935544
.4702422
.4484351
.4017428

LR test of model vs. saturated: chi2(15)

14.37, Prob > chi2 = 0.4976

Notes:
1. We use the correlated uniqueness model fit above to analyze a multitraitmultimethod (MTMM)
matrix. The MTMM matrix was developed by Campbell and Fiske (1959) to evaluate construct
validity of measures. Each trait is measured using different methods, and the correlation matrix
produced is used to evaluate whether measures that are related in theory are related in fact
(convergent validity) and whether measures that are not intended to be related are not related in
fact (discriminant validity).
In this example, the traits are the latent variables Par, Szt, and Szd.
The observed variables are the methodtrait combinations.
The observed traits are the personality disorders paranoid (par), schotypal (szt), and schizoid
(szd). The methods used to measure them are self-report ( i), clinical interview ( c), and
observer rating ( o). Thus variable par i is paranoid (par) measured by self-report ( i).
2. Note our use of the covstructure() option, which we abbreviated to covstr(). We used this
option instead of cov() to save typing; see Correlated uniqueness model in [SEM] intro 5.
3. Large values of the factor loadings (path coefficients) indicate convergent validity.
4. Small correlations between latent variables indicate discriminant validity.

Fitting the model with the Builder

Use the diagram above for reference.
There are many ways to draw this diagram. This one produces the cleanest diagram most easily.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/sem_cu1

example 17 Correlated uniqueness model

241

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Change the size of the observed variables rectangles.
a. In the SEM Builder menu, select Settings > Variables > All Observed....
b. In the resulting dialog box, change the first size to .44 and click on OK.
4. Create the nine measurement variables.
, and then click in the far left of the diagram,
Select the Add Observed Variables Set tool,
about one-third of the way up from the bottom.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. use the Variables control to select the nine variables in this order: par i, szt i, szd i,
par c, szt c, szd c, par o, szt o, and szd o;
c. select Horizontal in the Orientation control;
d. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
Be sure you select the observed variables in the order indicated above; otherwise, the instructions
below for creating covariances among the errors will not be correct.
5. Create the set of three latent variables.
Select the Add Latent Variables Set tool,
, and then click above the szt i rectangle, about
one-third of the way down from the top of the diagram.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. type the three latent variable names Par, Szt, and Szd into the Variables control;
c. select Horizontal in the Orientation control;
d. select the Distances tab;
e. select 1 (inch) from the Distance between variables control;
f. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
We could have forgone specifying the distances between variables and simply dragged each
variable where we wished after it was created.
6. Create the factor-loading paths.
a. Select the Add Path tool,

b. Click in the bottom of the Par oval (it will highlight when you hover over it), and drag a
path to the top of the par i rectangle (it will highlight when you can release to connect
the path).

242

example 17 Correlated uniqueness model

c. Continuing with the

tool, create the following paths by clicking first in the bottom of
the latent variable and dragging it to the top of the observed (measurement) variable:
Szt
Szd
Par
Szt
Szd
Par
Szt
Szd

->
->
->
->
->
->
->
->

szt_i
szd_i
par_c
szt_c
szd_c
par_o
szt_o
szd_o

7. Clean up the direction of the errors.

We want all the errors to be below the measurement variables.
a. Choose the Select tool,

b. Click on the rectangle of any measurement variable whose associated error is not below it.
c. Click on one of the Error Rotation buttons,
is below the measurement variable.

, in the Contextual Toolbar until the error

Repeat this for all errors that are not below the measurement variables.
8. Correlate the errors within the self-report, clinical interview, and observer rating groups.
a. Select the Add Covariance tool,

b. Click in the 2 circle (it will highlight when you hover over it), and drag a covariance to
the 1 circle (it will highlight when you can release to connect the covariance).
c. Continue with the Add Covariance tool,
, to create eight more covariances by clicking
the first-listed error and dragging it to the second-listed error.

3
3
5
6
6
8
9
9

->
->
->
->
->
->
->
->

2
1
4
5
4
7
8
7

The order in which we create the covariances is unimportant. We dragged each covariance from
right to left because the bow of the covariance is outward when we drag in a clockwise direction
and inward when we drag in a counterclockwise direction. Had we connected the opposite way,
we would have needed to use the Contextual Toolbar to mirror the bow of the covariances.
9. Correlate the latent factors.
a. Select the Add Covariance tool,

b. Click in the Par oval and drag a covariance to the Szt oval.
c. Click in the Szt oval and drag a covariance to the Szd oval.
d. Click in the Par oval and drag a covariance to the Szd oval.

example 17 Correlated uniqueness model

243

10. Clean up.

If you do not like where a covariance has been connected to its variable, use the Select tool,
, to click on the covariance, and then simply click on where it connects to an oval and drag
the endpoint. You can also change the bow of the covariance by dragging the control point that
extends from one end of the selected covariance.
11. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
SEM estimation options dialog box.
12. Show standardized estimates.
From the SEM Builder menu, select View > Standardized Estimates.

Reference
Campbell, D. T., and D. W. Fiske. 1959. Convergent and discriminant validation by the multitrait-multimethod matrix.
Psychological Bulletin 56: 81105.

Also see
[SEM] sem Structural equation model estimation command
[SEM] sem and gsem option covstructure( ) Specifying covariance restrictions

Title
example 18 Latent growth model

Description

Remarks and examples

Reference

Also see

Description
To demonstrate a latent growth model, we use the following data:
. use http://www.stata-press.com/data/r13/sem_lcm
. describe
Contains data from http://www.stata-press.com/data/r13/sem_lcm.dta
obs:
359
vars:
4
25 May 2013 11:08
size:
5,744
(_dta has notes)

variable name
lncrime0
lncrime1
lncrime2
lncrime3

storage
type
float
float
float
float

display
format

value
label

%9.0g
%9.0g
%9.0g
%9.0g

variable label
ln(crime
ln(crime
ln(crime
ln(crime

rate)
rate)
rate)
rate)

in
in
in
in

Jan
Mar
May
Jul

&
&
&
&

Feb
Apr
Jun
Aug

Sorted by:
. notes
_dta:
1. Data used in Bollen, Kenneth A. and Patrick J. Curran, 2006, _Latent
Curve Models: A Structural Equation Perspective_. Hoboken, New Jersey:
John Wiley & Sons
2. Data from 1995 Uniform Crime Reports for 359 communities in New York
state.

See Latent growth models in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Fitting the model
Fitting the model with the Builder

244

example 18 Latent growth model

Fitting the model

We fit the following model:

. sem (lncrime0 <- Intercept@1 Slope@0 _cons@0)

>
(lncrime1 <- Intercept@1 Slope@1 _cons@0)
>
(lncrime2 <- Intercept@1 Slope@2 _cons@0)
>
(lncrime3 <- Intercept@1 Slope@3 _cons@0),
>
latent(Intercept Slope)
>
var(e.lncrime0@var e.lncrime1@var
>
e.lncrime2@var e.lncrime3@var)
>
means(Intercept Slope)
Endogenous variables
Measurement: lncrime0 lncrime1 lncrime2 lncrime3
Exogenous variables
Latent:
Intercept Slope
Fitting target model:
Iteration 0:
log likelihood =
Iteration 1:
log likelihood =
Iteration 2:
log likelihood =
Iteration 3:
log likelihood =
Structural equation model
Estimation method = ml
Log likelihood
= -1033.9037
( 1) [lncrime0]Intercept = 1
( 2) [lncrime1]Intercept = 1
( 3) [lncrime1]Slope = 1
( 4) [lncrime2]Intercept = 1
( 5) [lncrime2]Slope = 2
( 6) [lncrime3]Intercept = 1
( 7) [lncrime3]Slope = 3
( 8) [var(e.lncrime0)]_cons ( 9) [var(e.lncrime1)]_cons (10) [var(e.lncrime2)]_cons (11) [lncrime0]_cons = 0
(12) [lncrime1]_cons = 0
(13) [lncrime2]_cons = 0
(14) [lncrime3]_cons = 0

-1034.1038
-1033.9044
-1033.9037
-1033.9037
Number of obs

[var(e.lncrime3)]_cons = 0
[var(e.lncrime3)]_cons = 0
[var(e.lncrime3)]_cons = 0

359

245

246

example 18 Latent growth model

Coef.

OIM
Std. Err.

Measurement
lncri~0 <Intercept
_cons

1
0

(constrained)
(constrained)

lncri~1 <Intercept
Slope
_cons

1
1
0

(constrained)
(constrained)
(constrained)

lncri~2 <Intercept
Slope
_cons

1
2
0

(constrained)
(constrained)
(constrained)

lncri~3 <Intercept
Slope
_cons

1
3
0

(constrained)
(constrained)
(constrained)

mean(Inter~t)
mean(Slope)

5.337915
.1426952

.0407501
.0104574

var(e.lncr~0)
var(e.lncr~1)
var(e.lncr~2)
var(e.lncr~3)
var(Interc~t)
var(Slope)

.0981956
.0981956
.0981956
.0981956
.527409
.0196198

.0051826
.0051826
.0051826
.0051826
.0446436
.0031082

cov(Inter~t,
Slope)

-.034316

.0088848

LR test of model vs. saturated: chi2(8)

130.99
13.65

-3.86
=

P>|z|

[95% Conf. Interval]

0.000
0.000

5.258047
.1221992

5.417784
.1631912

.0885457
.0885457
.0885457
.0885457
.4467822
.0143829

.1088972
.1088972
.1088972
.1088972
.6225858
.0267635

-.0517298

-.0169022

0.000

16.25, Prob > chi2 = 0.0390

Notes:
1. In this example, we have repeated measures of the crime rate in 1995. We will assume that the
underlying rate grows linearly.
2. As explained in Latent growth models in [SEM] intro 5, we assume
lncrimei = Intercept + i Slope
3. sem does not usually report the means of latent exogenous variables because sem automatically
includes the identifying constraint that the means are 0; see How sem (gsem) solves the problem
for you in [SEM] intro 4 and see Default normalization constraints in [SEM] sem.
In this case, sem did not constrain the means to be 0 because we specified sems means() option.
In particular, we specified means(Intercept Slope), which said not to constrain the means of
those two exogenous latent variables and to report the estimated result.
Our model was identified even without the usual 0 constraints on Intercept and Slope because
we specified enough other constraints.
4. We estimate the Intercept to have mean 5.34 and the mean Slope to be 0.14 per two months.
Remember, we have measured crime rates as log base e crime rates.

example 18 Latent growth model

247

5. It might help some to think of this as a mixed model:

. generate id = _n
. reshape long lncrime, i(id) j(year)
. mixed lncrime year || id:year, cov(unstructured) mle var

The mean Intercept and Slope are what mixed would refer to as the coefficients in the
fixed-effects part of the model.
Acock (2013, chap. 4) discusses the use of sem to fit latent growth-curve models in more detail.
Acock demonstrates extensions to the basic model we fit here, such as including time-varying and
time-invariant covariates in the model.

Fitting the model with the Builder

Use the diagram above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/sem_lcm

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Create the four repeated measurements for log crime rate.
Select the Add Observed Variables Set tool, , and then click in the diagram about one-fourth
of the way in from the left and about one-third of the way up from the bottom.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. use the Variables control to select the variables in this order: lncrime0, lncrime1,
lncrime2, and lncrime3;
c. select Horizontal in the Orientation control;
d. select the Distances tab;
e. select .25 (inch) from the Distance between variables control;
f. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
4. Create the two latent variables.
Select the Add Latent Variables Set tool,
, and then click in the diagram about one-third of
the way down from the top and above the left side of the lncrime0 rectangle.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. type the latent variable names Intercept and Slope into the Variables control;
c. select Horizontal in the Orientation control;
d. select the Distances tab;

248

example 18 Latent growth model

e. select 1 (inch) from the Distance between variables control;

f. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
We could have forgone specifying the distances between variables and simply dragged each
variable where we wished after it was created.
5. Create the factor-loading paths.
a. Select the Add Path tool,

b. Click in the bottom-left quadrant of the Intercept oval (it will highlight when you hover
over it), and drag a path to the top of the lncrime0 rectangle (it will highlight when you
can release to connect the path).
c. Continuing with the
tool, create the following paths by clicking first in the bottom of
the latent variable and dragging it to the top of the observed (measurement) variable:
Intercept -> lncrime1
Intercept -> lncrime2
Intercept -> lncrime3
Slope -> lncrime0
Slope -> lncrime1
Slope -> lncrime2
Slope -> lncrime3
6. Clean up the direction of the errors.
We want all the errors to be below the measurement variables.
a. Choose the Select tool,

b. Click on the rectangle of any measurement variable whose associated error is not below it.
c. Click on one of the Error Rotation buttons,
is below the measurement variable.

, in the Contextual Toolbar until the error

Repeat this for all errors that are not below the measurement variables.
7. Create the covariance between Intercept and Slope.
a. Select the Add Covariance tool,

b. Click in the top-right quadrant of the Intercept oval, and drag a covariance to the top left
of the Slope oval.
8. Clean up paths and covariance.
If you do not like where a path has been connected to its variables, use the Select tool,
, to
click on the path, and then simply click on where it connects to a rectangle and drag the endpoint.
Similarly, you can change where the covariance connects to the latent variables by clicking on
the covariance and dragging the endpoint. You can also change the bow of the covariance by
clicking on the covariance and dragging the control point that extends from one end of the
selected covariance.
9. Constrain the intercepts of the measurements to 0.
a. Choose the Select tool,

example 18 Latent growth model

b. Click on the rectangle for lncrime0. In the Contextual Toolbar, type 0 in the
press Enter.

249

box and

c. Repeat this process to add the 0 constraint on the intercept for lncrime1, lncrime2, and
lncrime3.
10. Set constraints on the paths from Intercept to the measurements.
a. Continue with the Select tool,

b. Click on the path from Intercept to lncrime0. In the Contextual Toolbar, type 1 in the
box and press Enter.
c. Repeat this process to add the 1 constraint on the following paths:
Intercept -> lncrime1
Intercept -> lncrime2
Intercept -> lncrime3
11. Set constraints on the paths from Slope to the measurements.
a. Continue with the Select tool,

b. Click on the path from Slope to lncrime0. In the Contextual Toolbar, type 0 in the
box and press Enter.
c. Click on the path from Slope to lncrime1. In the Contextual Toolbar, type 1 in the
box and press Enter.
d. Click on the path from Slope to lncrime2. In the Contextual Toolbar, type 2 in the
box and press Enter.
e. Click on the path from Slope to lncrime3. In the Contextual Toolbar, type 3 in the
box and press Enter.
12. Set equality constraints on the error variances.
a. Continue with the Select tool,

b. Click in the 1 circle, which is the error term for lncrime0. In the Contextual Toolbar,
type var in the
box and press Enter.
c. Repeat this process to add the var constraint on the three remaining error variances: 2 , 3 ,
and 4 .
13. Clean up placement of the constraints.
From the SEM Builder menu, select Settings > Connections > Paths....
In the resulting dialog box, do the following:
a. Click on the Results tab.
b. Click on the Result 1... button at the bottom left.
c. In the Appearance of result 1 - paths dialog box that opens, choose 20 (%) in the Distance
between nodes control.
d. Click on OK on the Appearance of result 1 - paths dialog box.
e. Click on OK on the Connection settings - paths dialog box.

250

example 18 Latent growth model

14. Specify that the means of Intercept and Slope are to be estimated.
a. Choose the Select tool,

b. Double-click on the Intercept oval.

c. In the resulting dialog box, check the Estimate mean box.
d. Click on OK.
e. In the same manner, double-click on Slope and check the Estimate mean box.
15. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
SEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem sem_lcm

Reference
Acock, A. C. 2013. Discovering Structural Equation Modeling Using Stata. Rev. ed. College Station, TX: Stata Press.

Also see
[SEM] sem Structural equation model estimation command

Title
example 19 Creating multiple-group summary statistics data

Description

Remarks and examples

Reference

Also see

Description
The data analyzed in [SEM] example 20 are summary statistics data (SSD) and contain summary
statistics on two groups of subjects, those from grade 4 and those from grade 5. Below we show how
we created this summary statistics dataset.
See [SEM] intro 11 for background on SSD.

Remarks and examples

See [SEM] example 2 for creating a single-group dataset from published covariances. In this
example, we will create a two-group dataset from published correlations, standard deviations, and
means.
Marsh and Hocevar (1985) publish lots of SSD, of which we will enter the data for students in
grade 4 and grade 5 found on pages 579581. In that source, the authors published the correlations,
standard deviations, and means of their variables.
We will set the data for the first group, declare that we have groups and wish to add another, and
set the data for the second group.
Starting with the first group, we will issue the following commands:
.
.
.
.
.

ssd
ssd
ssd
ssd
ssd

init
set obs
set means
set sd
set corr

variable names
values
values
values
values

We will first set the end-of-line delimiter to a semicolon because we are going to have some long
lines. We will be entering SSD for 16 variables!
. #delimit ;
delimiter now ;
. ssd init phyab1
phyab2
phyab3
>
appear1 appear2 appear3
>
peerrel1 peerrel2 peerrel3
>
parrel1 parrel2 parrel3
Summary statistics data initialized.

phyab4
appear4
peerrel4
parrel4 ;
Next use, in any order,

ssd set observations (required)

It is best to do this first.

ssd set means (optional)

Default setting is 0.

ssd set variances or ssd set sd (optional)

Use this only if you have set or will set correlations and, even
then, this is optional but highly recommended. Default setting is 1.

ssd set covariances or ssd set correlations (required)

251

252

example 19 Creating multiple-group summary statistics data

. ssd set obs 134 ;
(value set)
Status:
observations:
set
means: unset
variances or sd: unset
covariances or correlations: unset (required
. ssd set means
>
8.34 8.34 8.37 8.40 7.51 7.22 7.03 7.13
>
8.44 7.62 7.06 7.89 9.32 9.39 8.69 9.13 ;
(values set)
Status:
observations:
set
means:
set
variances or sd: unset
covariances or correlations: unset (required
. ssd set sd
>
1.90 1.75 2.06 1.88 2.30 2.63 2.71 2.42
>
2.05 2.22 2.38 2.12 1.21 1.21 1.71 1.32 ;
(values set)
Status:
observations:
set
means:
set
variances or sd:
set
covariances or correlations: unset (required
. ssd set corr
>
1.0 \
>
.50 1.0 \
>
.59 .46 1.0 \
>
.58 .43 .66 1.0 \
>
.30 .27 .35 .46 1.0 \
>
.32 .34 .38 .39 .71 1.0 \
>
.38 .41 .43 .53 .68 .67 1.0 \
>
.23 .29 .33 .43 .61 .63 .73 1.0 \
>
.43 .32 .40 .42 .36 .34 .45 .42 1.0 \
>
.38 .40 .38 .49 .53 .61 .69 .59 .59 1.0 \
>
.27 .24 .41 .37 .43 .46 .57 .57 .61 .59 1.0 \
>
.43 .41 .37 .47 .51 .45 .63 .61 .59 .58 .65 1.0
>
.20 .14 .15 .18 .22 .21 .13 .03 .15 .19 .12 .14
>
.29 .18 .26 .20 .25 .29 .17 .25 .35 .23 .23 .28
>
.37 .14 .34 .37 .34 .34 .35 .33 .42 .36 .39 .39
>
.13 .10 .16 .21 .33 .28 .23 .22 .23 .25 .23 .28
(values set)
Status:
observations:
set
means:
set
variances or sd:
set
covariances or correlations:
set
. #delimit cr
delimiter now cr

to be set)

to be set)

to be set)

\
1.0
.25
.53
.46

\
1.0 \
.50 1.0 \
.43 .59 1.0 ;

We have now entered the data for the first group, and ssd reports that we have a fully set dataset.

example 19 Creating multiple-group summary statistics data

253

Next we are going to add a second group by typing

. ssd addgroup grade
(new group grade==2 added)
The ssd set commands now modify the new group grade==2. If you need to
modify data for grade==1, place a 1 right after the set. For example,
. ssd set 1 means ...
would modify the means for group grade==1.

The ssd set command now modifies the new group grade==2. If we needed to modify data for
grade==1, we would place a 1 right after the set. For example,
. ssd set 1 means ...

We are not modifying data; however, we are now adding data for the second group. The procedure
for entering the second group is the same as the procedure for entering the first group:
.
.
.
.

ssd
ssd
ssd
ssd

set
set
set
set

obs
means
sd
corr

values
values
values
values

We do that below.
. #delimit ;
delimiter now ;
. ssd set obs 251 ;
(value set for group grade==2)
Status for group grade==2:
observations:
set
means: unset
variances or sd: unset
covariances or correlations: unset (required to be set)
. ssd set corr
>
1.0 \
>
.31 1.0 \
>
.52 .45 1.0 \
>
.54 .46 .70 1.0 \
>
.15 .33 .22 .21 1.0 \
>
.14 .28 .21 .13 .72 1.0 \
>
.16 .32 .35 .31 .59 .56 1.0 \
>
.23 .29 .43 .36 .55 .51 .65 1.0 \
>
.24 .13 .24 .23 .25 .24 .24 .30 1.0 \
>
.19 .26 .22 .18 .34 .37 .36 .32 .38 1.0 \
>
.16 .24 .36 .30 .33 .29 .44 .51 .47 .50 1.0 \
>
.16 .21 .35 .24 .31 .33 .41 .39 .47 .47 .55 1.0 \
>
.08 .18 .09 .12 .19 .24 .08 .21 .21 .19 .19 .20 1.0 \
>
.01 -.01 .03 .02 .10 .13 .03 .05 .26 .17 .23 .26 .33 1.0 \
>
.06 .19 .22 .22 .23 .24 .20 .26 .16 .23 .38 .24 .42 .40 1.0 \
>
.04 .17 .10 .07 .26 .24 .12 .26 .16 .22 .32 .17 .42 .42 .65 1.0 ;
(values set for group grade==2)
Status for group grade==2:
observations:
means:
variances or sd:
covariances or correlations:

set
unset
unset
set

254

example 19 Creating multiple-group summary statistics data

. ssd set sd
1.84 1.94 2.07 1.82 2.34 2.61 2.48 2.34
>
1.71 1.93 2.18 1.94 1.31 1.57 1.77 1.47 ;
(values set for group grade==2)
Status for group grade==2:
observations:
set
means: unset
variances or sd:
set
covariances or correlations:
set
. ssd set means 8.20 8.23 8.17 8.56 7.41 7.00 7.17 7.40
>
8.81 7.94 7.52 8.29 9.35 9.13 8.67 9.00 ;
(values set for group grade==2)
Status for group grade==2:
observations:
set
means:
set
variances or sd:
set
covariances or correlations:
set
. #delimit cr
delimiter now cr

We could stop here and save the data in a Stata dataset. We might type
. save sem_2fmmby

However, we intend to use these data as an example in this manual and online. Here is what you
would see if you typed ssd describe:
. ssd describe
Summary statistics data
obs:
385
vars:
16
variable name

variable label

phyab1
phyab2
phyab3
phyab4
appear1
appear2
appear3
appear4
peerrel1
peerrel2
peerrel3
peerrel4
parrel1
parrel2
parrel3
parrel4
Group variable:
Obs. by group:

grade (2 groups)
134, 251

example 19 Creating multiple-group summary statistics data

255

We are going to label these data so that ssd describe can provide more information:
. label data "two-factor CFA"
. label var phyab1
"Physical ability 1"
. label var phyab2
"Physical ability 2"
. label var phyab3
"Physical ability 3"
. label var phyab4
"Physical ability 4"
. label var appear1
"Appearance 1"
. label var appear2
"Appearance 2"
. label var appear3
"Appearance 3"
. label var appear4
"Appearance 4"
. label var peerrel1 "Relationship w/ peers 1"
. label var peerrel2 "Relationship w/ peers 2"
. label var peerrel3 "Relationship w/ peers 3"
. label var peerrel4 "Relationship w/ peers 4"
. label var parrel1
"Relationship w/ parent 1"
. label var parrel2
"Relationship w/ parent 2"
. label var parrel3
"Relationship w/ parent 3"
. label var parrel4
"Relationship w/ parent 4"
. #delimit ;
delimiter now ;
. notes: Summary statistics data from
>
Marsh, H. W. and Hocevar, D., 1985,
>
"Application of confirmatory factor analysis to the study of
>
self-concept: First- and higher order factor models and their
>
invariance across groups", _Psychological Bulletin_, 97: 562-582. ;
. notes: Summary statistics based on
>
134 students in grade 4 and
>
251 students in grade 5
>
from Sydney, Australia. ;
. notes: Group 1 is grade 4, group 2 is grade 5. ;
. notes: Data collected using the Self-Description Questionnaire
>
and includes sixteen subscales designed to measure
>
nonacademic traits: four intended to measure physical
>
ability, four intended to measure physical appearance,
>
four intended to measure relations with peers, and four
>
intended to measure relations with parents. ;
. #delimit cr
delimiter now cr

We would now save the dataset.

To see ssd describes output with the data labeled, see [SEM] example 20.

Reference
Marsh, H. W., and D. Hocevar. 1985. Application of confirmatory factor analysis to the study of self-concept: Firstand higher order factor models and their invariance across groups. Psychological Bulletin 97: 562582.

Also see
[SEM] ssd Making summary statistics data (sem only)
[SEM] example 20 Two-factor measurement model by group

Title
example 20 Two-factor measurement model by group

Description

Remarks and examples

Reference

Also see

Description
Below we demonstrate sems group() option, which allows fitting models in which path coefficients
and covariances differ across groups of the data, such as for males and females. We use the following
data:
. use http://www.stata-press.com/data/r13/sem_2fmmby
(two-factor CFA)
. ssd describe
Summary statistics data from
http://www.stata-press.com/data/r13/sem_2fmmby.dta
obs:
385
two-factor CFA
vars:
16
25 May 2013 11:11
(_dta has notes)
variable name

variable label

phyab1
phyab2
phyab3
phyab4
appear1
appear2
appear3
appear4
peerrel1
peerrel2
peerrel3
peerrel4
parrel1
parrel2
parrel3
parrel4

Physical ability 1
Physical ability 2
Physical ability 3
Physical ability 4
Appearance 1
Appearance 2
Appearance 3
Appearance 4
Relationship w/ peers 1
Relationship w/ peers 2
Relationship w/ peers 3
Relationship w/ peers 4
Relationship w/ parent 1
Relationship w/ parent 2
Relationship w/ parent 3
Relationship w/ parent 4

Group variable:
Obs. by group:
. notes

grade (2 groups)
134, 251

_dta:
1. Summary statistics data from Marsh, H. W. and Hocevar, D., 1985,
"Application of confirmatory factor analysis to the study of
self-concept: First- and higher order factor models and their invariance
across groups", _Psychological Bulletin_, 97: 562-582.
2. Summary statistics based on 134 students in grade 4 and 251 students in
grade 5 from Sydney, Australia.
3. Group 1 is grade 4, group 2 is grade 5.
4. Data collected using the Self-Description Questionnaire and includes
sixteen subscales designed to measure nonacademic status: four intended
to measure physical ability, four intended to measure physical
appearance, four intended to measure relations with peers, and four
intended to measure relations with parents.

256

example 20 Two-factor measurement model by group

257

Remarks and examples

Remarks are presented under the following headings:
Background
Fitting the model with all the data
Fitting the model with the group() option
Fitting the model with the Builder

Background
See [SEM] intro 6 for background on sems group() option.
We will fit the model

peerrel1

parrel1

peerrel2

parrel2

Peer

Par

peerrel3

parrel3

peerrel4

parrel4

which in command syntax can be written as

(Peer -> peerrel1 peerrel2 peerrel3 peerrel4)
(Par -> parrel1 parrel2 parrel3 parrel4)

///

We are using the same data used in [SEM] example 15, but we are using more of the data and
fitting a different model. To remind you, those data were collected from students in grade 5. The
dataset we are using, however, has data for students from grade 4 and from grade 5, which was
created in [SEM] example 19. We have the following observed variables:
1. Four measures of physical ability.
2. Four measures of appearance.
3. Four measures of quality of relationship with peers.
4. Four measures of quality of relationship with parents.
In this example, we will consider solely the measurement problem, and include only the measurement
variables for the two kinds of relationship quality. We are going to treat quality of relationship with
peers as measures of underlying factor Peer and quality of relationship with parents as measures of
underlying factor Par.
Below we will
1. Fit the model with all the data. This amounts to assuming that the students in grades 4 and 5
are identical in terms of this measurement problem.
2. Fit the model with sems group() option, which will constrain some parameters to be the same
for students in grades 4 and 5 and leave free of constraint the others.

258

example 20 Two-factor measurement model by group

Fitting the model with all the data

Throughout this example, we want you to appreciate that we are using SSD and that matters not
at all. Not one command would have a different syntax or option, or produce a different result, if we
had the real data.
We begin by fitting the model with all the data:
. sem (Peer -> peerrel1 peerrel2 peerrel3 peerrel4)
>
(Par -> parrel1 parrel2 parrel3 parrel4)
Endogenous variables
Measurement: peerrel1 peerrel2 peerrel3 peerrel4 parrel1 parrel2 parrel3 parrel4
Exogenous variables
Latent:
Peer Par
Fitting target model:
Iteration 0:
log likelihood = -5559.545
Iteration 1:
log likelihood = -5558.609
Iteration 2:
log likelihood = -5558.6017
Iteration 3:
log likelihood = -5558.6017
Structural equation model
Number of obs
=
385
Estimation method = ml
Log likelihood
= -5558.6017
( 1) [peerrel1]Peer = 1
( 2) [parrel1]Par = 1

Coef.

OIM
Std. Err.

P>|z|

[95% Conf. Interval]

0.000

8.497534

8.864908

Measurement
peerr~1 <Peer
_cons

1
8.681221

peerr~2 <Peer
_cons

1.113865
7.828623

.09796
.1037547

11.37
75.45

0.000
0.000

.9218666
7.625268

1.305863
8.031979

peerr~3 <Peer
_cons

1.42191
7.359896

.114341
.1149905

12.44
64.00

0.000
0.000

1.197806
7.134519

1.646014
7.585273

peerr~4 <Peer
_cons

1.204146
8.150779

.0983865
.1023467

12.24
79.64

0.000
0.000

1.011312
7.950183

1.39698
8.351375

parrel1 <Par
_cons

1
9.339558

(constrained)
.0648742
143.96

0.000

9.212407

9.46671

parrel2 <Par
_cons

1.112383
9.220494

.1378687
.0742356

8.07
124.21

0.000
0.000

.8421655
9.074994

1.382601
9.365993

parrel3 <Par
_cons

2.037924
8.676961

.204617
.088927

9.96
97.57

0.000
0.000

1.636882
8.502667

2.438966
8.851255

parrel4 <Par
_cons

1.52253
9.045247

.1536868
.0722358

9.91
125.22

0.000
0.000

1.221309
8.903667

1.82375
9.186826

(constrained)
.0937197
92.63

example 20 Two-factor measurement model by group

var(e.peer~1)
var(e.peer~2)
var(e.peer~3)
var(e.peer~4)
var(e.parr~1)
var(e.parr~2)
var(e.parr~3)
var(e.parr~4)
var(Peer)
var(Par)

1.809309
2.193804
1.911874
1.753037
1.120333
1.503003
.9680081
.8498834
1.572294
.5000022

.1596546
.194494
.214104
.1749613
.0899209
.1200739
.1419777
.0933687
.2255704
.093189

cov(Peer,Par)

.4226706

.0725253

LR test of model vs. saturated: chi2(19)

5.83
=

0.000

1.521956
1.843884
1.535099
1.441575
.9572541
1.285162
.7261617
.685245
1.186904
.3469983

2.150916
2.610129
2.381126
2.131792
1.311193
1.757769
1.290401
1.054078
2.082822
.7204709

.2805236

.5648176

259

28.19, Prob > chi2 = 0.0798

Note:
1. We are using SSD with data for two separate groups. There is no hint of that in the output above
because sem combined the summary statistics and produced overall results just as if we had the
real data.

Fitting the model with the group() option

. sem (Peer -> peerrel1 peerrel2 peerrel3 peerrel4)
>
(Par -> parrel1 parrel2 parrel3 parrel4), group(grade)
Endogenous variables
Measurement: peerrel1 peerrel2 peerrel3 peerrel4 parrel1 parrel2 parrel3
parrel4
Exogenous variables
Latent:
Peer Par
Fitting target model:
Iteration 0:
log likelihood =
Iteration 1:
log likelihood =
Iteration 2:
log likelihood =
Iteration 3:
log likelihood =
Iteration 4:
log likelihood =
Iteration 5:
log likelihood =
Iteration 6:
log likelihood =
Iteration 7:
log likelihood =
Iteration 8:
log likelihood =
Iteration 9:
log likelihood =
Iteration 10: log likelihood =
Iteration 11: log likelihood =
Iteration 12: log likelihood =
Iteration 13: log likelihood =
Iteration 14: log likelihood =
Iteration 15: log likelihood =
Iteration 16: log likelihood =
Iteration 17: log likelihood =
Iteration 18: log likelihood =
Structural equation model
Grouping variable = grade
Estimation method = ml
Log likelihood
= -5542.6774

-13049.77
-10819.682
-8873.4568
-6119.7114
-5949.354
-5775.6085
-5713.9178
-5638.1208
-5616.6335
-5595.7507
-5589.9802
-5578.8701
-5574.0162
-5568.0786
-5551.7349
-5544.0052
-5542.7113
-5542.6775
-5542.6774

(not
(not
(not
(not
(not
(not
(not
(not
(not
(not
(not
(not
(not

concave)
concave)
concave)
concave)
concave)
concave)
concave)
concave)
concave)
concave)
concave)
concave)
concave)

Number of obs
Number of groups

=
=

385
2

260

example 20 Two-factor measurement model by group

( 1)
( 2)
( 3)
( 4)
( 5)
( 6)
( 7)
( 8)
( 9)
(10)
(11)
(12)
(13)
(14)
(15)
(16)
(17)
(18)
(19)
(20)

[peerrel1]1bn.grade#c.Peer = 1
[peerrel2]1bn.grade#c.Peer - [peerrel2]2.grade#c.Peer = 0
[peerrel3]1bn.grade#c.Peer - [peerrel3]2.grade#c.Peer = 0
[peerrel4]1bn.grade#c.Peer - [peerrel4]2.grade#c.Peer = 0
[parrel1]1bn.grade#c.Par = 1
[parrel2]1bn.grade#c.Par - [parrel2]2.grade#c.Par = 0
[parrel3]1bn.grade#c.Par - [parrel3]2.grade#c.Par = 0
[parrel4]1bn.grade#c.Par - [parrel4]2.grade#c.Par = 0
[peerrel1]1bn.grade - [peerrel1]2.grade = 0
[peerrel2]1bn.grade - [peerrel2]2.grade = 0
[peerrel3]1bn.grade - [peerrel3]2.grade = 0
[peerrel4]1bn.grade - [peerrel4]2.grade = 0
[parrel1]1bn.grade - [parrel1]2.grade = 0
[parrel2]1bn.grade - [parrel2]2.grade = 0
[parrel3]1bn.grade - [parrel3]2.grade = 0
[parrel4]1bn.grade - [parrel4]2.grade = 0
[peerrel1]2.grade#c.Peer = 1
[parrel1]2.grade#c.Par = 1
[mean(Peer)]1bn.grade = 0
[mean(Par)]1bn.grade = 0

Coef.
Measurement
peerr~1 <Peer
[*]
_cons
[*]

OIM
Std. Err.

P>|z|

[95% Conf. Interval]

(constrained)

8.466539

.1473448

57.46

0.000

8.177748

8.755329

peerr~2 <Peer
[*]
_cons
[*]

1.109234

.0975279

11.37

0.000

.9180833

1.300385

7.589872

.1632145

46.50

0.000

7.269977

7.909766

peerr~3 <Peer
[*]
_cons
[*]

1.409361

.1138314

12.38

0.000

1.186256

1.632467

7.056996

.1964299

35.93

0.000

6.672001

7.441992

peerr~4 <Peer
[*]
_cons
[*]

1.195982

.0980272

12.20

0.000

1.003852

1.388112

7.89358

.169158

46.66

0.000

7.562036

8.225123

parrel1 <Par
[*]
_cons
[*]

parrel2 <Par
[*]
_cons
[*]

(constrained)

9.368654

.0819489

114.32

0.000

9.208037

9.529271

1.104355

.1369365

8.06

0.000

.8359649

1.372746

9.287629

.0903296

102.82

0.000

9.110587

9.464672

example 20 Two-factor measurement model by group

parrel3 <Par
[*]
_cons
[*]
parrel4 <Par
[*]
_cons
[*]
mean(Peer)
1
2
mean(Par)
1
2
var(e.peer~1)
1
2
var(e.peer~2)
1
2
var(e.peer~3)
1
2
var(e.peer~4)
1
2
var(e.parr~1)
1
2
var(e.parr~2)
1
2
var(e.parr~3)
1
2
var(e.parr~4)
1
2
var(Peer)
1
2
var(Par)
1
2
cov(Peer,Par)
1
2

2.05859

.2060583

9.99

0.000

1.654723

2.462457

8.741898

.136612

63.99

0.000

8.474144

9.009653

1.526706

.1552486

9.83

0.000

1.222424

1.830987

9.096609

.1061607

85.69

0.000

8.888538

9.30468

0
.3296841

(constrained)
.1570203
2.10

0.036

.02193

.6374382

0
-.0512439

(constrained)
.0818255
-0.63

0.531

-.211619

.1091313

1.824193
1.773813

.2739446
.1889104

1.359074
1.439644

2.448489
2.185549

2.236974
2.165228

.3310875
.2321565

1.673699
1.75484

2.989817
2.671589

1.907009
1.950679

.3383293
.2586196

1.346908
1.504298

2.700023
2.529516

1.639881
1.822448

.272764
.2151827

1.18367
1.445942

2.271925
2.296992

.9669121
1.213159

.1302489
.1192634

.7425488
1.000547

1.259067
1.470949

.9683878
1.79031

.133192
.1747374

.7395628
1.478596

1.268012
2.167739

.8377567
1.015707

.1986089
.1713759

.526407
.7297073

1.333258
1.4138

.8343032
.8599648

.1384649
.1165865

.6026352
.6592987

1.15503
1.121706

2.039297
1.307976

.3784544
.2061581

1.41747
.9603661

2.933912
1.781406

.4492996
.5201696

.1011565
.1029353

.2889976
.3529413

.6985183
.7666329

.5012091
.3867156

.1193333
.079455

.2673201
.2309867

.7350982
.5424445

4.20
4.87

0.000
0.000

Note: [*] identifies parameter estimates constrained to be equal across

groups.
LR test of model vs. saturated: chi2(50) =
61.91, Prob > chi2 = 0.1204

261

262

example 20 Two-factor measurement model by group

Notes:
1. In Which parameters vary by default, and which do not in [SEM] intro 6, we wrote that, generally
speaking, when we specify group(groupvar), the measurement part of the model is constrained
by default to be the same across the groups, whereas the remaining parts will have separate
parameters for each group.
More precisely, we revealed that sem classifies each parameter into one of nine classes, which
are the following:
Class description

Class name

1. structural coefficients
2. structural intercepts

scoef
scons

3. measurement coefficients
4. measurement intercepts

mcoef
mcons

5. covariances of structural errors

6. covariances of measurement errors
7. covariances between structural and measurement errors

serrvar
merrvar
smerrcov

8. means of exogenous variables

9. covariances of exogenous variables

meanex
covex

(*)
(*)

10. all the above

all
(*)
11. none of the above
none
(*) Be aware that classes 8, 9, and 10 (meanex, covex, and all) exclude the
observed exogenous variablesinclude only the latent exogenous
variablesunless you specify option noxconditional or the
noxconditional option is otherwise implied; see
[SEM] sem option noxconditional. This is what you would desire
in most cases.
By default, classes 3 and 4 are constrained to be equal and the rest are allowed to vary.
2. Thus you might expect that most of the parameters of our model would have been left unconstrained
until you remember that we are fitting a measurement model. That is why sem listed 20 constraints
at the top of the estimation results. Some of the constraints are substantive and some are
normalization.
3. In the output, paths listed with an asterisk are constrained to be equal across groups.
Paths labeled with group 1 and group 2 are group specific (unconstrained).
In our data, group 1 corresponds with students in grade 4, and group 2 corresponds with students
in grade 5.
4. It may surprise you that the output contains estimates for the means of the latent variables.
Usually, sem does not report this.
Usually, you are running on only one group of data and those means cannot be estimated, at least
not without additional identifying constraints. When you are running on two or more groups, the
means for all the groups except one can be estimated.
In [SEM] example 21, we use estat ggof to evaluate goodness of fit group by group.
In [SEM] example 22, we use estat ginvariant to test whether parameters that are constrained
across groups should not be and whether parameters that are not constrained could be.

example 20 Two-factor measurement model by group

263

In [SEM] example 23, we show how to constrain the parameters we choose to be equal across
groups.

Fitting the model with the Builder

Use the diagram above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/sem_2fmmby

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Create the measurement component for relationships with peers.
Select the Add Measurement Component tool, , and then click in the diagram about halfway
down from the top and about one-third of the way in from the left.
In the resulting dialog box,
a. change the Latent variable name to Peer;
b. select peerrel1, peerrel2, peerrel3, and peerrel4 by using the Measurement variables
control;
c. select Left in the Measurement direction control;
d. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
4. Create the measurement component for relationships with parents.
Select the Add Measurement Component tool, , and then click in the diagram about halfway
down from the top and about one-third of the way in from the right.
In the resulting dialog box,
a. change the Latent variable name to Par;
b. select parrel1, parrel2, parrel3, and parrel4 by using the Measurement variables
control;
c. select Right in the Measurement direction control;
d. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
5. Correlate the latent variables.
a. Select the Add Covariance tool,

b. Click in the upper-right quadrant of the Peer oval (it will highlight when you hover over
it), and drag a covariance to the upper-left quadrant of the Par oval (it will highlight when
you can release to connect the covariance).

264

example 20 Two-factor measurement model by group

6. Clean up.
If you do not like where a covariance has been connected to its variable, use the Select tool,
, to click on the covariance, and then simply click on where it connects to an oval and drag
the endpoint. You can also change the bow of the covariance by dragging the control point that
extends from one end of the selected covariance.
7. Estimate.
Click on the Estimate button,

, in the Standard Toolbar.

In the resulting dialog box, do the following:

a. Select the Group tab.
b. Select the Group analysis radio button. The variable grade should appear in the Group
variable control.
c. Click on OK.
d. In the Standard Toolbar, use the Group control to toggle between results for group 1 and
group 2.
You can open a completed diagram in the Builder by typing
. webgetsem sem_2fmmby

Reference
Acock, A. C. 2013. Discovering Structural Equation Modeling Using Stata. Rev. ed. College Station, TX: Stata Press.

Also see
[SEM] example 3 Two-factor measurement model
[SEM] example 19 Creating multiple-group summary statistics data
[SEM] example 21 Group-level goodness of fit
[SEM] example 22 Testing parameter equality across groups
[SEM] example 23 Specifying parameter constraints across groups

[SEM] intro 6 Comparing groups (sem only)

[SEM] sem Structural equation model estimation command
[SEM] sem group options Fitting models on different groups

Title
example 21 Group-level goodness of fit
Description

Remarks and examples

Also see

Description
Below we demonstrate the estat ggof command, which may be used after sem with the group()
option. estat ggof displays group-by-group goodness-of-fit statistics.
We pick up where [SEM] example 20 left off:
. use http://www.stata-press.com/data/r13/sem_2fmmby
. sem (Peer -> peerrel1 peerrel2 peerrel3 peerrel4)
///
(Par -> parrel1 parrel2 parrel3 parrel4), group(grade)

Remarks and examples

. estat ggof
Group-level fit statistics

grade
1
2

SRMR

CD

134
251

0.088
0.056

0.969
0.955

Note: group-level chi-squared are not

reported because of constraints between
groups.

Notes:
1. Reported are the goodness-of-fit tests that estat gof, stats(residuals) would report. The
difference is that they are reported for each group rather than overall.
2. If the fit is good, then SRMR (standardized root mean squared residual) will be close to 0 and
CD (the coefficient of determination) will be near 1.
It is also appropriate to run estat gof to obtain overall results:
. estat gof, stats(residuals)
Fit statistic

Value

Description

Size of residuals
SRMR
CD

0.074
0.958

Standardized root mean squared residual

Coefficient of determination

Also see
[SEM] example 20 Two-factor measurement model by group
[SEM] example 4 Goodness-of-fit statistics

[SEM] estat ggof Group-level goodness-of-fit statistics

[SEM] estat gof Goodness-of-fit statistics
265

Title
example 22 Testing parameter equality across groups

Description

Remarks and examples

Also see

Description
Below we demonstrate estat ginvariant to test parameters across groups.
We pick up where [SEM] example 20 left off:
. use http://www.stata-press.com/data/r13/sem_2fmmby
. sem (Peer -> peerrel1 peerrel2 peerrel3 peerrel4)
///
(Par -> parrel1 parrel2 parrel3 parrel4), group(grade)

Remarks and examples

We use estat ginvariant to test whether parameters that are constrained to be equal across
groups should not be and whether parameters that are not constrained across groups could be.

266

example 22 Testing parameter equality across groups

267

. estat ginvariant
Tests for group invariance of parameters

chi2

Wald Test
df

p>chi2

chi2

Score Test
df

p>chi2

Measurement
peerr~1 <Peer
_cons

.
.

.
.

.
.

2.480
0.098

1
1

0.1153
0.7537

peerr~2 <Peer
_cons

.
.

.
.

.
.

0.371
0.104

1
1

0.5424
0.7473

peerr~3 <Peer
_cons

.
.

.
.

.
.

2.004
0.002

1
1

0.1568
0.9687

peerr~4 <Peer
_cons

.
.

.
.

.
.

0.239
0.002

1
1

0.6246
0.9611

parrel1 <Par
_cons

.
.

.
.

.
.

0.272
0.615

1
1

0.6019
0.4329

parrel2 <Par
_cons

.
.

.
.

.
.

0.476
3.277

1
1

0.4903
0.0703

parrel3 <Par
_cons

.
.

.
.

.
.

3.199
1.446

1
1

0.0737
0.2291

parrel4 <Par
_cons

.
.

.
.

.
.

2.969
0.397

1
1

0.0849
0.5288

var(e.peer~1)
var(e.peer~2)
var(e.peer~3)
var(e.peer~4)
var(e.parr~1)
var(e.parr~2)
var(e.parr~3)
var(e.parr~4)
var(Peer)
var(Par)

0.024
0.033
0.011
0.294
1.981
14.190
0.574
0.022
4.583
0.609

1
1
1
1
1
1
1
1
1
1

0.8772
0.8565
0.9152
0.5879
0.1593
0.0002
0.4486
0.8813
0.0323
0.4350

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.

cov(Peer,Par)

0.780

0.3772

Notes:
1. In the output above, score tests are reported for parameters that were constrained. The null
hypothesis is that the constraint is valid. None of the tests reject a valid constraint.
2. Wald tests are reported for parameters that were not constrained. The null hypothesis is that a
constraint would be valid. Only in two cases does it appear that grade 4 differs from grade 5,
namely, the variance of e.parrel2 and the variance of Peer.

268

example 22 Testing parameter equality across groups

3. We remind you that these tests are marginal tests. That is, each test is intended to be interpreted
separately. These are not joint tests of simultaneous imposition or relaxation of constraints. If you
want simultaneous tests, you must do them yourself by using, for instance, the test command.
If joint tests of parameter classes are desired, the class option can be used.
These results imply that none of the constraints we impose should be relaxed, and that perhaps we
could constrain all the variances and covariances to be equal across groups except for the variances
of e.parrel2 and Peer. We do that in [SEM] example 23.

Also see
[SEM] example 20 Two-factor measurement model by group
[SEM] example 23 Specifying parameter constraints across groups

[SEM] estat ginvariant Tests for invariance of parameters across groups

Title
example 23 Specifying parameter constraints across groups
Description

Remarks and examples

Also see

Description
Below we demonstrate how to constrain the parameters we want constrained to be equal across
groups when using sem with the group() option.
We pick up where [SEM] example 22 left off:
. use http://www.stata-press.com/data/r13/sem_2fmmby
. sem (Peer -> peerrel1 peerrel2 peerrel3 peerrel4)
///
(Par -> parrel1 parrel2 parrel3 parrel4), group(grade)
. estat ginvariant

The estat ginvariant command implied that perhaps we could constrain all the variances and
covariances to be equal across groups except for the variances of e.parrel2 and Peer.

Remarks and examples

Remarks are presented under the following headings:
Background
Fitting the constrained model

Background
We can specify which parameters we wish to allow to vary. Remember that sems group() option
classifies the parameters of the model as follows:
Class description

Class name

1. structural coefficients
2. structural intercepts

scoef
scons

3. measurement coefficients
4. measurement intercepts

mcoef
mcons

5. covariances of structural errors

6. covariances of measurement errors
7. covariances between structural and measurement errors

serrvar
merrvar
smerrcov

8. means of exogenous variables

9. covariances of exogenous variables

meanex
covex

(*)
(*)

10. all the above

all
(*)
11. none of the above
none
(*) Exogenous variables means just the latent exogenous variables unless you specify
sem option noxconditional or you specify option method(mlmv)
(which implies option noxconditional); see [SEM] sem option noxconditional.
269

270

example 23 Specifying parameter constraints across groups

When fitting a model with the group() option,

. sem ..., ... group(varname)

you may also specify the ginvariant() option:

. sem ..., ... group(varname) ginvariant(class names)

You may specify any of the class names as being ginvariant(). You may specify as many class
names as you wish. When you specify ginvariant(), sem cancels its default actions on which
parameters vary and which do not, and uses the information you specify. All classes that you do not
mention as being ginvariant() are allowed to vary across groups.
By using ginvariant(), you can constrain, or free by your silence, whole classes of parameters.
For instance, you could type
. sem ..., group(mygroup) ginvariant(mcoef mcons serrvar)

and you are constraining those parameters to be equal across groups and leaving unconstrained scoef,
scons, merrvar, smerrcov, meanex, and covex.
In addition, if a class is constrained, you can still unconstrain individual coefficients. Consider the
model
. sem ... (x1<-L) ...

If you typed
. sem ... (1: x1<-L@a1) (2: x1<-L@a2) ..., group(mygroup) ginvariant(all)

then all estimated parameters would be the same across groups except for the path x1<-L, and it
would be free to vary in groups 1 and 2.
By the same token, if a class is unconstrained, you can still constrain individual coefficients. If
you typed
. sem ... (1: x1<-L@a) (2: x1<-L@a) ..., group(mygroup) ginvariant(none)

then you would leave unconstrained all parameters except the path x1<-L, and it would be constrained
to be equal in groups 1 and 2.
This is all discussed in [SEM] intro 6, including how to constrain and free variance and covariance
parameters.

Fitting the constrained model

In our case, we wish to fit our model:
. sem (Peer -> peerrel1 peerrel2 peerrel3 peerrel4)
(Par -> parrel1 parrel2 parrel3 parrel4),
group(grade)

///
///

example 23 Specifying parameter constraints across groups

271

We impose constraints on all parameters except the variances of e.parrel2 and Peer. We can do
that by typing
. sem (Peer -> peerrel1 peerrel2 peerrel3 peerrel4)
>
(Par -> parrel1 parrel2 parrel3 parrel4),
>
group(grade)
>
ginvariant(all)
>
var(1: e.parrel2@v1)
>
var(2: e.parrel2@v2)
>
var(1: Peer@v3)
>
var(2: Peer@v4)
Endogenous variables
Measurement:

peerrel1 peerrel2 peerrel3 peerrel4 parrel1 parrel2 parrel3

parrel4
Exogenous variables
Latent:
Peer Par
Fitting target model:
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration

0:
1:
2:
3:
4:
5:

log
log
log
log
log
log

likelihood
likelihood
likelihood
likelihood
likelihood
likelihood

=
=
=
=
=
=

-5560.9934
-5552.3122
-5549.5391
-5549.3528
-5549.3501
-5549.3501

Structural equation model

Number of obs
Grouping variable = grade
Number of groups
Estimation method = ml
Log likelihood
= -5549.3501
( 1) [peerrel1]1bn.grade#c.Peer = 1
( 2) [peerrel2]1bn.grade#c.Peer - [peerrel2]2.grade#c.Peer = 0
( 3) [peerrel3]1bn.grade#c.Peer - [peerrel3]2.grade#c.Peer = 0
( 4) [peerrel4]1bn.grade#c.Peer - [peerrel4]2.grade#c.Peer = 0
( 5) [parrel1]1bn.grade#c.Par = 1
( 6) [parrel2]1bn.grade#c.Par - [parrel2]2.grade#c.Par = 0
( 7) [parrel3]1bn.grade#c.Par - [parrel3]2.grade#c.Par = 0
( 8) [parrel4]1bn.grade#c.Par - [parrel4]2.grade#c.Par = 0
( 9) [var(e.peerrel1)]1bn.grade - [var(e.peerrel1)]2.grade = 0
(10) [var(e.peerrel2)]1bn.grade - [var(e.peerrel2)]2.grade = 0
(11) [var(e.peerrel3)]1bn.grade - [var(e.peerrel3)]2.grade = 0
(12) [var(e.peerrel4)]1bn.grade - [var(e.peerrel4)]2.grade = 0
(13) [var(e.parrel1)]1bn.grade - [var(e.parrel1)]2.grade = 0
(14) [var(e.parrel3)]1bn.grade - [var(e.parrel3)]2.grade = 0
(15) [var(e.parrel4)]1bn.grade - [var(e.parrel4)]2.grade = 0
(16) [cov(Peer,Par)]1bn.grade - [cov(Peer,Par)]2.grade = 0
(17) [var(Par)]1bn.grade - [var(Par)]2.grade = 0
(18) [peerrel1]1bn.grade - [peerrel1]2.grade = 0
(19) [peerrel2]1bn.grade - [peerrel2]2.grade = 0
(20) [peerrel3]1bn.grade - [peerrel3]2.grade = 0
(21) [peerrel4]1bn.grade - [peerrel4]2.grade = 0
(22) [parrel1]1bn.grade - [parrel1]2.grade = 0
(23) [parrel2]1bn.grade - [parrel2]2.grade = 0
(24) [parrel3]1bn.grade - [parrel3]2.grade = 0
(25) [parrel4]1bn.grade - [parrel4]2.grade = 0
(26) [peerrel1]2.grade#c.Peer = 1
(27) [parrel1]2.grade#c.Par = 1

=
=

385
2

272

example 23 Specifying parameter constraints across groups

Coef.
Measurement
peerr~1 <Peer
[*]
_cons
[*]

OIM
Std. Err.

P>|z|

[95% Conf. Interval]

(constrained)

8.708274

.0935844

93.05

0.000

8.524852

8.891696

peerr~2 <Peer
[*]
_cons
[*]

1.112225

.0973506

11.42

0.000

.9214217

1.303029

7.858713

.1035989

75.86

0.000

7.655663

8.061763

peerr~3 <Peer
[*]
_cons
[*]

1.416486

.113489

12.48

0.000

1.194052

1.638921

7.398217

.1147474

64.47

0.000

7.173316

7.623118

peerr~4 <Peer
[*]
_cons
[*]

1.196494

.0976052

12.26

0.000

1.005191

1.387796

8.183148

.1021513

80.11

0.000

7.982936

8.383361

parrel1 <Par
[*]
_cons
[*]

(constrained)

9.339558

.0648742

143.96

0.000

9.212407

9.46671

parrel2 <Par
[*]
_cons
[*]

1.100315

.1362999

8.07

0.000

.8331722

1.367458

9.255299

.0725417

127.59

0.000

9.11312

9.397478

parrel3 <Par
[*]
_cons
[*]

2.051278

.2066714

9.93

0.000

1.64621

2.456347

8.676961

.088927

97.57

0.000

8.502667

8.851255

parrel4 <Par
[*]
_cons
[*]

1.529938

.154971

9.87

0.000

1.2262

1.833675

9.045247

.0722358

125.22

0.000

8.903667

9.186826

example 23 Specifying parameter constraints across groups

var(e.peer~1)
[*]
var(e.peer~2)
[*]
var(e.peer~3)
[*]
var(e.peer~4)
[*]
var(e.parr~1)
[*]
var(e.parr~2)
1
2
var(e.parr~3)
[*]
var(e.parr~4)
[*]
var(Peer)
1
2
var(Par)
[*]
cov(Peer,Par)
[*]

1.799133

.159059

1.512898

2.139523

2.186953

.193911

1.838086

2.602035

1.915661

.2129913

1.54056

2.382094

1.767354

.1746104

1.45622

2.144965

1.125082

.0901338

.9615942

1.316366

.9603043
1.799668

.13383
.1747351

.730775
1.487807

1.261927
2.176898

.9606889

.1420406

.7190021

1.283617

.8496935

.0933448

.6850966

1.053835

1.951555
1.361431

.3387796
.2122853

1.388727
1.002927

2.742489
1.848084

.4952527

.0927994

.3430288

.7150281

.4096197

.0708726

.2707118

.5485275

5.78

0.000

273

Note: [*] identifies parameter estimates constrained to be equal across

groups.
LR test of model vs. saturated: chi2(61) =
75.25, Prob > chi2 = 0.1037

Notes:
1. In [SEM] example 20, we previously fit this model by typing
. sem (...) (...), group(grade)

This time, we typed

. sem (...) (...), group(grade)
ginvariant(all)
var(1: e.parrel2@v1)
var(2: e.parrel2@v2)
var(1: Peer@v3)
var(2: Peer@v4)

///
///
///
///
///

2. Previously, sem, group() mentioned 20 constraints that it imposed because of normalization or

because of assumed ginvariant(mcoef mcons).
This time, sem, group() mentioned 27 constraints. It applied more constraints because we
specified ginvariant(all).
3. After the ginvariant(all) option, we relaxed the following constraints:
var(1:
var(2:
var(1:
var(2:

e.parrel2@v1)
e.parrel2@v2)
Peer@v3)
Peer@v4)

ginvariant(all) specified, among other constraints, that

var(1: e.parrel2) == var(2: e.parrel2)
var(1: Peer) == var(2: Peer)

274

example 23 Specifying parameter constraints across groups

ginvariant(all) did that by secretly issuing the options

var(1:
var(2:
var(1:
var(2:

e.parrel2@secretname1)
e.parrel2@secretname1)
Peer@secretname2)
Peer@secretname2)

because that is how you impose equality constraints with the path notation. When we specified
var(1:
var(2:
var(1:
var(2:

e.parrel2@v1)
e.parrel2@v2)
Peer@v3)
Peer@v4)

our new constraints overrode the secretly issued constraints. It would not have worked to leave
off the symbolic names; see [SEM] sem path notation extensions. We specified the symbolic
names v1, v2, v3, and v4. v1 and v2 overrode secretname1, and thus the constraint that
var(e.parrel2) be equal across the two groups was relaxed. v3 and v4 overrode secretname2,
and thus the constraint that var(Peer) be equal across groups was relaxed.

Also see
[SEM] example 20 Two-factor measurement model by group
[SEM] example 22 Testing parameter equality across groups

[SEM] intro 6 Comparing groups (sem only)

[SEM] sem group options Fitting models on different groups

Title
example 24 Reliability

Description

Remarks and examples

Also see

Description
Below we demonstrate sems reliability() option with the following data:
. use http://www.stata-press.com/data/r13/sem_rel
(measurement error with known reliabilities)
. summarize
Variable

Obs

Mean

y
x1
x2

1234
1234
1234

701.081
100.278
100.2066

Std. Dev.

Min

Max

71.79378
14.1552
14.50912

487
51
55

943
149
150

. notes
_dta:
1. Fictional data.
2. Variables x1 and x2 each contain a test score designed to measure X.
test is scored to have mean 100.
3. Variables x1 and x2 are both known to have reliability 0.5.
4. Variable y is the outcome, believed to be related to X.

See [SEM] sem and gsem option reliability( ) for background.

Remarks and examples

Remarks are presented under the following headings:
Baseline model (reliability ignored)
Model with reliability
Model with two measurement variables and reliability

275

The

276

example 24 Reliability

Baseline model (reliability ignored)

. sem (y <- x1)
Endogenous variables
Observed: y
Exogenous variables
Observed: x1
Fitting target model:
Iteration 0:
log likelihood = -11629.745
Iteration 1:
log likelihood = -11629.745
Structural equation model
Estimation method = ml
Log likelihood
= -11629.745

Coef.
Structural
y <x1
_cons
var(e.y)

OIM
Std. Err.

3.54976
345.1184

.1031254
10.44365

2627.401

105.7752

LR test of model vs. saturated: chi2(0)

Number of obs

34.42
33.05

1234

P>|z|

[95% Conf. Interval]

0.000
0.000

3.347637
324.6492

3.751882
365.5876

2428.053

2843.115

0.00, Prob > chi2 =

Notes:
1. In these data, variable x1 is measured with error.
2. If we ignore that, we obtain a path coefficient for y<-x1 of 3.55.
3. We also ran this model for y<-x2. We obtained a path coefficient of 3.48.

example 24 Reliability

277

Model with reliability

. sem (x1<-X) (y<-X), reliability(x1 .5)
Endogenous variables
Measurement: x1 y
Exogenous variables
Latent:
X
Fitting target model:
Iteration 0:
log likelihood = -11745.845
Iteration 1:
log likelihood = -11661.626
Iteration 2:
log likelihood = -11631.469
Iteration 3:
log likelihood = -11629.755
Iteration 4:
log likelihood = -11629.745
Iteration 5:
log likelihood = -11629.745
Structural equation model
Estimation method = ml
Log likelihood
= -11629.745
( 1) [x1]X = 1
( 2) [var(e.x1)]_cons = 100.1036

Coef.
Measurement
x1 <X
_cons

1
100.278

OIM
Std. Err.

Number of obs

(constrained)
.4027933
248.96

P>|z|

1234

[95% Conf. Interval]

0.000

99.4885

101.0674

0.000
0.000

6.408705
697.077

7.790335
705.0851

2.152334
85.48963

5086.411
117.2157

y <X
_cons
var(e.x1)
var(e.y)
var(X)

7.09952
701.081
100.1036
104.631
100.1036

.352463
2.042929

20.14
343.17

(constrained)
207.3381
8.060038

LR test of model vs. saturated: chi2(0)

0.00, Prob > chi2 =

Notes:
1. We wish to estimate the effect of y<-x1 when x1 is measured with error (0.50 reliability). To
do that, we introduce latent variable X and write our model as (x1<-X) (y<-X).
2. When we ignored the measurement error of x1, we obtained a path coefficient for y<-x1 of
3.55. Taking into account the measurement error, we obtain a coefficient of 7.1.

278

example 24 Reliability

Model with two measurement variables and reliability

. sem (x1 x2<-X) (y<-X), reliability(x1 .5
Endogenous variables

x2 .5)

Measurement: x1 x2 y
Exogenous variables
Latent:
X
Fitting target model:
Iteration 0:
log likelihood = -16258.636
Iteration 1:
log likelihood = -16258.401
Iteration 2:
log likelihood =
-16258.4
Structural equation model
Estimation method = ml
Log likelihood
=
-16258.4
( 1) [x1]X = 1
( 2) [var(e.x1)]_cons = 100.1036
( 3) [var(e.x2)]_cons = 105.1719

Coef.
Measurement
x1 <X
_cons

1
100.278

OIM
Std. Err.

Number of obs

(constrained)
.4037851
248.34

1234

P>|z|

[95% Conf. Interval]

0.000

99.48655

101.0694

x2 <X
_cons

1.030101
100.2066

.0417346
.4149165

24.68
241.51

0.000
0.000

.9483029
99.39342

1.1119
101.0199

X
_cons

7.031299
701.081

.2484176
2.042928

28.30
343.17

0.000
0.000

6.544409
697.077

7.518188
705.0851

39.31868
87.67509

590.1553
116.5591

y <-

var(e.x1)
var(e.x2)
var(e.y)
var(X)

100.1036
105.1719
152.329
101.0907

(constrained)
(constrained)
105.26
7.343656

LR test of model vs. saturated: chi2(2)

0.59, Prob > chi2 = 0.7430

Notes:
1. We wish to estimate the effect of y<-X. We have two measures of Xx1 and x2both measured
with error (0.50 reliability).
2. In the previous section, we used just x1. We obtained path coefficient 7.1 with standard error
0.4. Using both x1 and x2, we obtain path coefficient 7.0 and standard error 0.2.
3. We at StataCorp created these fictional data. The true coefficient is 7.

Also see
[SEM] sem and gsem option reliability( ) Fraction of variance not due to measurement error
[SEM] example 1 Single-factor measurement model

Title
example 25 Creating summary statistics data from raw data

Description

Remarks and examples

Also see

Description
Below we show how to create summary statistics data (SSD) from raw data. We use auto2.dta:
. use http://www.stata-press.com/data/r13/auto2
(1978 Automobile Data)
. describe
(output omitted )
. summarize
(output omitted )

Remarks and examples

Remarks are presented under the following headings:
Preparing data for conversion
Converting to summary statistics form
Publishing SSD
Creating SSD with multiple groups

We are going to create SSD containing the variables price, mpg, weight, displacement, and
foreign.

Preparing data for conversion

Before building the SSD, prepare the data to be converted:
1. Drop variables that you do not intend to include in the SSD. Dropping variables is not a requirement,
but it will be easier to spot problems if you begin by eliminating the irrelevant variables.
2. Verify that you have no string variables in the resulting data. Summary statistics datasets cannot
contain string values.
3. Verify that there are no missing values. If there are, be aware that observations containing one
or more variables with missing values will be omitted from the SSD.
4. Verify that all variables are on a reasonable scale. We recommend that the means of variables be
only 3 or 4 orders of magnitude different from each other. This will help to preserve numerical
accuracy when the SSD are used.
5. Create any new variables containing transformations of existing variables that might be useful
later. Once the data are converted to summary statistics form, you will not be able to create such
variables.
6. Place the variables in a logical order. That will help the user of the SSD understand the data.
7. Save the resulting prepared data. Probably you will never need the prepared data, but one never
knows for sure.
279

280

example 25 Creating summary statistics data from raw data

We take our own advice below:

.
.
.
.
.
.
.
.
.
.
.

* ------------------------------------------------------------------------* Suggestion 1: Keep relevant variables:

*
keep price mpg weight displacement foreign
* ------------------------------------------------------------------------* Suggestion 2: Check for string variables
* Suggestion 3: Verify no missing values
* Suggestion 4: Verify variables on a reasonable scale:
*
summarize
Variable
Obs
Mean
Std. Dev.
Min
Max

price
74
6165.257
2949.496
mpg
74
21.2973
5.785503
74
3019.459
777.1936
weight
74
197.2973
91.83722
displacement
foreign
74
.2972973
.4601885
.
. * We will rescale weight and price:
. replace
weight = weight/1000
weight was int now float
(74 real changes made)
. replace
price = price/1000
price was int now float
(74 real changes made)
. label var weight "Weight (1000s lbs.)"
.
.
.
.

3291
12
1760
79
0

15906
41
4840
425
1

Min

Max

label var price "Price ($1,000s)"

* and now we check our work:
*
summarize
Variable

Obs

Mean

Std. Dev.

price
74
6.165257
2.949496
3.291
15.906
74
21.2973
5.785503
12
41
mpg
weight
74
3.019459
.7771936
1.76
4.84
displacement
74
197.2973
91.83722
79
425
foreign
74
.2972973
.4601885
0
1
.
. * ------------------------------------------------------------------------. * Suggestion 5: Create useful transformations:
. *
. gen
gpm
= 1/mpg
. label var gpm
"Gallons per mile"
.

example 25 Creating summary statistics data from raw data

.
.
.
.
.

281

* ------------------------------------------------------------------------* Suggestion 6: Place variables in logical order:

*
order price mpg gpm
summarize
Obs
Mean
Std. Dev.
Min
Max
Variable

price
mpg
gpm
weight
displacement

74
74
74
74
74

6.165257
21.2973
.0501928
3.019459
197.2973

2.949496
5.785503
.0127986
.7771936
91.83722

3.291
12
.0243902
1.76
79

15.906
41
.0833333
4.84
425

foreign
74
.2972973
.4601885
0
1
.
. * ------------------------------------------------------------------------. * Suggestion 7: save prepared data
. *
. save auto_raw
file auto_raw.dta saved
. * -------------------------------------------------------------------------

Converting to summary statistics form

To create the summary statistics dataset, you just need to type sud build and the names of the
variables to be included. If you have previously kept the relevant variables, you can type ssd build
all.
We recommend the following steps:
1. Convert data to summary statistics form:
. ssd build _all

2. Review the result:

. ssd describe
. notes
. ssd list

3. Digitally sign the data:

. datasignature set

4. Save the data:

. save auto_ss

We follow our advice below. After that, we will show you the advantages of digitally signing the
data.
. * ------------------------------------------------------------------------. * Convert data:
. *
. ssd build _all
(data in memory now summary statistics data; you can use ssd describe and
ssd list to describe and list results.)

282

example 25 Creating summary statistics data from raw data

.
.
.
.

* ------------------------------------------------------------------------* Review results:

*
ssd describe
Summary statistics data
obs:
74
vars:
6
(_dta has notes)
variable name

variable label

price
mpg
gpm
weight
displacement
foreign

Price ($1,000s)
Mileage (mpg)
Gallons per mile
Weight (1000s lbs.)
Displacement (cu. in.)
Car type

. notes
_dta:
1. summary statistics data built from auto_raw.dta on 30 Jun 2012 15:32:33
using -ssd build _all. ssd list
Observations = 74
Means:
price
6.1652567

mpg
21.297297

gpm
.0501928

weight
3.0194595

displacement
197.2973

foreign
.2972973
Variances implicitly defined; they are the diagonal of the covariance
matrix.
Covariances:
price
8.6995258
-7.9962828
.02178417
1.2346748
134.06705
.06612809

mpg

gpm

weight

displacement

33.472047
-.06991586
-3.6294262
-374.92521
1.0473899

.0001638
.00849897
.90648519
-.00212897

.60402985
63.87345
-.21202888

8434.0748
-25.938912

foreign
.21177342
.
.
.
.

* ------------------------------------------------------------------------* Digitally sign:

*
datasignature set
8:8(102846):1810957193:2605039657
(data signature set)

. * ------------------------------------------------------------------------. * Save:
. *
. save auto_ss
file auto_ss.dta saved
. * -------------------------------------------------------------------------

We recommend digitally signing the data. This way, anyone can verify later that the data are
unchanged:
. datasignature confirm
(data unchanged since 30jun2012 15:32)

example 25 Creating summary statistics data from raw data

283

Let us show you what would happen if the data had changed:
. replace mpg = mpg+.0001 in 5
(1 real change made)
. datasignature confirm
data have changed since 30jun2012 15:34
r(9);

There is no reason for you or anyone else to change the SSD after it has been created, so we
recommend that you digitally sign the data. With regular datasets, users do make changes, if only by
adding variables.
Be aware that the data signature is a function of the variable names, so if you rename a variable
something you are allowed to dothe signature will change and datasignature will report, for
example, data have changed since 30jun2012 15:34. Solutions to that problem are discussed in
[SEM] ssd.

Publishing SSD
The summary statistics dataset you have just created can obviously be sent to and used by any
Stata user. If you wish to publish your data in printed form, use ssd describe and ssd list to
describe and list the data.

Creating SSD with multiple groups

The process for creating SSD containing multiple groups is nearly the same as for creating singlegroup data. The only differences are that you do not drop the group variable during preparation and
that rather than typing
. ssd build _all

you type
. ssd build _all, group(varname)

Below we build the automobile SSD again, but this time, we specify group(rep78):
. ssd build _all, group(rep78)

If you think carefully about this, you may be worried that all includes rep78 and thus we will be
including the grouping variable among the summary statistics. ssd build knows to omit the group
variable:
. * ------------------------------------------------------------------------. * Suggestion 1: Keep relevant variables:
. *
. webuse auto2
(1978 Automobile Data)
. keep price mpg weight displacement foreign rep78

284

example 25 Creating summary statistics data from raw data

.
.
.
.
.
.

* ------------------------------------------------------------------------* Suggestion 2: Check for string variables

* Suggestion 3: Verify no missing values
* Suggestion 4: Verify variables on a reasonable scale:
*
summarize
Variable
Obs
Mean
Std. Dev.
Min
Max

price
mpg
rep78
weight
displacement

74
74
69
74
74

6165.257
21.2973
3.405797
3019.459
197.2973

2949.496
5.785503
.9899323
777.1936
91.83722

foreign
74
.2972973
.4601885
. drop if rep78 >= .
(5 observations deleted)
. * We will rescale weight and price:
. replace
weight = weight/1000
weight was int now float
(69 real changes made)
. replace
price = price/1000
price was int now float
(69 real changes made)
. label var weight "Weight (1000s lbs.)"
. label var price "Price ($1,000s)"
. * and now we check our work:
. summarize
Variable
Obs
Mean
Std. Dev.

3291
12
1
1760
79

15906
41
5
4840
425

Min

Max

price
mpg
rep78
weight
displacement

69
69
69
69
69

6.146043
21.28986
3.405797
3.032029
198

2.91244
5.866408
.9899323
.7928515
93.14789

3.291
12
1
1.76
79

15.906
41
5
4.84
425

foreign

69

.3043478

.4635016

.
.
.
.
.
.
.
.
.
.
.
.

* ------------------------------------------------------------------------* Suggestion 5: Create useful transformations:

*
gen
gpm
= 1/mpg
label var gpm
"Gallons per mile"
* ------------------------------------------------------------------------* Suggestion 6: Place variables in logical order:
*
order price mpg gpm
summarize
Obs
Mean
Std. Dev.
Min
Max
Variable
price
mpg
gpm
rep78
weight

69
69
69
69
69

6.146043
21.28986
.0502584
3.405797
3.032029

2.91244
5.866408
.0128353
.9899323
.7928515

3.291
12
.0243902
1
1.76

15.906
41
.0833333
5
4.84

displacement
foreign

69
69

198
.3043478

93.14789
.4635016

79
0

425
1

example 25 Creating summary statistics data from raw data

. * ------------------------------------------------------------------------. * Suggestion 7: save prepared data
. *
. save auto_group_raw
file auto_group_raw.dta saved
. * ------------------------------------------------------------------------.
. * ------------------------------------------------------------------------. * Convert data:
. *
. ssd build _all, group(rep78)
(data in memory now summary statistics data; you can use ssd describe and
ssd list to describe and list results.)
. * ------------------------------------------------------------------------. * Review results:
. *
. ssd describe
Summary statistics data
obs:
69
vars:
6
(_dta has notes)
variable name

variable label

price
mpg
gpm
weight
displacement
foreign

Price ($1,000s)
Mileage (mpg)
Gallons per mile
Weight (1000s lbs.)
Displacement (cu. in.)
Car type

Group variable: rep78 (5 groups)

Obs. by group: 2, 8, 30, 18, 11
. notes
_dta:
1. summary statistics data built from auto_group_raw.dta on 30 Jun 2012
15:32:33 using -ssd build _all, group(rep78). ssd list
Group rep78==Poor:
(output omitted )
Group rep78==Fair:
(output omitted )
Group rep78==Average:
(output omitted )
Group rep78==Good:
(output omitted )
Group rep78==Excellent:
Observations = 11
Means:
price
mpg
5.913
27.363636
foreign
.81818182

gpm
.04048131

weight
2.3227273

displacement
111.09091

285

286

example 25 Creating summary statistics data from raw data

Variances implicitly defined; they are the diagonal of the covariance
matrix.
Covariances:
price
mpg
gpm
weight displacement
6.8422143
-15.608899
76.254545
.02750797
-.1184875
.00019114
.956802
-3.0610912
.00510833
.16856184
55.493298
-201.03636
.33150758
9.9577283
648.09091
.34169998
-.92727273
.00182175
.07254547
3.8181818
foreign
.16363636
.
.
.
.

* ------------------------------------------------------------------------* Digitally sign:

*
datasignature set
40:8(34334):2463066775:1212422464
(data signature set)

. * ------------------------------------------------------------------------. * Save:
. *
. save auto_group_ss
file auto_group_ss.dta saved
. * -------------------------------------------------------------------------

Also see
[SEM] ssd Making summary statistics data (sem only)

Title
example 26 Fitting a model with data missing at random

Description

Remarks and examples

Also see

Description
sem method(mlmv) is demonstrated using
. use http://www.stata-press.com/data/r13/cfa_missing
(CFA MAR data)
. summarize
Variable

Obs

Mean

id
test1
test2
test3
test4

500
406
413
443
417

250.5
97.37475
98.04501
100.9699
99.56815

Std. Dev.
144.4818
13.91442
13.84145
13.4862
14.25438

Min

Max

1
56.0406
62.25496
65.51753
53.8719

500
136.5672
129.3881
137.3046
153.9779

taken
500
3.358
.6593219
. notes
_dta:
1. Fictional data on 500 subjects taking four tests.
2. Tests results M.A.R. (missing at random).
3. 230 took all 4 tests
4. 219 took 3 of the 4 tests
5. 51 took 2 of the 4 tests
6. All tests have expected mean 100, s.d. 14.

See [SEM] intro 4 for background.

Remarks and examples

Remarks are presented under the following headings:
Fitting the model with method(ml)
Fitting the model with method(mlmv)
Fitting the model with the Builder

287

288

example 26 Fitting a model with data missing at random

Fitting the model with method(ml)

We fit a single-factor measurement model.

test1

test2

test3

test4

. sem (test1 test2 test3 test4 <- X), nolog

(270 observations with missing values excluded)
Endogenous variables
Measurement: test1 test2 test3 test4
Exogenous variables
Latent:
X
Structural equation model
Estimation method = ml
Log likelihood
= -3464.3099
( 1) [test1]X = 1

Coef.
Measurement
test1 <X
_cons

1
96.76907

OIM
Std. Err.

Number of obs

(constrained)
.8134878
118.96

230

P>|z|

[95% Conf. Interval]

0.000

95.17467

98.36348

test2 <X
_cons

1.021885
92.41248

.1183745
.8405189

8.63
109.95

0.000
0.000

.789875
90.7651

1.253895
94.05987

.5084673
94.12958

.0814191
.7039862

6.25
133.71

0.000
0.000

.3488889
92.7498

.6680457
95.50937

.5585651
92.2556

.0857772
.7322511

6.51
125.99

0.000
0.000

.3904449
90.82042

.7266853
93.69079

55.86083
61.88092
89.07839
93.26508
96.34453

10.85681
11.50377
8.962574
9.504276
16.28034

38.16563
42.985
73.13566
76.37945
69.18161

81.76028
89.08338
108.4965
113.8837
134.1725

test3 <X
_cons
test4 <X
_cons
var(e.test1)
var(e.test2)
var(e.test3)
var(e.test4)
var(X)

LR test of model vs. saturated: chi2(2)

0.39, Prob > chi2 = 0.8212

example 26 Fitting a model with data missing at random

289

Notes:
1. This model was fit using 230 of the 500 observations in the dataset. Unless you use sems
method(mlmv), observations are casewise omitted, meaning that if there is a single variable with
a missing value among the variables being used, the observation is ignored.
2. The coefficients for test3 and test4 are 0.51 and 0.56. Because we at StataCorp manufactured
these data, we can tell you that the true coefficients are 1.
3. The error variance for e.test1 and e.test2 are understated. These data were manufactured
with an error variance of 100.
4. These data are missing at random (MAR), not missing completely at random (MCAR). In MAR
data, which values are missing can be a function of the observed values in the data. MAR data
can produce biased estimates if the missingness is ignored, as we just did. MCAR data do not
bias estimates.

Fitting the model with method(mlmv)

. sem (test1 test2 test3 test4 <- X), method(mlmv) nolog
Endogenous variables
Measurement:

test1 test2 test3 test4

Exogenous variables
Latent:

(output omitted )
Structural equation model
Estimation method = mlmv
Log likelihood
= -6592.9961
( 1)

Number of obs

500

[test1]X = 1

Coef.
Measurement
test1 <X
_cons

1
98.94386

OIM
Std. Err.

(constrained)
.6814418
145.20

P>|z|

[95% Conf. Interval]

0.000

97.60826

100.2795

test2 <X
_cons

1.069952
99.84218

.1079173
.6911295

9.91
144.46

0.000
0.000

.8584378
98.48759

1.281466
101.1968

.9489025
101.0655

.0896098
.6256275

10.59
161.54

0.000
0.000

.7732706
99.83928

1.124534
102.2917

1.021626
99.64509

.0958982
.6730054

10.65
148.06

0.000
0.000

.8336687
98.32603

1.209583
100.9642

101.1135
95.45572
95.14847
101.0943
94.04629

10.1898
10.79485
9.053014
10.0969
13.96734

82.99057
76.47892
78.9611
83.12124
70.29508

123.1941
119.1413
114.6543
122.9536
125.8225

test3 <X
_cons
test4 <X
_cons
var(e.test1)
var(e.test2)
var(e.test3)
var(e.test4)
var(X)

LR test of model vs. saturated: chi2(2)

2.27, Prob > chi2 = 0.3209

290

example 26 Fitting a model with data missing at random

Notes:
1. The model is now fit using all 500 observations in the dataset.
2. The coefficients for test3 and test4previously 0.51 and 0.56are now 0.95 and 1.02.
3. Error variance estimates are now consistent with the true value of 100.
4. Standard errors of path coefficients are mostly smaller than reported in the previous model.
5. method(mlmv) requires that the data be MCAR or MAR.
6. method(mlmv) requires that the data be multivariate normal.

Fitting the model with the Builder

Use the diagram above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/cfa_missing

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Create the measurement component for X.
Select the Add Measurement Component tool, , and then click in the diagram about one-third
of the way down from the top and about halfway in from the left.
In the resulting dialog box,
a. change the Latent variable name to X;
b. select test1, test2, test3, and test4 by using the Measurement variables control;
c. select Down in the Measurement direction control;
d. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
4. Estimate.
Click on the Estimate button,

, in the Standard Toolbar. In the resulting dialog box,

a. select the Model tab;

b. select the Maximum likelihood with missing values radio button;
c. click on OK.
You can open a completed diagram in the Builder by typing
. webgetsem cfa_missing

Also see
[SEM] intro 4 Substantive concepts
[SEM] sem option method( ) Specifying method and calculation of VCE

Title
example 27g Single-factor measurement model (generalized response)
Description

Remarks and examples

Also see

Description
This is the first example in the g series. The g means that the example focuses exclusively on the
gsem command. If you are interested primarily in standard linear SEMs, you may want to skip the
remaining examples. If you are especially interested in generalized SEMs, we suggest you read the
remaining examples in order.
gsem provides two features not provided by sem: the ability to fit SEMs containing generalized
linear response variables and the ability to fit multilevel mixed SEMs. These two features can be used
separately or together.
Generalized response variables means that the response variables can be specifications from the
generalized linear model (GLM). These include probit, logistic regression, ordered probit and logistic
regression, multinomial logistic regression, and more.
Multilevel mixed models refer to the simultaneous handling of group-level effects, which can be
nested or crossed. Thus you can include unobserved and observed effects for subjects, subjects within
group, group within subgroup, . . . , or for subjects, group, subgroup, . . . .
Below we demonstrate a single-factor measure model with pass/fail (binary outcome) responses
rather than continuous responses. This is an example of a generalized linear response variable. We
use the following data:
. use http://www.stata-press.com/data/r13/gsem_1fmm
(single-factor pass/fail measurement model)
. summarize
Obs
Mean
Std. Dev.
Variable
x1
x2
x3
x4
s4

123
123
123
123
123

.4065041
.4065041
.4227642
.3495935
690.9837

.4931897
.4931897
.4960191
.4787919
77.50737

Min

Max

0
0
0
0
481

1
1
1
1
885

. notes
_dta:
1. Fictional data.
2. The variables x1, x2, x3, and x4 record 1=pass, 0=fail.
3. Pass/fail for x1, x2, x3: score > 100
4. Pass/fail for x4: score > 725
5. Variable s4 contains actual score for test 4.

See Single-factor measurement models in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Single-factor pass/fail measurement model
Single-factor pass/fail + continuous measurement model
Fitting the model with the Builder

291

292

example 27g Single-factor measurement model (generalized response)

Single-factor pass/fail measurement model

Below we fit the following model:

Bernoulli

Bernoulli

Bernoulli

Bernoulli

x1

x2

x3

x4

probit

probit

probit

probit

The measurement variables we have (x1, . . . , x4) are not continuous. They are pass/fail, coded as 1
(pass) and 0 (fail). To account for that, we use probit (also known as family Bernoulli, link probit).
The equations for this model are
Pr(x1 = 1) = (1 + X1 )
Pr(x2 = 1) = (2 + X2 )
Pr(x3 = 1) = (3 + X3 )
Pr(x4 = 1) = (4 + X4 )
where () is the N (0, 1) cumulative distribution.
One way to think about this is to imagine a test that is scored on a continuous scale. Lets imagine
the scores were s1 , s2 , s3 , and s4 and distributed N (i , i2 ), as test scores often are. Lets further
imagine that for each test, a cutoff ci is chosen and the student passes the test if si ci .
If we had the test scores, we would fit this as a linear model. We would posit

si = i + Xi + i
where i N (0, i2 ). However, we do not have test scores in our data; we have only the pass/fail
results xi = 1 if si > ci .

example 27g Single-factor measurement model (generalized response)

293

So lets consider the pass/fail problem. The probability that a student passes test i is determined
by the probability that the student scores above the cutoff:
Pr(si > ci ) = Pr(i + Xi + i > ci )

= Pr{ i > ci (i + Xi ) }
= Pr{ i ci + (i + Xi ) }
= Pr{ i (i ci ) + Xi }
= Pr{ i /i (i ci )/i + Xi /i }
= { (i ci )/i + Xi /i }
The last equation is the probit model. In fact, we just derived the probit model, and now we know
the relationship between the parameters we will be able to estimate with our pass/fail data: i and
i . We also now know the parameters we could have estimated if we had the continuous test scores:
i and i . The relationship is
i = (i ci )/i

i = i /i
Notice that the right-hand sides of both equations are divided by i , the standard deviation of the
error term from the linear model for the ith test score. In pass/fail data, we lose the original scale of
the score, and the slope coefficient we will be able to estimate is the slope coefficient from the linear
model divided by the standard deviation of the error term. Meanwhile, the intercept we will be able
to estimate is just the difference of the continuous models intercept and the cutoff for passing the
test, sans scale.
The command to fit the model and the results are
. gsem (x1 x2 x3 x4 <-X), probit
Fitting fixed-effects model:
Iteration 0:
Iteration 1:
Iteration 2:

log likelihood = -329.82091

log likelihood = -329.57665
log likelihood = -329.57664

Refining starting values:

Grid node 0:

log likelihood = -273.75437

Fitting full model:

Iteration
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration
Iteration

0:
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:

log
log
log
log
log
log
log
log
log
log
log
log
log
log
log
log
log
log
log

likelihood
likelihood
likelihood
likelihood
likelihood
likelihood
likelihood
likelihood
likelihood
likelihood
likelihood
likelihood
likelihood
likelihood
likelihood
likelihood
likelihood
likelihood
likelihood

=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=

-273.75437
-264.3035
-263.37815
-262.305
-261.69025
-261.42132
-261.35508
-261.3224
-261.3133
-261.30783
-261.30535
-261.30405
-261.30337
-261.30302
-261.30283
-261.30272
-261.30267
-261.30264
-261.30263

294

example 27g Single-factor measurement model (generalized response)

Generalized structural equation model
Log likelihood = -261.30263
( 1) [x1]X = 1
Coef.

Number of obs

Std. Err.

123

P>|z|

[95% Conf. Interval]

0.053

-.738437

.0050844

x1 <X
_cons

1
-.3666763

(constrained)
.1896773
-1.93

X
_cons

1.33293
-.4470271

.4686743
.2372344

2.84
-1.88

0.004
0.060

.4143455
-.911998

2.251515
.0179438

X
_cons

.6040478
-.2276709

.1908343
.1439342

3.17
-1.58

0.002
0.114

.2300195
-.5097767

.9780761
.0544349

X
_cons

9.453342
-4.801027

5.151819
2.518038

1.83
-1.91

0.067
0.057

-.6440375
-9.736291

19.55072
.1342372

var(X)

2.173451

1.044885

.847101

5.576536

x2 <-

x3 <-

x4 <-

Notes:
1. In the path diagrams, x1, . . . , x4 are shown as being family Bernoulli, link probit. On the
command line, we just typed probit although we could have typed family(bernoulli)
link(probit). In the command language, probit is a synonym for family(bernoulli)
link(probit).
2. Variable X is latent exogenous and thus needs a normalizing constraint. The variable is anchored
to the first observed variable, x1, and thus the path coefficient is constrained to be 1. See
Identification 2: Normalization constraints (anchoring) in [SEM] intro 4.
3. The path coefficients for X->x1, X->x2, and X->x3 are 1, 1.33, and 0.60. Meanwhile, the path
coefficient for X->x4 is 9.45. This is not unexpected; we at StataCorp generated these fictional
data, and we made the x4 effect large and less precisely estimable.

Single-factor pass/fail + continuous measurement model

In the above example, all equations were probit. Different equations within a single SEM can have
a different family and link.
Below we refit our model with the continuous measure for test 4 (variable s4) in place of the
pass/fail measure (variable x4). We continue to use the pass/fail measures for tests 1, 2, and 3.
. gsem (x1 x2 x3 <-X, probit)
Fitting fixed-effects model:
Iteration 0:
log likelihood
Iteration 1:
log likelihood
Iteration 2:
log likelihood
Refining starting values:
Grid node 0:
log likelihood
Fitting full model:
Iteration 0:
log likelihood
Iteration 1:
log likelihood

(s4<-X)
= -959.23492
= -959.09499
= -959.09499
= -905.14944
= -905.14944
= -872.33773

(not concave)

example 27g Single-factor measurement model (generalized response)

Iteration 2:
log likelihood =
Iteration 3:
log likelihood =
Iteration 4:
log likelihood =
Iteration 5:
log likelihood =
Generalized structural equation
Log likelihood = -869.6892
( 1) [x1]X = 1
Coef.

-869.83144
-869.69578
-869.68928
-869.6892
model

Std. Err.

Number of obs

P>|z|

295

123

[95% Conf. Interval]

x1 <X
_cons

1
-.4171085

(constrained)
.1964736
-2.12

X
_cons

1.298311
-.4926357

.3280144
.2387179

X
_cons

.682969
-.2942021

X
_cons

0.034

-.8021896

-.0320274

3.96
-2.06

0.000
0.039

.6554142
-.9605142

1.941207
-.0247573

.1747328
.1575014

3.91
-1.87

0.000
0.062

.3404989
-.6028992

1.025439
.0144949

55.24829
690.9837

12.19904
6.960106

4.53
99.28

0.000
0.000

31.3386
677.3422

79.15798
704.6253

var(X)

1.854506

.7804393

.812856

4.230998

var(e.s4)

297.8565

408.64

20.24012

4383.299

x2 <-

x3 <-

s4 <-

Notes:
1. We obtain similar coefficients for x1, . . . , x3.
2. We removed x4 (a pass/fail variable) and substituted s4 (the actual test score). s4 turns out to
be more significant than x4. This suggests a poor cutoff was set for passing test 4.
3. The log-likelihood values for the two models we have fit are strikingly different: 261 in the
previous model and 870 in the current model. The difference has no meaning. Log-likelihood
values are dependent on the model specified. We changed the fourth equation from a probit
specification to a continuous (linear-regression) specification, and just doing that changes the
metric of the log-likelihood function. Comparisons of log-likelihood values are only meaningful
when they share the same metric.

Fitting the model with the Builder

Use the diagram in Single-factor pass/fail measurement model above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_1fmm

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

296

example 27g Single-factor measurement model (generalized response)

4. Create the measurement component for X.

Select the Add Measurement Component tool, , and then click in the diagram about one-third
of the way down from the top and slightly left of the center.
In the resulting dialog box,
a. change the Latent variable name to X;
b. select x1, x2, x3, and x4 by using the Measurement variables control;
c. check Make measurements generalized;
d. select Bernoulli, Probit in the Family/Link control;
e. select Down in the Measurement direction control;
f. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
5. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
6. To fit the model in Single-factor pass/fail + continuous measurement model, modify the diagram
created in the previous steps.
a. Use the Select tool,

, to click on the x4 rectangle.

b. In the Contextual Toolbar, select s4 in the Variable control.

c. In the Contextual Toolbar, select Gaussian, Identity in the Family/Link control.
7. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_1fmm

Also see
[SEM] example 1 Single-factor measurement model
[SEM] example 28g One-parameter logistic IRT (Rasch) model
[SEM] example 29g Two-parameter logistic IRT model
[SEM] example 30g Two-level measurement model (multilevel, generalized response)
[SEM] example 31g Two-factor measurement model (generalized response)

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
example 28g One-parameter logistic IRT (Rasch) model

Description

Remarks and examples

References

Also see

Description
To demonstrate a one-parameter logistic IRT (Rasch) model, we use the following data:
. use http://www.stata-press.com/data/r13/gsem_cfa
(Fictional math abilities data)
. summarize
Variable

Obs

Mean

school
id
q1
q2
q3

500
500
500
500
500

q4
q5
q6
q7
q8

Std. Dev.

Min

Max

10.5
50681.71
.506
.394
.534

5.772056
29081.41
.5004647
.4891242
.4993423

1
71
0
0
0

20
100000
1
1
1

500
500
500
500
500

.424
.49
.434
.52
.494

.4946852
.5004006
.4961212
.5001002
.5004647

0
0
0
0
0

1
1
1
1
1

att1
att2
att3
att4
att5

500
500
500
500
500

2.946
2.948
2.84
2.91
3.086

1.607561
1.561465
1.640666
1.566783
1.581013

1
1
1
1
1

5
5
5
5
5

test1
test2
test3
test4
. notes

500
500
500
500

75.548
80.556
75.572
74.078

5.948653
4.976786
6.677874
8.845587

55
65
50
43

93
94
94
96

_dta:
1. Fictional data on math ability and attitudes of 500 students from 20
schools.
2. Variables q1-q8 are incorrect/correct (0/1) on individual math questions.
3. Variables att1-att5 are items from a Likert scale measuring each
students attitude toward math.
4. Variables test1-test4 are test scores from tests of four different
aspects of mathematical abilities. Range of scores: 0-100.

These data record results from a fictional instrument measuring mathematical ability. Variables q1
through q8 are the items from the instrument.
For discussions of Rasch models, IRT models, and their extensions, see Embretson and Reise (2000),
van der Linden and Hambleton (1997), Skrondal and Rabe-Hesketh (2004), Andrich (1988), Bond
and Fox (2007), and Fischer and Molenaar (1995). Although not demonstrated in this example, many
of the extensions discussed in these books can be fit with gsem as well.
See Itemresponse theory (IRT) models in [SEM] intro 5 for background.

297

298

example 28g One-parameter logistic IRT (Rasch) model

Remarks and examples

Remarks are presented under the following headings:
1-PL IRT model with unconstrained variance
1-PL IRT model with variance constrained to 1
Obtaining itemcharacteristic curves
Fitting the model with the Builder

1-PL IRT model with unconstrained variance

Mechanically speaking, one-parameter logistic (1-PL) IRT models are similar to the probit measurement model we demonstrated in [SEM] example 27g. The differences are that we will use logit
rather than probit and that we will place various constraints on the logit model to obtain results that
will allow us to judge the difficulty of the individual questions.
The model we wish to fit is

MathAb

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

q1

q2

q3

q4

q5

q6

q7

q8

logit

logit

logit

logit

logit

logit

logit

logit

In the 1-PL model, we place constraints that all coefficients, the factor loadings, are equal to 1. The
negative of the intercept for each question will then represent the difficulty of the question:

example 28g One-parameter logistic IRT (Rasch) model

. gsem (MathAb -> (q1-q8)@1), logit
Fitting fixed-effects model:
Iteration 0:
log likelihood = -2750.3114
Iteration 1:
log likelihood = -2749.3709
Iteration 2:
log likelihood = -2749.3708
Refining starting values:
Grid node 0:
log likelihood = -2653.2353
Fitting full model:
Iteration 0:
log likelihood = -2653.2353
Iteration 1:
log likelihood = -2651.2171
Iteration 2:
log likelihood = -2650.9117
Iteration 3:
log likelihood = -2650.9116
Generalized structural equation model
Log likelihood = -2650.9116
( 1) [q1]MathAb = 1
( 2) [q2]MathAb = 1
( 3) [q3]MathAb = 1
( 4) [q4]MathAb = 1
( 5) [q5]MathAb = 1
( 6) [q6]MathAb = 1
( 7) [q7]MathAb = 1
( 8) [q8]MathAb = 1
Coef.

Std. Err.

Number of obs

P>|z|

500

[95% Conf. Interval]

q1 <MathAb
_cons

1
.0293252

(constrained)
.1047674
0.28

0.780

-.1760152

.2346656

MathAb
_cons

1
-.5025012

(constrained)
.1068768
-4.70

0.000

-.7119759

-.2930264

MathAb
_cons

1
.1607425

(constrained)
.104967
1.53

0.126

-.044989

.3664739

MathAb
_cons

1
-.3574951

(constrained)
.105835
-3.38

0.001

-.564928

-.1500623

MathAb
_cons

1
-.0456599

(constrained)
.1047812
-0.44

0.663

-.2510274

.1597075

MathAb
_cons

1
-.3097521

(constrained)
.1055691
-2.93

0.003

-.5166637

-.1028404

MathAb
_cons

1
.09497

(constrained)
.1048315
0.91

0.365

-.1104959

.300436

MathAb
_cons

1
-.0269104

(constrained)
.1047691
-0.26

0.797

-.232254

.1784332

.6154407

1.02171

q2 <-

q3 <-

q4 <-

q5 <-

q6 <-

q7 <-

q8 <-

var(MathAb)

.7929701

.1025406

299

300

example 28g One-parameter logistic IRT (Rasch) model

Notes:
1. We had to use gsem and not sem to fit this model because the response variables were 0/1 and
not continuous and because we wanted to use logit and not a continuous model.
2. To place the constraints that all coefficients are equal to 1, in the diagram we placed 1s along
the path from the underlying latent factor MathAb to each of the questions. In the command
language, we added @1 to our command:
gsem (MathAb -> (q1-q8)@1), logit

Had we omitted the @1, we would have obtained coefficients about how well each question
measured math ability.
There are several ways we could have asked that the model above be fit. They include the
following:
gsem (MathAb -> q1@1 q2@1 q3@1 q4@1 q5@1 q6@1 q7@1 q8@1), logit
gsem (MathAb -> (q1 q2 q3 q4 q5 q6 q7 q8)@1), logit
gsem (MathAb -> (q1-q8)@1), logit

Similarly, for the shorthand logit, we could have typed family(bernoulli) link(logit).
3. The negative of the reported intercept represents the difficulty of the item. The most difficult is
q2, and the least difficult is q3.

1-PL IRT model with variance constrained to 1

The goal of the 1-PL model is in fact to constrain the loadings to be equal. In the previous model,
that was achieved by constraining them to be 1 and letting the variance of the latent variable float. An
alternative with perhaps easier-to-interpret results would constrain the variance of the latent variable
to be 1giving it a standard-normal interpretationand constrain the loadings to be merely equal:

example 28g One-parameter logistic IRT (Rasch) model

. gsem (MathAb -> (q1-q8)@b), logit var(MathAb@1)
Fitting fixed-effects model:
Iteration 0:
log likelihood = -2750.3114
Iteration 1:
log likelihood = -2749.3709
Iteration 2:
log likelihood = -2749.3708
Refining starting values:
Grid node 0:
log likelihood = -2645.8536
Fitting full model:
Iteration 0:
log likelihood = -2656.1973
Iteration 1:
log likelihood = -2650.9139
Iteration 2:
log likelihood = -2650.9116
Iteration 3:
log likelihood = -2650.9116
Generalized structural equation model
Number of obs
Log likelihood = -2650.9116
( 1) [q1]MathAb - [q8]MathAb = 0
( 2) [q2]MathAb - [q8]MathAb = 0
( 3) [q3]MathAb - [q8]MathAb = 0
( 4) [q4]MathAb - [q8]MathAb = 0
( 5) [q5]MathAb - [q8]MathAb = 0
( 6) [q6]MathAb - [q8]MathAb = 0
( 7) [q7]MathAb - [q8]MathAb = 0
( 8) [var(MathAb)]_cons = 1
Coef.

Std. Err.

P>|z|

500

[95% Conf. Interval]

q1 <MathAb
_cons

.8904887
.0293253

.0575755
.1047674

15.47
0.28

0.000
0.780

.7776429
-.1760151

1.003335
.2346657

MathAb
_cons

.8904887
-.5025011

.0575755
.1068768

15.47
-4.70

0.000
0.000

.7776429
-.7119758

1.003335
-.2930264

MathAb
_cons

.8904887
.1607425

.0575755
.104967

15.47
1.53

0.000
0.126

.7776429
-.044989

1.003335
.366474

MathAb
_cons

.8904887
-.3574951

.0575755
.105835

15.47
-3.38

0.000
0.001

.7776429
-.5649279

1.003335
-.1500622

MathAb
_cons

.8904887
-.0456599

.0575755
.1047812

15.47
-0.44

0.000
0.663

.7776429
-.2510273

1.003335
.1597076

MathAb
_cons

.8904887
-.309752

.0575755
.1055691

15.47
-2.93

0.000
0.003

.7776429
-.5166637

1.003335
-.1028403

MathAb
_cons

.8904887
.0949701

.0575755
.1048315

15.47
0.91

0.000
0.365

.7776429
-.1104959

1.003335
.300436

MathAb
_cons

.8904887
-.0269103

.0575755
.1047691

15.47
-0.26

0.000
0.797

.7776429
-.232254

1.003335
.1784333

q2 <-

q3 <-

q4 <-

q5 <-

q6 <-

q7 <-

q8 <-

var(MathAb)

(constrained)

301

302

example 28g One-parameter logistic IRT (Rasch) model

Notes:
1. The log-likelihood values of both models is 2650.9116. The models are equivalent.
2. Intercepts are unchanged.

Obtaining itemcharacteristic curves

Itemcharacteristic curves graph the conditional probability of a particular response given the latent
trait. In our case, this simply amounts to graphing the probability of a correct answer against math
ability. After estimation, we can obtain the predicted probabilities of a correct answer by typing
. predict pr*, pr
(using 7 quadrature points)

We can obtain the predicted value of the latent variable by typing

. predict ability, latent(MathAb)
(using 7 quadrature points)

and thus we can obtain the itemcharacteristic curves for all eight questions by typing

.2

.4

.6

.8

. twoway line pr1 pr2 pr3 pr4 pr5 pr6 pr7 pr8 ability, sort xlabel(-1.5(.5)1.5)

1.5

.5
0
.5
empirical Bayes means for MathAb
Predicted mean (q1 correct)
Predicted mean (q2 correct)
Predicted mean (q3 correct)
Predicted mean (q4 correct)
Predicted mean (q5 correct)
Predicted mean (q6 correct)
Predicted mean (q7 correct)
Predicted mean (q8 correct)

1.5

example 28g One-parameter logistic IRT (Rasch) model

303

A less busy graph might show merely the most difficult and least difficult questions:

.2

.4

.6

.8

. twoway line pr2 pr3 ability, sort xlabel(-1.5(.5)1.5)

1.5

.5
0
.5
empirical Bayes means for MathAb

1.5

Predicted mean (q2 correct)

Predicted mean (q3 correct)

The slopes of each curve are identical because we have constrained them to be identical. Thus we
just see the shift between difficulties with the lower items having higher levels of difficulty.

Fitting the model with the Builder

Use the diagram in 1-PL IRT model with unconstrained variance above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_cfa

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the measurement component for MathAb.

Select the Add Measurement Component tool, , and then click in the diagram about one-third
of the way down from the top and slightly left of the center.
In the resulting dialog box,
a. change the Latent variable name to MathAb;
b. select q1, q2, q3, q4, q5, q6, q7, and q8 by using the Measurement variables control;
c. check Make measurements generalized;
d. select Bernoulli, Logit in the Family/Link control;
e. select Down in the Measurement direction control;
f. click on OK.
If you wish, move the component by clicking on any variable and dragging it.

304

example 28g One-parameter logistic IRT (Rasch) model

5. Constrain all path coefficients to 1.

a. Choose the Select tool,

b. Click on the path from MathAb to q1. In the Contextual Toolbar, type 1 in the
press Enter.

box and

c. Repeat this process to add the 1 constraint on the paths from MathAb to each of the other
measurement variables.
6. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
7. To fit the model in 1-PL IRT model with variance constrained to 1, change the constraints in the
diagram created above.
a. From the SEM Builder menu, select Estimation > Clear Estimates to clear results from the
previous model.
b. Choose the Select tool,

c. Click on the path from MathAb to q1. In the Contextual Toolbar, type b in the
press Enter.

box and

d. Repeat this process to add the b constraint on the paths from MathAb to each of the other
measurement variables.
e. With
, click on the oval for MathAb. In the Contextual Toolbar, type 1 in the
and press Enter.

box

8. Estimate again.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder for the first model by typing
. webgetsem gsem_irt1

You can open a completed diagram in the Builder for the second model by typing
. webgetsem gsem_irt2

References
Andrich, D. 1988. Rasch Models for Measurement. Newbury Park, CA: Sage.
Bond, T. G., and C. M. Fox. 2007. Applying the Rasch Model: Fundamental Measurement in the Human Sciences.
2nd ed. Mahwah, NJ: Lawrence Erlbaum.
Embretson, S. E., and S. P. Reise. 2000. Item Response Theory for Psychologists. Mahwah, NJ: Lawrence Erlbaum.
Fischer, G. H., and I. W. Molenaar, ed. 1995. Rasch Models: Foundations, Recent Developments, and Applications.
New York: Springer.
Rasch, G. 1960. Probabilistic Models for Some Intelligence and Attainment Tests. Copenhagen: Danish Institute of
Educational Research.
. 1980. Probabilistic Models for Some Intelligence and Attainment Tests (Expanded ed.). Chicago: University of
Chicago Press.
Skrondal, A., and S. Rabe-Hesketh. 2004. Generalized Latent Variable Modeling: Multilevel, Longitudinal, and
Structural Equation Models. Boca Raton, FL: Chapman & Hall/CRC.

example 28g One-parameter logistic IRT (Rasch) model

305

van der Linden, W. J., and R. K. Hambleton, ed. 1997. Handbook of Modern Item Response Theory. New York:
Springer.

Also see
[SEM] example 27g Single-factor measurement model (generalized response)
[SEM] example 29g Two-parameter logistic IRT model

[SEM] gsem Generalized structural equation model estimation command

[SEM] predict after gsem Generalized linear predictions, etc.
[SEM] intro 5 Tour of models

Title
example 29g Two-parameter logistic IRT model

Description

Remarks and examples

References

Also see

Description
We demonstrate a two-parameter logistic (2-PL) IRT model with the same data used in [SEM] example 28g:
. use http://www.stata-press.com/data/r13/gsem_cfa
(Fictional math abilities data)
. summarize
Variable

Obs

Mean

Std. Dev.

Min

Max

school
id
q1
q2
q3

500
500
500
500
500

10.5
50681.71
.506
.394
.534

5.772056
29081.41
.5004647
.4891242
.4993423

1
71
0
0
0

20
100000
1
1
1

q4
q5
q6
q7
q8

500
500
500
500
500

.424
.49
.434
.52
.494

.4946852
.5004006
.4961212
.5001002
.5004647

0
0
0
0
0

1
1
1
1
1

att1
att2
att3
att4
att5

500
500
500
500
500

2.946
2.948
2.84
2.91
3.086

1.607561
1.561465
1.640666
1.566783
1.581013

1
1
1
1
1

5
5
5
5
5

test1
500
75.548
5.948653
55
93
test2
500
80.556
4.976786
65
94
500
75.572
6.677874
50
94
test3
test4
500
74.078
8.845587
43
96
. notes
_dta:
1. Fictional data on math ability and attitudes of 500 students from 20
schools.
2. Variables q1-q8 are incorrect/correct (0/1) on individual math questions.
3. Variables att1-att5 are items from a Likert scale measuring each
students attitude toward math.
4. Variables test1-test4 are test scores from tests of four different
aspects of mathematical abilities. Range of scores: 0-100.

These data record results from a fictional instrument measuring mathematical ability. Variables q1
through q8 are the items from the instrument.
For discussions of IRT models and their extensions, see Embretson and Reise (2000), van der
Linden and Hambleton (1997), Skrondal and Rabe-Hesketh (2004), and Rabe-Hesketh, Skrondal, and
Pickles (2004).
See Itemresponse theory (IRT) models in [SEM] intro 5 for background.

306

example 29g Two-parameter logistic IRT model

307

Remarks and examples

Remarks are presented under the following headings:
Fitting the 2-PL IRT model
Obtaining predicted difficulty and discrimination
Using coeflegend to obtain the symbolic names of the parameters
Graphing itemcharacteristic curves
Fitting the model with the Builder

Fitting the 2-PL IRT model

When we fit the 1-PL model, we commented that it was similar to the probit measure model we
demonstrated in [SEM] example 27g. The 1-PL model differed in that it used logit rather than probit,
and it placed constraints on the loadings to judge the difficulty of the individual questions.
The 2-PL model is even more similar to [SEM] example 27g. We still substitute logit for probit, but
we only constrain the variance (the latent variable) to be 1we leave the loadings unconstrainedand
we constrain the variance to be 1 merely to aid interpretation. Compared with the 1-PL example, this
time we will measure not just difficulty but discrimination as well.
The model we wish to fit is

MathAb
1

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

q1

q2

q3

q4

q5

q6

q7

q8

logit

logit

logit

logit

logit

logit

logit

logit

308

example 29g Two-parameter logistic IRT model

The results are

. gsem (MathAb -> q1-q8), logit
Fitting fixed-effects model:
Iteration 0:
log likelihood =
Iteration 1:
log likelihood =
Iteration 2:
log likelihood =
Refining starting values:
Grid node 0:
log likelihood =
Fitting full model:
Iteration 0:
log likelihood =
Iteration 1:
log likelihood =
Iteration 2:
log likelihood =
Iteration 3:
log likelihood =
Generalized structural equation
Log likelihood = -2637.3759
( 1) [var(MathAb)]_cons = 1
Coef.

var(MathAb@1)
-2750.3114
-2749.3709
-2749.3708
-2645.8536
-2645.8536
-2637.4315
-2637.3761
-2637.3759
model

Number of obs

500

Std. Err.

P>|z|

[95% Conf. Interval]

q1 <MathAb
_cons

1.466636
.0373363

.2488104
.1252274

5.89
0.30

0.000
0.766

.9789765
-.208105

1.954296
.2827776

MathAb
_cons

.5597118
-.4613391

.1377584
.0989722

4.06
-4.66

0.000
0.000

.2897102
-.6553211

.8297134
-.2673571

MathAb
_cons

.73241
.1533363

.1486818
.1006072

4.93
1.52

0.000
0.127

.440999
-.0438503

1.023821
.3505228

MathAb
_cons

.4839501
-.3230667

.1310028
.0957984

3.69
-3.37

0.000
0.001

.2271893
-.5108281

.7407109
-.1353054

MathAb
_cons

1.232244
-.0494684

.2075044
.1163093

5.94
-0.43

0.000
0.671

.8255426
-.2774304

1.638945
.1784937

MathAb
_cons

.946535
-.3147231

.1707729
.1083049

5.54
-2.91

0.000
0.004

.6118262
-.5269969

1.281244
-.1024493

MathAb
_cons

1.197317
.1053405

.2029485
.1152979

5.90
0.91

0.000
0.361

.7995449
-.1206393

1.595088
.3313203

MathAb
_cons

.8461858
-.026705

.1588325
.1034396

5.33
-0.26

0.000
0.796

.5348799
-.2294429

1.157492
.1760329

q2 <-

q3 <-

q4 <-

q5 <-

q6 <-

q7 <-

q8 <-

var(MathAb)

(constrained)

Notes:
1. In the above model, we constrain the variance MathAb to be 1 by typing var(MathAb@1).
2. Had we not constrained var(MathAb@1), the path coefficient from MathAb to q1 would
have automatically constrained to be 1 to set the latent variables scale. When we applied

example 29g Two-parameter logistic IRT model

309

var(MathAb@1), the automatic constraint was automatically released. Setting the variance of a
latent variable is another way of setting its scale.
3. We set var(MathAb@1) to ease interpretation. Our latent variable, MathAb, is now N (0, 1).
4. Factor loadings, which are the slopes, are estimated above for each question.
5. The slopes reveal how discriminating each question is in regard to mathematical ability. Question 1
is the most discriminating, and question 4 is the least discriminating.
6. In the 1-PL model, the negative of the intercept is a measure of difficulty if we constrain the
slopes to be equal to each other. To measure difficulty in the 2-PL model, we divide the negative
of the intercept by the unconstrained slope. If you do the math, you will discover that question 2
is the most difficult and question 3 is the least difficult. It will be easier, however, merely to
continue reading; in the next section, we show an easy way to calculate the discrimination and
difficulty for all the questions.

Obtaining predicted difficulty and discrimination

For each question, discrimination is defined as the questions slope coefficient.
For each question, difficulty is defined as the negative of the questions intercept divided by its
slope.
Here is how we quickly obtain all the discrimination and difficulty values in a single, easy-to-read
table:
. preserve
. drop _all
. set obs 8
obs was 0, now 8
. gen str question = "q" + strofreal(_n)
. gen diff = .
(8 missing values generated)
. gen disc
= .
(8 missing values generated)
. forvalues i = 1/8 {
2.
replace diff = -_b[qi:_cons] / _b[qi:MathAb] in i
3.
replace disc = _b[qi:MathAb] in i
4. }
(1 real change made)
(1 real change made)
(1 real change made)
(1 real change made)
(1 real change made)
(1 real change made)
(1 real change made)
(1 real change made)
(1 real change made)
(1 real change made)
(1 real change made)
(1 real change made)
(1 real change made)
(1 real change made)
(1 real change made)
(1 real change made)
. format diff disc %9.4f
. egen rank_diff = rank(diff)

310

example 29g Two-parameter logistic IRT model

. egen rank_disc = rank(disc)
. list
question

diff

disc

rank_d~f

rank_d~c

1.
2.
3.
4.
5.

q1
q2
q3
q4
q5

-0.0255
0.8242
-0.2094
0.6676
0.0401

1.4666
0.5597
0.7324
0.4840
1.2322

3
8
1
7
5

8
2
3
1
7

6.
7.
8.

q6
q7
q8

0.3325
-0.0880
0.0316

0.9465
1.1973
0.8462

6
2
4

5
6
4

. restore

Notes:
1. Our goal in the Stata code above is to create a dataset containing one observation for each
question. The dataset will contain the following variables: question containing q1, q2, . . . ;
diff and disc containing each questions difficulty and discrimination values; and rank disc
and rank diff containing the ranks of those discrimination and difficulty values.
2. We first preserved the current data before tossing out the data in memory. Later, after making
and displaying our table, we restored the original contents.
3. We then made an 8-observation, 0-variable dataset (set obs 8) and added variables to it. We
created string variable question containing q1, q2, . . . .
4. We were ready to create variables diff and disc. They are defined in terms of estimated
coefficients, and we had no idea what the names of those coefficients were. To find out, we
typed gsem, coeflegend (output shown below). We quickly learned that the slope coefficients
had names like b[q1:MathAb], b[q2:MathAb], . . . , and the intercepts had names like
b[q1: cons], b[q2: cons], . . . .
5. We created new variables diff and disc containing missing values and then created a forvalues
loop to fill in the new variables. Notice the odd-looking i inside the loop. i is the way
that you say substitute the value of (local macro) i here.
6. We put a display format on new variables diff and disc so that when we listed them, they
would be easier to read.
7. We created the rank of each variable by using the egen command.
8. We listed the results. So now you do not have to do the math to see that question 2 is the
most difficult (it has rank diff = 8) and question 3 is the least (it has rank diff = 1).
9. We typed restore, bringing our original data back into memory and leaving ourselves in a
position to continue with this example.

Using coeflegend to obtain the symbolic names of the parameters

In the section above, we did not retype coefficient values to obtain discrimination and difficulty.
After estimation, coefficient values are stored in b[name]. To find out what the names are, type
gsem, coeflegend. Here are the results:

example 29g Two-parameter logistic IRT model

. gsem, coeflegend
Generalized structural equation model
Log likelihood = -2637.3759
( 1) [var(MathAb)]_cons = 1
Coef.

Number of obs

311

500

Legend

q1 <MathAb
_cons

1.466636
.0373363

_b[q1:MathAb]
_b[q1:_cons]

MathAb
_cons

.5597118
-.4613391

_b[q2:MathAb]
_b[q2:_cons]

MathAb
_cons

.73241
.1533363

_b[q3:MathAb]
_b[q3:_cons]

MathAb
_cons

.4839501
-.3230667

_b[q4:MathAb]
_b[q4:_cons]

MathAb
_cons

1.232244
-.0494684

_b[q5:MathAb]
_b[q5:_cons]

MathAb
_cons

.946535
-.3147231

_b[q6:MathAb]
_b[q6:_cons]

MathAb
_cons

1.197317
.1053405

_b[q7:MathAb]
_b[q7:_cons]

MathAb
_cons

.8461858
-.026705

_b[q8:MathAb]
_b[q8:_cons]

q2 <-

q3 <-

q4 <-

q5 <-

q6 <-

q7 <-

q8 <-

var(MathAb)

_b[var(MathAb):_cons]

Graphing itemcharacteristic curves

We showed you the itemcharacteristic curves in [SEM] example 28g, so we will show them to
you again. Graphs of itemcharacteristic curves plot the probability of a correct answer against the
latent trait, which in this case is math ability.
We obtain the probabilities of a correct answer (the values of the latent variable) just as we did
previously,
. predict pr2pl*, pr
(using 7 quadrature points)
. predict ability2pl, latent(MathAb)
(using 7 quadrature points)

312

example 29g Two-parameter logistic IRT model

and we graph the curves just as we did previously, too. Here are all eight curves on one graph:

.2

.4

.6

.8

. twoway line pr2pl* ability2pl, sort xlabel(-1.5(.5)1.5)

1.5

.5
0
.5
empirical Bayes means for MathAb

1.5

Predicted mean (q1 correct)

Predicted mean (q2 correct)
Predicted mean (q3 correct)
Predicted mean (q4 correct)
Predicted mean (q5 correct)
Predicted mean (q6 correct)
Predicted mean (q7 correct)
Predicted mean (q8 correct)

In [SEM] example 28g, we showed a graph for the most and least difficult questions. This time
we show a graph for the most and least discriminating questions:

.2

.4

.6

.8

. twoway line pr2pl1 pr2pl4 ability2pl, sort xlabel(-1.5(.5)1.5)

1.5

.5
0
.5
empirical Bayes means for MathAb

1.5

Predicted mean (q1 correct)

Predicted mean (q4 correct)

Here the curves are not parallel because the discrimination has not been constrained to be equal across
the questions. Question 1 has a steeper slope, so it is more discriminating.

Fitting the model with the Builder

Use the diagram in Fitting the 2-PL IRT model above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_cfa

example 29g Two-parameter logistic IRT model

313

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the measurement component for MathAb.

Select the Add Measurement Component tool, , and then click in the diagram about one-third
of the way down from the top and slightly left of the center.
In the resulting dialog box,
a. change the Latent variable name to MathAb;
b. select q1, q2, q3, q4, q5, q6, q7, and q8 by using the Measurement variables control;
c. check Make measurements generalized;
d. select Bernoulli, Logit in the Family/Link control;
e. select Down in the Measurement direction control;
f. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
5. Constrain the variance of MathAb to 1.
a. Choose the Select tool,

b. Click on the oval for MathAb. In the Contextual Toolbar, type 1 in the
Enter.

box and press

6. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_irt3

References
Embretson, S. E., and S. P. Reise. 2000. Item Response Theory for Psychologists. Mahwah, NJ: Lawrence Erlbaum.
Rabe-Hesketh, S., A. Skrondal, and A. Pickles. 2004. Generalized multilevel structural equation modeling. Psychometrika
69: 167190.
Skrondal, A., and S. Rabe-Hesketh. 2004. Generalized Latent Variable Modeling: Multilevel, Longitudinal, and
Structural Equation Models. Boca Raton, FL: Chapman & Hall/CRC.
van der Linden, W. J., and R. K. Hambleton, ed. 1997. Handbook of Modern Item Response Theory. New York:
Springer.

Also see
[SEM] example 27g Single-factor measurement model (generalized response)
[SEM] example 28g One-parameter logistic IRT (Rasch) model

[SEM] gsem Generalized structural equation model estimation command

[SEM] predict after gsem Generalized linear predictions, etc.
[SEM] intro 5 Tour of models

Title
example 30g Two-level measurement model (multilevel, generalized response)

Description

Remarks and examples

References

Also see

Description
We demonstrate a multilevel measurement model with the same data used in [SEM] example 29g:
. use http://www.stata-press.com/data/r13/gsem_cfa
(Fictional math abilities data)
. summarize
Variable

Obs

Mean

school
id
q1
q2
q3

500
500
500
500
500

q4
q5
q6
q7
q8

Std. Dev.

Min

Max

10.5
50681.71
.506
.394
.534

5.772056
29081.41
.5004647
.4891242
.4993423

1
71
0
0
0

20
100000
1
1
1

500
500
500
500
500

.424
.49
.434
.52
.494

.4946852
.5004006
.4961212
.5001002
.5004647

0
0
0
0
0

1
1
1
1
1

att1
att2
att3
att4
att5

500
500
500
500
500

2.946
2.948
2.84
2.91
3.086

1.607561
1.561465
1.640666
1.566783
1.581013

1
1
1
1
1

5
5
5
5
5

test1
test2
test3
test4
. notes

500
500
500
500

75.548
80.556
75.572
74.078

5.948653
4.976786
6.677874
8.845587

55
65
50
43

93
94
94
96

_dta:
1. Fictional data on math ability and attitudes of 500 students from 20
schools.
2. Variables q1-q8 are incorrect/correct (0/1) on individual math questions.
3. Variables att1-att5 are items from a Likert scale measuring each
students attitude toward math.
4. Variables test1-test4 are test scores from tests of four different
aspects of mathematical abilities. Range of scores: 0-100.

These data record results from a fictional instrument measuring mathematical ability. Variables q1
through q8 are the items from the instrument.
For discussions of multilevel measurement models, including extensions beyond the example we
present here, see Mehta and Neale (2005) and Skrondal and Rabe-Hesketh (2004).
See Single-factor measurement models and Multilevel mixed-effects models in [SEM] intro 5 for
background.

314

example 30g Two-level measurement model (multilevel, generalized response)

315

Remarks and examples

Remarks are presented under the following headings:
Fitting the two-level model
Fitting the variance-components model
Fitting the model with the Builder

Fitting the two-level model

We extend the measurement model fit in [SEM] example 29g to better account for our (fictional)
data. In the data, students are nested within school, but we have ignored that so far. In this example,
we include a latent variable at the school level to account for possible school-by-school effects.
The model we wish to fit is

MathAb

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

q1

q2

q3

q4

q5

q6

q7

q8

logit

logit

logit

logit

logit

logit

logit

logit

school1

The double-ringed school1 is new. That new component of the path diagram is saying, I am a latent
variable at the school levelmeaning I am constant within school and vary across schoolsand
I correspond to a latent variable named M1; see Specifying generalized SEMs: Multilevel mixed
effects (2 levels) in [SEM] intro 2. This new variable will account for the effect, if any, of the identity
of the school.
To fit this model without this new, school-level component in it, we would type
. gsem (MathAb -> q1-q8), logit

316

example 30g Two-level measurement model (multilevel, generalized response)

To include the new school-level component, we add M1[school] to the exogenous variables:
. gsem (MathAb M1[school] -> q1-q8), logit
Fitting fixed-effects model:
Iteration 0:
log likelihood = -2750.3114
Iteration 1:
log likelihood = -2749.3709
Iteration 2:
log likelihood = -2749.3708
Refining starting values:
Grid node 0:
log likelihood = -2649.0033
Fitting full model:
Iteration 0:
log likelihood = -2649.0033
Iteration 1:
log likelihood = -2645.0613
Iteration 2:
log likelihood = -2641.9755
Iteration 3:
log likelihood = -2634.3857
Iteration 4:
log likelihood = -2631.1111
Iteration 5:
log likelihood = -2630.7898
Iteration 6:
log likelihood = -2630.2477
Iteration 7:
log likelihood = -2630.2402
Iteration 8:
log likelihood = -2630.2074
Iteration 9:
log likelihood = -2630.2063
Iteration 10: log likelihood = -2630.2063
Generalized structural equation model
Log likelihood = -2630.2063
( 1) [q1]M1[school] = 1
( 2) [q2]MathAb = 1
Coef.

Std. Err.

(not concave)
(not concave)
(not concave)

Number of obs

500

P>|z|

[95% Conf. Interval]

q1 <M1[school]

MathAb
_cons

2.807515
.0388021

.9468682
.1608489

2.97
0.24

0.003
0.809

.9516878
-.276456

4.663343
.3540602

q2 <M1[school]

.6673925

.3058328

2.18

0.029

.0679712

1.266814

MathAb
_cons

1
-.4631159

(constrained)
.1201227
-3.86

0.000

-.698552

-.2276798

q3 <M1[school]

.3555867

.3043548

1.17

0.243

-.2409377

.9521111

MathAb
_cons

1.455529
.1537831

.5187786
.1070288

2.81
1.44

0.005
0.151

.4387416
-.0559894

2.472316
.3635556

q4 <M1[school]

.7073241

.3419273

2.07

0.039

.037159

1.377489

MathAb
_cons

.8420897
-.3252735

.3528195
.1202088

2.39
-2.71

0.017
0.007

.1505762
-.5608784

1.533603
-.0896686

q5 <M1[school]

.7295553

.3330652

2.19

0.028

.0767595

1.382351

MathAb
_cons

2.399529
-.0488674

.8110973
.1378015

2.96
-0.35

0.003
0.723

.8098079
-.3189533

3.989251
.2212185

(constrained)

example 30g Two-level measurement model (multilevel, generalized response)

q6 <M1[school]

.484903

.2844447

1.70

0.088

-.0725983

1.042404

MathAb
_cons

1.840627
-.3139302

.5934017
.1186624

3.10
-2.65

0.002
0.008

.6775813
-.5465042

3.003673
-.0813563

q7 <M1[school]

.3677241

.2735779

1.34

0.179

-.1684787

.903927

MathAb
_cons

2.444023
.1062164

.8016872
.1220796

3.05
0.87

0.002
0.384

.8727449
-.1330552

4.015301
.3454881

q8 <M1[school]

.5851299

.3449508

1.70

0.090

-.0909612

1.261221

MathAb
_cons

1.606287
-.0261962

.5367614
.1189835

2.99
-0.22

0.003
0.826

.5542541
-.2593995

2.65832
.2070071

.2121216
.2461246

.1510032
.1372513

.052558
.0825055

.8561121
.7342217

var(
M1[school])
var(MathAb)

317

Notes:
1. The variance of M1[school] is estimated to be 0.21.
2. So how important is M1[school]? The variance of MathAb is estimated to be 0.25, so math
ability and school have roughly the same variance, and both of course have mean 0. The math
ability coefficients, meanwhile, are largeroften much largerthan the school coefficients in
every case, so math ability is certainly more important than school in explaining whether questions
were answered correctly. At this point, we are merely exploring the magnitude of effect.
3. You could also include a school-level latent variable for each question. For instance, you could
type
. gsem (MathAb
(MathAb
(MathAb
(MathAb
(MathAb
(MathAb
(MathAb
(MathAb

M1[school]
M1[school]
M1[school]
M1[school]
M1[school]
M1[school]
M1[school]
M1[school]

N1[school]
N2[school]
N3[school]
N4[school]
N5[school]
N6[school]
N7[school]
N8[school]

->
->
->
->
->
->
->
->

q1) ///
q2) ///
q3) ///
q4) ///
q5) ///
q6) ///
q7) ///
q8), logit

You will sometimes see such effects included in multilevel measurement models in theoretical discussions of models. Be aware that estimation of models with many latent variables is
problematic, requiring both time and luck.

Fitting the variance-components model

In a variance-components model, school would affect math ability which would affect correctness
of answers to questions. The model might be drawn like this:

318

example 30g Two-level measurement model (multilevel, generalized response)

school1

e1

MathAb

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

q1

q2

q3

q4

q5

q6

q7

q8

logit

logit

logit

logit

logit

logit

logit

logit

The above is a great way to draw the model. Sadly, gsem cannot understand it. The problem from
gsems perspective is that one latent variable is affecting another and the two latent variables are at
different levels.
So we have to draw the model differently:

MathAb

c2

c4

c3

c5

c6

c8

c7

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

q1

q2

q3

q4

q5

q6

q7

q8

logit

logit

logit

logit

logit

logit

logit

logit

c2

c3

c4

c5

c6

c7

c8

school1

The models may look different, but constraining the coefficients along the paths from math ability
and from school to each question is identical in effect to the model above.

example 30g Two-level measurement model (multilevel, generalized response)

The result of fitting the model is

. gsem (MathAb M1[school] ->
>
q1@1 q2@c2 q3@c3 q4@c4 q5@c5 q6@c6 q7@c7 q8@c8), logit
Fitting fixed-effects model:
Iteration 0:
log likelihood = -2750.3114
Iteration 1:
log likelihood = -2749.3709
Iteration 2:
log likelihood = -2749.3708
Refining starting values:
Grid node 0:
log likelihood = -2642.8248
Fitting full model:
Iteration 0:
log likelihood = -2651.7239 (not concave)
Iteration 1:
log likelihood = -2644.4937
Iteration 2:
log likelihood =
-2634.92
Iteration 3:
log likelihood = -2633.9336
Iteration 4:
log likelihood = -2633.5924
Iteration 5:
log likelihood = -2633.5922
Generalized structural equation model
Number of obs
=
Log likelihood = -2633.5922
( 1) [q1]M1[school] = 1
( 2) [q1]MathAb = 1
( 3) [q2]M1[school] - [q2]MathAb = 0
( 4) [q3]M1[school] - [q3]MathAb = 0
( 5) [q4]M1[school] - [q4]MathAb = 0
( 6) [q5]M1[school] - [q5]MathAb = 0
( 7) [q6]M1[school] - [q6]MathAb = 0
( 8) [q7]M1[school] - [q7]MathAb = 0
( 9) [q8]M1[school] - [q8]MathAb = 0
Coef.

Std. Err.

q1 <M1[school]

MathAb
_cons

1
.0385522

q2 <M1[school]

.3876281

.1156823

MathAb
_cons

.3876281
-.4633143

q3 <M1[school]

P>|z|

500

[95% Conf. Interval]

(constrained)
(constrained)
.1556214
0.25

0.804

-.2664601

.3435646

3.35

0.001

.1608951

.6143612

.1156823
.1055062

3.35
-4.39

0.001
0.000

.1608951
-.6701028

.6143612
-.2565259

.4871164

.1295515

3.76

0.000

.2332001

.7410328

MathAb
_cons

.4871164
.1533212

.1295515
.1098068

3.76
1.40

0.000
0.163

.2332001
-.0618962

.7410328
.3685386

q4 <M1[school]

.3407151

.1058542

3.22

0.001

.1332446

.5481856

MathAb
_cons

.3407151
-.3246936

.1058542
.1011841

3.22
-3.21

0.001
0.001

.1332446
-.5230108

.5481856
-.1263763

q5 <M1[school]

.8327426

.1950955

4.27

0.000

.4503624

1.215123

MathAb
_cons

.8327426
-.0490579

.1950955
.1391324

4.27
-0.35

0.000
0.724

.4503624
-.3217524

1.215123
.2236365

319

320

example 30g Two-level measurement model (multilevel, generalized response)

q6 <M1[school]

.6267415

.1572247

3.99

0.000

.3185868

.9348962

MathAb
_cons

.6267415
-.3135398

.1572247
.1220389

3.99
-2.57

0.000
0.010

.3185868
-.5527317

.9348962
-.074348

q7 <M1[school]

.7660343

.187918

4.08

0.000

.3977219

1.134347

MathAb
_cons

.7660343
.1039102

.187918
.1330652

4.08
0.78

0.000
0.435

.3977219
-.1568927

1.134347
.3647131

q8 <M1[school]

.5600833

.1416542

3.95

0.000

.2824462

.8377203

MathAb
_cons

.5600833
-.0264193

.1416542
.1150408

3.95
-0.23

0.000
0.818

.2824462
-.2518951

.8377203
.1990565

.1719347
2.062489

.1150138
.6900045

.0463406
1.070589

.6379187
3.973385

var(
M1[school])
var(MathAb)

1. Note that for each question, the coefficient on MathAb is identical to the coefficient on M1[school].
2. We estimate separate variances for M1[school] and MathAb. They are 0.17 and 2.06. Now that
the coefficients are the same on school and ability, we can directly compare these variances. We
see that math ability has a much larger affect than does school.

Fitting the model with the Builder

Use the diagram in Fitting the two-level model above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_cfa

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the measurement component for MathAb.

Select the Add Measurement Component tool, , and then click in the diagram about one-fourth
of the way down from the top and slightly left of the center.
In the resulting dialog box,
a. change the Latent variable name to MathAb;
b. select q1, q2, q3, q4, q5, q6, q7, and q8 by using the Measurement variables control;
c. check Make measurements generalized;
d. select Bernoulli, Logit in the Family/Link control;

example 30g Two-level measurement model (multilevel, generalized response)

321

e. select Down in the Measurement direction control;

f. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
5. Create the school-level latent variable.
, and click about one-fourth of the way
a. Select the Add Multilevel Latent Variable tool,
up from the bottom and slightly left of the center.
b. In the Contextual Toolbar, click on the

button.

c. Select the nesting level and nesting variable by selecting 2 from the Nesting depth control
and selecting school > Observations in the next control.
d. Specify M1 as the Base name.
e. Click on OK.
6. Create the factor-loading paths for the multilevel latent variable.
.

a. Select the Add Path tool,

b. Click in the top-left quadrant of the double oval for school1 (it will highlight when you
hover over it), and drag a path to the bottom of the q1 rectangle (it will highlight when you
can release to connect the path).
c. Continuing with the

tool, draw paths from school1 to each of the remaining rectangles.

7. Clean up paths.
, to
If you do not like where a path has been connected to its variables, use the Select tool,
click on the path, and then simply click on where it connects to a rectangle or oval and drag the
endpoint.
8. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
9. To fit the model in Fitting the variance-components model, add constraints to the diagram created
above.
a. From the SEM Builder menu, select Estimation > Clear Estimates to clear results from the
previous model.
b. Choose the Select tool,

c. Click on the path from MathAb to q1. In the Contextual Toolbar, type 1 in the
press Enter.

box and

d. Click on the path from school1 to q1. In the Contextual Toolbar, type 1 in the
and press Enter.

box

c. Click on the path from MathAb to q2. In the Contextual Toolbar, type c2 in the
and press Enter.

box

d. Click on the path from school1 to q2. In the Contextual Toolbar, type c2 in the
and press Enter.

box

e. Repeat this process to add the c3 constraint on both paths to q3, the c4 constraint on both
paths to q4, . . . , and the c8 constraint on both paths to q8.

322

example 30g Two-level measurement model (multilevel, generalized response)

10. Estimate again.

Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder for the first model by typing
. webgetsem gsem_mlcfa1

You can open a completed diagram in the Builder for the second model by typing
. webgetsem gsem_mlcfa2

References
Mehta, P. D., and M. C. Neale. 2005. People are variables too: Multilevel structural equations modeling. Psychological
Methods 10: 259284.
Skrondal, A., and S. Rabe-Hesketh. 2004. Generalized Latent Variable Modeling: Multilevel, Longitudinal, and
Structural Equation Models. Boca Raton, FL: Chapman & Hall/CRC.

Also see
[SEM] example 27g Single-factor measurement model (generalized response)
[SEM] example 29g Two-parameter logistic IRT model

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
example 31g Two-factor measurement model (generalized response)

Description

Remarks and examples

Also see

Description
We demonstrate a two-factor generalized linear measurement model with the same data used in
[SEM] example 29g:
. use http://www.stata-press.com/data/r13/gsem_cfa
(Fictional math abilities data)
. describe
Contains data from http://www.stata-press.com/data/r13/gsem_cfa.dta
obs:
500
Fictional math abilities data
vars:
19
21 Mar 2013 10:38
size:
18,500
(_dta has notes)

variable name

storage
type

display
format

value
label

school
id
q1
q2
q3
q4
q5
q6
q7
q8
att1

byte
long
byte
byte
byte
byte
byte
byte
byte
byte
float

%9.0g
%9.0g
%9.0g
%9.0g
%9.0g
%9.0g
%9.0g
%9.0g
%9.0g
%9.0g
%26.0g

result
result
result
result
result
result
result
result
agree

att2

float

%26.0g

agree

att3

float

%26.0g

agree

att4

float

%26.0g

agree

att5

float

%26.0g

agree

test1
test2
test3
test4

byte
byte
byte
byte

%9.0g
%9.0g
%9.0g
%9.0g

variable label
School id
Student id
q1 correct
q2 correct
q3 correct
q4 correct
q5 correct
q6 correct
q7 correct
q8 correct
Skills taught in math class will
help me get a better job.
Math is important in everyday
life
Working math problems makes me
anxious.
Math has always been my worst
subject.
I am able to learn new math
concepts easily.
Score, math test 1
Score, math test 2
Score, math test 3
Score, math test 4

Sorted by:
. notes
_dta:
1. Fictional data on math ability and attitudes of 500 students from 20
schools.
2. Variables q1-q8 are incorrect/correct (0/1) on individual math questions.
3. Variables att1-att5 are items from a Likert scale measuring each
students attitude toward math.
4. Variables test1-test4 are test scores from tests of four different
aspects of mathematical abilities. Range of scores: 0-100.

323

324

example 31g Two-factor measurement model (generalized response)

These data record results from a fictional instrument measuring mathematical ability. Variables q1
through q8 are the items from the instrument.
In this example, we will also be using variables att1 through att5. These are five Likert-scale
questions measuring each students attitude toward math.
See Multiple-factor measurement models in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Fitting the two-factor model
Fitting the model with the Builder

Fitting the two-factor model

We extend the measurement model fit in [SEM] example 29g from one factor, math ability, to two
factors, math ability and attitude. The model we wish to fit is

ordinal

ordinal

ordinal

ordinal

ordinal

att1

att2

att3

att4

att5

logit

logit

logit

logit

logit

MathAtt

MathAb

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

q1

q2

q3

q4

q5

q6

q7

q8

logit

logit

logit

logit

logit

logit

logit

logit

In this model, mathematical ability affects the correctness of the answers to the items just as
previously. The new component, attitude toward mathematics, is correlated with math ability. We
expect this correlation to be positive, but that is yet to be determined.
What is important about the attitudinal questions is that the responses are ordinal, that is, the
ordering of the possible answers is significant. In other cases, we might have a categorical variable
taking on, say, five values; even if the values are 1, 2, 3, 4, and 5, there is no case in which answer
5 is greater than answer 4, answer 4 is greater than answer 3, and so on.

example 31g Two-factor measurement model (generalized response)

325

For our attitude measures, however, response 5 signifies strong agreement with a statement and
1 signifies strong disagreement. We handle the ordinal property by specifying that the attitudinal
responses are family ordinal, link logit, also known as ordered logit or ordinal logistic regression,
and also known in Stata circles as ologit.
In the command language, to fit a one-factor measurement model with math ability, we would type
gsem (MathAb -> q1-q8), logit

To include the second factor, attitude correlated with math ability, we would type
gsem (MathAb -> q1-q8,
logit)
(MathAtt -> att1-att5, ologit)

///

The covariance between MathAtt and MathAb does not even appear in the command! That is because
latent exogenous variables are assumed to be correlated in the command language unless you specify
otherwise; in path diagrams, such variables are correlated only if a curved path is drawn between
them.
There is another, minor difference in syntax between the one-factor and two-factor models that
is worth your attention. Notice that the logit was outside the parentheses in the command to fit
the one-factor model, but it is inside the parentheses in the command to fit the two-factor model.
Actually, logit could have appeared inside the parentheses to fit the one-factor model. When options
appear inside parentheses, they affect only what is specified inside the parentheses. When they appear
outside parentheses, they affect all parenthetical specifications.
To obtain the estimates of the two-factor model, we type
. gsem (MathAb -> q1-q8, logit)
>
(MathAtt -> att1-att5, ologit)
Fitting fixed-effects model:
Iteration 0:
log likelihood = -6629.7253
Iteration 1:
log likelihood = -6628.7848
Iteration 2:
log likelihood = -6628.7848
Refining starting values:
Grid node 0:
log likelihood = -6457.4584
Fitting full model:
Iteration 0:
log likelihood = -6457.4584
Iteration 1:
log likelihood = -6437.9594
Iteration 2:
log likelihood = -6400.2731
Iteration 3:
log likelihood = -6396.3795
Iteration 4:
log likelihood = -6394.5787
Iteration 5:
log likelihood = -6394.4019
Iteration 6:
log likelihood = -6394.3923
Iteration 7:
log likelihood = -6394.3923
Generalized structural equation model
Log likelihood = -6394.3923
( 1) [q1]MathAb = 1
( 2) [att1]MathAtt = 1
Coef.

Std. Err.

Number of obs

P>|z|

500

[95% Conf. Interval]

q1 <MathAb
_cons

1
.0446118

MathAb
_cons

.3446081
-.4572215

(constrained)
.1272964
0.35

0.726

-.2048845

.2941082

0.001
0.000

.1387601
-.6492911

.5504562
-.265152

q2 <.1050264
.0979965

3.28
-4.67

326

example 31g Two-factor measurement model (generalized response)

q3 <MathAb
_cons

.5445245
.1591406

.1386993
.1033116

3.93
1.54

0.000
0.123

.272679
-.0433464

.8163701
.3616276

MathAb
_cons

.2858874
-.3196648

.0948553
.0947684

3.01
-3.37

0.003
0.001

.0999743
-.5054075

.4718004
-.1339222

MathAb
_cons

.8174803
-.04543

.1867024
.116575

4.38
-0.39

0.000
0.697

.4515504
-.2739127

1.18341
.1830527

MathAb
_cons

.6030448
-.309992

.1471951
.1070853

4.10
-2.89

0.000
0.004

.3145478
-.5198753

.8915419
-.1001086

MathAb
_cons

.72084
.1047265

.1713095
.1116494

4.21
0.94

0.000
0.348

.3850796
-.1141023

1.056601
.3235552

MathAb
_cons

.5814761
-.0250442

.1426727
.1045134

4.08
-0.24

0.000
0.811

.3018428
-.2298868

.8611094
.1797983

att1 <MathAtt

att2 <MathAtt

.3788714

.0971223

3.90

0.000

.1885152

.5692276

att3 <MathAtt

-1.592717

.3614859

-4.41

0.000

-2.301216

-.8842173

att4 <MathAtt

-.8100107

.153064

-5.29

0.000

-1.11001

-.5100108

att5 <MathAtt

.5225423

.1170141

4.47

0.000

.2931988

.7518858

/cut1
/cut2
/cut3
/cut4

-1.10254
-.2495339
.2983261
1.333053

.1312272
.1160385
.1164414
.1391907

-8.40
-2.15
2.56
9.58

0.000
0.032
0.010
0.000

-1.359741
-.4769651
.0701052
1.060244

-.8453396
-.0221027
.5265471
1.605861

/cut1
/cut2
/cut3
/cut4

-1.055791
-.1941211
.3598488
1.132624

.1062977
.0941435
.0952038
.1082204

-9.93
-2.06
3.78
10.47

0.000
0.039
0.000
0.000

-1.264131
-.378639
.1732528
.9205156

-.8474513
-.0096032
.5464448
1.344732

/cut1
/cut2
/cut3
/cut4

-1.053519
-.0491073
.5570671
1.666859

.1733999
.1442846
.1538702
.2135554

-6.08
-0.34
3.62
7.81

0.000
0.734
0.000
0.000

-1.393377
-.3318999
.2554871
1.248298

-.7136614
.2336853
.8586471
2.08542

q4 <-

q5 <-

q6 <-

q7 <-

q8 <-

(constrained)

att1

att2

att3

example 31g Two-factor measurement model (generalized response)

327

att4
/cut1
/cut2
/cut3
/cut4

-1.07378
-.2112462
.406347
1.398185

.1214071
.1076501
.1094847
.1313327

-8.84
-1.96
3.71
10.65

0.000
0.050
0.000
0.000

-1.311734
-.4222366
.191761
1.140778

-.8358264
-.0002559
.620933
1.655593

/cut1
/cut2
/cut3
/cut4

-1.244051
-.336135
.2137776
.9286849

.1148443
.0986678
.0978943
.107172

-10.83
-3.41
2.18
8.67

0.000
0.001
0.029
0.000

-1.469142
-.5295203
.0219084
.7186316

-1.018961
-.1427498
.4056468
1.138738

var(MathAb)
var(MathAtt)

2.300652
1.520854

.7479513
.4077674

1.216527
.8992196

4.350909
2.572228

cov(MathAtt,
MathAb)

.8837681

.2204606

.4516733

1.315863

att5

4.01

0.000

Notes:
1. The estimated covariance between math attitude and ability is 0.88.
2. We can calculate the correlation from the estimated covariance; the formula is bxy =
bxy /(b
x
by ).
The estimated values are
bxy = 0.8838,
bx2 = 2.301, and
by2 = 1.521. Thus the estimated
correlation between attitude and ability is 0.4724.
3. There is something new in the output, namely, things labeled /cut1, . . . , /cut4. These appear
for each of the five attitudinal measures. These are the ordered logits cutpoints, the values on
the logits distribution that separate attitude 1 from attitude 2, attitude 2 from attitude 3, and so
on. The four cutpoints map the continuous distribution into five ordered, categorical groups.
4. Theres something interesting hiding in the MathAtt coefficients: the coefficients for two of
the paths, att3 att4 <- MathAtt, are negative! If you look back to the description of the
data, you will find that the sense of these two questions was reversed from those of the other
questions. Strong agreement on these two questions was agreement with a negative feeling about
mathematics.

Fitting the model with the Builder

Use the diagram in Fitting the two-factor model above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_cfa

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the measurement component for MathAb.

Select the Add Measurement Component tool, , and then click in the diagram about one-third
of the way up from the bottom and slightly left of the center.

328

example 31g Two-factor measurement model (generalized response)

In the resulting dialog box,

a. change the Latent variable name to MathAb;
b. select q1, q2, q3, q4, q5, q6, q7, and q8 by using the Measurement variables control;
c. check Make measurements generalized;
d. select Bernoulli, Logit in the Family/Link control;
e. select Down in the Measurement direction control;
f. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
5. Create the measurement component for MathAtt.
Select the Add Measurement Component tool, , and then click in the diagram about one-third
of the way down from the top and slightly left of the center.
In the resulting dialog box,
a. change the Latent variable name to MathAtt;
b. select att1, att2, att3, att4, and att5 by using the Measurement variables control;
c. check Make measurements generalized;
d. select Ordinal, Logit in the Family/Link control;
e. select Up in the Measurement direction control;
f. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
6. Create the covariance between MathAtt and MathAb.
a. Select the Add Covariance tool,

b. Click in the top-left quadrant of the MathAb oval, and drag a covariance to the bottom left
of the MathAtt oval.
7. Clean up.
If you do not like where a covariance has been connected to its variable, use the Select tool,
, to simply click on the covariance, and then click on where it connects to an oval and drag
the endpoint. You can also change the bow of the covariance by dragging the control point that
extends from one end of the selected covariance.
8. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_2fmm

example 31g Two-factor measurement model (generalized response)

Also see
[SEM] example 27g Single-factor measurement model (generalized response)
[SEM] example 29g Two-parameter logistic IRT model
[SEM] example 32g Full structural equation model (generalized response)

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

329

Title
example 32g Full structural equation model (generalized response)

Description

Remarks and examples

Also see

Description
To demonstrate a structural model with a measurement component, we use the same data used in
[SEM] example 31g:
. use http://www.stata-press.com/data/r13/gsem_cfa
(Fictional math abilities data)
. summarize
Variable

Obs

Mean

Std. Dev.

Min

Max

school
id
q1
q2
q3

500
500
500
500
500

10.5
50681.71
.506
.394
.534

5.772056
29081.41
.5004647
.4891242
.4993423

1
71
0
0
0

20
100000
1
1
1

q4
q5
q6
q7
q8

500
500
500
500
500

.424
.49
.434
.52
.494

.4946852
.5004006
.4961212
.5001002
.5004647

0
0
0
0
0

1
1
1
1
1

att1
att2
att3
att4
att5

500
500
500
500
500

2.946
2.948
2.84
2.91
3.086

1.607561
1.561465
1.640666
1.566783
1.581013

1
1
1
1
1

5
5
5
5
5

test1
500
75.548
5.948653
55
93
test2
500
80.556
4.976786
65
94
500
75.572
6.677874
50
94
test3
test4
500
74.078
8.845587
43
96
. notes
_dta:
1. Fictional data on math ability and attitudes of 500 students from 20
schools.
2. Variables q1-q8 are incorrect/correct (0/1) on individual math questions.
3. Variables att1-att5 are items from a Likert scale measuring each
students attitude toward math.
4. Variables test1-test4 are test scores from tests of four different
aspects of mathematical abilities. Range of scores: 0-100.

These data record results from a fictional instrument measuring mathematical ability. Variables q1
through q8 are the items from the instrument.
In this example, we will also be using variables att1 through att5. These are five Likert-scale
questions measuring each students attitude toward math.
See Structural models 8: Unobserved inputs, outputs, or both in [SEM] intro 5 for background.

330

example 32g Full structural equation model (generalized response)

331

Remarks and examples

Remarks are presented under the following headings:
Structural model with measurement component
Fitting the model with the Builder

Structural model with measurement component

We wish to fit the following model:
ordinal

ordinal

ordinal

ordinal

ordinal

att1

att2

att3

att4

att5

logit

logit

logit

logit

logit

MathAtt

e1

MathAb

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

Bernoulli

q1

q2

q3

q4

q5

q6

q7

q8

logit

logit

logit

logit

logit

logit

logit

logit

This is the same model we fit in [SEM] example 31g, except that rather than a correlation (curved
path) between MathAtt and MathAb, this time we assume a direct effect and so allow a straight path.
If you compare the two path diagrams, in addition to the new substitution of the direct path for
the curved path signifying correlation, there is now an error variable on MathAb. In the previous
diagram, MathAb was exogenous. In this diagram, it is endogenous and thus requires an error term.
In the Builder, the error term is added automatically.
To fit this model in the command language, we type
. gsem (MathAb -> q1-q8, logit)
>
(MathAtt -> att1-att5, ologit)
>
(MathAtt -> MathAb)
Fitting fixed-effects model:
Iteration 0:
Iteration 1:
Iteration 2:

log likelihood = -6629.7253

log likelihood = -6628.7848
log likelihood = -6628.7848

Refining starting values:

Grid node 0:

log likelihood = -6429.1636

332

example 32g Full structural equation model (generalized response)

Fitting full model:
Iteration 0:
log likelihood =
Iteration 1:
log likelihood =
Iteration 2:
log likelihood =
Iteration 3:
log likelihood =
Iteration 4:
log likelihood =
Iteration 5:
log likelihood =
Generalized structural equation
Log likelihood = -6394.3923
( 1) [q1]MathAb = 1
( 2) [att1]MathAtt = 1
Coef.

-6429.1636
-6396.7471
-6394.6197
-6394.3949
-6394.3923
-6394.3923
model

Std. Err.

Number of obs

500

P>|z|

[95% Conf. Interval]

0.726

-.204885

.294109

q1 <MathAb
_cons

1
.044612

(constrained)
.1272967
0.35

MathAb
_cons

.3446066
-.4572215

.1050261
.0979965

3.28
-4.67

0.001
0.000

.1387593
-.6492911

.550454
-.2651519

MathAb
_cons

.5445222
.1591406

.1386992
.1033116

3.93
1.54

0.000
0.123

.2726767
-.0433465

.8163677
.3616276

MathAb
_cons

.2858862
-.3196648

.0948549
.0947684

3.01
-3.37

0.003
0.001

.099974
-.5054075

.4717984
-.1339222

MathAb
_cons

.8174769
-.04543

.1867022
.116575

4.38
-0.39

0.000
0.697

.4515473
-.2739129

1.183406
.1830528

MathAb
_cons

.6030423
-.3099919

.1471949
.1070853

4.10
-2.89

0.000
0.004

.3145457
-.5198754

.8915389
-.1001085

MathAb
_cons

.7208369
.1047264

.171309
.1116494

4.21
0.94

0.000
0.348

.3850774
-.1141024

1.056597
.3235553

MathAb
_cons

.5814736
-.0250443

.1426725
.1045135

4.08
-0.24

0.000
0.811

.3018406
-.2298869

.8611067
.1797984

att1 <MathAtt

att2 <MathAtt

.3788715

.0971234

3.90

0.000

.1885131

.5692299

att3 <MathAtt

-1.592717

.3614956

-4.41

0.000

-2.301236

-.8841989

att4 <MathAtt

-.8100108

.1530675

-5.29

0.000

-1.110017

-.510004

att5 <MathAtt

.5225425

.1170166

4.47

0.000

.2931942

.7518907

q2 <-

q3 <-

q4 <-

q5 <-

q6 <-

q7 <-

q8 <-

(constrained)

example 32g Full structural equation model (generalized response)

MathAb <MathAtt

.581103

.14776

3.93

0.000

.2914987

.8707072

/cut1
/cut2
/cut3
/cut4

-1.10254
-.2495339
.2983261
1.333052

.131228
.1160385
.1164415
.1391919

-8.40
-2.15
2.56
9.58

0.000
0.032
0.010
0.000

-1.359742
-.4769652
.070105
1.060241

-.8453377
-.0221025
.5265472
1.605864

/cut1
/cut2
/cut3
/cut4

-1.055791
-.1941211
.3598488
1.132624

.1062977
.0941435
.0952038
.1082204

-9.93
-2.06
3.78
10.47

0.000
0.039
0.000
0.000

-1.264131
-.378639
.1732528
.9205156

-.8474513
-.0096032
.5464448
1.344732

/cut1
/cut2
/cut3
/cut4

-1.053519
-.0491074
.5570672
1.666859

.1734001
.1442846
.1538702
.2135557

-6.08
-0.34
3.62
7.81

0.000
0.734
0.000
0.000

-1.393377
-.3319
.2554871
1.248297

-.7136612
.2336853
.8586472
2.08542

/cut1
/cut2
/cut3
/cut4

-1.07378
-.2112462
.406347
1.398185

.1214071
.1076501
.1094847
.1313327

-8.84
-1.96
3.71
10.65

0.000
0.050
0.000
0.000

-1.311734
-.4222365
.191761
1.140778

-.8358264
-.0002559
.620933
1.655593

/cut1
/cut2
/cut3
/cut4

-1.244051
-.336135
.2137776
.9286849

.1148443
.0986678
.0978943
.107172

-10.83
-3.41
2.18
8.67

0.000
0.001
0.029
0.000

-1.469142
-.5295203
.0219084
.7186316

-1.018961
-.1427498
.4056468
1.138738

1.787117
1.520854

.5974753
.4077885

.9280606
.8991947

3.441357
2.572298

333

att1

att2

att3

att4

att5

var(e.MathAb)
var(MathAtt)

Notes:
1. In the model fit in [SEM] example 31g, we estimated a correlation between MathAtt and MathAb
of 0.4724.
2. Theoretically speaking, the model fit above and the model in [SEM] example 31g are equivalent.
Both posit a linear relationship between the latent variables and merely choose to parameterize
the relationship differently. In [SEM] example 31g, it was parameterized as a covariance. In this
example, it is parameterized as causal. People often use structural equation modeling to confirm
a proposed hypothesis. It is important that the causal model you specify be based on theory or
that you have some other justification. You need something other than empirical results to rule
out competing but equivalent models such as the covariance model. Distinguishing causality from
correlation is always problematic.
3. Practically speaking, note that the log-likelihood values for this model and the model in [SEM] example 31g are equal at 6394.3923. Also note that the estimated variances of math attitude,
var(MathAtt), are also equal at 1.520854.

334

example 32g Full structural equation model (generalized response)

Fitting the model with the Builder

Use the diagram in Structural model with measurement component above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_cfa

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the measurement component for MathAb.

Select the Add Measurement Component tool, , and then click in the diagram about one-third
of the way up from the bottom and slightly left of the center.
In the resulting dialog box,
a. change the Latent variable name to MathAb;
b. select q1, q2, q3, q4, q5, q6, q7, and q8 by using the Measurement variables control;
c. check Make measurements generalized;
d. select Bernoulli, Logit in the Family/Link control;
e. select Down in the Measurement direction control;
f. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
5. Create the measurement component for MathAtt.
Select the Add Measurement Component tool, , and then click in the diagram about one-third
of the way down from the top and slightly left of the center.
In the resulting dialog box,
a. change the Latent variable name to MathAtt;
b. select att1, att2, att3, att4, and att5 by using the Measurement variables control;
c. check Make measurements generalized;
d. select Ordinal, Logit in the Family/Link control;
e. select Up in the Measurement direction control;
f. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
6. Create path from MathAtt to MathAb.
a. Select the Add Path tool,

b. Click in the bottom of the MathAtt oval (it will highlight when you hover over it), and drag
a path to the top of the MathAb oval (it will highlight when you can release to connect the
path).

example 32g Full structural equation model (generalized response)

335

7. Clean up the direction of the error.

The error on MathAb has likely been created below the oval instead of to the right of the oval.
If so, choose the Select tool, , and then click in the MathAb oval. Click on one of the Error
, in the Contextual Toolbar until the error is where you want it.
Rotation buttons,
8. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_sem

Also see
[SEM] example 9 Structural model with measurement component
[SEM] example 31g Two-factor measurement model (generalized response)

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
example 33g Logistic regression

Description

Remarks and examples

Reference

Also see

Description
In this example, we demonstrate with gsem how to fit a standard logistic regression, which is
often referred to as the logit model in generalized linear model (GLM) framework.
. use http://www.stata-press.com/data/r13/gsem_lbw
(Hosmer & Lemeshow data)
. describe
Contains data from http://www.stata-press.com/data/r13/gsem_lbw.dta
obs:
189
Hosmer & Lemeshow data
vars:
11
21 Mar 2013 12:28
size:
2,646
(_dta has notes)

variable name
id
low
age
lwt
race
smoke
ptl
ht
ui
ftv
bwt

storage
type
int
byte
byte
int
byte
byte
byte
byte
byte
byte
int

display
format
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%9.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g

value
label

race
smoke

variable label
subject id
birth weight < 2500g
age of mother
weight, last menstrual period
race
smoked during pregnancy
premature labor history (count)
has history of hypertension
presence, uterine irritability
# physician visits, 1st trimester
birth weight (g)

Sorted by:
. notes
_dta:
1. Data from Hosmer, D. W., Jr., S. A. Lemeshow, and R. X. Sturdivant. 2013.
"Applied Logistic Regression". 3rd ed. Hoboken, NJ: Wiley.
2. Data from a study of risk factors associated with low birth weights.

See Structural models 3: Binary-outcome models in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Fitting the logit model
Obtaining odds ratios
Fitting the model with the Builder

336

example 33g Logistic regression

337

Fitting the logit model

The model we wish to fit is

age

lwt

1b.race

2.race
Bernoulli

3.race

low
logit

smoke

ptl

ht

ui

That is, we wish to fit a model in which low birthweight is determined by a history of hypertension
(ht), mothers age (age), mothers weight at last menstrual period (lwt), mothers race (white, black,
or other; race), whether the mother smoked during pregnancy (smoke), the number of premature
babies previously born to the mother (ptl), and whether the mother has suffered from the presence
of uterine irritability (ui).
The path diagram matches the variable names listed in parentheses above except for race, where
the path diagram contains not one but three boxes filled in with 1b.race, 2.race, and 3.race.
This is because in our dataset, race is coded 1, 2, or 3, meaning white, black, or other. We want to
include indicator variables for race so that we have a separate coefficient for each race. Thus we
need three boxes.
In Stata, 1.race means an indicator for race equaling 1. Thus it should not surprise you if you
filled in the boxes with 1.race, 2.race, and 3.race, and that is almost what we did. The difference
is that we filled in the first box with 1b.race rather than 1.race. We use the b to specify the base
category, which we specified as white. If we wanted the base category to be black, we would have
specified 2b.race and left 1.race alone.

338

example 33g Logistic regression

The above is called factor-variable notation. See [SEM] intro 3 for details on using factor-variable
notation with gsem.
In the command language, we could type
. gsem (low <- age lwt 1b.race 2.race 3.race smoke ptl ht ui), logit

to fit the model. Written that way, there is a one-to-one correspondence to what we would type and
what we would draw in the Builder. The command language, however, has a feature that will allow
us to type i.race instead of 1b.race 2.race 3.race. To fit the model, we could type
. gsem (low <- age lwt i.race smoke ptl ht ui), logit

i.varname is a command-language shorthand for specifying indicators for all the levels of a
variable and using the first level as the base category. You can use i.varname in the command
language but not in path diagrams because boxes can contain only one variable. In the Builder,
however, you will discover a neat feature so that you can type i.race and the Builder will create
however many boxes are needed for you, filled in, and with the first category marked as the base.
We will explain below how you do that.
The result of typing our estimation command is
. gsem (low <- age lwt i.race smoke ptl ht ui), logit
Iteration 0:
log likelihood = -101.0213
Iteration 1:
log likelihood = -100.72519
Iteration 2:
log likelihood =
-100.724
Iteration 3:
log likelihood =
-100.724
Generalized structural equation model
Number of obs
Log likelihood =
-100.724
Coef.

Std. Err.

P>|z|

189

[95% Conf. Interval]

low <age
lwt

-.0271003
-.0151508

.0364504
.0069259

-0.74
-2.19

0.457
0.029

-.0985418
-.0287253

.0443412
-.0015763

race
black
other

1.262647
.8620792

.5264101
.4391532

2.40
1.96

0.016
0.050

.2309024
.0013548

2.294392
1.722804

smoke
ptl
ht
ui
_cons

.9233448
.5418366
1.832518
.7585135
.4612239

.4008266
.346249
.6916292
.4593768
1.20459

2.30
1.56
2.65
1.65
0.38

0.021
0.118
0.008
0.099
0.702

.137739
-.136799
.4769494
-.1418484
-1.899729

1.708951
1.220472
3.188086
1.658875
2.822176

Obtaining odds ratios

Some of you are looking at the output above, nodding your heads, and thinking to yourselves,
Yes, thats right. Others are shaking your heads sadly and thinking, Where are the exponentiated
coefficients, the odds ratios? Researchers from different backgrounds are used to seeing logit results
presented in two different ways.

example 33g Logistic regression

339

If you want to see the odds ratios, type estat eform after fitting the model:
. estat eform
low

exp(b)

Std. Err.

age
lwt

.9732636
.9849634

.0354759
.0068217

race
white
black
other

1
3.534767
2.368079

smoke
ptl
ht
ui
_cons

2.517698
1.719161
6.249602
2.1351
1.586014

P>|z|

[95% Conf. Interval]

-0.74
-2.19

0.457
0.029

.9061578
.9716834

1.045339
.9984249

(empty)
1.860737
1.039949

2.40
1.96

0.016
0.050

1.259736
1.001356

9.918406
5.600207

1.00916
.5952579
4.322408
.9808153
1.910496

2.30
1.56
2.65
1.65
0.38

0.021
0.118
0.008
0.099
0.702

1.147676
.8721455
1.611152
.8677528
.1496092

5.523162
3.388787
24.24199
5.2534
16.8134

Whichever way you look at the results above, they are identical to the results that would be produced
by typing
. logit low age lwt i.race smoke ptl ht ui

or
. logistic low age lwt i.race smoke ptl ht ui

which are two other ways that Stata can fit logit models. logit, like gsem, reports coefficients by
default. logistic reports odds ratios by default.

Fitting the model with the Builder

Use the diagram in Fitting the logit model above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_lbw

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Enlarge the size of the canvas to accommodate the height of the diagram.
Click on the Adjust Canvas Size button,
5 (inches), and then click on OK.

, in the Standard Toolbar, change the second size to

5. Create the logistic regression component for low.

Select the Add Regression Component tool,
, and then click in the diagram about one-third
of the way in from the left and halfway down.
In the resulting dialog box,
a. select low in the Dependent variable control;

340

example 33g Logistic regression

b. check Make measurements generalized;

c. select Bernoulli, Logit in the Family/Link control;
d. select the Select variables radio button (it may already be selected);
e. use the Independent variables control to select the variables age and lwt;
button next to the
f. include the levels of the factor variable race by clicking on the
Independent variables control. In the resulting dialog box, select the Factor variable radio
button, select Main effect in the Specification control, and select race in the Variables
control for Variable 1. Click on Add to varlist, and then click on OK;
g. continue with the Independent variables control to select the variables smoke, ptl, ht, and
ui;
h. select Left in the Independent variables direction control;
i. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
6. Clean up.
The box for low is created closer to the independent variables than it is in the example diagram.
Use the Select tool, , and click on the box for low. Drag it to the right to allow more space
for results along the paths.
7. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_logit

Reference
Hosmer, D. W., Jr., S. A. Lemeshow, and R. X. Sturdivant. 2013. Applied Logistic Regression. 3rd ed. Hoboken,
NJ: Wiley.

Also see
[SEM] example 34g Combined models (generalized responses)
[SEM] example 35g Ordered probit and ordered logit
[SEM] example 37g Multinomial logistic regression

[SEM] gsem Generalized structural equation model estimation command

[SEM] estat eform Display exponentiated coefficients
[SEM] intro 3 Learning the language: Factor-variable notation (gsem only)
[SEM] intro 5 Tour of models

Title
example 34g Combined models (generalized responses)

Description

Remarks and examples

Reference

Also see

Description
We demonstrate how to fit a combined model with one Poisson regression and one logit regression
by using the following data:
. use http://www.stata-press.com/data/r13/gsem_lbw
(Hosmer & Lemeshow data)
. describe
Contains data from http://www.stata-press.com/data/r13/gsem_lbw.dta
obs:
189
Hosmer & Lemeshow data
vars:
11
21 Mar 2013 12:28
size:
2,646
(_dta has notes)

variable name
id
low
age
lwt
race
smoke
ptl
ht
ui
ftv
bwt

storage
type
int
byte
byte
int
byte
byte
byte
byte
byte
byte
int

display
format
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%9.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g

value
label

race
smoke

variable label
subject id
birth weight < 2500g
age of mother
weight, last menstrual period
race
smoked during pregnancy
premature labor history (count)
has history of hypertension
presence, uterine irritability
# physician visits, 1st trimester
birth weight (g)

Sorted by:
. notes
_dta:
1. Data from Hosmer, D. W., Jr., S. A. Lemeshow, and R. X. Sturdivant. 2013.
"Applied Logistic Regression". 3rd ed. Hoboken, NJ: Wiley.
2. Data from a study of risk factors associated with low birth weights.

See Structural models 7: Dependencies between response variables in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Fitting the combined model
Obtaining odds ratios and incidence-rate ratios
Fitting the model with the Builder

341

342

example 34g Combined models (generalized responses)

Fitting the combined model

The model we wish to fit is

age
smoke
Poisson

ht

ptl
log

lwt
1b.race
Bernoulli

2.race

low
logit

3.race
ui
This model has one logit equation and one Poisson regression equation, with the Poisson response
variable also being an explanatory variable in the logit equation.
Because the two equations are recursive, it is not necessary to fit these models together. We could
draw separate diagrams for each equation and fit each separately. Even so, many researchers often do
fit recursive models together, and sometimes, it is just the first step before placing constraints across
models or introducing a common latent variable. The latter might be likely in this case because neither
generalized linear response has an error that could be correlated and so the only way to correlate
these two responses in gsem is to add a shared latent variable affecting each.
Our purpose here is to show that you can mix models with generalized response variables of
different types.

example 34g Combined models (generalized responses)

343

To fit the model in the command language, we type

. gsem (low <- ptl age smoke ht lwt i.race ui, logit)
>
(ptl <- age smoke ht, poisson)
Iteration 0:
log likelihood = -322.96738
Iteration 1:
log likelihood = -200.5818
Iteration 2:
log likelihood = -198.58086
Iteration 3:
log likelihood = -198.56179
Iteration 4:
log likelihood = -198.56178
Generalized structural equation model
Number of obs
Log likelihood = -198.56178
Coef.

Std. Err.

P>|z|

189

[95% Conf. Interval]

low <ptl
age
smoke
ht
lwt

.5418366
-.0271003
.9233448
1.832518
-.0151508

.346249
.0364504
.4008266
.6916292
.0069259

1.56
-0.74
2.30
2.65
-2.19

0.118
0.457
0.021
0.008
0.029

-.136799
-.0985418
.137739
.4769494
-.0287253

1.220472
.0443412
1.708951
3.188086
-.0015763

race
black
other

1.262647
.8620791

.5264101
.4391532

2.40
1.96

0.016
0.050

.2309023
.0013548

2.294392
1.722803

ui
_cons

.7585135
.4612238

.4593768
1.20459

1.65
0.38

0.099
0.702

-.1418484
-1.899729

1.658875
2.822176

age
smoke
ht
_cons

.0370598
.9602534
-.1853501
-2.985512

.0298752
.3396867
.7271851
.7842174

1.24
2.83
-0.25
-3.81

0.215
0.005
0.799
0.000

-.0214946
.2944796
-1.610607
-4.52255

.0956142
1.626027
1.239906
-1.448474

ptl <-

Obtaining odds ratios and incidence-rate ratios

As mentioned in [SEM] example 33g, some researchers prefer to see exponentiated coefficients.
In both odds ratios and incidence-rate ratios, exponentiation is meaningful. Exponentiated logit
coefficients are odds ratios, and exponentiated Poisson regression coefficients are incidence-rate ratios.
To obtain exponentiated coefficients for both equations, we type

344

example 34g Combined models (generalized responses)

. estat eform low ptl
exp(b)

Std. Err.

P>|z|

[95% Conf. Interval]

low
ptl
age
smoke
ht
lwt

1.719161
.9732636
2.517698
6.249602
.9849634

.5952579
.0354759
1.00916
4.322407
.0068217

1.56
-0.74
2.30
2.65
-2.19

0.118
0.457
0.021
0.008
0.029

.8721455
.9061578
1.147676
1.611152
.9716834

3.388787
1.045339
5.523162
24.24199
.9984249

race
white
black
other

1
3.534767
2.368079

(empty)
1.860737
1.039949

2.40
1.96

0.016
0.050

1.259736
1.001356

9.918406
5.600207

ui
_cons

2.1351
1.586014

.9808153
1.910496

1.65
0.38

0.099
0.702

.8677528
.1496092

5.2534
16.8134

age
smoke
ht
_cons

1.037755
2.612358
.8308134
.0505137

.0310032
.8873835
.6041551
.0396137

1.24
2.83
-0.25
-3.81

0.215
0.005
0.799
0.000

.9787348
1.342428
.1997664
.0108613

1.100334
5.083638
3.45529
.2349286

ptl

Had we merely typed estat eform without the two equation names, we would have obtained
exponentiated coefficients for the first equation only.
Equation names are easily found on the output or the path diagrams. Equations are named after
the dependent variable.

Fitting the model with the Builder

Use the diagram in Fitting the combined model above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_lbw

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the independent variables.

Select the Add Observed Variables Set tool,
about one-third of the way in from the left.

, and then click at the bottom of the diagram

In the resulting dialog box,

a. select the Select variables radio button (it may already be selected);
b. use the Variables control to select the variables age, smoke, ht, and lwt in this order;
c. include the levels of the factor variable race by clicking on the
button next to the
Variables control. In the resulting dialog box, select the Factor variable radio button, select

example 34g Combined models (generalized responses)

345

Main effect in the Specification control, and select race in the Variables control for
Variable 1. Click on Add to varlist, and then click on OK;
d. continue with the Variables control and select the variable ui;
e. select Vertical in the Orientation control;
f. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
5. Create the generalized response for premature labor history.
a. Select the Add Generalized Response Variable tool,

b. Click about one-third of the way in from the right side of the diagram, to the right of ht.
c. In the Contextual Toolbar, select Poisson, Log in the Family/Link control.
d. In the Contextual Toolbar, select ptl in the Variable control.
6. Create the generalized response for low birthweight.
a. Select the Add Generalized Response Variable tool,

b. Click about one-third of the way in from the right side of the diagram, to the right of
2.race.
c. In the Contextual Toolbar, select Bernoulli, Logit in the Family/Link control.
d. In the Contextual Toolbar, select low in the Variable control.
7. Create paths from the independent variables to the dependent variables.
a. Select the Add Path tool,

b. Click in the right side of the age rectangle (it will highlight when you hover over it), and
drag a path to the left side of the ptl rectangle (it will highlight when you can release to
connect the path).
tool, create the following paths by clicking first in the right side of
c. Continuing with the
the rectangle for the independent variable and dragging it to the left side of the rectangle
for the dependent variable:
smoke -> ptl
ht -> ptl
age -> low
smoke -> low
ht -> low
lwt -> low
1b.race -> low
2.race -> low
3.race -> low
ui -> low
d. Continuing with the
tool, create the path from ptl to low by clicking in the bottom of
the ptl rectangle and dragging the path to the top of the low rectangle.

346

example 34g Combined models (generalized responses)

8. Clean up.
If you do not like where a path has been connected to its variables, use the Select tool,
,
to click on the path, and then simply click on where it connects to a rectangle and drag the
endpoint.
9. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_comb

Reference
Hosmer, D. W., Jr., S. A. Lemeshow, and R. X. Sturdivant. 2013. Applied Logistic Regression. 3rd ed. Hoboken,
NJ: Wiley.

Also see
[SEM] example 33g Logistic regression
[SEM] example 45g Heckman selection model
[SEM] example 46g Endogenous treatment-effects model

[SEM] gsem Generalized structural equation model estimation command

[SEM] estat eform Display exponentiated coefficients
[SEM] intro 5 Tour of models

Title
example 35g Ordered probit and ordered logit

Description

Remarks and examples

Reference

Also see

Description
Below we demonstrate ordered probit and ordered logit in a measurement-model context. We are
not going to illustrate every family/link combination. Ordered probit and logit, however, are unique
in that a single equation is able to predict a set of ordered outcomes. The unordered alternative,
mlogit, requires k 1 equations to fit k (unordered) outcomes.
To demonstrate ordered probit and ordered logit, we use the following data:
. use http://www.stata-press.com/data/r13/gsem_issp93
(Selection from ISSP 1993)
. describe
Contains data from http://www.stata-press.com/data/r13/gsem_issp93.dta
obs:
871
Selection for ISSP 1993
vars:
8
21 Mar 2013 16:03
size:
7,839
(_dta has notes)
storage
type

display
format

value
label

id
y1

int
byte

%9.0g
%26.0g

agree5

y2
y3
y4

byte
byte
byte

%26.0g
%26.0g
%26.0g

agree5
agree5
agree5

sex
age
edu

byte
byte
byte

%9.0g
%9.0g
%20.0g

sex
age
edu

variable name

variable label
respondent identifier
too much science, not enough
feelings & faith
science does more harm than good
any change makes nature worse
science will solve environmental
problems
sex
age (6 categories)
education (6 categories)

Sorted by:
. notes
_dta:
1. Data from Greenacre, M. and J Blasius, 2006, _Multiple Correspondence
Analysis and Related Methods_, pp. 42-43, Boca Raton: Chapman & Hall.
Data is a subset of the International Social Survey Program (ISSP) 1993.
2. Full text of y1: We believe too often in science, and not enough in
feelings and faith.
3. Full text of y2: Overall, modern science does more harm than good.
4. Full text of y3: Any change humans cause in nature, no matter how
scientific, is likely to make things worse.
5. Full text of y4: Modern science will solve our environmental problems
with little change to our way of life.

See Structural models 5: Ordinal models in [SEM] intro 5 for background.

347

348

example 35g Ordered probit and ordered logit

Remarks and examples

Remarks are presented under the following headings:
Ordered probit
Ordered logit
Fitting the model with the Builder

Ordered probit
For the measurement model, we focus on variables y1 through y4. Each variable contains 15,
with 1 meaning strong disagreement and 5 meaning strong agreement with a statement about science.
Ordered probit produces predictions about the probabilities that a respondent gives response 1,
response 2, . . . , response k . It does this by dividing up the domain of an N (0, 1) distribution into
k categories defined by k 1 cutpoints, c1 , c2 , . . . , ck1 . Individual respondents are assumed to
have a score s = X + , where  N (0, 1), and then that score is used along with the cutpoints
to produce probabilities for each respondent producing response 1, 2, . . . , k .
Pr(response is i | X) = Pr(ci1 < X +  ci )
where c0 = ; ck = +; and c1 , c2 , . . . , ck1 and are parameters of the model to be fit. This
ordered probit model has long been known in Stata circles as oprobit.
We have a set of four questions designed to determine the respondents attitude toward science,
each question with k = 5 possible answers ranging on a Likert scale from 1 to 5. With ordered probit
in hand, we have a way to take a continuous variable, say, a latent variable we will call SciAtt, and
produce predicted categorical responses.
The measurement model we want to fit is

SciAtt

ordinal

ordinal

ordinal

ordinal

y1

y2

y3

y4

probit

probit

probit

probit

example 35g Ordered probit and ordered logit

We fit the model in the command language by typing

. gsem (y1 y2 y3 y4 <- SciAtt),
Fitting fixed-effects model:
Iteration 0:
log likelihood =
Iteration 1:
log likelihood =
Refining starting values:
Grid node 0:
log likelihood =
Fitting full model:
Iteration 0:
log likelihood =
Iteration 1:
log likelihood =
Iteration 2:
log likelihood =
Iteration 3:
log likelihood =
Iteration 4:
log likelihood =
Iteration 5:
log likelihood =
Iteration 6:
log likelihood =
Iteration 7:
log likelihood =
Generalized structural equation
Log likelihood = -5039.823
( 1) [y1]SciAtt = 1
Coef.

oprobit
-5227.8743
-5227.8743
-5230.8106
-5230.8106
-5132.1849
-5069.5037
-5040.4779
-5040.2397
-5039.8242
-5039.823
-5039.823
model

Std. Err.

(not concave)
(not concave)

Number of obs

871

P>|z|

[95% Conf. Interval]

y1 <SciAtt

(constrained)

SciAtt

1.424366

.2126574

6.70

0.000

1.007565

1.841167

SciAtt

1.283359

.1797557

7.14

0.000

.931044

1.635674

SciAtt

-.0322354

.0612282

-0.53

0.599

-.1522405

.0877697

/cut1
/cut2
/cut3
/cut4

-1.343148
.0084719
.7876538
1.989985

.0726927
.0521512
.0595266
.0999181

-18.48
0.16
13.23
19.92

0.000
0.871
0.000
0.000

-1.485623
-.0937426
.6709837
1.794149

-1.200673
.1106863
.9043238
2.18582

/cut1
/cut2
/cut3
/cut4

-1.997245
-.8240241
.0547025
1.419923

.1311972
.0753839
.0606036
.1001258

-15.22
-10.93
0.90
14.18

0.000
0.000
0.367
0.000

-2.254387
-.9717738
-.0640784
1.22368

-1.740104
-.6762743
.1734834
1.616166

/cut1
/cut2
/cut3
/cut4

-1.271915
.1249493
.9752553
2.130661

.0847483
.0579103
.0745052
.1257447

-15.01
2.16
13.09
16.94

0.000
0.031
0.000
0.000

-1.438019
.0114472
.8292277
1.884206

-1.105812
.2384515
1.121283
2.377116

/cut1
/cut2
/cut3
/cut4

-1.484063
-.4259356
.1688777
.9413113

.0646856
.0439145
.0427052
.0500906

-22.94
-9.70
3.95
18.79

0.000
0.000
0.000
0.000

-1.610844
-.5120065
.0851771
.8431356

-1.357281
-.3398647
.2525782
1.039487

.5265523

.0979611

.3656637

.7582305

y2 <-

y3 <-

y4 <-

y1

y2

y3

y4

var(SciAtt)

349

350

example 35g Ordered probit and ordered logit

Notes:
1. The cutpoints c1 , . . . , c4 are labeled /cut1, . . . , /cut4 in the output. We have a separate
cutpoint for each of the four questions y1, . . . , y4. Look at the estimated cutpoints for y1, which
are 1.343, 0.008, 0.788, and 1.99. The probabilities that a person with SciAtt = 0 (its mean)
would give the various responses are
Pr(response 1) = normal(-1.343) = 0.090
Pr(response 2) = normal(0.008) normal(1.343) = 0.414
Pr(response 3) = normal(0.788) normal(0.008) = 0.281
Pr(response 4) = normal(1.99) normal(0.788) = 0.192
Pr(response 5) = 1 normal(1.99) = 0.023
2. The path coefficients (y1 y2 y3 y4 <- SciAtt) measure the effect of the latent variable we
called science attitude on each of the responses.
3. The estimated path coefficients are 1, 1.42, 1.28, and 0.03 for the four questions.
4. If you read the questionsthey are listed aboveyou will find that in all but the fourth question,
agreement signifies a negative attitude toward science. Thus SciAtt measures a negative attitude
toward science because the loadings on negative questions are positive and the loading on the
single positive question is negative.
5. The direction of the meanings of latent variables is always a priori indeterminate and is set by
the identifying restrictions we apply. We appliedor more correctly, gsem applied for usthe
constraint that y1 <- SciAtt has path coefficient 1. Because statement 1 was a negative statement
about science, that was sufficient to set the direction of SciAtt to be the opposite of what we
hoped for.
The direction does not matter. You simply must remember to interpret the latent variable correctly
when reading results based on it. In the models we fit, including more complicated models, the
signs of the coefficients will work themselves out to adjust for the direction of the variable.

Ordered logit
The description of the ordered logit model is identical to that of the ordered probit model except
that where we assumed a normal distribution in our explanation above, we now assume a logit
distribution. The distributions are similar.
To fit an ordered logit (ologit) model, the link function shown in the boxes merely changes from
probit to logit:

SciAtt

ordinal

ordinal

ordinal

ordinal

y1

y2

y3

y4

logit

logit

logit

logit

example 35g Ordered probit and ordered logit

We can fit the model in the command language by typing

. gsem (y1 y2 y3 y4 <- SciAtt), ologit
Fitting fixed-effects model:
Iteration 0:
log likelihood = -5227.8743
Iteration 1:
log likelihood = -5227.8743
Refining starting values:
Grid node 0:
log likelihood =
Fitting full model:
Iteration 0:
log likelihood =
Iteration 1:
log likelihood =
Iteration 2:
log likelihood =
Iteration 3:
log likelihood =
Iteration 4:
log likelihood =
Iteration 5:
log likelihood =
Generalized structural equation
Log likelihood = -5035.0352
( 1)

-5127.9026
-5127.9026
-5065.4679
-5035.9766
-5035.0943
-5035.0353
-5035.0352
model

(not concave)

Number of obs

871

[y1]SciAtt = 1
Coef.

Std. Err.

P>|z|

[95% Conf. Interval]

y1 <SciAtt

(constrained)

SciAtt

1.394767

.2065479

6.75

0.000

.9899406

1.799593

SciAtt

1.29383

.1845113

7.01

0.000

.9321939

1.655465

SciAtt

-.0412446

.0619936

-0.67

0.506

-.1627498

.0802606

/cut1
/cut2
/cut3
/cut4

-2.38274
-.0088393
1.326292
3.522017

.1394292
.0889718
.106275
.1955535

-17.09
-0.10
12.48
18.01

0.000
0.921
0.000
0.000

-2.656016
-.1832207
1.117997
3.138739

-2.109464
.1655422
1.534587
3.905295

/cut1
/cut2
/cut3
/cut4

-3.51417
-1.421711
.0963154
2.491459

.2426595
.135695
.1046839
.1840433

-14.48
-10.48
0.92
13.54

0.000
0.000
0.358
0.000

-3.989774
-1.687669
-.1088612
2.130741

-3.038566
-1.155754
.3014921
2.852178

/cut1
/cut2
/cut3
/cut4

-2.263557
.2024798
1.695997
3.828154

.1618806
.1012122
.1393606
.2464566

-13.98
2.00
12.17
15.53

0.000
0.045
0.000
0.000

-2.580838
.0041075
1.422855
3.345108

-1.946277
.400852
1.969138
4.3112

/cut1
/cut2
/cut3
/cut4

-2.606013
-.6866159
.268862
1.561921

.1338801
.0718998
.0684577
.0895438

-19.47
-9.55
3.93
17.44

0.000
0.000
0.000
0.000

-2.868413
-.8275369
.1346874
1.386419

-2.343613
-.5456949
.4030366
1.737424

1.715641

.3207998

1.189226

2.475077

y2 <-

y3 <-

y4 <-

y1

y2

y3

y4

var(SciAtt)

351

352

example 35g Ordered probit and ordered logit

Note:
1. Results are nearly identical to those reported for ordered probit.

Fitting the model with the Builder

Use the diagram in Ordered probit above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_issp93

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the measurement component for SciAtt.

Select the Add Measurement Component tool, , and then click in the diagram about one-third
of the way down from the top and slightly left of the center.
In the resulting dialog box,
a. change the Latent variable name to SciAtt;
b. select y1, y2, y3, and y4 by using the Measurement variables control;
c. check Make measurements generalized;
d. select Ordinal, Probit in the Family/Link control;
e. select Down in the Measurement direction control;
f. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
5. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
6. To fit the model in Ordered logit, change the type of generalized response for each of the
measurement variables.
a. Choose the Select tool,

b. Click on the y1 rectangle. In the Contextual Toolbar, select Ordinal, Logit in the
Family/Link control.
c. Repeat this process to change the family and link to Ordinal, Logit for y2, y3, and y4.
7. Estimate again.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram for the ordered probit model in the Builder by typing
. webgetsem gsem_oprobit

example 35g Ordered probit and ordered logit

353

You can open a completed diagram for the ordered logit model in the Builder by typing
. webgetsem gsem_ologit

Reference
Greenacre, M. J. 2006. From simple to multiple correspondence analysis. In Multiple Correspondence Analysis and
Related Methods, ed. M. J. Greenacre and J. Blasius. Boca Raton, FL: Chapman & Hall.

Also see
[SEM] example 1 Single-factor measurement model
[SEM] example 27g Single-factor measurement model (generalized response)
[SEM] example 33g Logistic regression
[SEM] example 36g MIMIC model (generalized response)
[SEM] example 37g Multinomial logistic regression

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
example 36g MIMIC model (generalized response)

Description

Remarks and examples

Reference

Also see

Description
To demonstrate a multiple-indicators multiple-causes (MIMIC) model with generalized indicators,
we use the same data used in [SEM] example 35g:
. use http://www.stata-press.com/data/r13/gsem_issp93
(Selection from ISSP 1993)
. describe
Contains data from http://www.stata-press.com/data/r13/gsem_issp93.dta
obs:
871
Selection for ISSP 1993
vars:
8
21 Mar 2013 16:03
size:
7,839
(_dta has notes)
storage
type

display
format

value
label

id
y1

int
byte

%9.0g
%26.0g

agree5

y2
y3
y4

byte
byte
byte

%26.0g
%26.0g
%26.0g

agree5
agree5
agree5

sex
age
edu

byte
byte
byte

%9.0g
%9.0g
%20.0g

sex
age
edu

variable name

variable label
respondent identifier
too much science, not enough
feelings & faith
science does more harm than good
any change makes nature worse
science will solve environmental
problems
sex
age (6 categories)
education (6 categories)

Sorted by:
. notes
_dta:
1. Data from Greenacre, M. and J Blasius, 2006, _Multiple Correspondence
Analysis and Related Methods_, pp. 42-43, Boca Raton: Chapman & Hall.
Data is a subset of the International Social Survey Program (ISSP) 1993.
2. Full text of y1: We believe too often in science, and not enough in
feelings and faith.
3. Full text of y2: Overall, modern science does more harm than good.
4. Full text of y3: Any change humans cause in nature, no matter how
scientific, is likely to make things worse.
5. Full text of y4: Modern science will solve our environmental problems
with little change to our way of life.

See Structural models 9: MIMIC models in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Fitting the MIMIC model
Fitting the model with the Builder

354

example 36g MIMIC model (generalized response)

355

Fitting the MIMIC model

For a refresher on MIMIC models, see [SEM] example 10. In the data above, we have ordered
categorical indicators. For an explanation on how we are about to treat these indicators with ordered
probit, see [SEM] example 35g.
We wish to fit the following model:

sex

SciAtt

ordinal

ordinal

ordinal

ordinal

y1

y2

y3

y4

probit

probit

probit

probit

We fit the model in the command language by typing

. gsem (y1 y2 y3 y4 <- SciAtt) (SciAtt <- sex), oprobit
Fitting fixed-effects model:
Iteration 0:
log likelihood = -5227.8743
Iteration 1:
log likelihood = -5227.8743
Refining starting values:
Grid node 0:
log likelihood = -5230.8106
Fitting full model:
Iteration 0:
log likelihood = -5230.8106 (not concave)
Iteration 1:
log likelihood = -5132.1065 (not concave)
Iteration 2:
log likelihood = -5066.8481
Iteration 3:
log likelihood = -5033.2807
Iteration 4:
log likelihood = -5032.7884
Iteration 5:
log likelihood = -5032.3837
Iteration 6:
log likelihood = -5032.3778
Iteration 7:
log likelihood = -5032.3778

356

example 36g MIMIC model (generalized response)

Generalized structural equation model
Log likelihood = -5032.3778
( 1)

Number of obs

871

[y1]SciAtt = 1
Coef.

Std. Err.

P>|z|

[95% Conf. Interval]

y1 <SciAtt

(constrained)

SciAtt

1.405732

.2089672

6.73

0.000

.9961641

1.8153

SciAtt

1.246449

.1710771

7.29

0.000

.911144

1.581754

SciAtt

-.0345517

.0602017

-0.57

0.566

-.1525449

.0834415

SciAtt <sex

-.2337427

.0644245

-3.63

0.000

-.3600124

-.1074729

/cut1
/cut2
/cut3
/cut4

-1.469615
-.10992
.6729334
1.879901

.0855651
.0615897
.0644695
.0996675

-17.18
-1.78
10.44
18.86

0.000
0.074
0.000
0.000

-1.63732
-.2306336
.5465755
1.684557

-1.301911
.0107937
.7992914
2.075246

/cut1
/cut2
/cut3
/cut4

-2.16739
-.9912152
-.1118914
1.252164

.1480596
.0943091
.075311
.0983918

-14.64
-10.51
-1.49
12.73

0.000
0.000
0.137
0.000

-2.457582
-1.176058
-.2594982
1.05932

-1.877199
-.8063727
.0357154
1.445008

/cut1
/cut2
/cut3
/cut4

-1.412372
-.0230879
.8209522
1.966042

.0977772
.0687432
.0771653
.1196586

-14.44
-0.34
10.64
16.43

0.000
0.737
0.000
0.000

-1.604012
-.1578221
.6697109
1.731515

-1.220733
.1116464
.9721935
2.200568

/cut1
/cut2
/cut3
/cut4

-1.47999
-.4218768
.172995
.9454906

.0650596
.0443504
.0432394
.0507422

-22.75
-9.51
4.00
18.63

0.000
0.000
0.000
0.000

-1.607505
-.508802
.0882473
.8460376

-1.352476
-.3349516
.2577427
1.044944

.5283629

.0978703

.3675036

.7596315

y2 <-

y3 <-

y4 <-

y1

y2

y3

y4

var(e.SciAtt)

Notes:
1. Our latent variable measures a negative attitude toward science, just as it did in [SEM] example 35g.
2. In this MIMIC model, we allow males and females to have different underlying attitudes toward
science.
3. The coefficient for SciAtt <- sex is 0.234. Variable sex is equal to 1 for females, thus
females have a lower mean value for SciAtt by 0.234. Because our SciAtt measure is reversed,
this means that females have a more positive attitude toward science. The effect is significant at
better than the 1% level.
4. The difference between males and females in SciAtt is 0.234. Is that big or small, practically
speaking?

example 36g MIMIC model (generalized response)

357

In the ordered probit specification, predicted agreement with questions is determined by an

index into a N (0, 1) distribution. The value of the index is then compared with the cutpoints to
determine the probability that the response is 1, 2, 3, 4, or 5.
For statement 1, the loading on SciAtt is 1, and therefore the average difference in the probit
index for males and females is 0.0234 1 = 0.0234 standard deviation units. Females are
0.234 standard deviations to the left of males on average.
For statement 2, the loading on SciAtt is 1.41, and therefore the average difference in the probit
index is 0.0234 1.41 = 0.033 standard deviation units.
You can work out the effect size for the other statements. We would say that the effect is medium
sized.

Fitting the model with the Builder

Use the diagram in Fitting the MIMIC model above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_issp93

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the measurement component for SciAtt.

Select the Add Measurement Component tool,
and slightly left of the center.

, and then click in the diagram halfway down

In the resulting dialog box,

a. change the Latent variable name to SciAtt;
b. select y1, y2, y3, and y4 by using the Measurement variables control;
c. check Make measurements generalized;
d. select Ordinal, Probit in the Family/Link control;
e. select Down in the Measurement direction control;
f. click on OK.
If you wish, move the component by clicking on any variable and dragging it.
5. Create the variable for the formative indicator of SciAtt.
a. Select the Add Observed Variable tool, , and then click in the diagram directly above the
oval for SciAtt and about one-fourth of the way down from the top. After adding it, you
can click inside the rectangle to move the variable if you wish.
b. In the Contextual Toolbar, select sex with the Variable control.

358

example 36g MIMIC model (generalized response)

6. Create the path from the formative indicator to SciAtt.

a. Select the Add Path tool,

b. Click in the bottom of the sex rectangle (it will highlight when you hover over it), and drag
a path to the top of the SciAtt oval (it will highlight when you can release to connect the
path).
7. Clean up the direction of the error.
The error on SciAtt is likely to have been created below the oval for SciAtt. Choose the
Select tool, , and then click in the SciAtt oval. Click on one of the Error Rotation buttons,
, in the Contextual Toolbar until the error is where you want it.
8. Clean up the location of the path.
If you do not like where the path between sex and SciAtt has been connected to its variables,
, to click on the path, and then simply click on where it connects to a
use the Select tool,
rectangle or oval and drag the endpoint.
9. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_mimic

Reference
Greenacre, M. J. 2006. From simple to multiple correspondence analysis. In Multiple Correspondence Analysis and
Related Methods, ed. M. J. Greenacre and J. Blasius. Boca Raton, FL: Chapman & Hall.

Also see
[SEM] example 10 MIMIC model
[SEM] example 35g Ordered probit and ordered logit

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
example 37g Multinomial logistic regression

Description

Remarks and examples

Reference

Also see

Description
With the data below, we demonstrate multinomial logistic regression, also known as multinomial
logit, mlogit, and family multinomial, link logit:
. use http://www.stata-press.com/data/r13/gsem_sysdsn1
(Health insurance data)
. describe
Contains data from http://www.stata-press.com/data/r13/gsem_sysdsn1.dta
obs:
644
Health insurance data
vars:
12
28 Mar 2013 13:46
size:
11,592
(_dta has notes)

variable name
site
patid
insure
age
male
nonwhite
noinsur0
noinsur1
noinsur2
ppd0
ppd1
ppd2

storage
type
byte
float
byte
float
byte
byte
byte
byte
byte
byte
byte
byte

Sorted by:

display
format
%9.0g
%9.0g
%9.0g
%10.0g
%8.0g
%9.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g

value
label

insure

variable label
study site (1-3)
patient id
insurance type
NEMC (ISCNRD-IBIRTHD)/365.25
NEMC PATIENT MALE
race
no insurance at baseline
no insurance at year 1
no insurance at year 2
prepaid at baseline
prepaid at year 1
prepaid at year 2

patid

. notes
_dta:
1. Data on health insurance available to 644 psychologically depressed
subjects.
2. Data from Tarlov, A.R., et al., 1989, "The Medical Outcomes Study. An
application of methods for monitoring the results of medical care." _J.
of the American Medical Association_ 262, pp. 925-930.
3. insure: 1=indemnity, 2=prepaid, 3=uninsured.

See Structural models 6: Multinomial logistic regression in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Simple multinomial logistic regression model
Multinomial logistic regression model with constraints
Fitting the simple multinomial logistic model with the Builder
Fitting the multinomial logistic model with constraints with the Builder

359

360

example 37g Multinomial logistic regression

Simple multinomial logistic regression model

In a multinomial logistic regression model, there are multiple unordered outcomes. In our case, these
outcomes are recorded in variable insure. This variable records three different outcomesindemnity,
prepaid, and uninsuredrecorded as 1, 2, and 3. The model we wish to fit is
multinomial

1b.insure
logit
multinomial

1.nonwhite

2.insure
logit
multinomial

3.insure
logit

The response variables are 1.insure, 2.insure, and 3.insure, meaning insure = 1 (code
for indemnity), insure = 2 (code for prepaid), and insure = 3 (code for uninsured). We specified
that insure = 1 be treated as the mlogit base category by placing a b on 1.insure to produce
1b.insure in the variable box.
Notice that there are no paths into 1b.insure. We could just as well have diagrammed the model
with a path arrow from the explanatory variable into 1b.insure. It would have made no difference.
In one sense, omitting the path is more mathematically appropriate, because multinomial logistic
base levels are defined by having all coefficients constrained to be 0.
In another sense, drawing the path would be more appropriate because, even with insure = 1
as the base level, we are not assuming that outcome insure = 1 is unaffected by the explanatory
variables. The probabilities of the three possible outcomes must sum to 1, and so any predictor that
increases one probability of necessity causes the sum of the remaining probabilities to decrease. If a
predictor x has positive effects (coefficients) for both 2.insure and 3.insure, then increases in x
must cause the probability of 1.insure to fall.
The choice of base outcome specifies that the coefficients associated with the other outcomes are
to be measured relative to that base. In multinomial logistic regression, the coefficients are logs of
the probability of the category divided by the probability of the base category, a mouthful also known
as the log of the relative-risk ratio.

example 37g Multinomial logistic regression

361

We drew the diagram one way, but we could just as well have drawn it like this:
multinomial

1b.insure
logit
multinomial

1.nonwhite

2.insure
logit
multinomial

3.insure
logit

In fact, we could just as well have chosen to indicate the base category by omitting it entirely from
our diagram, like this:
multinomial

2.insure
logit

1.nonwhite
multinomial

3.insure
logit

Going along with that, we could type three different commands, each exactly corresponding to
one of the three diagrams:
. gsem (1b.insure) (2.insure 3.insure <- i.nonwhite), mlogit
. gsem (1b.insure 2.insure 3.insure <- i.nonwhite), mlogit
. gsem (2.insure 3.insure <- i.nonwhite), mlogit

In the command language, however, we would probably just type

. gsem (i.insure <- i.nonwhite), mlogit

See [SEM] intro 3 for a complete description of factor-variable notation. It makes no difference
which diagram we draw or which command we type.

362

example 37g Multinomial logistic regression

This model can be fit using the command syntax by typing

. gsem (i.insure <- i.nonwhite), mlogit
Iteration 0:
log likelihood = -556.59502
Iteration 1:
log likelihood = -551.78935
Iteration 2:
log likelihood = -551.78348
Iteration 3:
log likelihood = -551.78348
Generalized structural equation model
Log likelihood = -551.78348
Coef.
1.insure

Std. Err.

Number of obs

P>|z|

616

[95% Conf. Interval]

(base outcome)

2.insure <1.nonwhite
_cons

.6608212
-.1879149

.2157321
.0937644

3.06
-2.00

0.002
0.045

.2379942
-.3716896

1.083648
-.0041401

.3779586
-1.941934

.407589
.1782185

0.93
-10.90

0.354
0.000

-.4209011
-2.291236

1.176818
-1.592632

3.insure <1.nonwhite
_cons

Notes:
1. The above results say that nonwhites are more likely to have insure = 2 relative to 1 than
whites, and that nonwhites are more likely to have insure = 3 relative to 1 than whites, which
obviously implies that whites are more likely to have insure = 1.
2. For a three-outcome multinomial logistic regression model with the first outcome set to be the
base level, the probability of each outcome is

P r(y = 1) = 1/D
P r(y = 2) = exp(X2 2 )/D
P r(y = 2) = exp(X3 3 )/D
where D = 1 + exp(X2 2 ) + exp(X3 3 ).
3. For whitesthat is, for 1.nonwhite = 0we have X2 2 = 0.1879 and X3 3 = 1.9419.
Thus D = 1.9721, and the probabilities for each outcome are 0.5071, 0.4202, and 0.0727.
Those probabilities sum to 1. You can make the similar calculations for nonwhitesthat is, for
1.nonwhite = 1for yourself.

example 37g Multinomial logistic regression

363

Multinomial logistic regression model with constraints

Using the same data, we wish to fit the following model:

age
multinomial

1.male

1b.insure
logit

1.nonwhite

multinomial

2.insure
1b.site

logit
multinomial

2.site

3.insure
logit

3.site

In the above, insure = 2 and insure = 3 have paths pointing to them from different sets of
predictors. They share predictor 1.nonwhite, but insure = 2 also has paths from age and 1.male,
whereas insure = 3 also has paths from the site variables. When we fit this model, we will not
obtain estimates of the coefficients on age and 1.male in the equation for insure = 3. This is
equivalent to constraining the coefficients for age and 1.male to 0 in this equation. In other words,
we are placing a constraint that the relative risk of choosing insure = 3 rather than insure = 1 is
the same for males and females and is the same for all ages.

364

example 37g Multinomial logistic regression

This model can be fit using command syntax by typing

. gsem (2.insure <- i.nonwhite age i.male)
>
(3.insure <- i.nonwhite i.site), mlogit
Iteration
Iteration
Iteration
Iteration
Iteration

0:
1:
2:
3:
4:

log
log
log
log
log

likelihood
likelihood
likelihood
likelihood
likelihood

=
=
=
=
=

-555.85446
-541.20487
-540.85219
-540.85164
-540.85164

Generalized structural equation model

Log likelihood = -540.85164
Coef.
1.insure

Std. Err.

Number of obs

P>|z|

615

[95% Conf. Interval]

(base outcome)

2.insure <1.nonwhite
age
1.male
_cons

.7219663
-.0101291
.5037961
.1249932

.2184994
.0058972
.1912717
.2743262

3.30
-1.72
2.63
0.46

0.001
0.086
0.008
0.649

.2937153
-.0216874
.1289104
-.4126763

1.150217
.0014292
.8786818
.6626627

1.nonwhite

.0569646

.4200407

0.14

0.892

-.7663001

.8802293

site
2
3

-1.273576
.0434253

.4562854
.3470773

-2.79
0.13

0.005
0.900

-2.167879
-.6368337

-.3792728
.7236843

_cons

-1.558258

.2540157

-6.13

0.000

-2.056119

-1.060396

3.insure <-

We could have gotten identical results from Statas mlogit command for both this example and
the previous one. To fit the first example, we would have typed
. mlogit insure i.nonwhite

To obtain the results for this second example, we would have been required to type a bit more:
.
.
.
.
.

constraint 1 [Uninsure]age = 0
constraint 2 [Uninsure]1.male = 0
constraint 3 [Prepaid]2.site = 0
constraint 4 [Prepaid]3.site = 0
mlogit insure i.nonwhite age i.male i.site, constraints(1/4)

Having mlogit embedded in gsem, of course, also provides the advantage that we can combine
the mlogit model with measurement models, multilevel models, and more. See [SEM] example 41g
for a two-level multinomial logistic regression with random effects.

Fitting the simple multinomial logistic model with the Builder

Use the first diagram in Simple multinomial logistic regression model above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_sysdsn1

example 37g Multinomial logistic regression

365

2. Open a new builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the rectangles for each possible outcome of the multinomial endogenous variable.
Select the Add Observed Variables Set tool,
, and then click in the diagram about one-third
of the way in from the right and one-fourth of the way up from the bottom.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. check Make variables generalized responses;
c. select Multinomial, Logit in the Family/Link control;
d. select insure in the Variable control;
e. select Vertical in the Orientation control;
f. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
5. Create the independent variable.
a. Select the Add Observed Variable tool,
2.insure.

, and then click in the diagram to the left of

b. In the Contextual Toolbar, type 1.nonwhite in the Variable control and press Enter.
6. Create the paths from the independent variable to the rectangles for outcomes insure = 2 and
insure = 3.
a. Select the Add Path tool,

b. Click in the right side of the 1.nonwhite rectangle (it will highlight when you hover over
it), and drag a path to the left side of the 2.insure rectangle (it will highlight when you
can release to connect the path).
c. Continuing with the
tool, click in the right side of the 1.nonwhite rectangle and drag
a path to the left side of the 3.insure rectangle.
7. Clean up the location of the paths.
If you do not like where the paths have been connected to the rectangles, use the Select tool,
, to click on the path, and then simply click on where it connects to a rectangle and drag the
endpoint.
8. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_mlogit1

366

example 37g Multinomial logistic regression

Fitting the multinomial logistic model with constraints with the Builder
Use the diagram in Multinomial logistic regression model with constraints above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_sysdsn1

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the rectangles for each possible outcome of the multinomial endogenous variable.
Select the Add Observed Variables Set tool,
, and then click in the diagram about one-third
of the way in from the right and one-fourth of the way up from the bottom.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. check Make variables generalized responses;
c. select Multinomial, Logit in the Family/Link control;
d. select insure in the Variable control;
e. select Vertical in the Orientation control;
f. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
5. Create the independent variables.
Select the Add Observed Variables Set tool,
, and then click in the diagram about one-third
from the left and one-fourth from the bottom.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. uncheck Make variables generalized responses;
c. use the Variables control and select age;
d. type 1.male 1.nonwhite in the Variables control after age (typing 1.varname rather
than using the
button to create them as i.varname factor variables prevents rectangles
corresponding to the base categories for these binary variables from being created);
e. include the levels of the factor variable site by clicking on the
button next to the
Variables control. In the resulting dialog box, select the Factor variable radio button, select
Main effect in the Specification control, and select site in the Variables control for
Variable 1. Click on Add to varlist, and then click on OK;
f. select Vertical in the Orientation control;
g. click on OK.

example 37g Multinomial logistic regression

367

6. Create the paths from the independent variables to the rectangles for outcomes insure = 2 and
insure = 3.
a. Select the Add Path tool,

b. Click in the right side of the age rectangle (it will highlight when you hover over it), and
drag a path to the left side of the 2.insure rectangle (it will highlight when you can release
to connect the path).
c. Continuing with the
tool, create the following paths by clicking first in the right side of
the rectangle for the independent variable and dragging it to the left side of the rectangle
for the given outcome of the dependent variable:
1.male -> 2.insure
1.nonwhite -> 2.insure
1.nonwhite -> 3.insure
1b.site -> 3.insure
2.site -> 3.insure
3.site -> 3.insure
7. Clean up the location of the paths.
If you do not like where the paths have been connected to the rectangles, use the Select tool,
, to click on the path, and then simply click on where it connects to a rectangle and drag the
endpoint.
8. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem sem_mlogit2

Reference
Tarlov, A. R., J. E. Ware, Jr., S. Greenfield, E. C. Nelson, E. Perrin, and M. Zubkoff. 1989. The medical outcomes
study. An application of methods for monitoring the results of medical care. Journal of the American Medical
Association 262: 925930.

Also see
[SEM] example 35g Ordered probit and ordered logit
[SEM] example 41g Two-level multinomial logistic regression (multilevel)

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
example 38g Random-intercept and random-slope models (multilevel)

Description

Remarks and examples

Reference

Also see

Description
Below we discuss random-intercept and random-slope models in the context of multilevel models, and specifically, 2-level models, although we could just as well use higher-level models (see
[SEM] example 39g). Some people refer to these models as random-effects models and as mixed-effects
models.
To demonstrate random-intercept and random-slope models, we will use the following data:
. use http://www.stata-press.com/data/r13/gsem_nlsy
(NLSY 1968)
. describe
Contains data from http://www.stata-press.com/data/r13/gsem_nlsy.dta
obs:
2,763
NLSY 1968
vars:
21
29 Mar 2013 11:30
size:
93,942
(_dta has notes)

variable name
idcode
year
birth_yr
age
race
msp
nev_mar
grade
collgrad
not_smsa
c_city
south
ind_code
occ_code
union
wks_ue
ttl_exp
tenure
hours
wks_work
ln_wage
Sorted by:

storage
type
int
int
byte
byte
byte
byte
byte
byte
byte
byte
byte
byte
byte
byte
byte
byte
float
float
int
int
float

idcode

display
format
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%9.0g
%9.0g
%8.0g
%8.0g
%9.0g

value
label

racelbl

variable label
NLS ID
interview year
birth year
age in current year
race
1 if married, spouse present
1 if never married
current grade completed
1 if college graduate
1 if not SMSA
1 if central city
1 if south
industry of employment
occupation
1 if union
weeks unemployed last year
total work experience
job tenure, in years
usual hours worked
weeks worked last year
ln(wage/GNP deflator)

year

. notes
_dta:
1. Data from National Longitudinal Survey of Young Women 14-27 years of age
in 1968 (NLSY), Center for Human Resource Research, Ohio State
University, first released in 1989.
2. This data was subsetted for purposes of demonstration.

368

example 38g Random-intercept and random-slope models (multilevel)

369

These 2-level data are recorded in long form, that is, each observation corresponds to a year within
a subject and the full set of data is spread across repeated observations.
. list id year ln_wage union grade in 1/20, sepby(idcode)
idcode

year

ln_wage

union

grade

1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.

1
1
1
1
1
1
1
1
1
1
1
1

1970
1971
1972
1973
1975
1977
1978
1980
1983
1985
1987
1988

1.451214
1.02862
1.589977
1.780273
1.777012
1.778681
2.493976
2.551715
2.420261
2.614172
2.536374
2.462927

.
.
1
.
.
0
.
1
1
1
1
1

12
12
12
12
12
12
12
12
12
12
12
12

13.
14.
15.
16.
17.
18.
19.
20.

2
2
2
2
2
2
2
2

1971
1972
1973
1975
1977
1978
1980
1982

1.360348
1.206198
1.549883
1.832581
1.726721
1.68991
1.726964
1.808289

0
.
.
.
1
1
1
1

12
12
12
12
12
12
12
12

In the repeated observations for a subject, some variables vary (they are at the observation level, such
as ln wage) and other variables do not vary (they are at the subject level, such as grade).
When using gsem, multilevel data must be recorded in the long form except in one case. The
exception is latent-growth curve models, which can be fit in the long or wide form. In the wide
form, there is one physical observation for each subject and multiple variables within subject, such
as ln wage 1970, ln wage 1971, and so on. Researchers from a structural equation modeling
background think about latent-growth models in the wide form; see [SEM] example 18.
In all other cases, if your data are in the wide form, use Statas reshape command (see [D] reshape)
to convert the data to long form.
See Structural models 1: Linear regression and Multilevel mixed-effects models in [SEM] intro 5
for background.

Remarks and examples

Remarks are presented under the following headings:
Random-intercept model, single-equation formulation
Random-intercept model, within-and-between formulation
Random-slope model, single-equation formulation
Random-slope model, within-and-between formulation
Fitting the random-intercept model with the Builder
Fitting the random-slope model with the Builder

370

example 38g Random-intercept and random-slope models (multilevel)

Random-intercept model, single-equation formulation

There are two formulations of the random-intercept model, which we call the single-equation
formulation and the within-and-between formulation. Results from both formulations are identical
unless you have observations with missing values, in which case the within-and-between formulation
will sometimes use more of the data.
We will show you both formulations, but the single-equation formulation makes a good starting
point. The model we wish to fit is

idcode1

grade

1.union

ln_wage

Figure 1
We use factor-variable notation in the diagram above; see [SEM] example 37g.
We are using multilevel data. ln wage and union (union membership) vary at the observation
level, while grade (school completion) varies at the subject level. We have used shading to emphasize
that.
In this model, we are including a random intercept (a random effect) at the subject level. Doubleringed idcode is saying, I am a latent variable at the idcode levelmeaning I am constant within
identification codes and vary across identification codesand I correspond to a latent variable named
M1. The M1 part of the statement came from the subscript 1; the M part is fixed.
Double-ringed idcode indicates a latent variable constant within idcodea random effect. And
the fact that the path from the latent variable is pointing to a box and not to another path means that
the latent variable is used as a random intercept rather than a random slope. By the way, variable
idcode in the data contains each subjects identification number.
Using command syntax, we can fit this model by typing
. gsem (ln_wage <- i.union grade M1[idcode])
Fitting fixed-effects model:
Iteration 0:
log likelihood = -925.06629
Iteration 1:
log likelihood = -925.06629
Refining starting values:
Grid node 0:
log likelihood = -763.3769
Fitting full model:
Iteration 0:
log likelihood = -763.3769
Iteration 1:
log likelihood = -622.04625
Iteration 2:
log likelihood = -613.54948
Iteration 3:
log likelihood = -607.56242
Iteration 4:
log likelihood = -607.49246
Iteration 5:
log likelihood = -607.49233
Iteration 6:
log likelihood = -607.49233

(backed up)

example 38g Random-intercept and random-slope models (multilevel)

Generalized structural equation model
Log likelihood = -607.49233
( 1) [ln_wage]M1[idcode] = 1
Coef.

Number of obs

Std. Err.

371

1904

P>|z|

[95% Conf. Interval]

0.000
0.000

.1191998
.0634791

.2082818
.0901046

0.000

.5997848

.955041

ln_wage <1.union
grade

.1637408
.0767919

.0227254
.0067923

7.21
11.31

M1[idcode]

_cons

.7774129

.0906282

var(
M1[idcode])

.080247

.0073188

.0671113

.0959537

var(e.ln_w~e)

.078449

.0028627

.0730342

.0842653

(constrained)
8.58

Notes:
1. The ln wage <- M1[idcode] coefficient is constrained to be 1. Such constraints are automatically
supplied by gsem to identify the latent variable. Our model is
ln wage = + 3 M1[idcode] + cons

= + 1 MI[idcode] + 0.7774
Thus M1[idcode] is being used as a random intercept.
Remember that the square bracketed [idcode] means that M1 is constant within idcode and
varies only across idcode.
2. The variance of our random intercept is estimated to be 0.0802, which is greater than the estimated
error variance of 0.0784.
3. Although it is obvious in this case that the latent variable (random intercept) has sufficient
variance that it cannot be ignored, we can test whether the variance is large enough that we could
not ignore it. The test will be up against a boundary (variances cannot be less than 0), and so
the test will be conservative. To perform the test, we would type
.
.
.
.
.

gsem (ln_wage <- i.union grade)

estimates store without
gsem (ln_wage <- i.union grade M1[idcode])
estimates store with
lrtest without with

// fit model w/o random intercept

// fit model w/ random intercept
// order of arguments unimportant

372

example 38g Random-intercept and random-slope models (multilevel)

Random-intercept model, within-and-between formulation

The other way to write the random-intercept model is

grade

idcode1

1.union

ln_wage

Figure 2
Do not read grade pointing to double-ringed idcode as grade being a predictor of idcode.
That would make no sense. Double rings indicate a latent variable, and grade is a predictor of a
latent variable. In particular, the subscript 1 on idcode indicates that the latent variable is named
M1. Thus grade is a predictor of M1. The idcode inside the double rings says that M1 is constant
within idcode. Thus grade, which itself does not vary within idcode, is a predictor of M1, which
does not vary within idcode; said more elegantly, grade is a predictor of M1 at the subject level.
It is logically required that grade vary at the same or higher level as M1[idcode], and gsem will
check that requirement for you.
In this model, M1[idcode] contains both the random intercept and the grade effect. There is now
an equation for M1[idcode], with an error associated with it, and it will be the variance of the error
term that will reflect the variance of the random intercept.

example 38g Random-intercept and random-slope models (multilevel)

373

To fit this within-and-between formulation of our model, we type

. gsem (ln_wage <- i.union M1[idcode]) (M1[idcode] <- grade)
Fitting fixed-effects model:
Iteration 0:
log likelihood = -1091.655
Iteration 1:
log likelihood = -1091.655
Refining starting values:
Grid node 0:
log likelihood = -886.50236
Fitting full model:
Iteration 0:
log likelihood = -886.50236 (not concave)
Iteration 1:
log likelihood = -683.7337
Iteration 2:
log likelihood = -630.91327
Iteration 3:
log likelihood = -608.14536
Iteration 4:
log likelihood = -607.49376
Iteration 5:
log likelihood = -607.49233
Iteration 6:
log likelihood = -607.49233
Generalized structural equation model
Number of obs
Log likelihood = -607.49233
( 1) [ln_wage]M1[idcode] = 1
Coef.

1904

Std. Err.

P>|z|

[95% Conf. Interval]

.0227254

7.21

0.000

.1191998

.2082818

ln_wage <1.union

.1637408

M1[idcode]

_cons

.7774129

.0906282

8.58

0.000

.5997848

.955041

M1[idco~] <grade

.0767919

.0067923

11.31

0.000

.0634791

.0901046

var(
e.M1[idcode])

.080247

.0073188

.0671113

.0959537

var(e.ln_w~e)

.078449

.0028627

.0730342

.0842653

(constrained)

Notes:
1. Results are identical to what we previously obtained.
2. The within-and-between formulation is equivalent to the single-equation formulation if there are
no missing values in the data.
3. In this simple model, the two formulations are also equivalent even in the presence of missing
values.
4. If M1[idcode] were also being used to predict another endogenous variable, then missing values
in grade would only cause the equation for the other endogenous variable to have to omit those
observations in the within-and-between formulation.

Random-slope model, single-equation formulation

Let us now turn to random slopes. Because it is generallynot alwaysa good idea to include
random intercepts with random slopes, we are going to include both. In addition, we are going to
complicate our model by adding an interaction term between union membership and grade so we can
make another point that we will not explain until the next section.

374

example 38g Random-intercept and random-slope models (multilevel)

The model we wish to fit is

idcode2

idcode1

1.union

ln_wage

grade

1.union#c.grade

Figure 3: Demonstrated formulation 1

Ignore the interaction between union membership and grade. We could have omitted it from this
example, but we want to show later the strikingly different way that interaction can be written in the
within-and-between formulation.
In this model, we have double-ringed idcode, just as we did in a previous section introducing a
random intercept into the model. This time, however, we also have a second double-ringed idcode
adding a random slope. When a path from a latent variable points to another path, it is specifying a
random slope. When specifying a random slope for a variable, you still include the variable in the
usual, fixed-slope way, and then you add the extra component.
The result of the extra component will be to add an interaction into the model. 1.union will
affect ln wage, just as we have drawn, and the extra component will add 1.union M2[idcode].
That is, the effect of 1.union in our model will be
ln wage = + 3 1.union + 4 1.union M2[idcode] +

= + 3 1.union + 4 M2[idcode] 1.union +

= + (3 + 4 M2[idcode]) 1.union +
The fixed-plus-random, total slope is 3 + 4 M2[idcode]. Latent variable M2 has an arbitrary
scale, and therefore coefficient 4 cannot be identified. So constraining 4 = 1 results in the total
slope being 3 + M2[idcode], which is exactly how you would expect a fixed-plus-random slope to
look. In other words, drawing the path from idcode2 to the path with coefficient 3 indicates that
M2[idcode] is added to 3 .

example 38g Random-intercept and random-slope models (multilevel)

To fit this model in the command language, we type

. gsem (ln_wage <- i.union grade i.union#c.grade M1[idcode] 1.union#M2[idcode])
Fitting fixed-effects model:
Iteration 0:
log likelihood = -925.06629
Iteration 1:
log likelihood = -925.06629
Refining starting values:
Grid node 0:
log likelihood =
Fitting full model:
Iteration 0:
log likelihood =
Iteration 1:
log likelihood =
Iteration 2:
log likelihood =
Iteration 3:
log likelihood =
Iteration 4:
log likelihood =
Iteration 5:
log likelihood =
Iteration 6:
log likelihood =
Iteration 7:
log likelihood =
Iteration 8:
log likelihood =
Iteration 9:
log likelihood =
Iteration 10: log likelihood =
Generalized structural equation
Log likelihood = -581.29003
( 1)
( 2)

-869.92256
-869.92256
-727.15806
-711.74718
-684.33867
-665.94123
-610.14526
-589.89989
-582.24119
-581.298
-581.29004
-581.29003
model

(not
(not
(not
(not
(not

concave)
concave)
concave)
concave)
concave)

Number of obs

1904

[ln_wage]M1[idcode] = 1
[ln_wage]1.union#M2[idcode] = 1
Coef.

Std. Err.

P>|z|

[95% Conf. Interval]

ln_wage <1.union
grade

.1199049
.0757883

.1508189
.0081803

0.80
9.26

0.427
0.000

-.1756946
.0597552

.4155045
.0918215

union#
c.grade
1

.0019983

.0113534

0.18

0.860

-.020254

.0242506

7.25

0.000

.574443

1.000334

M1[idcode]

(constrained)

union#
M2[idcode]
0
1

0
1

(omitted)
(constrained)

_cons

.7873884

.1086476

.0927931

.0088245

.0770136

.1118056

.0823065

.018622

.052826

.1282392

cov(
M2[idcode],
M1[idcode])

-.0549821

.0116103

-.077738

-.0322263

var(e.ln_w~e)

.0720873

.0027135

.0669603

.0776068

var(
M1[idcode])
var(
M2[idcode])

-4.74

0.000

375

376

example 38g Random-intercept and random-slope models (multilevel)

Notes:
1. M1[idcode] is the random intercept. The coefficient on it is constrained to be 1, just as previously
and just as we would expect.
2. The coefficient on 1.union is the fixed part of the slope of union.
3. M2[idcode] is the random part of the slope of union.
The coefficient on 1.union#M2[idcode] is constrained to be 1, just as we would expect.
1.union#M2[idcode] is the way Stata writes 1.union M2[idcode].
4. There is an unexpected term in the output, 0.union#M2[idcode], shown with coefficient 0.
The first thing to remember about unexpected terms is that they are irrelevant if their coefficients
are 0. gsem reports the coefficient as being 0 (omitted), which is gsems way of saying, Here
is a line that I did not even include in the model. There are a lot of terms gsem could tell us
about that were not included in the model, so why did gsem feel obligated to tell us about this
term? The term has to do with how Stata tracks base levels of factor variables.
There is a settingset showbaselevels offthat will prevent lines like that from being
displayed. There is also a settingset showbaselevels allthat will show even more of
them! The default is set showbaselevels on.
5. We specified the interaction as 1.union#M2[idcode] rather than i.union#M2[idcode]. Even
so, using #.union#M2[idcode] or i.union#M2[idcode] makes no difference because union
takes on two values. If union took on three values, however, think about how we would diagram
the model. We would have two latent variables, and we would want 1.union#M2[idcode] and
2.union#M3[idcode]. If union took on three or more values, typing i.union#M2[idcode]
simply would not produce the desired result.

Random-slope model, within-and-between formulation

To properly demonstrate the within-and-between formulation, we need a more complicated model;
we made one in the previous section when we included 1.union#c.grade and then told you to
ignore it.
Pretend that we had omitted 1.union#c.grade from the random-slope, single-equation formulation
model. It would have looked like this:

idcode2

idcode1

1.union

grade

ln_wage

Figure 4: Simple formulation 1

example 38g Random-intercept and random-slope models (multilevel)

377

Lets call this model Simple formulation 1 as compared with the model that we fit in the previous
section, which we will call Demonstrated formulation 1. We could have fit Simple formulation 1 by
typing
. gsem (ln_wage <- i.union grade M1[idcode] 1.union#M2[idcode])

The corresponding within-and-between model, which we will call Simple formulation 2, would look
something like this:

grade

idcode2

1.union

idcode1

ln_wage

Figure 5: Simple formulation 2

We say that it would look something like the above because Simple formulations 1 and 2 are
not identical models. In particular, there is nothing in Simple formulation 2 that corresponds to the
correlation between M1[idcode] and M2[idcode] in Simple formulation 1. Other than that, the models
are identical. In Simple formulation 1, grade directly affects ln wage. In Simple formulation 2,
grade affects M1[idcode] and M1[idcode] affects ln wage. The results are the same either way.
Many researchers prefer Simple formulation 2 because it visually separates the within (subject) and
between (subject) levels.
But as we said, the two formulations are not identical. They would be identical if we added a
curved covariance path between 2 and idcode2 , but Stata will not allow us to do that.
On the other hand, if you wanted to fit the simple model without the correlation, you could do it
in either formulation. For Simple formulation 1, you would type
. gsem (ln_wage <- i.union grade M1[idcode] 1.union#M2[idcode]), ///
cov(M1[idcode]*M2[idcode]@0)

For Simple formulation 2, you would type

. gsem (ln_wage <- i.union M1[idcode] 1.union#M2[idcode]) (M1[idcode] <- grade)

Either way, results would be identical.

378

example 38g Random-intercept and random-slope models (multilevel)

In more complicated models, you can correlate the random intercepts and slopes even in a second
formulation. Consider Demonstrated formulation 2:

grade

idcode2

1.union

idcode1

ln_wage

Figure 6: Demonstrated formulation 2

In this model, we allow correlation between 3 and 2 which Stata does allowand that is the
same as allowing correlation between M1[idcode] and M2[idcode] in the formulation-1 model.
The general rule is that formulation 2 can be used to correlate the random intercepts and random
slopes when both M1 and M2 are (latent) exogenous or (latent) endogenous.
Demonstrated formulation 2 corresponds to Demonstrated formulation 1. In Demonstrated formulation 1, we included ln wage <- i.union#c.grade. Do you see how that term is implied by
the path diagram above? Variable grade affects M2[idcode], which in turn affects the slope of
1.union. Affecting the slope is the same as an interaction.
And now you know why we complicated Demonstrated formulation 1. We did that so that we can
show you Demonstrated formulation 2:
. gsem (ln_wage <- i.union M1[idcode] 1.union#M2[idcode])
>
(M1[idcode] M2[idcode] <- grade), cov(e.M1[idcode]*e.M2[idcode])
Fitting fixed-effects model:
Iteration 0:
log likelihood = -1091.655
Iteration 1:
log likelihood = -1091.655
Refining starting values:
Grid node 0:
log likelihood = -982.18261
Fitting full model:
Iteration 0:
log likelihood = -982.18261
Iteration 1:
log likelihood = -690.60714
Iteration 2:
log likelihood = -677.13819
Iteration 3:
log likelihood = -661.26658
Iteration 4:
log likelihood = -615.66049
Iteration 5:
log likelihood = -585.69402
Iteration 6:
log likelihood = -581.34158
Iteration 7:
log likelihood = -581.29005
Iteration 8:
log likelihood = -581.29003

(not concave)
(not concave)
(not concave)

example 38g Random-intercept and random-slope models (multilevel)

Generalized structural equation model
Log likelihood = -581.29003
( 1) [ln_wage]M1[idcode] = 1
( 2) [ln_wage]1.union#M2[idcode] = 1
Coef.

Number of obs

379

1904

Std. Err.

P>|z|

[95% Conf. Interval]

.1508188

0.80

0.427

-.1756945

.4155045

ln_wage <1.union

.119905

M1[idcode]

(constrained)

union#
M2[idcode]
0
1

0
1

(omitted)
(constrained)

_cons

.7873884

.1086476

7.25

0.000

.5744431

1.000334

M1[idco~] <grade

.0757883

.0081803

9.26

0.000

.0597552

.0918215

M2[idco~] <grade

.0019983

.0113534

0.18

0.860

-.020254

.0242506

var(
e.M1[idcode])
var(
e.M2[idcode])

.092793

.0088245

.0770136

.1118055

.0823065

.018622

.052826

.1282392

cov(
e.M2[idcode],
e.M1[idcode])

-.0549822

.0116103

-.077738

-.0322263

var(e.ln_w~e)

.0720873

.0027135

.0669603

.0776068

-4.74

0.000

Note:
1. Results of the above are identical to the results of Demonstrated formulation 1.

Fitting the random-intercept model with the Builder

Use the diagram in Random-intercept model, single-equation formulation above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_nlsy

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

380

example 38g Random-intercept and random-slope models (multilevel)

4. Create the endogenous variable.

a. Select the Add Observed Variable tool,
, and then click in the diagram about one-third
of the way in from the right and one-third of the way up from the bottom. After adding it,
you can click inside the rectangle and move the variable if you wish.
b. In the Contextual Toolbar, select ln wage with the Variable control.
5. Create the observed exogenous variables.
Select the Add Observed Variables Set tool,
, and then click in the diagram about one-third
of the way in from the left and one-third of the way up from the bottom.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. use the Variables control and select grade;
c. type 1.union in the Variables control after grade (typing 1.union rather than using the
button to create i.union prevents the rectangle corresponding to the base category for
this binary variable from being created);
d. select Vertical in the Orientation control;
e. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
6. Create the multilevel latent variable corresponding to the random intercept.
a. Select the Add Multilevel Latent Variable tool,
b. In the Contextual toolbar, click on the

, and click above the rectangle for grade.

button.

c. Select the nesting level and nesting variable by selecting 2 from the Nesting depth control
and selecting idcode > Observations in the next control.
d. Specify M1 as the Base name.
e. Click on OK.
7. Create the paths from the exogenous variables to ln wage.
a. Select the Add Path tool,

b. Click in the right side of the 1.union rectangle (it will highlight when you hover over it),
and drag a path to the left side of the ln wage rectangle (it will highlight when you can
release to connect the path).
c. Continuing with the
tool, draw paths from the right side of the grade rectangle to the
left side of the ln wage rectangle and from the right side of the idcode1 double oval to
the left side of the ln wage rectangle.
8. Clean up the location of the paths.
If you do not like where the paths have been connected to the rectangles or oval, use the Select
tool, , to click on the path, and then simply click on where it connects to a rectangle or oval
and drag the endpoint.

example 38g Random-intercept and random-slope models (multilevel)

381

9. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_rint

Fitting the random-slope model with the Builder

Use the diagram in Random-slope model, single-equation formulation above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_nlsy

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Increase the width of the observed variable rectangles to accommodate the length of the name
of the interaction term.
From the SEM Builder menu, select Settings > Variables > All Observed....
In the resulting dialog box, change the first size to 1 and click on OK.
5. Create the endogenous variable.
, and then click in the diagram about one-third
a. Select the Add Observed Variable tool,
of the way in from the right and one-third of the way up from the bottom. After adding it,
you can click inside the rectangle and move the variable if you wish.
b. In the Contextual Toolbar, select ln wage with the Variable control.
6. Create the observed exogenous variables.
, and then click in the diagram about one-third
Select the Add Observed Variables Set tool,
of the way in from the left and one-third of the way up from the bottom.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
button to
b. type 1.union in the Variables control (typing 1.union rather than using the
create i.union prevents the rectangle corresponding to the base category for this binary
variable from being created);
c. use the Variables control and select grade;
d. type 1.union#c.grade in the Variables control after grade;
e. select Vertical in the Orientation control;
f. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.

382

example 38g Random-intercept and random-slope models (multilevel)

6. Create the multilevel latent variable corresponding to the random intercept.

a. Select the Add Multilevel Latent Variable tool,
b. In the Contextual Toolbar, click on the

, and click above the rectangle for ln wage.

button.

c. Select the nesting level and nesting variable by selecting 2 from the Nesting depth control
and selecting idcode > Observations in the next control.
d. Specify M1 as the Base name.
e. Click on OK.
7. Create the paths from the exogenous variables to ln wage.
a. Select the Add Path tool,

b. Click in the right side of the 1.union rectangle (it will highlight when you hover over it),
and drag a path to the left side of the ln wage rectangle (it will highlight when you can
release to connect the path).
c. Continuing with the
tool, draw paths from the right sides of the grade and
1.union#c.grade rectangles to the left side of the ln wage rectangle and from the
bottom of the idcode1 double oval to the top of the ln wage rectangle.
8. Create the random slope.
a. Select the Add Multilevel Latent Variable tool,
to ln wage.
b. In the Contextual Toolbar, click on the

, and click above the path from 1.union

button.

c. Select the nesting level and nesting variable by selecting 2 from the Nesting depth control
and selecting idcode > Observations in the next control.
d. Specify M2 as the Base name.
e. Click on OK.
f. Select the Add Path tool,

g. Click in the bottom of the idcode2 double oval, and drag a path to the path between
1.union and ln wage.
9. Create the covariance between the random slope and random intercept.
a. Select the Add Covariance tool,

b. Click in the top-right quadrant of the idcode2 double oval, and drag a covariance to the
top left of the idcode1 double oval.
10. Clean up paths and covariance.
If you do not like where a path has been connected to its variables, use the Select tool,
, to
click on the path, and then simply click on where it connects to a rectangle and drag the endpoint.
Similarly, you can change where the covariance connects to the latent variables by clicking on
the covariance and dragging the endpoint. You can also change the bow of the covariance by
clicking on the covariance and dragging the control point that extends from one end of the
selected covariance.

example 38g Random-intercept and random-slope models (multilevel)

383

11. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_rslope

Reference
Center for Human Resource Research. 1989. National Longitudinal Survey of Labor Market Experience, Young Women
1424 years of age in 1968. Columbus, OH: Ohio State University Press.

Also see
[SEM] example 39g Three-level model (multilevel, generalized response)
[SEM] example 42g One- and two-level mediation models (multilevel)

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
example 39g Three-level model (multilevel, generalized response)

Description

Remarks and examples

References

Also see

Description
To demonstrate three-level models, we use the following data:
. use http://www.stata-press.com/data/r13/gsem_melanoma
(Skin cancer (melanoma) data)
. describe
Contains data from http://www.stata-press.com/data/r13/gsem_melanoma.dta
obs:
354
Skin cancer (melanoma) data
vars:
6
25 Mar 2013 15:28
size:
4,956
(_dta has notes)
storage
type

display
format

value
label

nation
region
county

byte
byte
int

%12.0g
%9.0g
%9.0g

nation

deaths
expected
uv

int
float
float

%9.0g
%9.0g
%9.0g

variable name

variable label
Nation ID
Region ID: EEC level-I areas
County ID: EEC level-II/level-III
areas
No. deaths during 1971-1980
No. expected deaths
UV dose, mean-centered

Sorted by: nation region county

. notes
_dta:
1. Smans, M., C. S. Muir, and P. Boyle. 1992. _Atlas of Cancer Mortality in
the European Economic Community_. Lyon, France: IARC Scientific
Publications
2. Data on 7 nations, 3-95 regions w/i nation, 1-13 counties w/i region.
3. Variable deaths is # of deaths among males due to malignant melanoma,
1971-1980.
4. Variable expected contains # of expected male deaths based on crude rates
for the combined counties.

Rabe-Hesketh and Skrondal (2012, exercise 13.7) describe data from the Atlas of Cancer Mortality
in the European Economic Community (EEC) (Smans, Mair, and Boyle 1993). The data were analyzed
in Langford, Bentham, and McDonald (1998) and record the number of deaths among males due to
malignant melanoma during 19711980.
Data are stored in the long form. Observations are counties within regions within nation. These
data and some of the models fit below are also demonstrated in [ME] menbreg.
See Structural models 4: Count models and Multilevel mixed-effects models in [SEM] intro 5 for
background.

384

example 39g Three-level model (multilevel, generalized response)

385

Remarks and examples

Remarks are presented under the following headings:
Three-level negative binomial model
Three-level Poisson model
Testing for overdispersion
Fitting the models with the Builder

Three-level negative binomial model

The model we wish to fit is

nation1

region2

nbinomial mean

uv

deaths
log

Deaths due to malignant melanoma at the county level are modeled as being affected by ultraviolet
exposure with random region and nation effects.

386

example 39g Three-level model (multilevel, generalized response)

To fit this model, we type

. gsem (deaths <- uv M1[nation]
Fitting fixed-effects model:
Iteration 0:
log likelihood =
Iteration 1:
log likelihood =
Iteration 2:
log likelihood =
Iteration 3:
log likelihood =
Iteration 4:
log likelihood =
Iteration 5:
log likelihood =
Refining starting values:
Grid node 0:
log likelihood =
Fitting full model:

M2[nation>region]), nbreg exposure(expected)

-1361.855
-1230.0211
-1211.049
-1202.5641
-1202.5329
-1202.5329
-1209.6951

Iteration 0:
log likelihood = -1209.6951
Iteration 1:
log likelihood = -1195.0761
Iteration 2:
log likelihood = -1189.7235
Iteration 3:
log likelihood =
-1167.58
Iteration 4:
log likelihood = -1145.4325
Iteration 5:
log likelihood = -1138.4471
Iteration 6:
log likelihood = -1088.3882
Iteration 7:
log likelihood = -1086.7992
Iteration 8:
log likelihood = -1086.4085
Iteration 9:
log likelihood = -1086.3903
Iteration 10: log likelihood = -1086.3902
Iteration 11: log likelihood = -1086.3902
Generalized structural equation model
Log likelihood = -1086.3902
( 1) [deaths]M1[nation] = 1
( 2) [deaths]M2[nation>region] = 1
Coef.

Std. Err.

(not
(not
(not
(not
(not

concave)
concave)
concave)
concave)
concave)

Number of obs

354

P>|z|

[95% Conf. Interval]

0.003

-.055883

-.0113035

deaths <uv

-.0335933

M1[nation]
M2[nation>
region]

(constrained)

(constrained)

_cons
ln(expected)

-.0790606
1

.1295931
(exposure)

-0.61

0.542

-.3330583

.1749372

deaths
/lnalpha

-4.182603

.3415036

-12.25

0.000

-4.851937

-3.513268

.1283614

.0678971

.0455187

.3619758

.0401818

.0104855

.0240938

.067012

var(
M1[nation])
var(
M2[nation>
region])

.0113725

-2.95

Notes:
1. This is a three-level model of counties nested within region nested within nation, so we specified
the latent variables as M1[nation] M2[nation>region]. Actually, we did the same thing in the
diagram when we used the SEM Builder to define the latent variables, but the nesting information
does not show in the double rings.

example 39g Three-level model (multilevel, generalized response)

387

2. We fit this model by using negative binomial regression, also known as a mean-dispersion
model. In the command, we typed nbreg, which is shorthand for family(nbinomial mean)
link(log).
3. A negative binomial distribution can be regarded as a Gamma mixture of Poisson random variables,
where said Gamma distribution has mean 1 and variance . The estimated ln() is 4.183,
which is small; is estimated as 0.0153. The reported test statistic of 12.25 is significant at
better than the 1% level for ln() = 0, but this is a test for = 1, not 0.
4. Zero does not mean lack of overdispersion, because we are including random effects that also
allow for extra dispersion. For a discussion on these issues, see [ME] menbreg.
5. Notice that we specified exposure(expected), where variable expected contains the expected
number of deaths based on crude rates.
The exposure() option is allowed with Poisson and negative binomial models. If we specify
exposure(varname), we are usually saying that each observations time at risk is recorded in
variable varname. When we omit the option, we are saying that each observation has the same
time at risk. Obviously, if one observation had twice the time at risk of another observation, but
was otherwise identical, we would expect twice the number of events in the first observation.
In this case, however, we are using exposure() differently. We have a variable called expected containing the expected number of deaths from crude rates, and we are claiming
exposure(expected). What this is doing is saying that in two otherwise identical observations,
if the number of expected deaths differed, we would expect the number of deaths due to melanoma
to differ, too, and by the same proportion. See [SEM] gsem family-and-link options.

Three-level Poisson model

The same model, fit with Poisson, is

nation1

region2

Poisson

uv

deaths
log

388

example 39g Three-level model (multilevel, generalized response)

To fit the model in the command language, we type

. gsem (deaths <- uv M1[nation] M2[nation>region]), poisson exposure(expected)
Fitting fixed-effects model:
Iteration 0:
log likelihood = -2136.5847
Iteration 1:
log likelihood = -1723.8955
Iteration 2:
log likelihood = -1723.7727
Iteration 3:
log likelihood = -1723.7727
Refining starting values:
Grid node 0:
log likelihood = -1166.6536
Fitting full model:
Iteration 0:
log likelihood = -1166.6536
Iteration 1:
log likelihood = -1152.2741
Iteration 2:
log likelihood = -1146.3094
Iteration 3:
log likelihood = -1119.8479
Iteration 4:
log likelihood = -1108.0129
Iteration 5:
log likelihood = -1098.8067
Iteration 6:
log likelihood = -1095.7563
Iteration 7:
log likelihood = -1095.3164
Iteration 8:
log likelihood =
-1095.31
Iteration 9:
log likelihood =
-1095.31
Generalized structural equation model
Log likelihood =
-1095.31
( 1) [deaths]M1[nation] = 1
( 2) [deaths]M2[nation>region] = 1
Coef.

Std. Err.

(not
(not
(not
(not
(not

concave)
concave)
concave)
concave)
concave)

Number of obs

P>|z|

354

[95% Conf. Interval]

deaths <uv

-.0282041

M1[nation]
M2[nation>
region]

(constrained)

(constrained)

_cons
ln(expected)

-.0639672
1

.1335515
(exposure)

.1371732

.0483483

var(
M1[nation])
var(
M2[nation>
region])

.0113998

-2.47

0.013

-.0505473

-.0058608

0.632

-.3257234

.197789

.0723303

.048802

.3855676

.0109079

.0310699

.0752353

-0.48

Testing for overdispersion

The negative binomial model allows for overdispersion, or in a multilevel framework, allows for
conditional overdispersion. The Poisson model has no overdispersion, or in a multilevel model, no
overdispersion beyond that predicted by the latent variables. We can test whether there is dispersion
beyond what Poisson would predict:

example 39g Three-level model (multilevel, generalized response)

389

. gsem (deaths <- uv M1[nation] M2[nation>region]), nbreg exposure(expected)

(output omitted )
estimates store nbreg
. gsem (deaths <- uv M1[nation] M2[nation>region]), poisson exposure(expected)
(output omitted )
. estimates store poisson
. lrtest nbreg poisson
Likelihood-ratio test
LR chi2(1) =
17.84
(Assumption: poisson nested in nbreg)
Prob > chi2 =
0.0000

We can reject at any reasonable level that the Poisson model adequately accounts for the dispersion
in these data. Be aware that this test is conservative, because we are testing whether a variance goes
to 0. lrtest usually issues a warning in such cases, but lrtest does not know that the relationship
between negative binomial regression and Poisson regression involves a variance going to 0.

Fitting the models with the Builder

Use the diagram in Three-level negative binomial model above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_melanoma

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the generalized response variable.

a. Select the Add Generalized Response Variable tool,

b. Click in the diagram about one-third of the way in from the right and one-fourth of the way
up from the bottom.
c. In the Contextual Toolbar, select Nbinomial mean, Log in the Family/Link control.
d. In the Contextual Toolbar, select deaths in the Variable control.
5. Create the observed exogenous variable.
a. Select the Add Observed Variable tool,
, and then click in the diagram about one-third
of the way in from the right and one-fourth of the way up from the bottom.
b. In the Contextual Toolbar, select uv with the Variable control.
6. Create the level-three latent variable.
a. Select the Add Multilevel Latent Variable tool,
about one-fourth of the way down from the top.
b. In the Contextual Toolbar, click on the

, and click above the rectangle for uv

button.

c. Select the nesting level and nesting variable by selecting 2 from the Nesting depth control
and selecting nation > Observations in the next line.

390

example 39g Three-level model (multilevel, generalized response)

d. Specify M1 as the Base name.

e. Click on OK.
7. Create the level-two latent variable.
a. Select the Add Multilevel Latent Variable tool,
and the double oval for nation1 .
b. In the Contextual Toolbar, click on the

, and click between the rectangle for uv

button.

c. Select the nesting level and nesting variable by selecting 3 from the Nesting depth control
and selecting nation > region > Observations in the next control.
d. Specify M2 as the Base name.
e. Click on OK.
8. Create the paths from the exogenous variables to deaths.
a. Select the Add Path tool,

b. Click in the right side of the uv rectangle (it will highlight when you hover over it), and
drag a path to the left side of the deaths rectangle (it will highlight when you can release
to connect the path).
c. Continuing with the
tool, draw paths from the right side of the double ovals for nation1
and region2 to the left side of the deaths rectangle.
9. Specify the level of exposure.
Use the Select tool,
, and double-click in the deaths rectangle. In the resulting dialog box,
select expected in the Exposure control, and click on OK.
10. Clean up the location of the paths.
If you do not like where the paths have been connected to the rectangles or oval, use the Select
tool, , to click on the path, and then simply click on where it connects to a rectangle or oval
and drag the endpoint.
11. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_3lev

References
Langford, I. H., G. Bentham, and A. McDonald. 1998. Multi-level modelling of geographically aggregated health
data: A case study on malignant melanoma mortality and UV exposure in the European community. Statistics in
Medicine 17: 4157.
Rabe-Hesketh, S., and A. Skrondal. 2012. Multilevel and Longitudinal Modeling Using Stata. 3rd ed. College Station,
TX: Stata Press.
Smans, M., C. S. Mair, and P. Boyle. 1993. Atlas of Cancer Mortality in the European Economic Community. Lyon,
France: IARC Scientific Publications.

example 39g Three-level model (multilevel, generalized response)

Also see
[SEM] example 38g Random-intercept and random-slope models (multilevel)
[SEM] example 34g Combined models (generalized responses)

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

391

Title
example 40g Crossed models (multilevel)

Description

Remarks and examples

Reference

Also see

Description
To illustrate crossed models, we use
. use http://www.stata-press.com/data/r13/gsem_fifeschool
(School data from Fife, Scotland)
. describe
Contains data from http://www.stata-press.com/data/r13/gsem_fifeschool.dta
obs:
3,435
School data from Fife, Scotland
vars:
5
25 Mar 2013 16:17
size:
24,045
(_dta has notes)

variable name

storage
type

display
format

pid
sid
attain
vrq

int
byte
byte
int

%9.0g
%9.0g
%9.0g
%9.0g

sex

byte

%9.0g

Sorted by:
. notes

pid

value
label

variable label
Primary school ID
Secondary school ID
Attainment score at age 16
Verbal-reasoning score from final
year of primary school
1: female; 0: male

sid

_dta:
1. Paterson, L. 1991. "Socio-economic status and education attainment: A
multidimensional and multilevel study" in _Evaluation and Research in
Education_ 5: 97-121.
2. Each observation is a different student. Each student attended a primary
school (pid) and a secondary school (sid).
3. pid and sid are crossed, not nested. All combinations are possible.

Rabe-Hesketh and Skrondal (2012, 443460) give an introduction to crossed-effects models and
provide other examples of crossed-effects models by using the school data from Fife, Scotland.
See Structural models 1: Linear regression and Multilevel mixed-effects models in [SEM] intro 5
for background.

Remarks and examples

Remarks are presented under the following headings:
The crossed model
Fitting the model with the Builder

392

example 40g Crossed models (multilevel)

393

The crossed model

In these data, 3,435 students attended 148 different primary schools and 19 different secondary
schools. The model we wish to fit is

sid2

1.sex

attain

pid1

We include latent (random) effects for primary and secondary school because we think that school
identities may have an effect.

394

example 40g Crossed models (multilevel)

The command syntax to fit this model is

. gsem (attain <- i.sex M1[pid] M2[sid])
note: crossed random-effects model specified; option intmethod(laplace)
implied
Fitting fixed-effects model:
Iteration 0:
log likelihood = -8701.1372
Iteration 1:
log likelihood = -8701.1372
Refining starting values:
Grid node 0:
log likelihood = -8587.4875
Fitting full model:
Iteration 0:
log likelihood = -8587.4875
Iteration 1:
log likelihood = -8573.5619
Iteration 2:
log likelihood = -8562.9985
Iteration 3:
log likelihood = -8561.7425
Iteration 4:
log likelihood = -8561.7354
Iteration 5:
log likelihood = -8561.7354
Generalized structural equation model
Number of obs
=
Log likelihood = -8561.7354
( 1) [attain]M2[sid] = 1
( 2) [attain]M1[pid] = 1
Coef.

3435

Std. Err.

P>|z|

[95% Conf. Interval]

.0982634

5.07

0.000

.3060224

.6912079

28.99

0.000

4.901876

5.612867

attain <1.sex

.4986152

M1[pid]
M2[sid]

1
1

_cons

5.257372

.1813784

var(M1[pid])
var(M2[sid])

1.104316
.3457086

.2022595
.1608679

.7712451
.1388744

1.581226
.8605937

var(e.attain)

8.053437

.1990023

7.672694

8.453074

(constrained)
(constrained)

Notes:
1. These data are not nested, but the diagram above would look the same even if they were. The
fact that primary and secondary schools are crossed and not nested is, however, specified when
we enter the model into the SEM Builder and is implicit in the command syntax.
2. We typed attain <- i.sex M1[pid] M2[sid]. We would have typed attain <- i.sex
M1[pid] M2[sid<pid] had secondary school been nested within primary school.
3. gsem produced the following note when it began estimation: crossed random effects detected;
option intmethod(laplace) assumed. gsem provides four integration methods. The default is
mvaghermite, which stands for mean-variance adaptive GaussHermite quadrature. The others
are mcaghermite (mode-curvature adaptive GaussHermite quadrature); ghermite (nonadaptive
GaussHermite quadrature); and laplace (Laplacian approximation).
In general, the adaptive methods mvaghermite and mcaghermite are considered superior in
terms of accuracy to the nonadaptive method ghermite, which is considered superior to the
approximation method laplace. They also take longer.
Fitting crossed models can be difficult. You may specify intmethod() with one of the superior
methods, but be aware, convergence may not be achieved in finite time.

example 40g Crossed models (multilevel)

395

Fitting the model with the Builder

Use the diagram in The crossed model above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_fifeschool

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the endogenous variable.

, and then click in the diagram about one-third
a. Select the Add Observed Variable tool,
of the way in from the right and halfway down from the top. After adding it, you can click
inside the rectangle and move the variable if you wish.
b. In the Contextual Toolbar, select attain with the Variable control.
5. Create the observed exogenous variable.
a. Select the Add Observed Variable tool,
, and then click in the diagram about one-third
of the way in from the left and halfway down from the top.
b. In the Contextual Toolbar, type 1.sex in the Variable control (typing 1.sex rather than
button to create i.sex prevents the rectangle corresponding to the base category
using the
for this binary variable from being created);
6. Create the pid-level latent variable.
a. Select the Add Multilevel Latent Variable tool,
b. In the Contextual Toolbar, click on the

, and click below the rectangle for attain.

button.

c. Select the nesting level and nesting variable by selecting 2 from the Nesting depth control
and selecting pid > Observations in the next control.
d. Specify M1 as the Base name.
e. Click on OK.
7. Create the sid-level latent variable.
a. Select the Add Multilevel Latent Variable tool,
b. In the Contextual Toolbar, click on the

, and click above the rectangle for attain.

button.

c. Select the nesting level and nesting variable by selecting 2 from the Nesting depth control
and selecting sid > Observations in the next control.
d. Specify M2 as the Base name.
e. Click on OK.

396

example 40g Crossed models (multilevel)

8. Create the paths from the exogenous variables to attain.

a. Select the Add Path tool,

b. Click in the right side of the 1.sex rectangle (it will highlight when you hover over it), and
drag a path to the left side of the attain rectangle (it will highlight when you can release
to connect the path).
c. Continuing with the
tool, draw paths from bottom of the sid2 double oval to the top
of the attain rectangle and from the top of the pid1 double oval to the bottom of the
attain rectangle.
9. Clean up the location of the paths.
If you do not like where the paths have been connected to the rectangles or oval, use the Select
tool, , to click on the path, and then simply click on where it connects to a rectangle or oval
and drag the endpoint.
10. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_cross

Reference
Rabe-Hesketh, S., and A. Skrondal. 2012. Multilevel and Longitudinal Modeling Using Stata. 3rd ed. College Station,
TX: Stata Press.

Also see
[SEM] example 38g Random-intercept and random-slope models (multilevel)

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
example 41g Two-level multinomial logistic regression (multilevel)

Description

Remarks and examples

References

Also see

Description
We demonstrate two-level multinomial logistic regression with random effects by using the following
data:
. use http://www.stata-press.com/data/r13/gsem_lineup
(Fictional suspect identification data)
. describe
Contains data from http://www.stata-press.com/data/r13/gsem_lineup.dta
obs:
6,535
Fictional suspect
identification data
vars:
6
29 Mar 2013 10:35
size:
156,840
(_dta has notes)

variable name

storage
type

suspect
suswhite
violent
location
witmale
chosen

float
float
float
float
float
float

display
format
%9.0g
%9.0g
%9.0g
%14.0g
%9.0g
%9.0g

value
label

loc
choice

variable label
suspect id
suspect is white
violent crime
lineup location
witness is male
indvidual identified in linup by
witness

Sorted by: suspect

. notes
_dta:
1. Fictional data inspired by Wright, D.B and Sparks, A.T., 1994, "Using
multilevel multinomial regression to analyse line-up data", _Multilevel
Modeling Newsletter_, Vol. 6, No. 1
2. Data contain repeated values of variable suspect. Each suspect is viewed
by multiple witnesses and each witness (1) declines to identify a
suspect, (2) chooses a foil, or (3) chooses the suspect.
. tabulate location
lineup
location

Freq.

Percent

Cum.

police_station
suite_1
suite_2

2,228
1,845
2,462

34.09
28.23
37.67

34.09
62.33
100.00

Total

6,535

100.00

397

398

example 41g Two-level multinomial logistic regression (multilevel)

. tabulate chosen
indvidual
identified
in linup by
witness

Freq.

Percent

Cum.

none
foil
suspect

2,811
1,369
2,355

43.01
20.95
36.04

43.01
63.96
100.00

Total

6,535

100.00

In what follows, we re-create results similar to those of Wright and Sparks (1994), but we use
fictional data. These data resemble the real data used by the authors in proportion of observations
having each level of the outcome variable chosen, and the data produce results similar to those
presented by the authors.
See Structural models 6: Multinomial logistic regression and Multilevel mixed-effects models in
[SEM] intro 5 for background.
For additional discussion of fitting multilevel multinomial logistic regression models, see Skrondal
and Rabe-Hesketh (2003).

Remarks and examples

Remarks are presented under the following headings:
Two-level multinomial logistic model with shared random effects
Two-level multinomial logistic model with separate but correlated random effects
Fitting the model with the Builder

Two-level multinomial logistic model with shared random effects

We wish to fit the following model:

1b.location

2.location

multinomial

2.chosen
1

3.location

logit

suspect1
1.suswhite

multinomial

3.chosen
1.witmale

1.violent

logit

example 41g Two-level multinomial logistic regression (multilevel)

399

This model concerns who is chosen in a police lineup. The response variables are 1.chosen,
2.chosen, and 3.chosen, meaning chosen = 1 (code for not chosen), chosen = 2 (code for foil
chosen), and chosen = 3 (code for suspect chosen). A foil is a stand-in who could not possibly be
guilty of the crime.
We say the response variables are 1.chosen, 2.chosen, and 3.chosen, but 1.chosen does not
even appear in the diagram. By its omission, we are specifying that chosen = 1 be treated as the
base mlogit category. There are other ways we could have drawn this; see [SEM] example 37g.
In these data, each suspect was viewed by multiple witnesses. In the model, we include a random
effect at the suspect level, and we constrain the effect to be equal for chosen values 2 and 3 (selecting
the foil or the suspect).

400

example 41g Two-level multinomial logistic regression (multilevel)

We can fit this model with command syntax by typing

. gsem (i.chosen <- i.location i.suswhite i.witmale i.violent M1[suspect]@1),
> mlogit
Fitting fixed-effects model:
Iteration 0:
log likelihood = -6914.9098
Iteration 1:
log likelihood = -6696.7136
Iteration 2:
log likelihood = -6694.0006
Iteration 3:
log likelihood = -6693.9974
Iteration 4:
log likelihood = -6693.9974
Refining starting values:
Grid node 0:
log likelihood = -6705.0919
Fitting full model:
Iteration 0:
log likelihood = -6705.0919
Iteration 1:
log likelihood = -6654.5724
Iteration 2:
log likelihood = -6653.5717
Iteration 3:
log likelihood = -6653.5671
Iteration 4:
log likelihood = -6653.5671
Generalized structural equation model
Log likelihood = -6653.5671
( 1) [2.chosen]M1[suspect] = 1
( 2) [3.chosen]M1[suspect] = 1
Coef.
1.chosen

Std. Err.

(not concave)

Number of obs

6535

P>|z|

[95% Conf. Interval]

(base outcome)

2.chosen <location
suite_1
suite_2

.3867066
.4915675

.1027161
.0980312

3.76
5.01

0.000
0.000

.1853868
.2994299

.5880264
.6837051

1.suswhite
1.witmale
1.violent

-.0275501
-.0001844
.0356477

.0751664
.0680803
.0773658

-0.37
-0.00
0.46

0.714
0.998
0.645

-.1748736
-.1336193
-.1159864

.1197734
.1332505
.1872819

M1[suspect]

_cons

-1.002334

.099323

-10.09

0.000

-1.197003

-.8076643

location
suite_1
suite_2

-.2832042
.1391796

.0936358
.0863473

-3.02
1.61

0.002
0.107

-.4667271
-.0300581

-.0996814
.3084172

1.suswhite
1.witmale
1.violent

-.2397561
.1419285
-1.376579

.0643075
.059316
.0885126

-3.73
2.39
-15.55

0.000
0.017
0.000

-.3657965
.0256712
-1.55006

-.1137158
.2581857
-1.203097

M1[suspect]

_cons

.1781047

.0833393

0.033

.0147627

.3414468

.2538014

.0427302

.1824673

.3530228

(constrained)

3.chosen <-

var(
M1[suspect])

(constrained)
2.14

example 41g Two-level multinomial logistic regression (multilevel)

401

Notes:
1. We show the interpretation of mlogit coefficients in [SEM] example 37g.
2. The estimated variance of the random effect is 0.2538, implying a standard deviation of 0.5038.
Thus a 1-standard-deviation change in the random effect amounts to a exp(0.5038) = 1.655
change in the relative-risk ratio. The effect is both practically significant and, from the output,
statistically significant.
3. This is not the model fit by Wright and Sparks (1994). Those authors did not constrain the
random effect to be the same for chosen equal to 2 and 3. They included separate but correlated
random effects, and then took that even a step further.

Two-level multinomial logistic model with separate but correlated random effects
The model we wish to fit is

1b.location

2.location

multinomial

2.chosen
3.location

logit

1.suswhite

multinomial

3.chosen
1.witmale

suspect1

suspect2

logit

1.violent

This is one of the models fit by Wright and Sparks (1994), although remember that we are using
fictional data.
We can fit this model with command syntax by typing
. gsem (2.chosen <- i.location i.suswhite i.witmale i.violent M1[suspect]) ///
>
(3.chosen <- i.location i.suswhite i.witmale i.violent M2[suspect]), ///
>
mlogit

We did not even mention the assumed covariance between the random effects because latent exogenous
variables are assumed to be correlated in the command language. Even so, we can specify the cov()
option if we wish, and we might do so for emphasis or because we are unsure whether the parameter
would be included.
. gsem (2.chosen <- i.location i.suswhite i.witmale i.violent M1[suspect])
>
(3.chosen <- i.location i.suswhite i.witmale i.violent M2[suspect]),
>
cov(M1[suspect]*M2[suspect]) mlogit
Fitting fixed-effects model:
Iteration 0:

log likelihood = -6914.9098

402

example 41g Two-level multinomial logistic regression (multilevel)

Iteration 1:
log likelihood = -6696.7136
Iteration 2:
log likelihood = -6694.0006
Iteration 3:
log likelihood = -6693.9974
Iteration 4:
log likelihood = -6693.9974
Refining starting values:
Grid node 0:
log likelihood = -6793.4228
Fitting full model:
Iteration 0:
log likelihood = -6793.4228
Iteration 1:
log likelihood = -6717.7507
Iteration 2:
log likelihood = -6684.6592
Iteration 3:
log likelihood = -6660.1404
Iteration 4:
log likelihood = -6652.1368
Iteration 5:
log likelihood = -6651.7841
Iteration 6:
log likelihood = -6651.7819
Iteration 7:
log likelihood = -6651.7819
Generalized structural equation model
Log likelihood = -6651.7819
( 1) [2.chosen]M1[suspect] = 1
( 2) [3.chosen]M2[suspect] = 1
Coef.
1.chosen

Std. Err.

(not concave)
(not concave)

Number of obs

6535

P>|z|

[95% Conf. Interval]

(base outcome)

2.chosen <location
suite_1
suite_2

.3881676
.48938

.1004754
.0960311

3.86
5.10

0.000
0.000

.1912394
.3011625

.5850958
.6775974

1.suswhite
1.witmale
1.violent

-.0260152
-.0007652
.0369381

.0749378
.0679187
.0771594

-0.35
-0.01
0.48

0.728
0.991
0.632

-.1728906
-.1338833
-.1142915

.1208602
.132353
.1881677

M1[suspect]

_cons

-1.000382

.0992546

-10.08

0.000

-1.194918

-.8058469

location
suite_1
suite_2

-.2904225
.1364246

.0968578
.089282

-3.00
1.53

0.003
0.127

-.4802604
-.0385649

-.1005847
.3114142

1.suswhite
1.witmale
1.violent

-.2437654
.139826
-1.388013

.0647275
.0596884
.0891863

-3.77
2.34
-15.56

0.000
0.019
0.000

-.370629
.0228389
-1.562815

-.1169018
.256813
-1.213212

M2[suspect]

_cons

.1750622

.0851614

0.040

.008149

.3419754

.2168248

.0549321

.131965

.3562533

.2978104

.0527634

.2104416

.421452

.2329749

.0438721

.1469872

.3189627

(constrained)

3.chosen <-

var(
M1[suspect])
var(
M2[suspect])
cov(
M2[suspect],
M1[suspect])

(constrained)
2.06

5.31

0.000

example 41g Two-level multinomial logistic regression (multilevel)

403

Notes:
1. The estimated variances of the two random effects are 0.2168 and 0.2978, which as explained
in the second note of above example, are both practically and statistically significant.

2. The covariance is estimated to be 0.2300. Therefore, 0.2300/ 0.2168 0.2978 = 0.9052 is the
estimated correlation.
3. Wright and Sparks (1994) were interested in whether the location of the lineup mattered. They
found that it did, and that foils were more likely to be chosen at lineups outside of the police
station (at the two specialist suites). They speculated the cause might be that the police at the
station strongly warn witnesses against misidentification, or possibly because the specialist suites
had better foils.

Fitting the model with the Builder

Use the first diagram in Two-level multinomial logistic model with shared random effects above
for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_lineup

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the independent variables.

Select the Add Observed Variables Set tool,
about one-third of the way in from the left.

, and then click near the bottom of the diagram

In the resulting dialog box,

a. select the Select variables radio button (it may already be selected);
button next to the
b. include the levels of the factor variable location by clicking on the
Variables control. In the resulting dialog box, select the Factor variable radio button, select
Main effect in the Specification control, and select location in the Variables control for
Variable 1. Click on Add to varlist, and then click on OK;
c. type 1.suswhite 1.witmale 1.violent in the Variables control after i.location (typing
1.varname rather than using the
button to create them as i.varname factor variables
prevents rectangles corresponding to the base categories for these binary variables from being
created);
d. select Vertical in the Orientation control;
e. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
5. Create the rectangles for the possible outcomes of the multinomial endogenous variable.
Select the Add Observed Variables Set tool,
, and then click in the diagram about one-third
of the way in from the right and one-fourth of the way up from the bottom.

404

example 41g Two-level multinomial logistic regression (multilevel)

In the resulting dialog box,

a. select the Select variables radio button (it may already be selected);
b. check Make variables generalized responses;
c. select Multinomial, Logit in the Family/Link control;
d. select chosen in the Variable control;
e. under Levels, remove 1b to prevent the rectangle corresponding to the base category from
being created;
f. select Vertical in the Orientation control;
g. select the Distances tab;
h. select .5 (inch) from the from the Distance between variables control;
i. click on OK.
6. Create the paths from the independent variables to the rectangles for outcomes chosen = 2 and
chosen = 3.
a. Select the Add Path tool,

b. Click in the right side of the 1b.location rectangle (it will highlight when you hover over
it), and drag a path to the left side of the 2.chosen rectangle (it will highlight when you
can release to connect the path).
tool, click in the right side of each independent variable and drag a
c. Continuing with the
path to both the 2.chosen and 3.chosen rectangles.
7. Create the suspect-level latent variable.
a. Select the Add Multilevel Latent Variable tool,
, and click near the right side of the
diagram, vertically centered between 2.chosen and 3.chosen.
b. In the Contextual Toolbar, click on the

button.

c. Select the nesting level and nesting variable by selecting 2 from the Nesting depth control
and selecting suspect > Observations in the next control.
d. Specify M1 as the Base name.
e. Click on OK.
8. Create the paths from the multilevel latent variable to the rectangles for outcomes chosen = 2
and chosen = 3.
a. Select the Add Path tool,

b. Click in the upper-left quadrant of the suspect1 double oval, and drag a path to the right
side of the 2.chosen rectangle.
c. Continuing with the
tool, click in the lower-left quadrant of the suspect1 double oval,
and drag a path to the right side of the 3.chosen rectangle.
9. Place constraints on path coefficients from the multilevel latent variable.
Use the Select tool,
, to select the path from the suspect1 double oval to the 2.chosen
rectangle. Type 1 in the
box in the Contextual Toolbar and press Enter. Repeat this process
to constrain the coefficient on the path from the suspect1 double oval to the 3.chosen rectangle
to 1.

example 41g Two-level multinomial logistic regression (multilevel)

405

10. Clean up the location of the paths.

If you do not like where the paths have been connected to the rectangles, use the Select tool,
, to click on the path, and then simply click on where it connects to a rectangle and drag the
endpoint.
11. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
12. If you wish to fit the model described in Two-level multinomial logistic model with separate but
correlated random effects, use the Select tool to select the path from the suspect1 double oval
to the 3.chosen rectangle in the diagram created above. Select Object > Delete from the SEM
Builder menu.
Using the Select tool, select the suspect1 double oval and move it up so that it is parallel with
the rectangle for 2.chosen.
13. Create the multilevel latent variable corresponding to the random effects of suspect in the
3.chosen equation.
a. Select the Add Multilevel Latent Variable tool,
diagram, next to the 3.chosen rectangle.
b. In the Contextual Toolbar, click on the

, and click near the right side of the

button.

c. Select the nesting level and nesting variable by selecting 2 from the Nesting depth control
and selecting suspect > Observations in the next control.
d. Specify M2 as the Base name.
e. Click on OK.
14. Draw a path from the newly added suspect-level latent variable to 3.chosen.
Select the Add Path tool, click in the left of the suspect2 double oval, and drag a path to the
right side of the 3.chosen rectangle.
15. Create the covariance between the random effects.
a. Select the Add Covariance tool,

b. Click in the bottom-right quadrant of the suspect1 double oval, and drag a covariance to
the top right of the suspect2 double oval.
16. Clean up paths and covariance.
If you do not like where a path has been connected to its variables, use the Select tool,
, to
click on the path, and then simply click on where it connects to a rectangle and drag the endpoint.
Similarly, you can change where the covariance connects to the latent variables by clicking on
the covariance and dragging the endpoint. You can also change the bow of the covariance by
clicking on the covariance and dragging the control point that extends from one end of the
selected covariance.
17. Estimate again.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram for the first model in the Builder by typing
. webgetsem gsem_mlmlogit1

406

example 41g Two-level multinomial logistic regression (multilevel)

You can open a completed diagram for the second model in the Builder by typing
. webgetsem gsem_mlmlogit2

References
Skrondal, A., and S. Rabe-Hesketh. 2003. Multilevel logistic regression for polytomous data and rankings. Psychometrika
68: 267287.
Wright, D. B., and A. T. Sparks. 1994. Using multilevel multinomial regression to analyse line-up data. Multilevel
Modelling Newsletter 6: 810.

Also see
[SEM] example 37g Multinomial logistic regression
[SEM] example 38g Random-intercept and random-slope models (multilevel)

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
example 42g One- and two-level mediation models (multilevel)

Description

Remarks and examples

References

Also see

Description
To demonstrate linear mediation models, we use the following data:
. use http://www.stata-press.com/data/r13/gsem_multmed
(Fictional job-performance data)
. summarize
Variable

Obs

Mean

branch
support
satis
perform

1500
1500
1500
1500

38
.0084667
.0212
5.005317

Std. Dev.
21.65593
.5058316
.6087235
.8949845

Min

Max

1
-1.6
-1.6
2.35022

75
1.8
2
8.084294

. notes
_dta:
1. Fictional data on job performance, job satisfaction, and perceived
support from managers for 1,500 sales employees of a large department
store in 75 locations.
2. Variable support is average of Likert-scale questions, each question
scored from -2 to 2.
3. Variable satis is average of Likert-scale questions, each question scored
from -2 to 2.
4. Variable perform is job performance measured on continuous scale.

See Structural models 1: Linear regression and Multilevel mixed-effects models in [SEM] intro 5
for background.

Remarks and examples

Remarks are presented under the following headings:
One-level model with sem
One-level model with gsem
Two-level model with gsem
Fitting the models with the Builder

One-level model with sem

You can fit single-level mediation models with sem or gsem. You will be better off using sem
because then you can use estat teffects afterward to compute indirect and total effects.

407

408

example 42g One- and two-level mediation models (multilevel)

The model we wish to fit is the simplest form of a mediation model, namely,

satis

support

perform

We are interested in the effect of managerial support on job performance, but we suspect a portion
of the effect might be mediated through job satisfaction. In traditional mediation analysis, the model
would be fit by a series of linear regression models as described in Baron and Kenny (1986). That
approach is sufficient because the errors are not correlated. The advantage of using structural equation
modeling is that you can fit a single model and estimate the indirect and total effects, and you can
embed the simple mediation model in a larger model and even use latent variables to measure any
piece of the mediation model.
To fit this model with the command syntax, we type
. sem (perform <- satis support) (satis <- support)
Endogenous variables
Observed: perform satis
Exogenous variables
Observed: support
Fitting target model:
Iteration 0:
log likelihood = -3779.9224
Iteration 1:
log likelihood = -3779.9224
Structural equation model
Estimation method = ml
Log likelihood
= -3779.9224

Coef.

OIM
Std. Err.

Number of obs

1500

P>|z|

[95% Conf. Interval]

Structural
perform <satis
support
_cons

.8984401
.6161077
4.981054

.0251903
.0303143
.0150589

35.67
20.32
330.77

0.000
0.000
0.000

.849068
.5566927
4.951539

.9478123
.6755227
5.010569

satis <support
_cons

.2288945
.019262

.0305047
.0154273

7.50
1.25

0.000
0.212

.1691064
-.0109749

.2886826
.0494989

var(e.perf~m)
var(e.satis)

.3397087
.3569007

.0124044
.0130322

.3162461
.3322507

.364912
.3833795

LR test of model vs. saturated: chi2(0)

0.00, Prob > chi2 =

example 42g One- and two-level mediation models (multilevel)

409

Notes:
1. The direct effect of managerial support on job performance is measured by perform <- support
and is estimated to be 0.6161. The effect is small albeit highly statistically significant. The
standard deviations of performance and support are 0.89 and 0.51. A one standard deviation
increase in support improves performance by a third of a standard deviation.
2. The direct effect of job satisfaction on job performance is measured by perform <- satis and
is estimated to be 0.8984. That also is a moderate effect, practically speaking, and is highly
statistically significant.
3. The effect of managerial support on job satisfaction is measured by satis <- support and is
practically small but statistically significant.
4. What is the total effect of managerial support on performance? It is the direct effect (0.6161)
plus the indirect effect of support on satisfaction on performance (0.2289 0.8984 = 0.2056),
meaning the total effect is 0.8217. It would be desirable to put a standard error on that, but thats
more work.
We can use estat teffects after estimation to obtain the total effect and its standard error:
. estat teffects
Direct effects

Coef.

OIM
Std. Err.

P>|z|

[95% Conf. Interval]

Structural
perform <satis
support

.8984401
.6161077

.0251903
.0303143

35.67
20.32

0.000
0.000

.849068
.5566927

.9478123
.6755227

satis <support

.2288945

.0305047

7.50

0.000

.1691064

.2886826

OIM
Std. Err.

P>|z|

[95% Conf. Interval]

7.34

0.000

P>|z|

[95% Conf. Interval]

Indirect effects

Coef.
Structural
perform <satis
support

0
.205648

(no path)
.0280066

satis <support

(no path)

.150756

.26054

Total effects

Coef.

OIM
Std. Err.

Structural
perform <satis
support

.8984401
.8217557

.0251903
.0404579

35.67
20.31

0.000
0.000

.849068
.7424597

.9478123
.9010516

satis <support

.2288945

.0305047

7.50

0.000

.1691064

.2886826

410

example 42g One- and two-level mediation models (multilevel)

One-level model with gsem

We can fit the same model with gsem. The command is the same except that we substitute gsem
for sem, and results are identical:
. gsem (perform <- satis support) (satis <- support)
Iteration 0:
Iteration 1:

log likelihood = -2674.3421

log likelihood = -2674.3421

Generalized structural equation model

Log likelihood = -2674.3421
Coef.

Number of obs

Std. Err.

1500

P>|z|

[95% Conf. Interval]

perform <satis
support
_cons

.8984401
.6161077
4.981054

.0251903
.0303143
.0150589

35.67
20.32
330.77

0.000
0.000
0.000

.849068
.5566927
4.951539

.9478123
.6755227
5.010569

satis <support
_cons

.2288945
.019262

.0305047
.0154273

7.50
1.25

0.000
0.212

.1691064
-.0109749

.2886826
.0494989

var(e.perf~m)
var(e.satis)

.3397087
.3569007

.0124044
.0130322

.3162461
.3322507

.364912
.3833795

After gsem, however, we cannot use estat teffects:

. estat teffects
invalid subcommand teffects
r(321);

We can, however, calculate the indirect and total effects for ourselves and obtain the standard error
by using nlcom. Referring back to note 4 of the previous section, the formula for the indirect effect
and total effects are
indirect effect = 1 4
total effect = 2 + 1 4
where

1 = path coefficient for perform <- satis

4 = path coefficient for satis

<- support

2 = path coefficient for perform <- support

It turns out that we can access the coefficients by typing

1 = b[perform:satis]
4 = b[satis:support]
2 = b[perform:support]
which is most easily revealed by typing
. gsem, coeflegend
(output omitted )

example 42g One- and two-level mediation models (multilevel)

411

Thus we can obtain the indirect effect by typing

. nlcom _b[perform:satis]*_b[satis:support]
_nl_1: _b[perform:satis]*_b[satis:support]
Coef.
_nl_1

.205648

Std. Err.

P>|z|

.0280066

7.34

0.000

[95% Conf. Interval]

.150756

.26054

and we can obtain the total effect by typing

. nlcom _b[perform:support]+_b[perform:satis]*_b[satis:support]
_nl_1: _b[perform:support]+_b[perform:satis]*_b[satis:support]
Coef.
_nl_1

.8217557

Std. Err.
.0404579

z
20.31

P>|z|

[95% Conf. Interval]

0.000

.7424597

.9010516

Two-level model with gsem

It may be easier to use sem rather than gsem for fitting single-level models, but if you want to fit
multilevel models, you must use gsem. A variation on the model we just fit is

branch1

satis

support

branch2

perform

In this model, we include a random intercept in each equation at the branch (individual store)
level. The model above is one of many variations on two-level mediation models; see Krull and
MacKinnon (2001) for an introduction to multilevel mediation models, and see Preacher, Zyphur, and
Zhang (2010) for a discussion of fitting these models with structural equation modeling.

412

example 42g One- and two-level mediation models (multilevel)

To fit this model with the command syntax, we type

. gsem (perform <- satis support M1[branch]) (satis <- support M2[branch]),
> cov(M1[branch]*M2[branch]@0)
Fitting fixed-effects model:
Iteration 0:
log likelihood = -2674.3421
Iteration 1:
log likelihood = -2674.3421
Refining starting values:
Grid node 0:
log likelihood = -2132.1613
Fitting full model:
Iteration 0:
log likelihood = -2132.1613 (not concave)
Iteration 1:
log likelihood = -1801.3155
Iteration 2:
log likelihood = -1769.6421
Iteration 3:
log likelihood = -1705.1282
Iteration 4:
log likelihood = -1703.746
Iteration 5:
log likelihood = -1703.7141
Iteration 6:
log likelihood = -1703.714
Generalized structural equation model
Number of obs
=
1500
Log likelihood = -1703.714
( 1) [perform]M1[branch] = 1
( 2) [satis]M2[branch] = 1
Coef.

Std. Err.

P>|z|

[95% Conf. Interval]

0.000
0.000

.5383313
.6490687

.6701968
.7472364

perform <satis
support

.604264
.6981525

M1[branch]

_cons

4.986596

.0489465

101.88

0.000

4.890663

5.082529

satis <support

.2692633

.0179649

14.99

0.000

.2340528

.3044739

M2[branch]

_cons

.0189202

.0570868

0.740

-.0929678

.1308083

.1695962

.0302866

.119511

.2406713

.2384738

.0399154

.1717781

.3310652

.201053
.1188436

.0075451
.0044523

.1867957
.1104299

.2163985
.1278983

var(
M1[branch])
var(
M2[branch])
var(e.perf~m)
var(e.satis)

.0336398
.0250432

17.96
27.88

(constrained)

(constrained)
0.33

Notes:
1. In One-level model with sem above, we measured the direct effects on job performance of job
satisfaction and managerial support as 0.8984 and 0.6161. Now the direct effects are 0.6043 and
0.6982.
2. We can calculate the indirect and total effects just as we did in the previous section, which
we will do below. We mentioned earlier that there are other variations of two-level mediation
models, and how you calculate total effects depends on the model chosen.

example 42g One- and two-level mediation models (multilevel)

413

In this case, the indirect effect is

. nlcom _b[perform:satis]*_b[satis:support]
_nl_1: _b[perform:satis]*_b[satis:support]
Coef.
_nl_1

.1627062

Std. Err.
.0141382

z
11.51

P>|z|

[95% Conf. Interval]

0.000

.1349958

.1904165

and the total effect is

. nlcom _b[perform:support]+_b[perform:satis]*_b[satis:support]
_nl_1: _b[perform:support]+_b[perform:satis]*_b[satis:support]
Coef.
_nl_1

.8608587

Std. Err.
.0257501

z
33.43

P>|z|

[95% Conf. Interval]

0.000

.8103894

.911328

Fitting the models with the Builder

Use the diagram in One-level model with sem above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_multmed

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Create a regression component for the perform outcome.
Select the Add Regression Component tool,

, and then click in the center of the diagram.

In the resulting dialog box,

a. select perform in the Dependent variable control;
b. select support with the Independent variables control;
c. select Left in the Independent variables direction control;
d. click on OK.
e. Use the Select tool,
, to select only the perform rectangle, and drag it to the right to
increase the distance between the rectangles. (You can hold the Shift key while dragging to
ensure that the movement is directly to the right.)
4. Create the mediating variable.
a. Select the Add Observed Variable tool,
from support to perform.

, and then click in the diagram above the path

b. In the Contextual Toolbar, select satis with the Variable control.

414

example 42g One- and two-level mediation models (multilevel)

5. Create the paths to and from the mediating variable.

a. Select the Add Path tool,

b. Click in the upper right of the support rectangle (it will highlight when you hover over
it), and drag a path to the lower left of the satis rectangle (it will highlight when you can
release to connect the path).
c. Continuing with the
tool, draw a path from the lower right of the satis rectangle to
the upper left of the perform rectangle.
6. Clean up the direction of the error term.
We want the error for each of the endogenous variables to be to the right of the rectangle. The
error for satis may have been created in another direction. If so,
a. choose the Select tool,

b. click in the satis rectangle;

c. click on one of the Error Rotation buttons,
is to the right of the rectangle.

, in the Contextual Toolbar until the error

7. Clean up the location of the paths.

If you do not like where the paths have been connected to the rectangles, use the Select tool,
, to click on the path, and then simply click on where it connects to a rectangle and drag the
endpoint.
8. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
SEM estimation options dialog box.
9. To fit the model in Two-level model with gsem, continue with the previous diagram, and put the
builder in gsem mode by clicking on the
button.
10. Create the multilevel latent variable corresponding to the random intercept for satis.
a. Select the Add Multilevel Latent Variable tool,
b. In the Contextual Toolbar, click on the

, and click above the rectangle for satis.

button.

c. Select the nesting level and nesting variable by selecting 2 from the Nesting depth control
and selecting branch > Observations in the next control.
d. Specify M1 as the Base name.
e. Click on OK.
11. Create the multilevel latent variable corresponding to the random intercept for perform.
a. Select the Add Multilevel Latent Variable tool,
and to the right of the branch1 double oval.
b. In the Contextual Toolbar, click on the

, and click above the rectangle for satis

button.

c. Select the nesting level and nesting variable by selecting 2 from the Nesting depth control
and selecting branch > Observations in the next control.
d. Specify M2 as the Base name.
e. Click on OK.

example 42g One- and two-level mediation models (multilevel)

415

12. Draw paths from the multilevel latent variables to their corresponding endogenous variables.
a. Select the Add Path tool,

b. Click in the bottom of the branch1 double oval, and drag a path to the top of the satis
rectangle.
c. Continuing with the
tool, click in the bottom of the branch2 double oval, and drag a
path to the top of the perform rectangle.
13. Estimate again.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram for the first model in the Builder by typing
. webgetsem sem_med

You can open a completed diagram for the second model in the Builder by typing
. webgetsem gsem_mlmed

References
Baron, R. M., and D. A. Kenny. 1986. The moderatormediator variable distinction in social psychological research:
Conceptual, strategic, and statistical considerations. Journal of Personality and Social Psychology 51: 11731182.
Krull, J. L., and D. P. MacKinnon. 2001. Multilevel modeling of individual and group level mediated effects.
Multivariate Behavorial Research 36: 249277.
Preacher, K. J., M. J. Zyphur, and Z. Zhang. 2010. A general multilevel SEM framework for assessing multilevel
mediation. Psychological Methods 15: 209233.

Also see
[SEM] example 38g Random-intercept and random-slope models (multilevel)

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
example 43g Tobit regression

Description

Remarks and examples

Also see

Description
Tobit regression is demonstrated using auto.dta:
. use http://www.stata-press.com/data/r13/auto
(1978 Automobile Data)
. generate wgt = weight/1000

See Structural models 1: Linear regression in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Fitting tobit regression models
Fitting the model with the Builder

Fitting tobit regression models

The first example in [R] tobit is
. tobit mpg wgt, ll(17)

This model corresponds to

Gaussian

wgt

mpg

identity

Censoring information does not appear in the path diagram by default. It can be added to the path
diagram by customizing the appearance of mpg in the Builder. The Builder reports the censoring
information for mpg in the Details pane.

416

example 43g Tobit regression

417

To fit this model with gsem, we type

. gsem mpg <- wgt, family(gaussian, lcensored(17))
Refining starting values:
Grid node 0:
log likelihood = -170.32555
Fitting full model:
Iteration 0:
log likelihood = -170.32555
Iteration 1:
log likelihood = -164.66209
Iteration 2:
log likelihood = -164.25471
Iteration 3:
log likelihood = -164.25438
Iteration 4:
log likelihood = -164.25438
Generalized structural equation model
Number of obs
Log likelihood = -164.25438
Coef.

Std. Err.

P>|z|

74

[95% Conf. Interval]

mpg <wgt
_cons
var(e.mpg)

-6.87305
41.49856

.700257
2.058384

14.78942

2.817609

-9.82
20.16

0.000
0.000

-8.245529
37.4642

-5.500572
45.53291

10.18085

21.48414

Notes:
1. Reported coefficients match those reported by tobit.
2. Reported standard errors (SEs) differ very slightly from those reported by tobit.
2
3. gsem reports the point estimate of e.mpg as 14.78942. This
is an estimate of , the error
variance. tobit reports an estimated as 3.845701. And 14.78942 = 3.8457.

Fitting the model with the Builder

Use the diagram in Fitting tobit regression models above for reference.
1. Open the dataset and create the rescaled weight variable.
In the Command window, type
. use http://www.stata-press.com/data/r13/auto
. gen wgt = weight/1000

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the independent variable.

a. Select the Add Observed Variable tool,

b. Click in the diagram about one-fourth of the way in from the left and half of the way up
from the bottom.
c. In the Contextual Toolbar, use the Variable control to select the variable wgt.
5. Create the tobit response.
a. Select the Add Generalized Response Variable tool,

418

example 43g Tobit regression

b. Click about one-third of the way in from the right side of the diagram, to the right of the
wgt rectangle.
c. In the Contextual Toolbar, select Gaussian, Identity in the Family/Link control (it may
already be selected).
d. In the Contextual Toolbar, use the Variable control to select the variable mpg.
e. In the Contextual Toolbar, click on the Properties button.
f. In the resulting Variable properties dialog box, click on the Censoring button in the Variable
tab.
g. In the resulting Censoring dialog box, select the Left censored radio button. In the resulting
Left censoring box below, select the Constant radio button (it may already be selected), and
type 17 in the Constant control.
h. Click on OK in the Censoring dialog box, and then click on OK in the Variable properties
dialog box. The Details pane will now show the Censoring information for mpg.
6. Create a path from the independent variable to the dependent variable.
a. Select the Add Path tool,

b. Click in the right side of the wgt rectangle (it will highlight when you hover over it), and
drag a path to the left side of the mpg rectangle (it will highlight when you can release to
connect the path).
7. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_tobit

Also see
[SEM] example 38g Random-intercept and random-slope models (multilevel)
[SEM] example 44g Interval regression
[SEM] example 45g Heckman selection model
[SEM] example 46g Endogenous treatment-effects model

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
example 44g Interval regression
Description

Remarks and examples

Also see

Description
Interval regression is demonstrated using intregxmpl.dta:
. use http://www.stata-press.com/data/r13/intregxmpl
(Wages of women)

See Structural models 1: Linear regression in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
Fitting interval regression models
Fitting the model with the Builder

Fitting interval regression models

The first example in [R] intreg is
. intreg wage1 wage2 age c.age#c.age nev_mar rural school tenure

This model corresponds to

age

c.age#c.age
Gaussian

nev_mar

wage1

identity

rural

school

tenure

Interval measure information does not appear in the path diagram by default. It can be added to the
path diagram by customizing the appearance of wage1 in the Builder. The Builder reports the interval
measure information for wage1 in the Details pane.
419

420

example 44g Interval regression

To fit this model with gsem, we type

. gsem wage1 <- age c.age#c.age nev_mar rural school tenure,
> family(gaussian, udepvar(wage2))
Refining starting values:
Grid node 0:

log likelihood = -856.59446

Fitting full model:

Iteration
Iteration
Iteration
Iteration

0:
1:
2:
3:

log
log
log
log

likelihood
likelihood
likelihood
likelihood

=
=
=
=

-856.59446
-856.33321
-856.33293
-856.33293

Generalized structural equation model

Log likelihood = -856.33293
Coef.

Number of obs

Std. Err.

P>|z|

488

[95% Conf. Interval]

wage1 <age

.7914438

.4433604

1.79

0.074

-.0775265

1.660414

c.age#c.age

-.0132624

.0073028

-1.82

0.069

-.0275757

.0010509

nev_mar
rural
school
tenure
_cons

-.2075022
-3.043044
1.334721
.8000664
-12.70238

.8119581
.7757324
.1357873
.1045077
6.367117

-0.26
-3.92
9.83
7.66
-1.99

0.798
0.000
0.000
0.000
0.046

-1.798911
-4.563452
1.068583
.5952351
-25.1817

1.383906
-1.522637
1.600859
1.004898
-.2230584

53.28454

3.693076

46.51635

61.0375

var(e.wage1)

Notes:
1. Just like intreg, gsem requires two dependent variables for fitting interval regression models.
The udepvar() suboption in family(gaussian) allows you to specify the dependent variable
containing the upper-limit values for the interval regression. Consequently, the dependent variable
participating in the path specification necessarily contains the lower-limit values.
2. Reported coefficients match those reported by intreg.
3. Reported standard errors (SEs) match those reported by intreg.
2
4. gsem reports the point estimate of e.wage1 as 53.28454. This
is an estimate of , the error
variance. intreg reports an estimated as 7.299626. And 53.28454 = 7.299626.

Fitting the model with the Builder

Use the diagram in Fitting interval regression models above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/intregxmpl

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

example 44g Interval regression

421

4. Create the independent variables.

Select the Add Observed Variables Set tool, , and then click in the diagram about one-fourth
of the way in from the left and one-fourth of the way up from the bottom.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. type the following variable names in the Variables control: age, c.age#c.age, nev mar,
rural, school, and tenure;
c. select Vertical in the Orientation control; and
d. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
5. Create the interval responses.
a. Select the Add Generalized Response Variable tool,

b. Click about one-third of the way in from the right side of the diagram, to the right of the
nev mar rectangle.
c. In the Contextual Toolbar, select Gaussian, Identity in the Family/Link control (it may
already be selected).
d. In the Contextual Toolbar, use the Variable control to select the variable wage1.
e. In the Contextual Toolbar, click on the Properties button.
f. In the resulting Variable properties dialog box, click on the Censoring button in the Variable
tab.
g. In the resulting Censoring dialog box, select the Interval measured, depvar is lower boundary
radio button. In the resulting Interval measured box below, use the Upper bound control to
select the variable wage2.
h. Click on OK in the Censoring dialog box, and then click on OK in the Variable properties
dialog box. The Details pane will now show that wage1 is the lower bound and wage2 is
the upper bound of our interval measure.
6. Create paths from the independent variables to the dependent variable.
a. Select the Add Path tool,

b. Click in the right side of the age rectangle (it will highlight when you hover over it), and
drag a path to the left side of the wage1 rectangle (it will highlight when you can release
to connect the path).
tool, create the following paths by clicking first in the right side of
c. Continuing with the
the rectangle for the independent variable and dragging it to the left side of the rectangle
for the dependent variable:
c.age#c.age -> wage1
nev_mar -> wage1
rural -> wage1
school -> wage1
tenure -> wage1

422

example 44g Interval regression

7. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_intreg

Also see
[SEM] example 38g Random-intercept and random-slope models (multilevel)
[SEM] example 43g Tobit regression
[SEM] example 45g Heckman selection model
[SEM] example 46g Endogenous treatment-effects model

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
example 45g Heckman selection model

Description

Remarks and examples

References

Also see

Description
To demonstrate selection models, we will use the following data:
. use http://www.stata-press.com/data/r13/gsem_womenwk
(Fictional data on women and work)
. summarize
Variable

Obs

Mean

age
educ
married
children
wage

2000
2000
2000
2000
1343

36.208
13.084
.6705
1.6445
23.69217

Std. Dev.
8.28656
3.045912
.4701492
1.398963
6.305374

Min

Max

20
10
0
0
5.88497

59
20
1
5
45.80979

. notes
_dta:
1.
2.
3.
4.
5.
6.

Fictional data on 2,000 women, 1,343 of whom work.

age ....... age in years
educ ...... years of schooling
married ... 1 if married spouse present
children .. # of children under 12 years
wage ...... hourly wage (missing if not working)

See Structural models 7: Dependencies between response variables and Structural models 8:
Unobserved inputs, outputs, or both in [SEM] intro 5 for background.

Remarks and examples

Remarks are presented under the following headings:
The Heckman selection model as an SEM
Fitting the Heckman selection model as an SEM
Transforming results and obtaining rho
Fitting the model with the Builder

The Heckman selection model as an SEM

We demonstrate below how gsem can be used to fit the Heckman selection model (Gronau 1974;
Lewis 1974; Heckman 1976) and produce results comparable to those of Statas dedicated heckman
command; see [R] heckman.
Our purpose is not to promote gsem as an alternative to heckman. We have two other purposes.
One is to show that gsem can be used to generalize the Heckman selection model to response
functions other than linear and, in addition or separately, to include multilevel effects when such
effects are present.
The other is to show how Heckman selection models can be included in more complicated SEMs.
423

424

example 45g Heckman selection model

For those unfamiliar with this model, it deals with a continuous outcome that is observed only
when another equation determines that the observation is selected, and the errors of the two equations
are allowed to be correlated. Subjects often choose to participate in an event or medical trial or even
the labor market, and thus the outcome of interest might be correlated with the decision to participate.
Heckman won a Nobel Prize for this work.
The model is sometimes cast in terms of female labor supply, but it obviously has broader
application. Nevertheless, we will consider a female labor-supply example.
Women are offered employment at a wage of w,

w i = X i + i
Not all women choose to work, and w is observed only for those women who do work. Women
choose to work if
Zi + i > 0
where

i N (0, 2 )
i N (0, 1)
corr(, ) =

More generally, we can think of this model as applying to any continuously measured outcome
wi , which is observed only if Zi + i > 0. The important feature of the model is that the errors i
of the selection equation and the errors i of the observed-data equation are allowed to be correlated.
The Heckman selection model can be recast as a two-equation SEM one linear regression (for the
continuous outcome) and the other censored regression (for selection)and with a latent variable Li
added to both equations. The latent variable is constrained to have variance 1 and to have coefficient 1
in the selection equation, leaving only the coefficient in the continuous-outcome equation to be
estimated. For identification, the variance from the censored regression will be constrained to be equal
to that of the linear regression. The results of doing this are the following:
1. Latent variable Li becomes the vehicle for carrying the correlation between the two equations.
2. All the parameters given above, namely, , , 2 , and , can be recovered from the SEM
estimates.

3. If we call the estimated parameters in the SEM formulation , , and 2 , and let
denote the coefficient on Li in the continuous-outcome equation, then
=
= /

p
2 + 1

2 = 2 + 2
q
= / ( 2 + 2 )( 2 + 1)
This parameterization places no restriction on the range or sign of . See Skrondal and
Rabe-Hesketh (2004, 107108).

example 45g Heckman selection model

425

Fitting the Heckman selection model as an SEM

We wish to fit the following Heckman selection model:

1
married

Gaussian

selected
1

children

identity

L
1

educ
wage
age
2

What makes this a Heckman selection model is

1. the inclusion of latent variable L in both the continuous-outcome (wage) equation and the
censored-outcome selection equation;
2. constraining the selected <- L path coefficient to be 1;
3. constraining the variance of L to be 1; and
4. constraining the error variances to be equal.
Before we can fit this model, we need to create new variables selected and notselected. selected
will equal 0 if the woman works (wage is not missing) and missing otherwise. notselected is the
complement of selected: it equals 0 if the woman does not work (wage is missing) and missing
otherwise. selected and notselected will be used as the dependent variables in the censored
regression, providing the equivalent of a scaled probit regression.
. gen selected = 0 if wage < .
(657 missing values generated)
. gen notselected = 0 if wage >= .
(1343 missing values generated)
. tabulate selected notselected, missing
selected

notselected
0

Total

0
.

0
657

1,343
0

1,343
657

Total

657

1,343

2,000

Old-time Stata users may be worried that because wage is missing in so many observations, namely,
all those corresponding to nonworking women, there must be something special we need to do so
that gsem uses all the data. There is nothing special we need to do. gsem counts missing values on
an equation-by-equation basis, so it will use all the data for the censored regression part of the model
while simultaneously using only the working-woman subsample for the continuous-outcome (wage)

426

example 45g Heckman selection model

part of the model. We use all the data for the censored regression because gsem understands the
meaning of missing values in the censored dependent variables so long as one of them is nonmissing.
To fit this model in command syntax, we type
. gsem (wage <- educ age L)
> (selected <- married children educ age L@1,
>
family(gaussian, udepvar(notselected))),
> var(L@1 e.wage@a e.selected@a)
Fitting fixed-effects model:
Iteration 0:
log likelihood = -5568.1366
Iteration 1:
log likelihood = -5211.0882 (not concave)
Iteration 2:
log likelihood = -5209.4228 (not concave)
Iteration 3:
log likelihood = -5209.2214
Iteration 4:
log likelihood = -5209.1638
Iteration 5:
log likelihood = -5208.9052 (not concave)
Iteration 6:
log likelihood = -5208.9044 (not concave)
Iteration 7:
log likelihood = -5208.9042 (not concave)
Iteration 8:
log likelihood = -5208.904
Iteration 9:
log likelihood = -5208.9038
Refining starting values:
Grid node 0:
log likelihood = -5259.1366
Fitting full model:
Iteration 0:
log likelihood = -5557.2489 (not concave)
Iteration 1:
log likelihood = -5439.0882 (not concave)
Iteration 2:
log likelihood = -5285.2854
Iteration 3:
log likelihood = -5229.0964
Iteration 4:
log likelihood = -5179.3914
Iteration 5:
log likelihood = -5178.3235
Iteration 6:
log likelihood = -5178.3046
Iteration 7:
log likelihood = -5178.3046
Generalized structural equation model
Number of obs
Log likelihood = -5178.3046
( 1) [selected]L = 1
( 2) [var(e.selected)]_cons - [var(e.wage)]_cons = 0
( 3) [var(L)]_cons = 1
Coef.

Std. Err.

P>|z|

2000

[95% Conf. Interval]

wage <educ
age
L
_cons

.9899509
.213128
5.923733
.4859256

selected <married
children
educ
age
L
_cons

.624276
.615211
.0781544
.0511984
1
-3.493224

var(L)

var(e.sele~d)
var(e.wage)

.9664716
.9664716

.0532552
.020602
.1846827
1.076867

18.59
10.34
32.08
0.45

0.000
0.000
0.000
0.652

.8855727
.1727488
5.561761
-1.624696

1.094329
.2535073
6.285704
2.596547

.1054324
5.92
.0652008
9.44
.0162868
4.80
.006637
7.71
(constrained)
.3730411
-9.36

0.000
0.000
0.000
0.000

.4176322
.4874197
.0462328
.0381901

.8309197
.7430023
.110076
.0642067

0.000

-4.224371

-2.762077

.5601427
.5601427

1.667552
1.667552

(constrained)
.2689702
.2689702

example 45g Heckman selection model

427

Notes:
1. Some of the estimated coefficients and parameters above will match those reported by the heckman
command and others will not. The above parameters are in the transformed structural equation
modeling metric. That metric can be transformed back to the Heckman metric and results will
match. The relationship to the Heckman metric is
=

= / 2 + 1

2 = 2 + 2
q
= / ( 2 + 2 )( 2 + 1)
2. refers to the coefficients on the continuous-outcome (wage) equation. We can read those
coefficients directly, without transformation except that we ignore the wage <- L path:
wage = 0.9900 educ + 0.2131 age + 0.4859

3. refers to the selection equation, and because = / 2 + 1, we must divide the reported

coefficients by the square root of 2 + 1. What has happened here is that the scaled probit has
2
variance + 1, and we are merely transforming back to the standard probit model, which has
variance 1. The results are
Pr(selected = 0) =

(0.4452 married + 0.4387 children + 0.0557 educ + 0.0365 age 2.4910)

p

4. To calculate , we first calculate 2 = 2 + 2 and then calculate = / 2 ( 2 + 1) :

2 = 0.9664 + 5.92372 = 36.0571
p
= 5.9237/ 2 (.9664 + 1) = 0.7035
5. These transformed results match the results that would have been reported had we typed
. heckman wage educ age, select(married children educ age)
(output omitted )

6. There is an easier way to obtain the transformed results than by hand, and the easier way provides
standard errors. That is the subject of the next section.

Transforming results and obtaining rho

We can use Statas nlcom command to perform the transformations we made by hand above, and
we can obtain standard errors.
Lets start by obtaining 2 and . To remind you, the formulas are

2 = 2 + 2
q
= / 2 ( 2 + 1)
We must describe these two formulas in a way that nlcom can understand. The Stata notation for

parameters 2 and fit by gsem is

2 :
:

b[var(e.wage): cons]
b[wage:L]

428

example 45g Heckman selection model

We cannot remember that notation; however, we can type gsem, coeflegend to be reminded. We
now have all that we need to obtain the estimates of 2 and . Because heckman reports rather
than 2 , we will tell nlcom to report the sqrt( 2 ):
. nlcom (sigma: sqrt(_b[var(e.wage):_cons] +_b[wage:L]^2))
>
(rho: _b[wage:L]/(sqrt((_b[var(e.wage):_cons]+1)*(_b[var(e.wage):_cons]
> + _b[wage:L]^2))))
sigma: sqrt(_b[var(e.wage):_cons] +_b[wage:L]^2)
rho: _b[wage:L]/(sqrt((_b[var(e.wage):_cons]+1)*(_b[var(e.wage):_cons
> ] + _b[wage:L]^2)))
Coef.
sigma
rho

6.004755
.7034874

Std. Err.
.1656476
.0511867

z
36.25
13.74

P>|z|

[95% Conf. Interval]

0.000
0.000

5.680091
.6031633

6.329418
.8038116

The output above nearly matches what heckman reports. heckman does not report the test statistics
and p-values for these two parameters. In addition, the confidence interval that heckman reports for
will differ slightly from the above and is better. heckman uses a method that will not allow to be
outside of 1 and 1, whereas nlcom is simply producing a confidence interval for the calculation we
requested and in absence of the knowledge that the calculation corresponds to a correlation coefficient.
The same applies to the confidence interval for , where the bounds are 0 and infinity.
To obtain the coefficients and standard errors for the selection equation, we type
. nlcom (married: _b[selected:married]/sqrt(_b[var(e.wage):_cons]+1))
>
(children: _b[selected:children]/sqrt(_b[var(e.wage):_cons]+1))
>
(educ: _b[selected:educ]/sqrt(_b[var(e.wage):_cons]+1))
>
(age: _b[selected:age]/sqrt(_b[var(e.wage):_cons]+1))
married: _b[selected:married]/sqrt(_b[var(e.wage):_cons]+1)
children: _b[selected:children]/sqrt(_b[var(e.wage):_cons]+1)
educ: _b[selected:educ]/sqrt(_b[var(e.wage):_cons]+1)
age: _b[selected:age]/sqrt(_b[var(e.wage):_cons]+1)
Coef.
married
children
educ
age

.4451771
.4387128
.0557326
.0365101

Std. Err.
.0673953
.0277788
.0107348
.0041534

z
6.61
15.79
5.19
8.79

P>|z|

[95% Conf. Interval]

0.000
0.000
0.000
0.000

.3130847
.3842673
.0346927
.0283696

.5772694
.4931583
.0767725
.0446505

The above output matches what heckman reports.

Fitting the model with the Builder

Use the diagram in Fitting the Heckman selection model as an SEM above for reference.
1. Open the dataset and create the selection variable.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_womenwk
. gen selected = 0 if wage < .
. gen notselected = 0 if wage >= .

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.

example 45g Heckman selection model

3. Put the Builder in gsem mode by clicking on the

429

button.

4. Create the independent variables.

Select the Add Observed Variables Set tool, , and then click in the diagram about one-fourth
of the way in from the left and one-fourth of the way up from the bottom.
In the resulting dialog box,
a. select the Select variables radio button (it may already be selected);
b. use the Variables control to select the variables married, children, educ, and age in this
order;
c. select Vertical in the Orientation control;
d. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
5. Create the generalized response for selection.
a. Select the Add Generalized Response Variable tool,

b. Click about one-third of the way in from the right side of the diagram, to the right of the
married rectangle.
c. In the Contextual Toolbar, select Gaussian, Identity in the Family/Link control (it may
already be selected).
d. In the Contextual Toolbar, select selected in the Variable control.
e. In the Contextual Toolbar, click on the Properties button.
f. In the resulting Variable properties dialog box, click on the Censoring button in the Variable
tab.
g. In the resulting Censoring dialog box, select the Interval measured, depvar is lower boundary
radio button. In the resulting Interval measured box below, use the Upper bound control to
select the variable notselected.
h. Click on OK in the Censoring dialog box, and then click on OK in the Variable properties
dialog box. The Details pane will now show selected as the lower bound and notselected
as the upper bound of our interval measure.
6. Create the endogenous wage variable.
a. Select the Add Observed Variable tool,
, and then click about one-third of the way in
from the right side of the diagram, to the right of the age rectangle.
b. In the Contextual Toolbar, select wage with the Variable control.
7. Create paths from the independent variables to the dependent variables.
a. Select the Add Path tool,

b. Click in the right side of the married rectangle (it will highlight when you hover over it),
and drag a path to the left side of the selected rectangle (it will highlight when you can
release to connect the path).
c. Continuing with the
tool, create the following paths by clicking first in the right side of
the rectangle for the independent variable and dragging it to the left side of the rectangle
for the dependent variable:

430

example 45g Heckman selection model

children -> selected

educ -> selected
age -> selected
educ -> wage
age -> wage
8. Clean up the direction of the error terms.
We want the error for selected to be above the rectangle and the error for wage to be below
the rectangle, but it is likely they have been created in other directions.
a. Choose the Select tool,

b. Click in the selected rectangle.

c. Click on one of the Error Rotation buttons,
is above the rectangle.

, in the Contextual Toolbar until the error

d. Click in the wage rectangle.

e. Click on one of the Error Rotation buttons,
is below the rectangle.

, in the Contextual Toolbar until the error

9. Create the latent variable.

a. Select the Add Latent Variable tool,
, and then click at the far right of the diagram and
vertically centered between the selected and wage variables.
b. In the Contextual Toolbar, type L in the Name control and press Enter.
10. Draw paths from the latent variable to each endogenous variable.
.

a. Select the Add Path tool,

b. Click in the upper left quadrant of the L oval, and drag a path to the right side of the
selected rectangle.
c. Continuing with the
tool, create another path by clicking first in the lower-left quadrant
of the L oval and dragging a path to the right side of the wage rectangle.
11. Place constraints on the variances and on the path from L to selected.
a. Choose the Select tool,

b. Click on the L oval. In the Contextual Toolbar, type 1 in the

box and press Enter.

c. Click on the error oval attached to the wage rectangle. In the Contextual Toolbar, type a in
the
box and press Enter.
d. Click on the error oval attached to the selected rectangle. In the Contextual Toolbar, type
a in the
box and press Enter.
e. Click on the path from L to selected. In the Contextual Toolbar, type 1 in the
and press Enter.

box

12. Clean up the location of the paths.

,
If you do not like where a path has been connected to its variables, use the Select tool,
to click on the path, and then simply click on where it connects to a rectangle and drag the
endpoint.

example 45g Heckman selection model

431

13. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_select

References
Gronau, R. 1974. Wage comparisons: A selectivity bias. Journal of Political Economy 82: 11191143.
Heckman, J. 1976. The common structure of statistical models of truncation, sample selection and limited dependent
variables and a simple estimator for such models. Annals of Economic and Social Measurement 5: 475492.
Lewis, H. G. 1974. Comments on selectivity biases in wage comparisons. Journal of Political Economy 82: 11451155.
Skrondal, A., and S. Rabe-Hesketh. 2004. Generalized Latent Variable Modeling: Multilevel, Longitudinal, and
Structural Equation Models. Boca Raton, FL: Chapman & Hall/CRC.

Also see
[SEM] example 34g Combined models (generalized responses)
[SEM] example 46g Endogenous treatment-effects model

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
example 46g Endogenous treatment-effects model

Description

Remarks and examples

References

Also see

Description
To illustrate the treatment-effects model, we use the following data:
. use http://www.stata-press.com/data/r13/gsem_union3
(NLSY 1972)
. describe
Contains data from http://www.stata-press.com/data/r13/gsem_union3.dta
obs:
1,693
NLSY 1972
vars:
24
29 Mar 2013 11:39
size:
79,571
(_dta has notes)

variable name
idcode
year
birth_yr
age
race
msp
nev_mar
grade
collgrad
not_smsa
c_city
south
ind_code
occ_code
union
wks_ue
ttl_exp
tenure
hours
wks_work
ln_wage
wage
black
smsa

storage
type
int
int
byte
byte
byte
byte
byte
byte
byte
byte
byte
byte
byte
byte
byte
byte
float
float
int
int
float
double
float
byte

display
format
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%8.0g
%9.0g
%9.0g
%8.0g
%8.0g
%9.0g
%10.0g
%9.0g
%8.0g

value
label

racelbl

variable label
NLS ID
interview year
birth year
age in current year
race
1 if married, spouse present
1 if never married
current grade completed
1 if college graduate
1 if not SMSA
1 if central city
1 if south
industry of employment
occupation
1 if union
weeks unemployed last year
total work experience
job tenure, in years
usual hours worked
weeks worked last year
ln(wage/GNP deflator)
real wage
race black
1 if SMSA

Sorted by: idcode

. notes
_dta:
1. Data from National Longitudinal Survey of Young Women 14-27 years of age
(NLSY) in 1968, Center for Human Resource Research, Ohio State
University, first released in 1989.
2. This data was subsetted for purposes of demonstration. Among other
things, it contains data only for 1972.

See Structural models 7: Dependencies between response variables and Structural models 8:
Unobserved inputs, outputs, or both in [SEM] intro 5 for background.
432

example 46g Endogenous treatment-effects model

433

Remarks and examples

Remarks are presented under the following headings:
Fitting the treatment-effects model
Fitting the model with the Builder

Fitting the treatment-effects model

We wish to fit the following model:

1.south
1
1.black

Gaussian

llunion
tenure

identity

age

grade

wage

1.smsa

1.union

We wish to estimate the treatment effect of being a union member. That is, we speculate that
union membership has an effect on wages, and we want to measure that effect. The problem would
be easy if we had data on the same workers from two different but nearly identical universes, one in
which the workers were not union members and another in which they were.
The model above is similar to the Heckman selection model we fit in [SEM] example 45g. The
differences are that the continuous variable (wage) is observed in all cases and that we have a path
from the treatment indicator (previously selection, now treatment) to the continuous variable. Just as
with the Heckman selection model, we allow for correlation by introducing a latent variable with
model identification constraints.
Before we can fit this model, we need to create new variables llunion and ulunion. llunion
will equal 0 if union is 1 and missing otherwise. ulunion is the complement of llunion: it equals
0 if union is 0 and missing otherwise. llunion and ulunion will be used as the dependent variables
in the treatment equation, providing the equivalent of a scaled probit regression.
. gen llunion = 0 if union == 1
(1433 missing values generated)
. gen ulunion = 0 if union == 0
(709 missing values generated)

434

example 46g Endogenous treatment-effects model

We can now fit this model with command syntax by typing

. gsem (wage <- age grade i.smsa i.black tenure 1.union L)
>
(llunion <- i.black tenure i.south L@1,
>
family(gaussian, udepvar(ulunion))),
>
var(L@1 e.wage@a e.llunion@a)
Fitting fixed-effects model:
Iteration 0:
log likelihood = -3308.0497
Iteration 1:
log likelihood = -3062.8995 (not concave)
Iteration 2:
log likelihood = -3062.2307
Iteration 3:
log likelihood = -3061.5084 (not concave)
Iteration 4:
log likelihood = -3061.5029
Iteration 5:
log likelihood = -3061.497 (not concave)
Iteration 6:
log likelihood = -3061.497 (backed up)
Refining starting values:
Grid node 0:
log likelihood = -3104.2036
Fitting full model:
Iteration 0:
log likelihood =
-3178.91 (not concave)
Iteration 1:
log likelihood = -3112.9538
Iteration 2:
log likelihood = -3059.5126
Iteration 3:
log likelihood = -3051.6997
Iteration 4:
log likelihood = -3051.576
Iteration 5:
log likelihood = -3051.575
Iteration 6:
log likelihood = -3051.575
Generalized structural equation model
Number of obs
Log likelihood = -3051.575
( 1) [llunion]L = 1
( 2) [var(e.llunion)]_cons - [var(e.wage)]_cons = 0
( 3) [var(L)]_cons = 1
Coef.

Std. Err.

P>|z|

1210

[95% Conf. Interval]

wage <age
grade
1.smsa
1.black
tenure
1.union
L
_cons

.1487409
.4205658
.9117044
-.7882472
.1524015
2.945816
-1.706795
-4.351572

.0193291
.0293577
.1249041
.1367077
.0369595
.2749549
.1288024
.5283952

7.70
14.33
7.30
-5.77
4.12
10.71
-13.25
-8.24

0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000

.1108566
.3630258
.6668969
-1.056189
.0799621
2.406915
-1.959243
-5.387207

.1866252
.4781057
1.156512
-.520305
.2248408
3.484718
-1.454347
-3.315936

.148057
4.53
.0357986
3.58
.136439
-6.26
(constrained)
.1407538
-9.25

0.000
0.000
0.000

.3802185
.0580384
-1.121683

.9605913
.1983664
-.5868518

0.000

-1.578548

-1.026804

.7725324
.7725324

1.753298
1.753298

llunion <1.black
tenure
1.south
L
_cons

.6704049
.1282024
-.8542673
1
-1.302676

var(L)

var(e.llun~n)
var(e.wage)

1.163821
1.163821

(constrained)
.2433321
.2433321

Notes:
1. The treatment effect is measured by the coefficient on the path treatment variable>continuous variable or, in our case, 1.union->wage. It is estimated to be 2.9458, which
is practically large and statistically significant.

example 46g Endogenous treatment-effects model

435

2. The interpretation formulas are the same as for the Heckman selection model in [SEM] example 45g,
namely,
=
p

= / 2 + 1

2 = 2 + 2
q
= / ( 2 + 2 )( 2 + 1)
To remind you, are the coefficients in the continuous-outcome equation, are the coefficients
in the treatment equation, 2 is the variance of the error in the continuous-outcome equation,
and is the correlation between the errors in the treatment and continuous-outcome equations.

3. In the output above, 2 (var(e.wage)) is 1.1638 and (the path coefficient on wage<-L)
is 1.7068. In [SEM] example 45g, we calculated by hand and then showed how to use the
software to obtain the value and its standard error. This time, we will go right to the software.
After obtaining symbolic names by typing gsem, coeflegend, we type the following to obtain
:
. nlcom (rho: _b[wage:L]/(sqrt(_b[var(e.wage):_cons] + 1)*sqrt(_b[var(e.wage):_
> cons] + _b[wage:L]^2)))
rho: _b[wage:L]/(sqrt(_b[var(e.wage):_cons] + 1)*sqrt(_b[var(e.wage):
> _cons] + _b[wage:L]^2))
Coef.
rho

-.574648

Std. Err.
.060969

z
-9.43

P>|z|
0.000

[95% Conf. Interval]

-.6941451

-.455151

4. We can obtain the untransformed treatment coefficients just as we did in [SEM] example 45g.
5. Just as with the Heckman selection model, with gsem, the treatment-effects model can be applied
to generalized outcomes and include multilevel effects. See Skrondal and Rabe-Hesketh (2004,
chap. 14.5) for an example with a Poisson response function.

Fitting the model with the Builder

Use the diagram in Fitting the treatment-effects model above for reference.
1. Open the dataset.
In the Command window, type
. use http://www.stata-press.com/data/r13/gsem_union3
. gen llunion = 0 if union == 1
. gen ulunion = 0 if union == 0

2. Open a new Builder diagram.

Select menu item Statistics > SEM (structural equation modeling) > Model building and
estimation.
3. Put the Builder in gsem mode by clicking on the

button.

4. Create the independent variables.

Select the Add Observed Variables Set tool, , and then click in the diagram about one-fourth
of the way in from the left and one-fourth of the way up from the bottom.

436

example 46g Endogenous treatment-effects model

In the resulting dialog box,

a. select the Select variables radio button (it may already be selected);
b. type 1.south, 1.black, tenure, age, grade, 1.smsa, and 1.union in the Variables
button to create i.varname prevents the
control (typing 1.varname rather than using the
rectangle corresponding to the base category for these binary variables from being created);
c. select Vertical in the Orientation control; and
d. click on OK.
If you wish, move the set of variables by clicking on any variable and dragging it.
5. Create the generalized response for llunion.
a. Select the Add Generalized Response Variable tool,

b. Click about one-third of the way in from the right side of the diagram, to the right of the
1.black rectangle.
c. In the Contextual Toolbar, select Gaussian, Identity in the Family/Link control (it may
already be selected).
d. In the Contextual Toolbar, select llunion in the Variable control.
e. In the Contextual Toolbar, click on the Properties button.
f. In the resulting Variable properties dialog box, click on the Censoring button in the Variable
tab.
g. In the resulting Censoring dialog box, select the Interval measured, depvar is lower boundary
radio button. In the resulting Interval measured box below, use the Upper bound control to
select the variable ulunion.
h. Click on OK in the Censoring dialog box, and then click on OK in the Variable properties
dialog box. The Details pane will now show llunion as the lower bound and ulunion as
the upper bound for our interval measure.
6. Create the endogenous wage variable.
, and then click about one-third of the way in
a. Select the Add Observed Variable tool,
from the right side of the diagram, to the right of the grade rectangle.
b. In the Contextual Toolbar, select wage with the Variable control.
7. Create paths from the independent variables to the dependent variables.
a. Select the Add Path tool,

b. Click in the right side of the 1.south rectangle (it will highlight when you hover over it),
and drag a path to the left side of the llunion rectangle (it will highlight when you can
release to connect the path).
c. Continuing with the
tool, create the following paths by clicking first in the right side of
the rectangle for the independent variable and dragging it to the left side of the rectangle
for the dependent variable:

example 46g Endogenous treatment-effects model

437

1.black -> llunion

tenure -> llunion
1.black -> wage
tenure -> wage
age -> wage
grade -> wage
1.smsa -> wage
1.union -> wage
8. Clean up the direction of the error terms.
We want the error for llunion to be above the rectangle and the error for wage to be below
the rectangle, but it is likely they have been created in other directions.
a. Choose the Select tool,

b. Click in the llunion rectangle.

c. Click on one of the Error Rotation buttons,
is above the rectangle.

, in the Contextual Toolbar until the error

d. Click in the wage rectangle.

e. Click on one of the Error Rotation buttons,
is below the rectangle.

, in the Contextual Toolbar until the error

9. Create the latent variable.

a. Select the Add Latent Variable tool,
, and then click at the far right of the diagram and
vertically centered between the llunion and wage variables.
b. In the Contextual Toolbar, type L in the Name control and press Enter.
10. Draw paths from the latent variable to each endogenous variable.
a. Select the Add Path tool,

b. Click in the upper-left quadrant of the L oval, and drag a path to the right side of the
llunion rectangle.
tool, create another path by clicking first in the lower-left quadrant
c. Continuing with the
of the L oval and dragging a path to the right side of the wage rectangle.
11. Place constraints on the variances and on the path from L to llunion.
a. Choose the Select tool,

b. Click on the L oval. In the Contextual Toolbar, type 1 in the

box and press Enter.

c. Click on the error oval attached to the wage rectangle. In the Contextual Toolbar, type a in
the
box and press Enter.
d. Click on the error oval attached to the llunion rectangle. In the Contextual Toolbar, type
a in the
box and press Enter.
e. Click on the path from L to llunion. In the Contextual Toolbar, type 1 in the
press Enter.

box and

438

example 46g Endogenous treatment-effects model

12. Clean up the location of the paths.

If you do not like where a path has been connected to its variables, use the Select tool,
, to
click on the path, and then simply click on where it connects to a variable and drag the endpoint.
13. Estimate.
Click on the Estimate button, , in the Standard Toolbar, and then click on OK in the resulting
GSEM estimation options dialog box.
You can open a completed diagram in the Builder by typing
. webgetsem gsem_treat

References
Center for Human Resource Research. 1989. National Longitudinal Survey of Labor Market Experience, Young Women
1424 years of age in 1968. Columbus, OH: Ohio State University Press.
Skrondal, A., and S. Rabe-Hesketh. 2004. Generalized Latent Variable Modeling: Multilevel, Longitudinal, and
Structural Equation Models. Boca Raton, FL: Chapman & Hall/CRC.

Also see
[SEM] example 34g Combined models (generalized responses)
[SEM] example 45g Heckman selection model

[SEM] gsem Generalized structural equation model estimation command

[SEM] intro 5 Tour of models

Title
gsem Generalized structural equation model estimation command
Syntax
Remarks and examples

Menu
Stored results

Description
Also see

Options

Syntax
gsem paths

if

 

in

 

, options

where paths are the paths of the model in command-language path notation; see [SEM] sem and gsem
path notation.
options

Description

model description options

fully define, along with paths, the model to be fit

estimation options

method used to obtain estimation results

reporting options

reporting of estimation results

syntax options

controlling interpretation of syntax

Factor variables and time-series operators are allowed.

bootstrap, by, jackknife, permute, and statsby are allowed; see [U] 11.1.10 Prefix commands.
Also see [SEM] gsem postestimation for features available after estimation.

Menu
Statistics

>

SEM (structural equation modeling)

>

Model building and estimation

Description
gsem fits generalized SEMs. When you use the Builder in gsem mode, you are using the gsem
command.

Options
model description options describe the model to be fit. The model to be fit is fully specified by
pathswhich appear immediately after gsemand the options covariance(), variance(), and
means(). See [SEM] gsem model description options and [SEM] sem and gsem path notation.
estimation options control how the estimation results are obtained. These options control how the
standard errors (VCE) are obtained and control technical issues such as choice of estimation method.
See [SEM] gsem estimation options.
reporting options control how the results of estimation are displayed. See [SEM] gsem reporting
options.
syntax options control how the syntax that you type is interpreted. See [SEM] sem and gsem syntax
options.
439

440

gsem Generalized structural equation model estimation command

Remarks and examples

gsem provides important features not provided by sem and correspondingly omits useful features
provided by sem. The differences in capabilities are the following:
1. gsem allows generalized linear response functions as well as the linear response functions
allowed by sem.
2. gsem allows for multilevel models, something sem does not.
3. gsem allows Statas factor-variable notation to be used in specifying models, something sem
does not.
4. gsems method ML is sometimes able to use more observations in the presence of missing
values than can sems method ML. Meanwhile, gsem does not provide the MLMV method
provided by sem for explicitly handling missing values.
5. gsem lacks the group-comparison capabilities that sem provides.
6. gsem lacks the ability to produce estimates adjusted for survey sampling that sem provides.
7. gsem cannot produce standardized coefficients.
8. gsem cannot use summary statistic datasets (SSDs); sem can.
gsem has nearly identical syntax to sem. Differences in syntax arise because of differences in
capabilities. The resulting differences in syntax are the following:
1. gsem adds new syntax to paths to handle latent variables associated with multilevel modeling.
2. gsem adds new options to handle the family and link of generalized linear responses.
3. gsem deletes options related to features it does not have, such as group-comparison capabilities
and SSDs.
4. gsem adds technical options for controlling features not provided by sem, such as numerical
integration (quadrature choices), number of integration points, and a number of options
dealing with starting values, which are a more difficult proposition in the generalized SEM
framework.
For a readable explanation of what gsem can do and how to use it, see the intro sections. You
might start with [SEM] intro 1.
For examples of gsem in action, see the example sections. You might start with [SEM] example 1.
For detailed syntax and descriptions, see the references below.
Remarks on three advanced topics are presented under the following headings:
Default normalization constraints
Default covariance assumptions
How to solve convergence problems

Default normalization constraints

gsem applies the same rules as sem to identify models; see [SEM] sem and see [SEM] intro 4.
Everything said there about latent variables applies to multilevel latent variables such as M1[school]
and M2[school>teacher].

gsem Generalized structural equation model estimation command

441

Default covariance assumptions

gsem assumes the same covariance structures as does sem; see [SEM] sem and see [SEM] intro 4.
gsem, however, treats covariances between observed exogenous variables as given. Actually, so does
sem unless you specify an override. The override cannot be specified with gsem.

How to solve convergence problems

See [SEM] intro 12.

Stored results
gsem stores the following in e():
Scalars
e(N)
e(N clust)
e(k)
e(k cat#)
e(k dv)
e(k eq)
e(k out#)
e(k rc)
e(k rs)
e(ll)
e(n quad)
e(rank)
e(ic)
e(rc)
e(converged)
Macros
e(cmd)
e(cmdline)
e(depvar)
e(title)
e(clustvar)
e(family#)
e(link#)
e(offset#)
e(intmethod)
e(vce)
e(vcetype)
e(opt)
e(which)
e(method)
e(ml method)
e(user)
e(technique)
e(datasignature)
e(datasignaturevars)
e(properties)
e(estat cmd)
e(predict)
e(covariates)
e(footnote)
e(marginsnotok)

number of observations
number of clusters
number of parameters
number of categories for the #th depvar, ordinal
number of dependent variables
number of equations in e(b)
number of outcomes for the #th depvar, mlogit
number of covariances
number of variances
log likelihood
number of integration points
rank of e(V)
number of iterations
return code
1 if target model converged, 0 otherwise
gsem
command as typed
names of dependent variables
title in estimation output
name of cluster variable
family for the #th depvar
link for the #th depvar
offset for the #th depvar
integration method
vcetype specified in vce()
title used to label Std. Err.
type of optimization
max or min; whether optimizer is to perform maximization or minimization
estimation method: ml
type of ml method
name of likelihood-evaluator program
maximization technique
the checksum
variables used in calculation of checksum
b V
program used to implement estat
program used to implement predict
list of covariates
program used to implement the footnote display
predictions not allowed by margins

442

gsem Generalized structural equation model estimation command

Matrices
e(b)
e(b pclass)
e(cat#)
e(out#)
e(Cns)
e(ilog)
e(gradient)
e(V)
e(V modelbased)
Functions
e(sample)

parameter vector
parameter class
categories for the #th depvar, ordinal
outcomes for the #th depvar, mlogit
constraints matrix
iteration log (up to 20 iterations)
gradient vector
covariance matrix of the estimators
model-based variance
marks estimation sample

Also see
[SEM] intro 1 Introduction
[SEM] sem and gsem path notation Command syntax for path diagrams
[SEM] gsem path notation extensions Command syntax for path diagrams
[SEM] gsem model description options Model description options
[SEM] gsem estimation options Options affecting estimation
[SEM] gsem reporting options Options affecting reporting of results
[SEM] sem and gsem syntax options Options affecting interpretation of syntax
[SEM] gsem postestimation Postestimation tools for gsem
[SEM] methods and formulas for gsem Methods and formulas

Title
gsem estimation options Options affecting estimation

Syntax

Description

Options

Remarks and examples

Also see

Syntax
gsem paths . . . , . . . estimation options
estimation options

Description

method(ml)
vce(vcetype)

method used to obtain the estimated parameters; only one

method available with gsem
vcetype may be oim, opg, robust, or cluster clustvar

from(matname)
startvalues(svmethod)


startgrid (gridspec)
noestimate
intmethod(intmethod)
intpoints(#)
adaptopts(adaptopts)

specify starting values

method for obtaining starting values
perform a grid search to improve starting values
do not fit the model; show starting values instead
integration method
set the number of integration (quadrature) points
options for adaptive quadrature

listwise

apply sems (not gsems) rules for omitting observations

with missing values

dnumerical

use numerical derivative techniques

maximize options

control maximization process for specified model; seldom used

intmethod

Description

mvaghermite
mcaghermite
ghermite
laplace

mean-variance adaptive GaussHermite quadrature; the default

mode-curvature adaptive GaussHermite quadrature
nonadaptive GaussHermite quadrature
Laplacian approximation

adaptopts
 
no log

Description

iterate(#)
tolerance(#)

whether to display the iteration log for each numerical integral

calculation
set the maximum number of iterations of the adaptive technique;
default is iterate(1001)
set tolerance for determining convergence of the adaptive
parameters; default is tolerance(1e-8)

443

444

gsem estimation options Options affecting estimation

Description
These options control how results are obtained, from starting values, to numerical integration (also
known as quadrature), to how variance estimates are obtained.

Options
method(ml) is the default and is the only method available with gsem. This option is included for
compatibility with sem, which provides several methods; see [SEM] sem option method( ).
vce(vcetype) specifies the technique used to obtain the variancecovariance matrix of the estimates.
See [SEM] sem option method( ).
from(matname), startvalues(svmethod), and startgrid[(gridspec)] specify overriding starting
values, specify how other starting values are to be calculated, and provide the ability to improve
the starting values. All of this is discussed in [SEM] intro 12. Below we provide a technical
description.
from(matname) allows you to specify starting values. See [SEM] intro 12 and see [SEM] sem and
gsem option from( ). We show the syntax as from(matname), but from() has another, less useful
syntax, too. An alternative to from() is init() used in the path specifications; see [SEM] sem
and gsem path notation.
startvalues() specifies how starting values are to be computed. Starting values specified in
from() override the computed starting values, and starting values specified via init() override
both.
startvalues(zero) specifies that starting values are to be set to 0.
startvalues(constantonly) builds on startvalues(zero) by fitting a constant-only model
for each response to obtain estimates of intercept and scale parameters, and it substitutes 1 for the
variances of latent variables.
startvalues(fixedonly) builds on startvalues(constantonly) by fitting a full fixedeffects model for each response variable to obtain estimates of coefficients along with intercept
and scale parameters, and it continues to use 1 for the variances of latent variables.
startvalues(ivloadings) builds on startvalues(fixedonly) by using instrumental-variable
methods with the generalized residuals from the fixed-effects models to compute starting values
for latent variable loadings, and still uses 1 for the variances of latent variables.
startvalues(iv) builds on startvalues(ivloadings) by using instrumental-variable methods
with generalized residuals to obtain variances of latent variables.
startgrid() performs a grid search on variance components of latent variables to improve starting
values. This is well discussed in [SEM] intro 12. No grid search is performed by default unless the
starting values are found to be not feasible, in which case gsem runs startgrid() to perform a
minimal search involving L3 likelihood evaluations, where L is the number of latent variables.
Sometimes this resolves the problem. Usually, however, there is no problem and startgrid() is
not run by default. There can be benefits from running startgrid() to get better starting values
even when starting values are feasible.
noestimate specifies that the model is not to be fit. Instead, starting values are to be shown (as
modified by the above options if modifications were made), and they are to be shown using the
coeflegend style of output. An important use of this option is before you have modified starting
values at all; you can type the following:

gsem estimation options Options affecting estimation

.
.
.
.

445

gsem ..., ... noestimate

matrix b = e(b)
... (modify elements of b) ...
gsem ..., ... from(b)

intmethod(intmethod), intpoints(#), and adaptopts(adaptopts) affect how integration for the

latent variables is numerically calculated.
intmethod(intmethod) specifies the method and defaults to intmethod(mvaghermite).
We recommend this method, although sometimes the more computationally intensive intmethod(mcaghermite) works better for multilevel models that are failing to converge. Sometimes it is useful to fall back on the less computationally intensive and less accurate intmethod(ghermite) and intmethod(laplace) to get the model to converge and then perhaps
use one of the other more accurate methods. All of this is explained in [SEM] intro 12. intmethod(laplace) is the default when fitting crossed models. Crossed models are often difficult.
intpoints(#) specifies the number of integration points to use and defaults to intpoints(7).
Increasing the number increases accuracy but also increases computational time. Computational
time is roughly proportional to the number specified. See [SEM] intro 12.
adaptopts(adaptopts) affects the adaptive part of adaptive quadrature (another term for numerical
integration) and thus is relevant only for intmethod(mvaghermite), intmethod(mcaghermite),
and intmethod(laplace).
adaptopts() defaults to adaptopts(nolog iterate(1001) tolerance(1e-8)).
 
no log specifies whether iteration logs are shown each time a numerical integral is calculated.
iterate() specifies the maximum number of iterations of the adaptive technique.
tolerance() specifies the tolerance for determining convergence of the adaptive parameters.
Convergence is declared when the relative change in the log likelihood is less than or equal to the
tolerance.
listwise applies sems rules rather than gsems rules for omitting observations with missing values.
By default, gsem is sometimes able to use observations containing missing values for fitting parts
of the model. sem, meanwhile, applies a listwise-deletion rule unless it is using method(mlmv).
Specifying listwise allows us at StataCorp to verify that gsem and sem produce the same results.
We find that reassuring. Actually, automated tests verify that results are the same before shipping.
For your information, sem and gsem use different numerical machinery for obtaining results,
and thus the near equality of results is a strong test that each is coded correctly. You may find
listwise useful if you are reproducing results from another package that uses listwise deletion.
dnumerical specifies that during optimization, the gradient vector and Hessian matrix be computed
using numerical techniques instead of analytical formulas. By default, gsem uses analytical formulas
for computing the gradient and Hessian for all integration methods except intmethod(laplace).
maximize options specify the standard and rarely specified options for controlling the maximization process; see [R] maximize. The
options for gsem are difficult,
 relevant

technique(algorithm spec), iterate(#), no log, trace, gradient, showstep, hessian,
tolerance(#), ltolerance(#), nrtolerance(#) and nonrtolerance.

Remarks and examples

For more information on vce(), see [SEM] intro 8 and [SEM] intro 9.
For more information on the other options, see [SEM] intro 12.

446

gsem estimation options Options affecting estimation

Also see
[SEM] gsem Generalized structural equation model estimation command
[SEM] intro 8 Robust and clustered standard errors
[SEM] intro 9 Standard errors, the full story
[SEM] intro 12 Convergence problems and how to solve them

Title
gsem family-and-link options Family-and-link options
Syntax

Description

Options

Remarks and examples

Also see

Syntax
gsem paths . . . , . . . family and link options
family and link options

Description

family(family)
link(link)

distribution family; default is family(gaussian)

link function; default varies per family

cloglog
gamma
logit
nbreg
mlogit
ocloglog
ologit
oprobit
poisson
probit
regress

synonym
synonym
synonym
synonym
synonym
synonym
synonym
synonym
synonym
synonym
synonym

exposure(varnamee )
offset(varnameo )

include ln(varnamee ) with coefficient constrained to 1

include varnameo with coefficient constrained to 1

family

Description

for
for
for
for
for
for
for
for
for
for
for

family(bernoulli) link(cloglog)
family(gamma) link(log)
family(bernoulli) link(logit)
family(nbreg mean) link(log)
family(multinomial) link(logit)
family(ordinal) link(cloglog)
family(ordinal) link(logit)
family(ordinal) link(probit)
family(poisson) link(log)
family(bernoulli) link(probit)
family(gaussian) link(identity)

gaussian , options
Gaussian (normal); the default
bernoulli
Bernoulli

binomial # | varname
binomial; default number of binomial trials is 1
gamma
gamma
multinomial

 multinomial
nbinomial mean | constant negative binomial; default dispersion is mean
ordinal
ordinal
poisson
Poisson
link

Description

identity
log
logit
probit
cloglog

identity
log
logit
probit
complementary log-log
447

448

gsem family-and-link options Family-and-link options

options

Description

ldepvar(varname)
udepvar(varname)
lcensored(varname|#)
rcensored(varname|#)

lower
upper
lower
upper

depvar for interval-response data

depvar for interval-response data
limit for left-censoring
limit for right-censoring

Only allowed with family(gaussian) with link(identity).

If you specify both family() and link(), not all combinations make sense. You may choose from
the following combinations:

Gaussian
Bernoulli
binomial
family multinomial
gamma

identity

log

logit

probit

cloglog

D
D
D

x
x

x
x

negative binomial
ordinal
Poisson
D denotes the default.

D
D

Description
gsem not only allows models of the form yi = xi + ui , it also allows

g{E(yi )} = xi
yi F
where you can choose F and g() from a menu. F is called the family, and g() is called the link.
One set of choices is the Gaussian distribution for F and the identity function for g(). In that case,
gsem reproduces linear regression. Other combinations of g() and F produce other popular models,
including logit (also known as logistic regression), probit, multinomial logit, Poisson regression, and
more.

Options
family(family) and link(linkname) specify F and g(). If neither is specified, linear regression is
assumed.
Two of the families allow optional arguments:


family(binomial # | varname ) specifies that data are in binomial form, that is, that the
response variable records the number of successes from a series of Bernoulli trials. The number
of trials is given either as a constant number or as a varname that allows the number of trials
to vary over observations, or it is not given at all. In the last case, the number of trials is thus
equivalent to specifying family(bernoulli).

gsem family-and-link options Family-and-link options

449



family(nbinomial mean | constant ) specifies a negative binomial model, a Poisson model
with overdispersion. Be aware, however, that even Poisson models can have overdispersion if latent
variables are included in the model. Lets use the term conditional overdispersion to refer to
dispersion above and beyond that implied by latent variables, if any.
That conditional overdispersion can take one of two forms. In mean overdispersion, the conditional
overdispersion is a linear function of the conditional (predicted) mean. Constant overdispersion
refers to the conditional overdispersion being, of course, constant.
If you do not specify mean or constant, then mean is assumed.
cloglog, gamma, logit, nbreg, mlogit, ocloglog, ologit, oprobit, poisson, probit, and
regress are shorthands for specifying popular models.
exposure(varnamee ) and offset(varnameo ) are used only with families poisson and nbreg,
that is, they concern count models.
exposure() specifies a variable that reflects the amount of exposureusually measured in time
unitsfor each observation over which the responses were counted. If one observation was exposed
for twice the time of another, and the observations were otherwise identical, one would expect twice
as many events to be counted. To assume that, ln(varnamee ) is entered into xi with coefficient
constrained to be 1.
offset() enters varnameo into xi with coefficient constrained to be 1. offset() is just another
way of specifying exposure() where the offset variable is the log of amount of exposure.
If neither exposure() nor offset() is specified, observations are assumed to have equal amounts
of exposure.
ldepvar(varname) and udepvar(varname) specify that each observation can be point data, interval
data, left-censored data, or right-censored data. The type of data for a given observation is
determined by the values in yi and varname. The following specifications are equivalent:
depvar1 <- ... , family(gauss, udepvar(depvar2 )
depvar2 <- ... , family(gauss, ldepvar(depvar1 )
Thus only one of ldepvar() or udepvar() is allowed. In either case, depvar1 and depvar2
should have the following form:
Type of data
point data
a = [ a, a ]
interval data
[ a, b ]
left-censored data
( , b ]
right-censored data
[ a, + )

var1

var2

a
a
.
a

a
b
b
.

lcensored(varname|#) and rcensored(varname|#) indicate the lower and upper limits for censoring, respectively. You may specify only one.
lcensored(arg) specifies that observations with yi arg are left-censored and the remaining
observations are not.
rcensored(arg) specifies that observations with yi arg are right-censored and the remaining
observations are not.
Neither lcensored() nor rcensored() may not be combined with ldepvar() or udepvar().

450

gsem family-and-link options Family-and-link options

Remarks and examples

In the command language, the family-and-link options may be specified at the end of a command
among the global options:
. gsem ..., ... family(...) link(...) ...
. gsem ..., ... poisson exposure(time) ...

Specified that way, the options apply to all the response variables. Alternatively, they may be specified
inside paths to affect single equations:
. gsem (y1
. gsem (y1
(y2
. gsem (y1
. gsem (y1
(y2

<<<<<<-

x1
x1
x2
x1
x1
x2

x2,
x2,
L,
x2,
x2,
L,

family(...) poisson(...)) (y2 <- x2 L) ...

family(...) link(...))
///
family(...) link(...)) ...
poisson exposure(time)) (y2 <- x2 L) ...
poisson exposure(time))
///
family(...) link(...)) ...

On a different topic, it is worth nothing that you can fit exponential-regression models with
family(gamma) link(log) if you constrain the log of the scale parameter to be 0 with gsems
constraints() option. For instance, you might type
. constraint 1 _b[y_logs:_cons] = 0
. gsem (y <- x1 x2, gamma), constraints(1)

The name b[y logs: cons] changes according to the name of the dependent variable. Had y instead
been named waitingtime, the parameter would have been named b[waitingtime logs: cons].
Rather than remembering that, remember instead that the best way to discover the names of parameters
is to type
. gsem (waitingtime <- x1 x2, gamma), noestimate

and then look at the output to discover the names. See [SEM] sem and gsem option constraints( ).

gsem family-and-link options Family-and-link options

For examples of generalized response variables, see the following:

[SEM] example 27g: Single-factor measurement model (generalized response)
[SEM] example 28g: One-parameter logistic IRT (Rasch) model
[SEM] example 29g: Two-parameter logistic IRT model
[SEM] example 30g: Two-level measurement model (multilevel, generalized response)
[SEM] example 31g: Two-factor measurement model (generalized response)
[SEM] example 32g: Full structural equation model (generalized response)
[SEM] example 33g: Logistic regression
[SEM] example 34g: Combined models (generalized responses)
[SEM] example 35g: Ordered probit and ordered logit
[SEM] example 36g: MIMIC model (generalized response)
[SEM] example 37g: Multinomial logistic regression
[SEM] example 39g: Three-level model (multilevel, generalized response)
[SEM] example 41g: Two-level multinomial logistic regression (multilevel)
[SEM] example 43g: Tobit regression
[SEM] example 44g: Interval regression
[SEM] example 45g: Heckman selection model
[SEM] example 46g: Endogenous treatment-effects model

Also see
[SEM] gsem Generalized structural equation model estimation command
[SEM] intro 2 Learning the language: Path diagrams and command language
[SEM] sem and gsem path notation Command syntax for path diagrams
[SEM] gsem path notation extensions Command syntax for path diagrams

451

Title
gsem model description options Model description options

Syntax

Description

Options

Remarks and examples

Also see

Syntax
gsem paths . . . , . . . model description options
model description options

Description

family(), link(), . . .

see [SEM] gsem family-and-link options

covariance()

variance()

means()

path notation for treatment of covariances;

see [SEM] sem and gsem path notation
path notation for treatment of variances;
see [SEM] sem and gsem path notation
path notation for treatment of means;
see [SEM] sem and gsem path notation

covstructure()

alternative method to place restrictions on covariances;

see [SEM] sem and gsem option covstructure( )

collinear
noconstant
noasis
fvstandard
noanchor
forcenoanchor

keep collinear variables

do not fit intercepts
omit perfect predictors from Bernoulli models
interpret factor variables according to Stata standard
do not apply default anchoring
programmers option

reliability()

reliability of measurement variables;

see [SEM] sem and gsem option reliability( )

constraints()
from()

specify constraints; see [SEM] sem and gsem option constraints( )

specify starting values; see [SEM] sem and gsem option from( )

Option may be specified more than once.

Description
paths and the options above describe the model to be fit by gsem.

Options
family() and link() specify the distribution and link function, such as family(poisson)
link(log), for generalized linear responses. There are lots of synonyms, so you can specify, for example, just poisson. In addition, there are exposure() and offset() options. See
[SEM] gsem family-and-link options.
452

gsem model description options Model description options

453

covariance(), variance(), and means() fully describe the model to be fit. See [SEM] sem and
gsem path notation.
covstructure() provides a convenient way to constrain covariances in your model. Alternatively
or in combination, you can place constraints by using the standard path notation. See [SEM] sem
and gsem option covstructure( ).
collinear; see [R] estimation options.
noconstant specifies that all intercepts be constrained to 0. See [SEM] sem and gsem path notation.
This option is seldom specified.
noasis specifies that perfect-predictor variables be omitted from all family Bernoulli models. By
default, gsem does not omit the variable, so one can specify tricky models where an equation
contains perfect predictors that are still identified through other portions of the model.
fvstandard specifies that factor-variable notation be interpreted according to the Stata standard.
gsem interprets factor variables slightly differently than do other Stata commands and, given
how factor-variable notation is used in command mode, this usually makes no difference. This is
explained in [SEM] intro 3.
To be technical, when fvstandard is specified, all factor variables automatically are assigned a
base level among the specified or implied levels, and implied yet unspecified elements of interaction
terms will be left in the model specification.
noanchor specifies that gsem not check for lack of identification or fill in anchors where needed.
gsem is instead to issue an error message if anchors would be needed. Specify this option when
you believe you have specified the necessary normalization constraints and want to hear about it
if you are wrong. See Identification 2: Normalization constraints (anchoring) in [SEM] intro 4.
forcenoanchor is similar to noanchor except that rather than issue an error message, gsem proceeds
to estimation. There is no reason you should specify this option. forcenoanchor is used in testing
of gsem at StataCorp.
reliability() specifies the fraction of variance not due to measurement error for a variable. See
[SEM] sem and gsem option reliability( ).
constraints() specifies parameter constraints you wish to impose on your model; see [SEM] sem
and gsem option constraints( ). Constraints can also be specified as described in [SEM] sem and
gsem path notation, and they are usually more conveniently specified using the path notation.
from() specifies the starting values to be used in the optimization process; see [SEM] sem and gsem
option from( ). Starting values can also be specified using the init() suboption as described in
[SEM] sem and gsem path notation.

Remarks and examples

To use gsem successfully, you need to understand paths, covariance(), variance(), and
means(); see Using path diagrams to specify standard linear SEMs in [SEM] intro 2 and see
[SEM] sem and gsem path notation.
covstructure() is often convenient; see [SEM] sem and gsem option covstructure( ).

454

gsem model description options Model description options

Also see
[SEM] gsem Generalized structural equation model estimation command
[SEM] intro 2 Learning the language: Path diagrams and command language
[SEM] sem and gsem option constraints( ) Specifying constraints
[SEM] sem and gsem option covstructure( ) Specifying covariance restrictions
[SEM] sem and gsem option from( ) Specifying starting values
[SEM] sem and gsem option reliability( ) Fraction of variance not due to measurement error
[SEM] sem and gsem path notation Command syntax for path diagrams

Title
gsem path notation extensions Command syntax for path diagrams

Syntax

Description

Options

Remarks and examples

Also see

Syntax
gsem paths . . .

, covariance() variance() means()

paths specifies the direct paths between the variables of your model.
The model to be fit is fully described by paths, covariance(), variance(), and means().

Description
This entry concerns gsem only.
The command syntax for describing generalized SEMs is fully specified by paths, covariance(),
variance(), and means(); see [SEM] sem and gsem path notation.
With gsem, the notation is extended to allow for generalized linear response variables and to allow
for multilevel latent variables. That is the subject of this entry.

Options
covariance(), variance(), and means() are described in [SEM] sem and gsem path notation.

Remarks and examples

Remarks are presented under the following headings:
Specifying multilevel nested latent variables
Specifying multilevel crossed latent variables
Specifying family and link

Specifying multilevel nested latent variables

Latent variables are indicated by a name in which at least the first letter is capitalized. This generic
form of the name is often written Lname.
In regular latent variables, which we will call level-1 latent variables, the unobserved values vary
observation by observation. Level-1 latent variables are the more common kinds of latent variables.
gsem allows higher-level latent variables as well as the level-1 variables. Lets consider three-level
data: students at the observational level, teachers at the second level, and schools at the third. In these
data, each observation is a student. We have data on students nested within teacher nested within
school.
Lets assume that we correspondingly have three identification (ID) variables. We number the
following list with the nesting level of the data:
3. Variable school contains a school ID number. If two observations have the same value of
school, then both of those students attended the same school.
455

456

gsem path notation extensions Command syntax for path diagrams

2. Variable teacher contains a teacher ID number, or it contains a teacher-within-school ID

number. That is, we do not care whether different schools assigned teachers the same ID
number. It will be sufficient for us that the ID number is unique within school.
1. Variable student contains a student ID number, or it contains a student-within-school ID
number, or even a student-within-teacher-within-school ID number. That is, we do not care
whether different observations have the same student ID as long as they have different teacher
IDs or different school IDs.
Here is how you write latent-variable names at each level of the model:
3. Level 3 is the school level. Latent variables are written as
Lname[school]
An example would be SchQuality[school].
The unobserved values of Lname[school] vary across schools and are constant within
school.
If Lname[school] is endogenous, its error variable is e.Lname[school].
You must refer to Lname[school] without omitting the [school] part. Lname by itself
looks like another latent variable to gsem.
2. Level 2 is the teacher-within-school level. Latent variables are written as
Lname[school>teacher] or
Lname[teacher<school]
An example would be TeachQuality[school>teacher] or
TeachQuality[teacher<school].
To gsem, Lname[school>teacher] and Lname[teacher<school] mean the same thing.
You can even refer to TeachQuality[school>teacher] in one place and refer to
TeachQuality[teacher<school] in another, and there will be no confusion.
The unobserved values of Lname[school>teacher] vary across schools and teachers, and
they are constant within teacher.
If Lname[school>teacher] is endogenous, its error variable is
e.Lname[school>teacher] or, equivalently, e.Lname[teacher<school].
1. Level 1 is the student or observational level. Latent variables are written as
Lname[school>teacher>student] or
Lname[student<teacher<school] or
Lname
Everybody just writes Lname. These are the latent variables that correspond to the latent
variables that sem provides. Unobserved values within the latent variable vary observation
by observation.
If Lname is endogenous, its error variable is e.Lname.
You can use multilevel latent variables in paths and options just as you would use any other latent
variable; see [SEM] sem and gsem path notation. Remember, however, that you must type out the full
name of all but the first-level latent variables. You type, for instance, SchQual[school>teacher].
There is a real tendency to type just SchQual when the name is unique.

gsem path notation extensions Command syntax for path diagrams

457

Changing the subject, the names by which effects are referred to are a function of the top level.
We just discussed a three-level model. The three levels of the model were
(3) school
(2) school>teacher
(1) school>teacher>student
If we had a two-level model, the levels would be
(2) teacher
(1) teacher>student
Thus, if we had started with a two-level model and then wanted to add a third, higher level onto
it, latent variables that were previously referred to as, say, TeachQual[teacher] would now be
referred to as TeachQual[school>teacher].

Specifying multilevel crossed latent variables

In our previous example, we had a three-level nested model in which student was nested within
teacher, which was nested within school.
Lets consider data on employees that also have the characteristics of working in an occupation
and working in an industry. These variables are not nested. Just as before, we will assume we have
variable employee containing an employee ID, and variables industry and occupation. The latent
variables associated with this model could be
Level
occupation
industry
employee (observational)

Latent-variable name
Lname[occupation]
Lname[industry]
Lname

Specifying family and link

gsem fits not only linear models but also generalized linear models. There is a set of options for
specifying the specific model to be fit. These options are known as family-and-link options, which
include family() and link(), but those options are seldom used in favor of other family-and-link
shorthand options such as logit, which means family(bernoulli) and link(logit). These
options are explained in [SEM] gsem family-and-link options.
In the command language, you can specify these options among the shared options at the end of
a gsem command:
. gsem ..., ... logit ...

That is convenient, but only if all the equations in the model are using the same specific response
function. Many models include multiple equations with each using a different response function.
You can specify any of the family-and-link options within paths. For instance, typing
. gsem (y <- x1 x2), logit

has the same effect as typing

. gsem (y <- x1 x2, logit)

458

gsem path notation extensions Command syntax for path diagrams

Thus you can type

. gsem (y1 <- x1 L, logit) (y2 <- x2 L, poisson) ..., ...

The y1 equation would be logit, and the y2 equation would be Poisson. If you wanted y2 to be linear
regression, you could type
. gsem (y1 <- x1 L, logit) (y2 <- x2 L, regress) ..., ...

or you could be silent and let y2 default to linear regression,

. gsem (y1 <- x1 L, logit) (y2 <- x2 L) ..., ...

Also see
[SEM] gsem Generalized structural equation model estimation command
[SEM] sem and gsem path notation Command syntax for path diagrams
[SEM] intro 2 Learning the language: Path diagrams and command language

Title
gsem postestimation Postestimation tools for gsem

Description

Remarks and examples

Also see

Description
The following are the postestimation commands that you can use after estimation by gsem:
Command

Description

gsem, coeflegend

display b[ ] notation

estat eform
estat ic

display exponentiated coefficients

Akaikes and Schwarzs Bayesian information criteria (AIC and BIC)

lrtest
test
lincom
nlcom
testnl

likelihood-ratio tests
Wald tests
linear combination of parameters
nonlinear combination of parameters
Wald tests of nonlinear hypotheses

estat summarize
estat vce

summary statistics for the estimation sample

variancecovariance matrix of the estimators (VCE)

predict

generalized linear predictions, etc.

margins
contrast
pwcompare

marginal means, predictive margins, and marginal effects

contrasts and linear hypothesis tests
pairwise comparisons

estimates

cataloging estimation results

For a summary of postestimation features, see [SEM] intro 7.

Remarks and examples

This manual entry concerns gsem. For information on postestimation features available after sem,
see [SEM] sem postestimation.

Also see
[SEM] gsem reporting options Options affecting reporting of results

459

Title
gsem reporting options Options affecting reporting of results
Syntax

Description

Options

Remarks and examples

Also see

Syntax
gsem paths . . . , . . . reporting options
gsem, reporting options
reporting options

Description

level(#)
coeflegend
nocnsreport
noheader
notable

set confidence level; default is level(95)

display coefficient legend
do not display constraints
do not display header above parameter table
do not display parameter table

display options

control column formats, row spacing, line width, display of omitted

variables and base and empty cells, and factor-variable labeling

Description
These options control how gsem displays estimation results.

Options
level(#); see [R] estimation options.
coeflegend displays the legend that reveals how to specify estimated coefficients in b[ ] notation,
which you are sometimes required to type in specifying postestimation commands.
nocnsreport suppresses the display of the constraints. Fixed-to-zero constraints that are automatically
set by gsem are not shown in the report to keep the output manageable.
noheader suppresses the header above the parameter table, the display that reports the final loglikelihood value, number of observations, etc.
notable suppresses the parameter table.
display options: noomitted, vsquish, noemptycells, baselevels, allbaselevels, nofvlabel, fvwrap(#), fvwrapon(style), cformat(% fmt), pformat(% fmt), sformat(% fmt), and
nolstretch; see [R] estimation options.

Remarks and examples

Any of the above options may be specified when you fit the model or when you redisplay results,
which you do by specifying nothing but options after the gsem command:
460

gsem reporting options Options affecting reporting of results

. gsem (...) (...), ...
(original output displayed)
. gsem
(output redisplayed)
. gsem, coeflegend
(coefficient-name table displayed)
. gsem
(output redisplayed)

Also see
[SEM] gsem Generalized structural equation model estimation command
[SEM] example 29g Two-parameter logistic IRT model

461

Title
lincom Linear combinations of parameters
Syntax
Remarks and examples

Menu
Stored results

Description
Also see

Options

Syntax


lincom exp , options

Menu
Statistics

>

SEM (structural equation modeling)

>

Testing and CIs

>

Linear combinations of parameters

Description
lincom is a postestimation command for use after sem, gsem, and nearly all Stata estimation
commands.
lincom computes point estimates, standard errors, z statistics, p-values, and confidence intervals
for linear combinations of the estimated parameters.
After sem and gsem, you must use the b[ ] coefficient notation; you cannot refer to variables
by using shortcuts to obtain coefficients on variables.
See [R] lincom.

Options
See Options in [R] lincom.

Remarks and examples

lincom works in the metric of SEM, which is to say path coefficients, variances, and covariances.
If you want to frame your linear combinations in terms of standardized coefficients and correlations,
and you fit the model with sem, not gsem, then prefix lincom with estat stdize:; see [SEM] estat
stdize.

Stored results
See Stored results in [R] lincom.

Also see
[R] lincom Linear combinations of estimators
[SEM] estat stdize Test standardized parameters
[SEM] nlcom Nonlinear combinations of parameters
[SEM] test Wald test of linear hypotheses
462

Title
lrtest Likelihood-ratio test of linear hypothesis
Syntax
Stored results

Menu
Also see

Description

Remarks and examples

Syntax
{sem | gsem} . . . , . . .

(fit model 1)

estimates store modelname1

{sem | gsem} . . . , . . .

(fit model 2)

estimates store modelname2

(compare models)

lrtest modelname1 modelname2

where one of the models is constrained and the other is unconstrained. lrtest counts
parameters to determine which model is constrained and which is unconstrained, so it does
not matter which model is which.
Warning concerning use with sem: Do not omit variables, observed or latent, from the model.
Constrain their coefficients to be 0 instead. The models being compared must contain the same
set of variables. This is because the standard SEM likelihood value is a function of the variables
appearing in the model. Despite this fact, use of lrtest is appropriate under the relaxed conditional
normality assumption.
Note concerning gsem: The above warning does not apply to gsem just as it does not apply to other
Stata estimation commands. Whether you omit variables or constrain coefficients to 0, results will
be the same. The generalized SEM likelihood is conditional on the exogenous variables.

Menu
Statistics

>

SEM (structural equation modeling)

>

Testing and CIs

>

Likelihood-ratio test

Description
lrtest is a postestimation command for use after sem, gsem, and other Stata estimation commands.
lrtest performs a likelihood-ratio test comparing models. See [R] lrtest.

Remarks and examples

See [SEM] example 10 and [SEM] example 39g.
When using lrtest after sem, you must be careful that the models being compared have the
same observed and latent variables. For instance, the following is allowed:
463

464

lrtest Likelihood-ratio test of linear hypothesis

.
.
.
.
.

sem (L1 -> x1 x2 x3) (L1 <- x4 x5) (x1 <- x4) (x2 <- x5)
estimates store m1
sem (L1 -> x1 x2 x3) (L1 <- x4 x5)
estimates store m2
lrtest m1 m2

This is allowed because both models contain the same variables, namely, L1, x1, . . . , x5, even though
the second model omitted some paths.
The following would produce invalid results:
.
.
.
.
.

sem (L1 -> x1 x2 x3) (L1 <- x4 x5) (x1 <- x4) (x2 <- x5)
estimates store m1
sem (L1 -> x1 x2 x3) (L1 <- x4)
estimates store m2
lrtest m1 m2

It produces invalid results because the second model does not include variable x5 and the first model
does. To run this test correctly, you type
. sem (L1 -> x1 x2 x3) (L1 <- x4 x5) (x1 <- x4) (x2 <- x5)
.
.
.
.

estimates store m1
sem (L1 -> x1 x2 x3) (L1 <- x4 x5@0)
estimates store m2
lrtest m1 m2

If we were using gsem rather than sem, all the above would still be valid.

Stored results
See Stored results in [R] lrtest.

Also see
[SEM] example 10 MIMIC model
[SEM] example 39g Three-level model (multilevel, generalized response)

[R] lrtest Likelihood-ratio test after estimation

[SEM] test Wald test of linear hypotheses
[SEM] estat stdize Test standardized parameters
[SEM] estat eqtest Equation-level test that all coefficients are zero

Title
methods and formulas for gsem Methods and formulas
Description

Remarks and examples

References

Also see

Description
The methods and formulas for the gsem command are presented below.

Remarks and examples

Remarks are presented under the following headings:
Introduction
Families of distributions
The Bernoulli family
The binomial family
The ordinal family
The multinomial family
The Poisson family
The negative binomial family
The gamma family
The Gaussian family
Reliability
Link functions
The logit link
The probit link
The complementary log-log link
The log link
The identity link
The likelihood
GaussHermite quadrature
Adaptive quadrature
Laplacian approximation
Postestimation
Empirical Bayes
Other predictions

Introduction
gsem fits generalized linear models with latent variables via maximum likelihood. Here is a table
identifying the family/link combinations that gsem allows.
logit

probit

cloglog

Bernoulli
binomial

x
x

x
x

x
x

ordinal
multinomial
Poisson
negative binomial
gamma
Gaussian

x
x

465

log

identity

x
x
x
x

466

methods and formulas for gsem Methods and formulas

Log-likelihood calculations for fitting any model with latent variables require integrating out
the latent variables. One widely used modern method is to directly estimate the integral required to
calculate the log likelihood by GaussHermite quadrature or some variation thereof. gsem implements
four different methods for numerically evaluating the integrals.
1. GaussHermite quadrature (GHQ)
2. Mean-variance adaptive quadrature (MVAGH)
3. Mode-curvature adaptive quadrature (MCAGH)
4. Laplacian approximation
The default method is MVAGH. The numerical integration method for MVAGH is based on RabeHesketh, Skrondal, and Pickles (2005), and the other numerical integration methods described in this
manual entry are based on Skrondal and Rabe-Hesketh (2004, chap. 6.3).

Families of distributions
gsem implements the most commonly used distribution families associated with generalized linear
models. gsem also implements distributions for ordinal and multinomial outcomes.
In this manual entry, observed endogenous variables are also known as generalized responses
or generalized outcomes, but we will simply refer to them as responses or outcomes. The random
variable corresponding to a given response will be denoted by Y . An observed value of Y will be
denoted by y , and the expected value of Y by . For the ordinal and multinomial families, we will
refer to a linear prediction, denoted by z , instead of the expected value.

The Bernoulli family

The Bernoulli family is a binary response model. The response Y is assumed to take on the values
0 or 1; however, gsem allows any nonzero and nonmissing value to mean 1.
The log of the conditional probability mass function is

log f (y|) = y log + (1 y) log(1 )

where is also known as the probability of a success. The default link for the Bernoulli family is
the logit link.

The binomial family

The binomial family is a count response model and generalizes the Bernoulli family by taking
the sum of k independent Bernoulli outcomes. The response Y is assumed to take on the values
0, 1, . . . , k .

methods and formulas for gsem Methods and formulas

467

The log of the conditional probability mass function is

log f (y|) = log{(k + 1)} log{(y + 1)} log{(k y + 1)}

+ y log + (1 y) log(1 )
where is the expected value for a single Bernoulli outcome. The default link for the binomial family
is the logit link.

The ordinal family

The ordinal family is a discrete response model. The response Y is assumed to take on one of k
unique values. The actual values are irrelevant except that higher values are assumed to correspond to
higher outcomes. Without loss of generality, we will assume that Y takes on the values 1, . . . , k .
The ordinal family with k outcomes has cutpoints 0 , 1 , . . . , k , where 0 = , y < y+1 ,
and k = +.
Given a linear prediction z , the probability that a random response Y takes the value y is
Pr(Y = y|z) = Pr(Y < y z) Pr(Y < y1 z)
where Y is the underlying stochastic component for Y . The distribution for Y is determined by
the link function. gsem allows logit, probit, and cloglog for the ordinal family. The logit
link assigns Y the extreme value distribution that is synonymous with the logit link for Bernoulli
outcomes. The probit link assigns Y the standard normal distribution that is synonymous with the
probit link for Bernoulli outcomes. The cloglog link assigns Y the distribution that is synonymous
with the complementary log-log link for Bernoulli outcomes. The default link for the ordinal family
is the logit link.

The multinomial family

The multinomial family is a discrete response model. The response Y is assumed to take on one of
k unique values. The actual values are irrelevant and order does not matter; however, gsem requires
that the values are nonnegative integers. Without loss of generality, we will assume that Y takes
on the values 1, . . . , k . Each of the k outcomes has its own linear prediction. For the model to be
identified, one of the outcomes is chosen to be the base or reference. The linear prediction for the
base outcome is constrained to be 0 for all observations. Without loss of generality, we will assume
the base outcome is the first outcome. Let zi be the prediction for outcome i, where z1 = 0 for the
base outcome.
Given the k linear predictions z0 = (z1 , z2 , . . . , zk ), the log of the conditional probability mass
function is

(
log f (y|z) = zy log

k
X

)
exp(zi )

i=1

The only link allowed for the multinomial family is the logit link.

468

methods and formulas for gsem Methods and formulas

The Poisson family

The Poisson family is a count-data response model. The response Y is assumed to take on
nonnegative integer values.
The log of the conditional probability mass function is

log f (y|) = + y log log (y + 1)

The only link allowed for the Poisson family is the log link.
The negative binomial family

The negative binomial family is another count-data response model. It is commonly thought of as
a Poisson family with overdispersion. gsem allows two parameterizations for the dispersion in this
family: mean dispersion and constant dispersion.
The log of the conditional probability mass function is

log f (y|, ) = log{(y + m)} log{(y + 1)} log{(m)}

+ m log p + y log(1 p)
where m and p depend on the form of dispersion.
The only link allowed for the negative binomial family is the log link.
For mean dispersion, we have

m = 1/
1
p=
1 +
where is the expected value of Y and is the scale parameter. gsem fits in the log scale.
For constant dispersion, we have

m = exp(log log )
1
p=
1+
where is the expected value of Y and is the scale parameter. gsem fits in the log scale.
The gamma family

The gamma family is a continuous response model. The response Y is assumed to be a nonnegative
real value.
The log of the conditional probability density function is
2

log f (y|, s) = log (s

y
) log + 2 log s +

s2 + (s2 1) log y

where is the expected value of Y and s is the scale parameter. gsem fits s in the log scale.
The only link allowed for the gamma family is the log link.

methods and formulas for gsem Methods and formulas

469

The Gaussian family

The Gaussian family is a continuous response model and is synonymous with the normal distribution.
When the Gaussian family is specified with the identity link but no censoring, gsem fits this
family by using a single multivariate density function and allows the following two special features:
1. gsem can fit covariances between the Gaussian error variables.
2. gsem can fit paths between Gaussian responses, including nonrecursive systems.
The log of the conditional probability density function is

log f (y|, ) =

1
d log 2 + log || + (y )0 1 (y )
2

where d is the dimension of the observed response vector y, is the mean of the responses, and
is the variance matrix of their unexplained errors.
When the Gaussian family is specified with the log link or censoring, the two special features
described above no longer apply. In addition, the multivariate density function is no longer used.
Instead, for each response using the log link, the log of the conditional probability density function
corresponds to the formula above with d = 1. For censored responses, the log likelihood corresponds
to the one in the Methods and formulas for [R] intreg.
Reliability

For a given Gaussian response variable with the identity link, the reliability Y may be specified
as p or 100 p%. The variance of Y s associated error variable is then constrained to (1 p) times
the observed variance of Y .

Link functions
Except for the ordinal and multinomial families, the link function defines the transformation between
the mean and the linear prediction for a given response. If Y is the random variable corresponding
to an observed response variable y , then the link function performs the transformation

g() = z
where = E(Y ) and z is the linear prediction. In practice, the likelihood evaluator function uses
the inverse of the link function to map the linear prediction to the mean.
The logit link

The logit link is

g() = log log(1 )

and its inverse is

= g 1 (z) =

1
1 + ez

470

methods and formulas for gsem Methods and formulas

The probit link

The probit link is

g() = 1 ()
and its inverse is

= g 1 (z) = (z)
where () is the cumulative distribution function for the standard normal distribution and 1 () is
its inverse.
The complementary log-log link

The complementary log-log link is

g() = log{ log(1 )}

and its inverse is

= g 1 (z) = 1 exp{ exp(z)}

The log link

The log link is

g() = log
and its inverse is

= g 1 (z) = ez

The identity link

The identity link is g() = .

The likelihood
gsem fits generalized linear models with latent variables via maximum likelihood. The likelihood
for the specified model is derived under the assumption that each response variable is independent and
identically distributed across the estimation sample. The response variables are also assumed to be
independent of each other. These assumptions are conditional on the latent variables and the observed
exogenous variables.

methods and formulas for gsem Methods and formulas

471

The likelihood is computed by integrating out the latent variables. Let be the vector of model
parameters, y be the vector of observed response variables, x be the vector of observed exogenous
variables, and u be the r 1 vector of latent variables. Then the marginal likelihood looks something
like

Z
L() =
<r

f (y|x, u, )(u|u , u )u

where < denotes the set of values on the real line, <r is the analog in r-dimensional space, is a
vector of the unique model parameters, f () is the conditional probability density function for the
observed response variables, () is the multivariate normal density for u, u is the expected value of
u, and u is the covariance matrix for u. All auxiliary parameters are fit directly without any further
parameterization, so we simply acknowledge that the auxiliary parameters are among the elements of
.
The y variables are assumed to be independent, conditionally on x and u, so f () is the product of
the individual conditional densities. One exception to this is when y contains two or more Gaussian
response variables with the identity link, in which case the Gaussian responses are actually modeled
using a multivariate normal density to allow for correlated errors and nonrecursive systems among
Gaussian responses. This one exception does not change how the integral is numerically evaluated,
so we make no effort to represent this distinction in the formulas.
For a single-level model with n response variables, the conditional joint density function for a
given observation is

f (y|x, u, ) =

n
Y

fi (yi |x, u, )

i=1

For a two-level model, the likelihood is computed at the cluster level, so the conditional density is
also a product of the observation-level density contributions within a given cluster

f (y|x, u, ) =

n Y
t
Y

fi (yij |xj , u, )

i=1 j=1

where t is the number of individuals within the cluster. This extends to more levels by expanding
the products down to the observations nested within the hierarchical groups. Because the single-level
model is a special case of a two-level model where all the groups have a single observation, we will
now use the two-level notation and subscripts.
Except for the ordinal and multinomial families, we use the link function to map the conditional
mean
ij = E(yij |xj , u)
to the linear prediction

zij = x0j i + x0j i u

where i is the vector of the fixed-effect coefficients and i is the matrix of the latent loadings for
yij . For notational convenience, we will overload the definitions of f () and fi () so that they are
functions of the responses and model parameters through the linear predictions z0 = (z1 , . . . , zn ).
Thus f (y|x, u, ) is equivalently specified as f (y, z, ), and fi (yij |xj , u, ) is equivalently specified
as fi (yij , zij , ). In this new notation, the likelihood for a given cluster is

472

methods and formulas for gsem Methods and formulas

Z
L() =

f (y, z, )(u|u , u )u


Z
1
1
0 1
p
exp log f (y, z, ) (u u ) u (u u ) u
=
2
(2)r/2 |u | <r
<r

(1)

gsem allows nonrecursive systems between Gaussian response variables with the identity link, but
non-Gaussian responses and Gaussian responses with the log link are not allowed to participate in
any nonrecursive systems. This means that if a given response y is specified with a family other
than Gaussian or a link other than identity, then y cannot have a path that ultimately leads back to
itself. Any response may participate in a recursive system because the participating responses may
be treated as exogenous variables when predicting other responses in a recursive system.
The latent vector u consists of stacked collections of the latent variables from each level. Within
each level, the latent endogenous variables are stacked over the latent exogenous variables . Within
a given level, the latent exogenous variables and latent endogenous errors are assumed independent
and multivariate normal
N (, )
 N (0, )
so according to the linear relationship
= B + + Ax + 
we have that the latent variables are jointly multivariate normal. This linear relationship implies that
gsem allows latent variables to predict each other, but only within level. It also means that gsem
allows paths from observed variables to latent variables; however, the observed variable must be
constant within group if the path is to a group-level latent variable.
For our two-level model, we have

u N (u , u )
where


u =


u =

= (I B)1 ( + Ax)


0
= (I B)1 (0 + ) (I B)1
= (I B)1
The vector is therefore the set of unique model parameters taken from the following:
i is the vector of fixed-effect coefficients for yij .
i is the matrix of latent loadings for yij .

methods and formulas for gsem Methods and formulas

473

B is the matrix of latent endogenous coefficients.

is the matrix of latent exogenous coefficients.

A is the matrix of latent fixed-effect coefficients.

is the vector of latent exogenous means.
is the matrix of latent exogenous variances and covariances.
is the matrix of latent endogenous error variances and covariances.
Auxiliary parameters that result from some of the distribution families.
Each level of a multilevel model will have its own set of the following parameters: B, , A, , ,
and . For multilevel models, u is a block-diagonal matrix with a block for each level.
GaussHermite quadrature

The integral in (1) is generally not tractable, so we must use numerical methods. In the univariate
case, the integral of a function multiplied by the kernel of the standard normal distribution can be
approximated using GaussHermite quadrature (GHQ). For q -point GHQ, let the abscissa and weight
pairs be denoted by (ak , wk ), k = 1, . . . , q . The GHQ approximation is then

f (x) exp(x ) dx

q
X

wk f (ak )

k=1

Using the standard normal distribution yields the approximation

f (x)(x) dx

where ak =

q
X

wk f (ak )

k=1

2ak and wk = (wk )/ .

We can use a change-of-variables technique to transform the multivariate integral (1) into a set
of nested univariate integrals. Each univariate integral can then be evaluated using GHQ. Let v be
a random vector whose elements are independently standard normal, and let L be the Cholesky
decomposition of u , that is, u = LL0 . In the distribution, we have that u = u + Lv, and the
linear predictions vector as a function of v is

zij = x0j i + x0j i (u + Lv)

so the likelihood for a given cluster is
r/2

L() = (2)

1X 2
...
exp log f (y, z, )
vk
2

)
dv1 . . . dvr

(2)

k=1

where r is the number of latent variables.

Consider an r-dimensional quadrature grid containing q quadrature points in each dimension. Let
the vector of abscissas ak = (ak1 , . . . , akr )0 be a point in this grid, and let wk = (wk1 , . . . , wkr )0
be the vector of corresponding weights. The GHQ approximation to the likelihood for a given cluster
is

474

methods and formulas for gsem Methods and formulas

GHQ

() =

q
X

...

k1 =1

q
X

"

(
exp

n
X

)
log fi (yij , zijk , )

#
wks

s=1

i=1

kr =1

r
Y

where

zijk = x0j + x0j i (u + Lak )

Adaptive quadrature

This section sets the stage for mean-variance adaptive GaussHermite quadrature (MVAGH) and
mode-curvature adaptive GaussHermite quadrature (MCAGH).
Lets reconsider the likelihood in (2). If we fix the observed variables and the model parameters,
we see that the posterior density for v is proportional to

(v)f (y, z, )
It is reasonable to assume that this posterior density can be approximated by a multivariate normal
density with mean vector v and variance matrix v . Instead of using the prior density of v as the
weighting distribution in the integral, we can use our approximation for the posterior density,

Z
L() =
<r

f (y, z, )(v)
(v, v , v ) dv
(v, v , v )

The likelihood is then approximated with

L () =

q
X
k1 =1

...

q
X
kr =1

"

(
exp

n
X

log fi (yij , zijk

, )

i=1

r
Y

#
ks

s=1

where

zijk
= x0j + x0j i (u + Lk )

and k and the ks are the adaptive versions of the abscissas and weights after an orthogonalizing
transformation, which eliminates posterior covariances between the latent variables. k and the ks
are functions of ak and wk and the adaptive parameters v and v .
For MVAGH, v is the posterior mean and v is the posterior variance of v. They are computed
iteratively by updating the posterior moments by using the MVAGH approximation, starting with a 0
mean vector and identity variance matrix.
For MCAGH, v is the posterior mode for v and v is the curvature at the mode. They are computed
by optimizing the integrand in (2) with respect to v.

methods and formulas for gsem Methods and formulas

475

Laplacian approximation
Lets reconsider the likelihood in (1) and denote the argument in the exponential function by

1
h(u) = log f (y, z, ) (u u )0 1
u (u u )
2
n X
t
X
1
=
log fi (yij , zij , ) (u u )0 1
u (u u )
2
i=1 j=1
where

zij = x0j i + x0j i u

The Laplacian approximation is based on a second-order Taylor expansion of h(u) about the value
of u that maximizes it. The first and second partials with respect to u are
n

h(u) X X log fi (yij , zij , ) 0

=
h (u) =
i xj 1 (u u )
u
z
ij
i=1 j=1
0

H(u) =

2 h(u) X X 0 2 log fi (yij , zij , ) 0

xj i
=
i xj 1
uu0
z
z
ij
ij
i=1 j=1

b such that h0 (b
The maximizer of h(u) is u
u) = 0. The integral in (1) is proportional to the posterior
b is also the posterior mode.
density of u given the data, so u
The second-order Taylor approximation then takes the form

1
b )0 H(b
b)
u)(u u
h(u) h(b
u) + (u u
2

(3)

because the first order derivative term is 0. The integral is approximated by

1/2

exp{h(u)} du exp{h(b
u)}(2)r/2 |H(b
u)|

<r

because the second term in (3) is the kernel of a multivariate normal density once it is exponentiated.
The Laplacian approximation for the log likelihood is

1
1
log LLap () = log |u | log |H(b
u)| + h(b
u)
2
2

Postestimation
We begin by considering the prediction of the latent variables u for a given cluster in a two-level
model. Prediction of latent variables in multilevel generalized linear models involves assigning values
to the latent variables, and there are many methods for doing so; see Skrondal and Rabe-Hesketh (2009)
and Skrondal and Rabe-Hesketh (2004, chap. 7) for a comprehensive review. Stata offers two methods
of predicting latent variables: empirical Bayes means (also known as posterior means) and empirical
Bayes modes (also known as posterior modes). Below we provide more details about the two methods.

476

methods and formulas for gsem Methods and formulas

Empirical Bayes

Let b
denote the estimated model parameters. Empirical Bayes (EB) predictors of the latent variables
are the means or modes of the empirical posterior distribution with the parameter estimates replaced
with their estimates b
. The method is called empirical because b
is treated as known. EB combines
the prior information about the latent variables with the likelihood to obtain the conditional posterior
distribution of latent variables. Using Bayes theorem, the empirical conditional posterior distribution
of the latent variables for a given cluster is

b u)
f (y|u, x; b
) (u;
bu,
b u ) du
f (y|u, x; b
) (u;
bu,

(u|y, x; b
) = R
=

b u)
f (y|u, x; b
) (u;
bu,
L(b
)

The denominator is just the likelihood contribution of the given cluster.

EB mean predictions of latent variables,

e , also known as posterior means, are calculated as

u

Z
e=
u

u (u|y, x; b
) du
<r

e rather than u
b to distinguish predicted values from estimates. This
where we use the notation u
multivariate integral is approximated by MVAGH. If you have multiple latent variables within a level
or latent variables across levels, the calculation involves orthogonalizing transformations with the
Cholesky transformation because the latent variables are no longer independent under the posterior
distribution.
When all the response variables are normal, the posterior density is multivariate normal, and EB
means are also best linear unbiased predictors (BLUPs); see Skrondal and Rabe-Hesketh (2004, 227).
In generalized mixed-effects models, the posterior density tends to multivariate normal as cluster size
increases.
e such that
e
EB modal predictions can be approximated by solving for u

log (u|y, x; b
) e = 0
e
u=u
u
Because the denominator in () does not depend on u, we can omit it from the calculation to obtain
the EB mode. The calculation of EB modes does not require numerical integration, and for that reason
they are often used in place of EB means. As the posterior density gets closer to being multivariate
normal, EB modes get closer and closer to EB means.
Just like there are many methods of assigning values to the random effects, there exist many methods
of calculating standard errors of the predicted random effects; see Skrondal and Rabe-Hesketh (2009)
for a comprehensive review.
Stata uses the posterior standard deviation as the standard error of the posterior means predictor of
random effects. For a given level, the EB posterior covariance matrix of the random effects is given
by
Z
e )(u u
e )0 (u|y, x; b
Cov(e
u|y, x; b
) =
) du
(u u
<r

The posterior covariance matrix and the integrals are approximated by MVAGH.

methods and formulas for gsem Methods and formulas

477

Conditional standard errors for the estimated posterior modes are derived from standard theory of
e
e is the negative inverse
maximum likelihood, which dictates that the asymptotic variance matrix of u
of the Hessian matrix.
Other predictions

e , which
In what follows, we show formulas with the posterior means estimates of random effects u
e are simply
are used by default or if the means option is specified. If the modes option is specified, u
e
e in these formulas.
replaced with the posterior modes u
For the ith response in the j th observation within a given cluster in a two-level model, the linear
predictor is computed as

b + x0j
biu
e
zbij = x0j
The linear predictor includes the offset or exposure variable if one was specified during estimation,
unless the nooffset option is specified. If the fixedonly option is specified, the linear predictor
is computed as

b
zbij = x0j
The predicted mean, conditional on the predicted latent variables, is

bij = g 1 (b
zij )
where g 1 () is the inverse link function defined in Link functions above. For the ordinal and
multinomial families, the predicted mean is actually a probability, and gsem can produce a probability
for each outcome value as described in The ordinal family and The multinomial family above.

References
Rabe-Hesketh, S., A. Skrondal, and A. Pickles. 2005. Maximum likelihood estimation of limited and discrete dependent
variable models with nested random effects. Journal of Econometrics 128: 301323.
Skrondal, A., and S. Rabe-Hesketh. 2004. Generalized Latent Variable Modeling: Multilevel, Longitudinal, and
Structural Equation Models. Boca Raton, FL: Chapman & Hall/CRC.
. 2009. Prediction in multilevel generalized linear models. Journal of the Royal Statistical Society, Series A 172:
659687.

Also see
[SEM] gsem Generalized structural equation model estimation command

Title
methods and formulas for sem Methods and formulas for sem

Description

Remarks and examples

References

Also see

Description
The methods and formulas for the sem commands are presented below.

Remarks and examples

Remarks are presented under the following headings:
Variable notation
Model and parameterization
Summary data
Maximum likelihood
Weighted least squares
Groups
Fitted parameters
Standardized parameters
Reliability
Postestimation
Model framework
Goodness of fit
Group goodness of fit
Equation-level goodness of fit
Wald tests
Score tests
Residuals
Testing standardized parameters
Stability of nonrecursive systems
Direct, indirect, and total effects
Predictions

Variable notation
We will use the following convention to keep track of the five variable types recognized by the
sem estimation command:
1. Observed endogenous variables are denoted y .
2. Observed exogenous variables are denoted x.
3. Latent endogenous variables are denoted .
4. Latent exogenous variables are denoted .
5. Error variables are denoted with prefix e. on the associated endogenous variable.
a. Error variables for observed endogenous are denoted e.y .
b. Error variables for latent endogenous are denoted e. .
478

methods and formulas for sem Methods and formulas for sem

479

In any given analysis, there are typically several variables of each type. Vectors of the four main
variable types are denoted y, x, , and . The vector of all endogenous variables is
 
y
Y=

The vector of all exogenous variables is

X=
The vector of all error variables is

 
x


=

e.y
e.

Model and parameterization

sem can fit models of the form

Y = BY + X + +
where B = [ij ] is the matrix of coefficients on endogenous variables predicting other endogenous
variables, = [ij ] is the matrix of coefficients on exogenous variables, = [i ] is the vector of
intercepts for the endogenous variables, and is assumed to have mean 0 and
Cov(X, ) = 0
Let
= [j ] = E(X)
= [ij ] = Var(X)
= [ij ] = Var()
Then the mean vector of the endogenous variables is
Y = E(Y ) = (I B)1 ( + )
the variance matrix of the endogenous variables is


0
Y Y = Var(Y ) = (I B)1 (0 + ) (I B)1
and the covariance matrix between the endogenous variables and the exogenous variables is
Y X = Cov(Y, X ) = (I B)1
Let Z be the vector of all variables:


Z=

Y
X

Then its mean vector is


= E(Z) =

and its variance matrix is


= Var(Z) =

Y Y
0Y X

Y X

480

methods and formulas for sem Methods and formulas for sem

Summary data
Let zt be the vector of all observed variables for the tth observation,
 
yt
zt =
xt
and let wt be the corresponding weight value, where t = 1, . . . , N . If no weights were specified,
then wt = 1. Let w. be the sum of the weights; then the sample mean vector is

z=

N
1 X
wt zt
w. t=1

and the sample variance matrix is

N

S=

1 X
wt (zt z)(zt z)0
w. 1 t=1

Maximum likelihood
Let be the vector of unique model parameters, such as

vec(B)
vec()

vech()
=

vech()

Then under the assumption of the multivariate normal distribution, the overall log likelihood for is


w. 
log L() =
k log(2) + log {det(o )} + tr D1
o
2
where k is the number of observed variables, o is the submatrix of corresponding to the observed
variables, and
D = f S + (z o )(z o )0
where

1, if nm1 is specified
w. 1

, otherwise
w.
and o is the subvector of corresponding to the observed variables.
f=

For the BHHH optimization technique and when computing observation-level scores, the log
likelihood for is computed as

log L() =

N
X

wt 
k log(2) + log {det(o )} + (zt o )0 1
o (zt o )
2
t=1

and the nm1 option is ignored.

When method(mlmv) is specified, sem groups the data according to missing-value patterns. Each
missing-value pattern will have its own summary data: k , z, and S. The log likelihood for a missingvalue pattern is computed using this summary data and the corresponding elements of o and o .
The overall log likelihood is computed by summing the log-likelihood values from each missing-value
pattern.

methods and formulas for sem Methods and formulas for sem

481

Weighted least squares

Let v be the vector of unique sample moments of the observed variables and be the corresponding
vector of population moments. Then

z
v=
vech(f S)
and


=

o
o

The weighted least-squares (WLS) criterion function to minimize is the quadratic form

FWLS () = (v )0 W1 (v )
where W is the least-squares weight matrix. For unweighted least squares (ULS), the weight matrix
is the identity matrix W = I. Other weight matrices are mentioned in Bollen (1989).
The weight matrix implemented in sem is an estimate of the asymptotic covariance matrix of
v. This weight matrix is derived without any distributional assumptions and is often referred to as
derived from an arbitrary distribution function or is asymptotic distribution free (ADF), thus the option
method(adf).

Groups
When the group() option is specified, each group has its own summary data and model parameters.
The entire collection of model parameters is

1
2

=
...

G
where G is the number of groups. The group-level criterion values are combined to produce an overall
criterion value.
For method(ml) and method(mlmv), the overall log likelihood is

log L() =

G
X

log L(g )

g=1

For method(adf), the overall criterion function to minimize is

FWLS () =

G
X
g=1

FWLS (g )

482

methods and formulas for sem Methods and formulas for sem

Fitted parameters
sem fits the specified model by maximizing the log likelihood or minimizing the WLS criterion. If
is the vector of model parameters, then the fitted parameter vector is denoted by b
, and similarly
b , ,
b ,
b ,
b
b
for B
b,
b , ,
b , and their individual elements.

Standardized parameters
b Y Y . Then the standardized parameter estimates are
Let
bii be the ith diagonal element of
s

bii
eij = bij

bjj
s

eij =
bij

bii

bjj

bij / bii bjj , if i 6= j

eij =
p

bii , if i = j
bii /
bij
eij = q
bii bjj

bi

ei =

bii

bj

ej = q
bjj
The variance matrix of the standardized parameters is estimated using the delta method.

Reliability
For an observed endogenous variable, y , the reliability may be specified as p or 100 p%. The
variance of e.y is then constrained to (1 p) times the observed variance of y .

Postestimation
Model framework
estat framework reports the fitted parameters in their individual matrix forms as introduced in
Fitted parameters.
Goodness of fit
estat gof reports the following goodness-of-fit statistics.

methods and formulas for sem Methods and formulas for sem

483

Let the degrees of freedom for the specified model be denoted by df m . In addition to the specified
model, sem fits saturated and baseline models corresponding to the observed variables in the specified
model. The saturated model fits a full covariance matrix for the observed variables and has degrees
of freedom


p+q+1
df s =
+p+q
2
where p is the number of observed endogenous variables and q is the number of observed exogenous
variables in the model. The baseline model fits a reduced covariance matrix for the observed variables
depending on the presence of endogenous variables. If there are no endogenous variables, all variables
are uncorrelated in the baseline model; otherwise, only exogenous variables are correlated in the
baseline model. The degrees of freedom for the baseline model is

df b =

2q , if p = 0


q+1

2p
+
q
+
, if p > 0

For method(ml) and method(mlmv), let the saturated log likelihood be denoted by log Ls and
the baseline log likelihood be denoted by log Lb . The likelihood-ratio test of the baseline versus
saturated models is computed as

2bs = 2(log Ls log Lb )

with degrees of freedom df bs = df s df b . The likelihood-ratio test of the specified model versus
the saturated model is computed as

2ms = 2{log Ls log L(b

)}
with degrees of freedom df ms = df s df m .
For method(adf), the saturated criterion value is 0, Fs = 0. Let the baseline criterion value be
denoted by Fb . The chi-squared test of the baseline versus saturated models is computed as

2bs = N Fb
with degrees of freedom df bs = df s df b . The chi-squared test of the specified model versus the
saturated model is computed as
2ms = N FWLS (b
)
with degrees of freedom df ms = df s df m .
The Akaike information criterion (Akaike 1974) is defined as
AIC

= 2 log L(b
) + 2df m

The Bayesian information criterion (Schwarz 1978) is defined as

BIC

= 2 log L(b
) + N df m

See [R] BIC note for additional information on calculating and interpreting the BIC.

484

methods and formulas for sem Methods and formulas for sem

The overall coefficient of determination is computed as

CD

=1

b)
det(
b)
det(

This value is also referred to as the overall R2 in estat eqgof (see [SEM] estat eqgof).
The root mean squared error of approximation (Browne and Cudeck 1993) is computed as


RMSEA

The 90% confidence interval for RMSEA is

r

(2ms df ms )G
N df ms

GL
,
N df ms

GU
N df ms

1/2

where L and U are the noncentrality parameters corresponding to a noncentral chi-squared distribution
with df ms degrees of freedom in which the noncentral chi-squared random variable has a cumulative
distribution function equal to 0.95 and 0.05, respectively.
The Browne and Cudeck (1993) p-value for the test of close fit with null hypothesis

H0 : RMSEA 0.05
is computed as

p = 1 Pr(2 < 2ms |, df ms )

where 2 is distributed noncentral chi-squared with noncentrality parameter = (0.05)2 N df ms and

df ms degrees of freedom. This p-value is not computed when the group() option is specified.
The comparative fit index (Bentler 1990) is computed as


(2ms df ms )
CFI = 1
max{(2bs df bs ), (2ms df ms )}
The TuckerLewis index (Bentler 1990) is computed as
TLI

(2bs /df bs ) (2ms /df ms )

(2bs /df bs ) 1

Let k be the number of observed variables in the model. If means are not in the fitted model, the
standardized root mean squared residual is computed according to

SRMR

( Pk P
)1/2
2
2 i=1 ji rij
k(k + 1)G

(Hancock and Mueller 2006) where rij is the standardized covariance residual

rij =

sij

bij
p
sii sjj

bii
bjj

methods and formulas for sem Methods and formulas for sem

485

If means are in the fitted model, SRMR is computed as

SRMR

 1/2

P
P
2
2 ki=1 m2i + ji rij
k(k + 3)G

where mi is the standardized mean residual

zi

bi
mi =
sii

bii
These standardized residuals are not the same as those reported by estat residuals; see Residuals
below.
Group goodness of fit
estat ggof reports CD, SRMR, and model versus saturated 2 values for each group separately.
The group-level formulas are the same as those computed for a single group analysis; see Goodness
of fit above.
Equation-level goodness of fit
estat eqgof reports goodness-of-fit statistics for each endogenous variable in the specified model.
The coefficient of determination for the ith endogenous variable is computed as

Ri2 = 1

bii

bii

The BentlerRaykov (Bentler and Raykov 2000) squared multiple correlation for the ith endogenous
variable is computed as
d (yi , ybi )
Cov
mc2i = q
c (b

bii Var
yi )

b Var
c (b
where
bii is a diagonal element of ,
yi ) is a diagonal element of
n
o0 n
o n
o0
b ) = (I B)
b 1
b 1 + (I B)
b 1 I
b 1 I
b
b
b0 (I B)
b (I B)
c (Y
Var
d (yi , ybi ) is a diagonal element of
and Cov
n
o0
n
o0
b ) = (I B)
b 1
b 1 + (I B)
b 1
b 1 I
b
b
b0 (I B)
b (I B)
d (Y, Y
Cov

Wald tests
estat eqtest performs Wald tests on the coefficients for each endogenous equation in the model.
estat ginvariant computes a Wald test of group invariance for each model parameter that is free
to vary across all groups. See [R] test.

486

methods and formulas for sem Methods and formulas for sem

Score tests
estat mindices computes modification indices for each constrained parameter in the model,
including paths and covariances that were not even part of the model specification. Modification indices
are score tests, which are also known as Lagrange multiplier tests. estat scoretests performs a
score test for each user-specified linear constraint. estat ginvariant performs a score test of group
invariance for each model parameter that is constrained to be equal across all groups.
A score test compares a constrained model fit to the same model without one or more constraints.
The score test is computed as
2 = g(b
)0 V(b
)g(b
)
where b
is the fitted parameter vector from the constrained model, g() is the gradient vector function
for the unconstrained model, and V() is the variance matrix function computed from the expected
information matrix function for the unconstrained model. For method(ml) and method(mlmv),

log L ()

  2
1
log L ()
V() = E
0
g() =

where log L () is the log-likelihood function for the unconstrained model. For method(adf),

F ()
g() = WLS

  2
1
FWLS ()
V() = E
0
where F () is the WLS criterion function for the unconstrained model.
The score test is computed as described in Wooldridge (2010) when vce(robust) or vce(cluster
clustvar) is specified.
Residuals
estat residuals reports raw, normalized, and standardized residuals for means and covariances
of the observed variables.
The raw residual for the mean of the ith observed variable is

zi
bi
The raw residual for the covariance between the ith and j th observed variables is

Sij
bij
The normalized residual for the mean of the ith observed variable is

z
bi
qi
c (z i )
Var

methods and formulas for sem Methods and formulas for sem

487

where

Sii

, if the sample option is specified

N
c (z i ) =
Var

bii

, otherwise
N
The normalized residual for the covariance between the ith and j th observed variables is
S
bij
qij
c (Sij )
Var
where

Sii Sjj + Sij

, if the sample option is specified

N
c (Sij ) =
Var

bii
bjj +
bij

, otherwise
N
If the nm1 option is specified, the denominator in the variance estimates is N 1 instead of N .
The standardized residual for the mean of the ith observed variable is

zi
bi
q

c (z i
bi )
Var

where

c (z i
c (z i ) Var
c (b
Var
bi ) = Var
i )
c (b
and Var
i ) is computed using the delta method. Missing values are reported when the computed
c (b
c (z i ) is less than Var
i ). The standardized residual for the covariance between the ith
value of Var
and j th observed variables is
S
bij
q ij
c (Sij
Var
bij )
where

c (Sij
c (Sij ) Var
c (b
Var
bij ) = Var
ij )
c (b
and Var
ij ) is computed using the delta method. Missing values are reported when the computed
c (Sij ) is less than Var
c (b
value of Var
ij ). The variances of the raw residuals used in the standardized
residual calculations are derived in Hausman (1978).
Testing standardized parameters
estat stdize provides access to tests on the standardized parameter estimates. estat stdize
can be used as a prefix to lincom (see [R] lincom), nlcom (see [R] nlcom), test (see [R] test), and
testnl (see [R] testnl).
Stability of nonrecursive systems
estat stable reports a stability index for nonrecursive systems. The stability index is calculated
as the maximum of the modulus of the eigenvalues of B. The nonrecursive system is considered
stable if the stability index is less than 1.

488

methods and formulas for sem Methods and formulas for sem

Direct, indirect, and total effects

are

estat teffects reports direct, indirect, and total effects for the fitted model. The direct effects


b
b
Ed = B

the total effects are



b 1 I , (I B)
b 1
b
Et = (I B)
and the indirect effects are Ei = Et Ed . The standard errors of the effects are computed using
the delta method.
b
Let D be the diagonal matrix whose elements are the square roots of the diagonal elements of ,
and let DY be the submatrix of D associated with the endogenous variables. Then the standardized
effects are
e d = D1 Ed D
E
Y

e i = D1 Ei D
E
Y
e
Et = D1
Y Et D
Predictions
predict computes factor scores and linear predictions.
Factor scores are computed with a linear regression by using the mean vector and variance matrix
from the fitted model. For notational convenience, let
 
z
Z=
l
where

 
y
z=
x

and

 

l=

The fitted mean of Z is

bZ =

and fitted variance of Z is


bZ =

bz

bl

b zz

b 0zl

b zl

b ll

The factor scores are computed as

el =

 

e
b 0zl
b zz
=
bz +
bl
e

The linear prediction for the endogenous variables in the tth observation is computed as

bt = B
bY
et +
et +
bX
Y
b

methods and formulas for sem Methods and formulas for sem

where

yt

xt
e

et =
Y
and

et =
X

489

References
Akaike, H. 1974. A new look at the statistical model identification. IEEE Transactions on Automatic Control 19:
716723.
Bentler, P. M. 1990. Comparative fit indexes in structural models. Psychological Bulletin 107: 238246.
Bentler, P. M., and T. Raykov. 2000. On measures of explained variance in nonrecursive structural equation models.
Journal of Applied Psychology 85: 125131.
Bollen, K. A. 1989. Structural Equations with Latent Variables. New York: Wiley.
Browne, M. W., and R. Cudeck. 1993. Alternative ways of assessing model fit. Reprinted in Testing Structural
Equation Models, ed. K. A. Bollen and J. S. Long, pp. 136162. Newbury Park, CA: Sage.
Hancock, G. R., and R. O. Mueller, ed. 2006. Structural Equation Modeling: A Second Course. Charlotte, NC:
Information Age Publishing.
Hausman, J. A. 1978. Specification tests in econometrics. Econometrica 46: 12511271.
Schwarz, G. 1978. Estimating the dimension of a model. Annals of Statistics 6: 461464.
Wooldridge, J. M. 2010. Econometric Analysis of Cross Section and Panel Data. 2nd ed. Cambridge, MA: MIT Press.

Also see
[SEM] sem Structural equation model estimation command

Title
nlcom Nonlinear combinations of parameters
Syntax
Remarks and examples

Menu
Stored results

Description
Also see

Options

Syntax


nlcom exp , options

Menu
Statistics

>

SEM (structural equation modeling)

>

Testing and CIs

>

Nonlinear combinations of parameters

Description
nlcom is a postestimation command for use after sem, gsem, and other Stata estimation commands.
nlcom computes point estimates, standard errors, z statistics, p-values, and confidence intervals
for (possibly) nonlinear combinations of the estimated parameters. See [R] nlcom.

Options
See Options in [R] nlcom.

Remarks and examples

See [SEM] example 42g.
nlcom works in the metric of SEM, which is to say path coefficients, variances, and covariances. If
you want to frame your nonlinear combinations in terms of standardized coefficients and correlations
and you fit the model with sem, not gsem, then prefix nlcom with estat stdize:; see [SEM] estat
stdize.

Technical note
estat stdize: is, strictly speaking, unnecessary because everywhere you wanted a standardized
coefficient or correlation, you could just type the formula. If you did that, you would get the same
results except for numerical precision. The answer produced with the estat stdize: prefix will
be a little more accurate because estat stdize: is able to substitute an analytic derivative in one
part of the calculation where nlcom, doing the whole thing itself, would be forced to use a numeric
derivative.

490

nlcom Nonlinear combinations of parameters

Stored results
See Stored results in [R] nlcom.

Also see
[R] nlcom Nonlinear combinations of estimators
[SEM] estat stdize Test standardized parameters
[SEM] lincom Linear combinations of parameters
[SEM] test Wald test of linear hypotheses
[SEM] testnl Wald test of nonlinear hypotheses
[SEM] example 42g One- and two-level mediation models (multilevel)

491

Title
predict after gsem Generalized linear predictions, etc.
Syntax
Remarks and examples

Menu
Reference

Description
Also see

Options

Syntax
predict

type

 

stub* | newvarlist

if

 

in

 

, options obs endog

predict

type

 

stub* | newvarlist

if

 

in

 

, options obs endog options latent

The default is to predict observed endogenous variables with empirical Bayes means predictions of
the latent variables.
options obs endog
 
outcome(depvar # )1

Description

mu
pr
eta

calculate expected value of depvar

calculate probability (synonym for mu when is a probability)
calculate expected value of linear prediction of depvar

nooffset

make calculation ignoring Poisson or nbreg offset or exposure

fixedonly

make calculation setting latent variables to 0

or
use empirical Bayes means of latent variables; the default
use empirical Bayes modes of latent variables
# of integration points; default is intpoints(7) (means only)
tolerance; default is tolerance(1.0e-8) (means and modes)
maximum # of iterations; default is iterate(1001) (means and modes)

means
modes
intpoints(#)
tolerance(#)
iterate(#)
1

specify observed response variable (default all)

outcome(depvar #) is allowed only after mlogit, ologit, and oprobit. Predicting other generalized responses
requires specifying only outcome(depvar).
outcome(depvar #) may also be specified as outcome(#.depvar) or outcome(depvar ##).
outcome(depvar #3) means the third outcome value. outcome(depvar #3) would mean the same as
outcome(depvar 4) if outcomes were 1, 3, and 4.

492

predict after gsem Generalized linear predictions, etc.

493

options latent

Description

latent
latent(varlist)

calculate empirical Bayes prediction of all latent variables

calculate empirical Bayes prediction of specified latent variables

se(stub* | newvarlist)

calculate standard errors

means
modes
intpoints(#)
tolerance(#)
iterate(#)

use empirical Bayes means of latent variables; the default

use empirical Bayes modes of latent variables
# of integration points; default is intpoints(7) (means only)
tolerance; default is tolerance(1.0e-8) (means and modes)
maximum # of iterations; default is iterate(1001) (means and modes)

Either latent or latent() must be specified to obtain predictions of latent variables.

Menu
Statistics

>

SEM (structural equation modeling)

>

Predictions

Description
predict is a standard postestimation command of Stata. This entry concerns use of predict
after gsem. See [SEM] predict after sem if you fit your model with sem.
predict after gsem creates new variables containing observation-by-observation values of estimated
observed response variables, linear predictions of observed response variables, or endogenous or
exogenous latent variables.
Out-of-sample prediction is allowed in three cases:
1. if the prediction does not involve latent variables, or
2. if the prediction involves latent variables, directly or indirectly, option fixedonly is specified,
or
3. if the prediction involves latent variables, directly or indirectly, the model is multilevel and
no observational-level latent variables are involved.
predict has two ways of specifying the name(s) of the variable(s) to be created:
. predict stub*, . . .
or
. predict firstname secondname . . . , . . .
The first creates variables named stub1, stub2, . . . . The second creates variables named as you specify.
We strongly recommend using the stub* syntax when creating multiple variables because you have
no way of knowing the order in which to specify the individual variable names to correspond to the
order in which predict will make the calculations. If you use stub*, the variables will be labeled
and you can rename them.
The second syntax is useful when creating one variable and you specify either outcome() or
latent().

494

predict after gsem Generalized linear predictions, etc.

Options
 


outcome(depvar # ) and latent (varlist) determine what is to be calculated.
neither specified
outcome(depvar [#]) specified
latent specified
latent(varlist) specified

predict
predict
predict
predict

all observed response variables

specified observed response variable
all latent variables
specified latent variables

If you are predicting latent variables, both empirical Bayes means and modes are available; see
options means, modes, intpoints(#), tolerance(#), and iterate(#) below.

b ) or x;
b see options mu
If you are predicting observed response variables, you can obtain g 1 (x
and eta below. Predictions can include latent variables or treat them as 0; see option fixedonly.
If predictions include latent variables, then just as when predicting latent variables, both means and
modes are available; see options means, modes, intpoints(#), tolerance(#), and iterate(#).
b ) be calculated, the inverse-link of the expected value of the linear
mu and pr specify that g 1 (x
predictions. x by default contains predictions of latent variables. pr is a synonym for mu if response
variables are multinomial, ordinal, or Bernoulli. Otherwise, pr is not allowed.
b be calculated, the expected value of the linear prediction. x by default contains
eta specifies that x
predictions of latent variables.
fixedonly and nooffset are relevant only if observed response variables are being predicted.
fixedonly concerns predictions of latent variables used in the prediction of observed response
variables. fixedonly specifies latent variables be treated as 0, and thus only the fixed-effects part
of the model is used to produce the predictions.
nooffset is relevant only if option offset() or exposure() were specified at estimation time.
nooffset specifies that offset() or exposure() be ignored, thus producing predictions as if
all subjects had equal exposure.
means, modes, intpoints(#), tolerance(#), and iterate(#) specify what predictions of the
latent variables are to be calculated.
means and modes specify that empirical Bayes means or modes be used. Means are the default.
intpoints(#) specifies the number of numerical integration points and is relevant only in the
calculation of empirical Bayes means. intpoints() defaults to the number of integration points
specified at estimation time or to intpoints(7).
tolerance(#) is relevant for the calculation of empirical Bayes means and modes. It specifies the
convergence tolerance. It defaults to the value specified at estimation time with gsems adaptopts()
or to tolerance(1e-8).
iterate(#) is relevant for the calculation of empirical Bayes means and modes. It specifies the
maximum number of iterations to be performed in the calculation of each integral. It defaults to
the value specified at estimation time with gsems adaptopts() or to tolerance(1e-8).

Remarks and examples

See [SEM] intro 7, [SEM] example 28g, and [SEM] example 29g.

predict after gsem Generalized linear predictions, etc.

495

Reference
Skrondal, A., and S. Rabe-Hesketh. 2009. Prediction in multilevel generalized linear models. Journal of the Royal
Statistical Society, Series A 172: 659687.

Also see
[SEM] gsem Generalized structural equation model estimation command
[SEM] gsem postestimation Postestimation tools for gsem
[SEM] intro 7 Postestimation tests and predictions
[SEM] example 28g One-parameter logistic IRT (Rasch) model
[SEM] example 29g Two-parameter logistic IRT model

[SEM] methods and formulas for gsem Methods and formulas

Title
predict after sem Factor scores, linear predictions, etc.
Syntax
Remarks and examples

Menu
Reference

Description
Also see

Options

Syntax
predict

type

 

stub* | newvarlist

if

 

in

 

, options

options

Description

xb
xb(varlist)

linear prediction for all OEn variables; the default

linear prediction for specified OEn variables

xblatent
xblatent(varlist)

linear prediction for all LEn variables

linear prediction for specified LEn variables

latent
latent(varlist)

factor scores for all latent variables

factor scores for specified latent variables

scores

calculate first derivative of the log likelihood

Key: OEn = observed endogenous; LEn = latent endogenous

Menu
Statistics

>

SEM (structural equation modeling)

>

Predictions

Description
predict is a standard postestimation command of Stata. This entry concerns use of predict
after sem. See [SEM] predict after gsem if you fit your model with gsem.
predict after sem creates new variables containing observation-by-observation values of estimated
factor scores (meaning predicted values of latent variables) and predicted values for latent and observed
endogenous variables. Out-of-sample prediction is allowed.
When predict is used on a model fit by sem with the group() option, results are produced
with the appropriate group-specific estimates. Out-of-sample prediction is allowed; missing values are
filled in for groups not included at the time the model was fit.
predict allows two syntaxes. You can type
. predict stub*,

...

to create variables named stub1, stub2, . . . , or you can type

. predict var1 var2 . . . ,

...

to create variables named var1, var2, . . . .

predict may not be used with summary statistics data.
496

predict after sem Factor scores, linear predictions, etc.

497

Options
xb calculates the linear prediction for all observed endogenous variables in the model. xb is the
default if no option is specified.
xb(varlist) calculates the linear prediction for the variables specified, all of which must be observed
endogenous variables.
xblatent and xblatent(varlist) calculate the linear prediction for all or the specified latent
endogenous variables, respectively.
latent and latent(varlist) calculate the factor scores for all or the specified latent variables,
respectively. The calculation method is an analog of regression scoring; namely, it produces the
means of the latent variables conditional on the observed variables used in the model. If missing
values are found among the observed variables, conditioning is on the variables with observed
values only.
scores is for use by programmers. It provides the first derivative of the observation-level log likelihood
with respect to the parameters.
Programmers: In single-group sem, each parameter that is not constrained to be 0 has an associated
equation. As a consequence, the number of equations, and hence the number of score variables
created by predict, may be large.

Remarks and examples

See [SEM] example 14.
Factor scoring for latent variables can be interpreted as a form of missing-value imputationthink
of each latent variable as an observed variable that has only missing values.
When latent variables are present in the model, linear predictions from predict, xb are computed
by substituting the factor scores in place of each latent variable before computing the linear combination
of coefficients. This method will lead to inconsistent coefficient estimates when the factor score contains
measurement error; see Bollen (1989, 305306).

Reference
Bollen, K. A. 1989. Structural Equations with Latent Variables. New York: Wiley.

Also see
[SEM] example 14 Predicted values

[SEM] methods and formulas for sem Methods and formulas for sem
[SEM] sem postestimation Postestimation tools for sem

Title
sem Structural equation model estimation command

Syntax
Remarks and examples

Menu
Stored results

Description
Reference

Options
Also see

Syntax
sem paths

if

 

in

 

weight

 

, options

where paths are the paths of the model in command-language path notation; see [SEM] sem and gsem
path notation.
options

Description

model description options

fully define, along with paths, the model to be fit

group options

fit model for different groups

ssd options

for use with summary statistics data

estimation options

method used to obtain estimation results

reporting options

reporting of estimation results

syntax options

controlling interpretation of syntax

Time-series operators are allowed.

bootstrap, by, jackknife, permute, statsby, and svy are allowed; see [U] 11.1.10 Prefix commands.
Weights are not allowed with the bootstrap prefix; see [R] bootstrap.
vce() and weights are not allowed with the svy prefix; see [SVY] svy.
fweights, iweights, and pweights are allowed; see [U] 11.1.6 weight.
Also see [SEM] sem postestimation for features available after estimation.

Menu
Statistics

>

SEM (structural equation modeling)

>

Model building and estimation

Description
sem fits structural equation models. Even when you use the SEM Builder, you are using the sem
command.

Options
model description options describe the model to be fit. The model to be fit is fully specified by
pathswhich appear immediately after semand the options covariance(), variance(), and
means(). See [SEM] sem model description options and [SEM] sem and gsem path notation.
498

sem Structural equation model estimation command

499

group options allow the specified model to be fit for different subgroups of the data, with some
parameters free to vary across groups and other parameters constrained to be equal across groups.
See [SEM] sem group options.
ssd options allow models to be fit using summary statistics data (SSD), meaning data on means,
variances (standard deviations), and covariances (correlations). See [SEM] sem ssd options.
estimation options control how the estimation results are obtained. These options control how the
standard errors (VCE) are obtained and control technical issues such as choice of estimation method.
See [SEM] sem estimation options.
reporting options control how the results of estimation are displayed. See [SEM] sem reporting
options.
syntax options control how the syntax that you type is interpreted. See [SEM] sem and gsem syntax
options.

Remarks and examples

For a readable explanation of what sem can do and how to use it, see any of the intro sections.
You might start with [SEM] intro 1.
For examples of sem in action, see any of the example sections. You might start with [SEM] example 1.
For detailed syntax and descriptions, see the references below.
Remarks on three advanced topics are presented under the following headings:
Default normalization constraints
Default covariance assumptions
How to solve convergence problems

Default normalization constraints

sem applies the following rules as necessary to identify the model:
1. means(1: LatentExogenous@0)
sem constrains all latent exogenous variables to have mean 0. When the group() option is
specified, this rule is applied to the first group only.
2. (LatentEndogenous <- cons@0)
sem sets all latent endogenous variables to have intercept 0.
3. (first <- Latent@1)
If latent variable Latent is measured by observed endogenous variables, then sem sets the path
coefficient of (first<-Latent) to be 1; first is the first observed endogenous variable.
4. (First<-Latent@1)
If item 3 does not applyif latent variable Latent is measured by other latent endogenous
variables onlythen sem sets the path coefficient of First<-Latent to be 1; First is the first
latent variable.
The above constraints are applied as needed. Here are the available overrides:
1. To override the normalization constraints, specify your own constraints. Most normalization
constraints are added by sem as needed. See How sem (gsem) solves the problem for you under
Identification 2: Normalization constraints (anchoring) in [SEM] intro 4.

500

sem Structural equation model estimation command

2. To override means() constraints, you must use the means() option to free the parameter.
To override that the mean of latent exogenous variable MyLatent has mean 0, specify the
means(MyLatent) option. See [SEM] sem and gsem path notation.
3. To override constrained path coefficients from cons, such as (LatentEndogenous <cons@0), you must explicitly specify the path without a constraint (LatentEndogenous
<- cons). See [SEM] sem and gsem path notation.

Default covariance assumptions

sem assumes the following covariance structure:
1. covstructure( Ex, unstructured)
All exogenous variables (observed and latent) are assumed to be correlated with each other.
2. covstructure(e. En, diagonal)
The error variables associated with all endogenous variables are assumed to be uncorrelated with
each other.
You can override these assumptions by
1. Constraining or specifying the relevant covariance to allow e.y and e.Ly to be correlated (specify
the covariance(e.y*e.Ly) option); see [SEM] sem model description options.
2. Using the covstructure() option; see [SEM] sem and gsem option covstructure( ).

How to solve convergence problems

See [SEM] intro 12.

Stored results
sem stores the following in e():
Scalars
e(N)
e(N clust)
e(N groups)
e(N missing)
e(ll)
e(df m)
e(df b)
e(df s)
e(chi2 ms)
e(df ms)
e(p ms)
e(chi2 bs)
e(df bs)
e(p bs)
e(rank)
e(ic)
e(rc)
e(converged)
e(critvalue)
e(critvalue b)
e(critvalue s)
e(modelmeans)

number of observations
number of clusters
number of groups
number of missing values in the sample for method(mlmv)
log likelihood of model
model degrees of freedom
baseline model degrees of freedom
saturated model degrees of freedom
test of target model against saturated model
degrees of freedom for e(chi2 ms)
p-value for e(chi2 ms)
test of baseline model against saturated model
degrees of freedom for e(chi2 bs)
p-value for e(chi2 bs)
rank of e(V)
number of iterations
return code
1 if target model converged, 0 otherwise
log likelihood or discrepancy of fitted model
log likelihood or discrepancy of baseline model
log likelihood or discrepancy of saturated model
1 if fitting means and intercepts, 0 otherwise

sem Structural equation model estimation command

Macros
e(cmd)
e(cmdline)
e(data)
e(wtype)
e(wexp)
e(title)
e(clustvar)
e(vce)
e(vcetype)
e(method)
e(technique)
e(properties)
e(estat cmd)
e(predict)
e(lyvars)
e(oyvars)
e(lxvars)
e(oxvars)
e(groupvar)
e(xconditional)
Matrices
e(b)
e(b std)
e(b pclass)
e(admissible)
e(ilog)
e(gradient)
e(V)
e(V std)
e(V modelbased)
e(nobs)
e(groupvalue)
 
e(S # )

e(means # )
e(W)
Functions
e(sample)

501

sem
command as typed
raw or ssd if SSD were used
weight type
weight expression
title in estimation output
name of cluster variable
vcetype specified in vce()
title used to label Std. Err.
estimation method: ml, mlmv, or adf
maximization technique
b V
program used to implement estat
program used to implement predict
names of latent y variables
names of observed y variables
names of latent x variables
names of observed x variables
name of group variable
empty if noxconditional specified, xconditional otherwise
parameter vector
standardized parameter vector
parameter class
admissibility of , ,
iteration log (up to 20 iterations)
gradient vector
covariance matrix of the estimators
standardized covariance matrix of the estimators
model-based variance
vector with number of observations per group
vector of group values of e(groupvar)
sample covariance matrix of observed variables (for group #)
sample means of observed variables (for group #)
weight matrix for method(adf)
marks estimation sample (not with SSD)

Reference
Wiggins, V. L. 2011. Multilevel random effects in xtmixed and semthe long and wide of it. The Stata Blog: Not
Elsewhere Classified.
http://blog.stata.com/2011/09/28/multilevel-random-effects-in-xtmixed-and-sem-the-long-and-wide-of-it/.

502

sem Structural equation model estimation command

Also see
[SEM] intro 1 Introduction
[SEM] sem and gsem path notation Command syntax for path diagrams
[SEM] sem path notation extensions Command syntax for path diagrams
[SEM] sem model description options Model description options
[SEM] sem group options Fitting models on different groups
[SEM] sem ssd options Options for use with summary statistics data
[SEM] sem estimation options Options affecting estimation
[SEM] sem reporting options Options affecting reporting of results
[SEM] sem and gsem syntax options Options affecting interpretation of syntax
[SEM] sem postestimation Postestimation tools for sem
[SEM] methods and formulas for sem Methods and formulas for sem

Title
sem and gsem option constraints( ) Specifying constraints

Syntax

Description

Remarks and examples

Also see

Syntax


# ... ) ...




gsem . . . , . . . constraints(# # . . . ) . . .
sem . . .

, . . . constraints(#

where # are constraint numbers. Constraints are defined by the constraint command; see [R] constraint.

Description
Constraints refer to constraints to be imposed on the estimated parameters of a model. These
constraints usually come in one of three forms:
1. Constraints that a parameter such as a path coefficient or variance is equal to a fixed value such
as 1.
2. Constraints that two or more parameters are equal.
3. Constraints that two or more parameters are related by a linear equation.
It is usually easier to specify constraints with sems and gsems path notation; see [SEM] sem and
gsem path notation.
sems and gsems constraints() option provides an alternative way of specifying constraints.

Remarks and examples

Remarks are presented under the following headings:
Use with sem
Use with gsem

Also see [R] constraint.

Use with sem

There is only one case where constraints() might be easier to use with sem instead of specifying
constraints in the path notation. You wish to specify that two or more parameters are related and then
decide you would like to fix the value at which they are related.
For example, if you wanted to specify that parameters are equal, you could type
. sem ... (y1 <- x@c1) (y2 <- x@c1)

(y3 <- x@c1)

...

Using the path notation, you can specify more general relationships, too, such as
. sem ... (y1 <- x@c1) (y2 <- x@(2*c1)) (y3 <- x@(3*c1+1)) ...
503

504

sem and gsem option constraints( ) Specifying constraints

Say you now decide you want to fix c1 at value 1. Using the path notation, you modify what you
previously typed:
. sem ... (y1 <- x@1) (y2 <- x@2)

(y3 <- x@4)

...

Alternatively, you could do the following:

.
.
.
.
.

constraint 1 _b[y2:x] = 2*_b[y1:x]

constraint 2 _b[y3:x] = 3*_b[y1:x] + 1
sem ..., ... constraints(1 2)
constraint 3 _b[y1:x] = 1
sem .., ... constraints(1 2 3)

Use with gsem

Gamma regression can produce exponential regression estimates if you constrain the log of the
scale parameter to 0. Parameters associated with particular generalized linear families, such as scalar
parameters, cutpoints, and the like, cannot be constrained using the @ notation in paths. You must
use Statas constraints.
Say we wish to fit the model y <- x1 with exponential regression. We admit that we do not
remember the name under which gsem stores the scalar parameter, so first we type
. gsem (y <- x1, gamma), noestimate
From the output, we quickly discover that the log of the scale parameter is stored as
b[y logs: cons]. With that information, to obtain the constrained results, we type
. constraint 1 _b[y_logs:_cons] = 0
. gsem (y <- x1, gamma), constraints(1)

Also see
[SEM] sem Structural equation model estimation command
[SEM] gsem Generalized structural equation model estimation command
[SEM] sem and gsem path notation Command syntax for path diagrams
[SEM] gsem model description options Model description options
[SEM] sem model description options Model description options
[R] constraint Define and list constraints

Title
sem and gsem option covstructure( ) Specifying covariance restrictions

Syntax

Description

Option

Remarks and examples

Also see

Syntax
sem . . .

, . . . covstructure(variables, structure) . . .

sem . . .

, . . . covstructure(groupid: variables, structure) . . .

gsem . . .

, . . . covstructure(variables, structure) . . .

where variables is one of

1. a list of (a subset of the) exogenous variables (sem) or latent exogenous variables (gsem) in your
model, for instance,
. sem ..., ... covstruct(x1 x2, structure)
. sem ..., ... covstruct(L1 L2, structure)
. gsem ..., ... covstruct(L1 L2, structure)
2.
3.
4.

OEx, meaning all observed exogenous variables in your model (sem only)
LEx, meaning all latent exogenous variables in your model (including any multilevel latent
exogenous variables in the case of gsem)
Ex, meaning all exogenous variables in your model (sem only)

or where variables is one of

1. a list of (a subset of the) error variables in your model, for example,
. sem ..., ... covstruct(e.y1 e.y2 e.Aspect, structure)
2. e. OEn, meaning all error variables associated with observed endogenous variables in your model
3. e. LEn, meaning all error variables associated with latent endogenous variables in your model
4. e. En, meaning all error variables in your model

505

506

sem and gsem option covstructure( ) Specifying covariance restrictions

and where structure is

structure

Description

diagonal

all variances unrestricted

all covariances fixed at 0

unstructured

all variances unrestricted

all covariances unrestricted

identity

all variances equal

all covariances fixed at 0

exchangeable

all variances equal

all covariances equal

zero

all variances fixed at 0

all covariances fixed at 0

pattern(matname)

covariances (variances) unrestricted if

matname[i, j] .
covariances (variances) equal if
matname[i, j] = matname[k, l]

fixed(matname)

covariances (variances) unrestricted if

matname[i, j] .
covariances (variances) fixed at
matname[i, j] otherwise

Notes

(1)

(2)

Notes:
(1) Only elements in the lower triangle of matname are used. All values in matname are interpreted
as the floor() of the value if noninteger values appear. Row and column stripes of matname
are ignored.
(2) Only elements on the lower triangle of matname are used. Row and column stripes of matname
are ignored.
groupid may be specified only when the group() option is also specified, and even then it is optional;
see [SEM] sem group options.

Description
Option covstructure() provides a sometimes convenient way to constrain the covariances of
your model.
Alternatively or in combination, you can place constraints on the covariances by using the standard
path notation, such as
. sem ..., ... cov(name1*name2@c1 name3*name4@c1) ...
. gsem ..., ... cov(name1*name2@c1 name3*name4@c1) ...
See [SEM] sem and gsem path notation.

sem and gsem option covstructure( ) Specifying covariance restrictions

507

Option


covstructure( groupid: variables, structure) is used either to modify the covariance structure
among the exogenous variables of your model or to modify the covariance structure among the
error variables of your model. Optional groupid is available only with sem with option group()
specified; see [SEM] sem group options.
You may specify the covstructure() option multiple times.
The default covariance structure for exogenous variables is covstructure( Ex, unstructured)
for sem. There is no simple way in this notation to write the default for gsem.
The default covariance structure for error variables is covstructure(e. En, diagonal) for
sem and gsem.

Remarks and examples

See [SEM] example 17.
Standard linear structural equation modeling allows covariances among exogenous variables, both
latent and observed, and allows covariances among the error variables. Covariances between exogenous
variables and error variables are disallowed (assumed to be 0).
Some authors refer to the covariances among the exogenous variables in standard SEMs as matrix
and to the covariances among the error variables as matrix .

Also see
[SEM] sem Structural equation model estimation command
[SEM] gsem Generalized structural equation model estimation command
[SEM] sem and gsem path notation Command syntax for path diagrams
[SEM] example 17 Correlated uniqueness model

Title
sem and gsem option from( ) Specifying starting values

Syntax

Description

Option

Remarks and examples

Also see

Syntax
{sem | gsem} . . .




, . . . from(matname , skip ) . . .

{sem | gsem} . . .

, . . . from(svlist) . . .

where matname is the name of a Stata matrix and

where svlist is a space-separated list of the form
eqname:name = #

Description
See [SEM] intro 12 for a description of starting values.
Starting values are usually not specified. When there are convergence problems, it is often necessary
to specify starting values. You can specify starting values by using
1. suboption init() as described in [SEM] sem and gsem path notation, or by using
2. option from() as described here.
Option from() is especially convenient for using the solution of one model as starting values for
another.

Option
skip is an option of from(matname). It specifies to ignore any parameters in matname that do not
appear in the model being fit. If this option is not specified, the existence of such parameters
causes sem (gsem) to issue an error message.

Remarks and examples

Remarks are presented under the following headings:
Syntax 1, especially useful when dealing with convergence problems
Syntax 2, seldom used

Option from() can be used with sem or gsem. We illustrate below using sem.

508

sem and gsem option from( ) Specifying starting values

509

Syntax 1, especially useful when dealing with convergence problems

Say you are attempting to fit
. sem your_full_model, ...
and are having difficulty with convergence. Following the advice in [SEM] intro 12, you have simplified
your model,
. sem your_simple_model, ...
and that does converge. You now want to use those values as starting values for the full model. Lets
imagine that there are 47 estimated parameters in the simple model.
Using the standard init() method for specifying starting values, you now have a real job in front
of you. You have to type your full model, find all the places where you want to add starting values,
and add an init() suboption. Just a piece of your full model might read
... (y<-L1 L2) (L1->x1 x2) (L2->x3 L4) ...
and you need to modify that to read
... (y<-(L1, init(14.283984)) L2)
(L1->(x1, init(2.666532)) (x2, init(-6.39499)))
(L2->x3 L4) ...

///
///

That change handles 3 of the 47 parameters you need to specify.

There is an easier way. Type
. sem your_simple_model, ...
. matrix b = e(b)
. sem your_full_model, ... from(b)
Here is how this works:
1. You fit the simple model. sem stores the resulting parameters in e(b).
2. You store the fitted parameters in Stata matrix b.
3. To fit your full model, type the model just as you would usually, and add option from(b).
That option tells sem to get any starting values it can from Stata matrix b. sem does that
and then follows its usual logic for producing starting values for the remaining parameters.
Just because you use from(b) does not mean you cannot specify starting values with init().
You can even specify starting values for some of the same parameters. Starting values specified by
init() take precedence over those obtained from from().

Syntax 2, seldom used

In syntax 2, you specify
. sem ..., ... from(eqname:name = # eqname:name = # ...)
For instance, you could type
. sem ..., ... from(var(X):_cons=10)
or you could type
. sem ..., ... var(X, init(10))
It is usually easier to type the second. See [SEM] sem and gsem path notation.

510

sem and gsem option from( ) Specifying starting values

You may combine the two notations. If starting values are specified for a parameter both ways,
those specified by init() take precedence.

Also see
[SEM] sem Structural equation model estimation command
[SEM] gsem Generalized structural equation model estimation command
[SEM] sem and gsem path notation Command syntax for path diagrams
[SEM] gsem model description options Model description options
[SEM] sem model description options Model description options
[SEM] gsem estimation options Options affecting estimation
[R] maximize Details of iterative maximization

Title
sem and gsem option reliability( ) Fraction of variance not due to measurement error
Syntax

Description

Option

Remarks and examples

Also see

Syntax
{sem | gsem} . . .

, . . . reliability(varname #

varname #

...

 
)

where varname is the name of an observed endogenous variable and # is the fraction or percentage
of variance not due to measurement error:
.
.

{sem | gsem} ..., ... reliability(x1 .8 x2 .9)

{sem | gsem} ..., ... reliability(x1 80% x2 90%)

Description
Option reliability() allows you to specify the fraction of variance not due to measurement
error for measurement variables.

Option


reliability(varname # . . . ) specifies the reliability for variable varname. Reliability is bounded
by 0 and 1 and is equal to
noise variance
1
total variance
The reliability is assumed to be 1 when not specified.

Remarks and examples

See [SEM] example 24.
Remarks are presented under the following headings:
Background
Dealing with measurement error of exogenous variables
Dealing with measurement error of endogenous variables
What can go wrong

Background
Option reliability() may be used with sem and may be used with gsem for Gaussian response
variables with the identity link but only in the absence of censoring. We illustrate using sem, but we
could just as well have used gsem.
Variables measured with error have attenuated path coefficients. If we had the model
. sem (y<-x)

511

512

sem and gsem option reliability( ) Fraction of variance not due to measurement error

and x were measured with error, then the estimated path coefficient would be biased toward 0. The
usual solution to such measurement problems is to find multiple measurements and develop a latent
variable from them:
. sem (x1 x2 x3<-X) (y<-X)

Another solution is available if we know the reliability of x. In that case, we can fit the model
. sem (x<-X) (y<-X), reliability(x .9)

The two solutions can even be combined:

. sem (x1 x2 x3<-X) (y<-X), reliability(x1 .9

x2 .8

x3 .9)

Even if you do not know the reliability, you can experiment using different but reasonable values
for the reliability and thus determine the sensitivity of your estimation results to the measurement
problem.

Dealing with measurement error of exogenous variables

Measurement error is most important when it occurs in exogenous variables, yet the reliability()
option deals with measurement error of endogenous variables only. By creation of a latent variable,
reliability() can deal with the measurement error of exogenous variables.
To fit the model (y<-x) where x is measured with error, you must introduce a latent variable
corresponding to x measured without error. That is, the model (y<-x) can be converted into the
model (x<-X) and (y<-X):

x = 0 + 0 X + e.x

y = 1 + 1 X + e.y
To fit this model, you type
. sem (x<-X) (y<-X), reliability(x .9)

sem will introduce a normalization constraint, namely, that the path coefficient 0 for x<-X is 1, but
that is of no importance. What is important is that the estimate of that path coefficient 1 of y<-X
is the coefficient that would be obtained from y<-x were x measured without error.
In the above, we specified the measurement part of the model first. Be sure to do that. You might
think you could equally well reverse the two terms so that, rather than writing
(x<-X) (y<-X)

(correct)

you could write

(y<-X) (x<-X)

(incorrect)

But you cannot do that unless you write

(y<-X) (x<-X@1)

(correct)

because otherwise results are as if you typed

(y<-X@1) (x<-X)

(incorrect)

sem and gsem option reliability( ) Fraction of variance not due to measurement error

513

All of that is because sem places its normalization constraint from the latent variable to the first
observed endogenous variable. There is no real error if the terms are interchanged except that you
will be surprised by the coefficient of 1 for y<-X and (the reciprocal of) the coefficient of interest
will be on x<-X.
See How sem (gsem) solves the problem for you in [SEM] intro 4 and see Default normalization
constraints in [SEM] sem.

Dealing with measurement error of endogenous variables

When a variable would already be endogenous before you add the reliability() option, it really
makes little difference whether you add the reliability() option. That is because endogenous
variables are assumed to contain error, and if some of that error is measurement error, it is still just
an error. Coefficients will be unchanged by the inclusion of the reliability() option.
Some variances and covariances will change, but the changes are offsetting in the calculation of
path coefficients.
What will change are the standardized coefficients should you ask to see them. That is because
the variances are changed.

What can go wrong

Consider a model of y on x. Say we fit the model with linear regression. If the R2 of the fit is
0.6, then we know the reliability must be greater than 0.6. R2 measures the fraction of variance of
y that is explained by x, and the reliability of x measures the fraction of the variance of x that is
not due to measurement error. Measurement error is assumed to be pure noise. It is just not possible
that we could explain 0.6 of the variance of y by using a variable with reliability of, say, 0.5.
Well, in fact, it is because there is always a chance that the pure noise will correlate with y, too.
Asymptotically, that probability vanishes, but in finiteespecially smallsamples, it could happen.
Even so, the calculation of the corrected SEM estimates blows up.
If you have convergence problems, you need to check for this. Specify a larger value for the
reliability. The problem is you cannot specify a value of 1, and large values such as 0.99999 can lead
to a lack of identification. In most cases, you will be able to find a value in between, but the only
way to be sure is to remove the reliability() option and, if necessary, simplify your model by
removing any intermediary latent variables you had to add because of reliability.
If your model converges without reliability, then your measure of reliability is too low. At this
point, we have little useful advice for you. Check whether you have the right value, of course. If you
do, then there are two possibilities: either the experts who provided that estimate are wrong or you
got unlucky in that the measurement error did just happen to correlate with the rest of your data. You
will need to evaluate the chances of that for yourself. In any case, you can experiment with higher
values of the reliability and at least provide an idea of the sensitivity of your estimates to differing
assumptions.

Also see
[SEM] sem Structural equation model estimation command
[SEM] gsem Generalized structural equation model estimation command
[SEM] gsem model description options Model description options
[SEM] sem model description options Model description options
[SEM] example 24 Reliability

Title
sem and gsem path notation Command syntax for path diagrams

Syntax

Description

Options

Remarks and examples

Also see

Syntax
sem paths . . .

, covariance() variance() means()

gsem paths . . .

, covariance() variance() means()

paths specifies the direct paths between the variables of your model.
The model to be fit is fully described by paths, covariance(), variance(), and means().

Description
The command syntax for describing your SEM is fully specified by paths, covariance(),
variance(), and means(). How this works is described below.
If you are using sem, also see [SEM] sem path notation extensions for documentation of the
group() option for comparing different groups in the data. The syntax of the elements described
below is modified when group() is specified.
If you are using gsem, also see [SEM] gsem path notation extensions for documentation on
specification of family-and-link for generalized (nonlinear) response variables and for specification of
multilevel latent variables.
Either way, read this section first.

Options
covariance() is used to
1. specify that a particular covariance path of your model that usually is assumed to be 0 be
estimated,
2. specify that a particular covariance path that usually is assumed to be nonzero is not to be
estimated (to be constrained to be 0),
3. constrain a covariance path to a fixed value, such as 0, 0.5, 1, etc., and
4. constrain two or more covariance paths to be equal.
variance() does the same as covariance() except it does it with variances.
means() does the same as covariance() except it does it with means.

Remarks and examples

Path notation is used by the sem and gsem commands to specify the model to be fit, for example,
. sem (x1 x2 x3 x4 <- X)
. gsem (L1 -> x1 x2 x3 x4 x5, logit) (L2 -> x6 x7 x8 x9 x10)

514

sem and gsem path notation Command syntax for path diagrams

515

In the path notation,

1. Latent variables are indicated by a name in which at least the first letter is capitalized.
2. Observed variables are indicated by a name in which at least the first letter is lowercased.
Observed variables correspond to variable names in the dataset.
3. Error variables, while mathematically a special case of latent variables, are considered in a class
by themselves. For sem, every endogenous variable (whether observed or latent) automatically has
an error variable associated with it. For gsem, the same is true of Gaussian endogenous variables
(and latent variables, which are Gaussian). The error variable associated with endogenous variable
name is e.name.
4. Paths between variables are written as
(name1 <- name2)
or
(name2 -> name1)
There is no significance to which coding is used.
5. Paths between the same variables can be combined: The paths
(name1 <- name2) (name1 <- name3)
can be combined as
(name1 <- name2 name3)
or as
(name2 name3 -> name1)
The paths
(name1 <- name3) (name2 <- name3)
can be combined as
(name1 name2 <- name3)
or as
(name3 -> name1 name2)
The paths
(name1 <- name2 name3)
(name4 <- name2 name3)
may be written as
(name1 name4 <- name2 name3)
or as
(name2 name3 -> name1 name4)

516

sem and gsem path notation Command syntax for path diagrams

6. Variances and covariances (curved paths) between variables are indicated by options. Variances
are indicated by
..., ... var(name1)
Covariances are indicated by
..., ... cov(name1*name2)
..., ... cov(name2*name1)
There is no significance to the order of the names.
The actual names of the options are variance() and covariance(), but they are invariably
abbreviated as var() and cov(), respectively.
The var() and cov() options are the same option, so a variance can be typed as
..., ... cov(name1)
and a covariance can be typed as
..., ... var(name1*name2)
7. Variances may be combined, covariances may be combined, and variances and covariances may
be combined.
If you have
..., ... var(name1) var(name2)
you may code this as
..., ... var(name1 name2)
If you have
..., ... cov(name1*name2) cov(name2*name3)
you may code this as
..., ... cov(name1*name2 name2*name3)
All the above combined can be coded as
..., ... var(name1 name2 name1*name2 name2*name3)
or as
..., ... cov(name1 name2 name1*name2 name2*name3)
8. All variables except endogenous variables are assumed to have a variance; it is only necessary
to code the var() option if you wish to place a constraint on the variance or specify an initial
value. See items 11, 12, 13, and 16 below. (In gsem, the variance and covariances of observed
exogenous variables are not estimated and thus var() cannot be used with them.)
Endogenous variables have a variance, of course, but that is the variance implied by the model. If
name is an endogenous variable, then var(name) is invalid. The error variance of the endogenous
variable is var(e.name).
9. Variables mostly default to being correlated:
a. All exogenous variables are assumed to be correlated with each other, whether observed or
latent.
b. Endogenous variables are never directly correlated, although their associated error variables
can be.
c. All error variables are assumed to be uncorrelated with each other.

sem and gsem path notation Command syntax for path diagrams

517

You can override these defaults on a variable-by-variable basis with the cov() option.
To assert that two variables are uncorrelated that otherwise would be assumed to be correlated,
constrain the covariance to be 0:
..., ... cov(name1*name2@0)
To allow two variables to be correlated that otherwise would be assumed to be uncorrelated,
simply specify the existence of the covariance:
..., ... cov(name1*name2)
This latter is especially commonly done with errors:
..., ... cov(e.name1*e.name2)
(In gsem, you may not use the cov() option with observed exogenous variables. You also may
not use cov() with error terms associated with family Gaussian, link log.)
10. Means of variables are indicated by the following option:
..., ... means(name)
Variables mostly default to having nonzero means:
a. All observed exogenous variables are assumed to have nonzero means. In sem, the means can
be constrained using the means() option, but only if you are performing noxconditional
estimation; [SEM] sem option noxconditional.
b. Latent exogenous variables are assumed to have mean 0. Means of latent variables are not
estimated by default. If you specify enough normalization constraints to identify the mean
of a latent exogenous variable, you can specify means(name) to indicate that the mean
should be estimated in either.
c. Endogenous variables have no separate mean. Their means are those implied by the model.
The means() option may not be used with endogenous variables.
d. Error variables have mean 0 and this cannot be modified. The means() option may not be
used with error variables.
To constrain the mean to a fixed value, such as 57, code
..., ... means(name@57)
Separate means() options may be combined:
..., ... means(name1@57 name2@100)
11. Fixed-value constraints may be specified for a path, variance, covariance, or mean by using @
(the at symbol). For example,
(name1 <- name2@1)
(name1 <- name2@1 name3@1)
..., ... var(name@100)
..., ... cov(name1*name2@223)
..., ... cov(name1@1 name2@1 name1*[email protected])
..., ... means(name@57)

518

sem and gsem path notation Command syntax for path diagrams

12. Symbolic constraints may be specified for a path, variance, covariance, or mean by using @ (the
at symbol). For example,
(name1 <- name2@c1) (name3 <- name4@c1)
..., ... var(name1@c1 name2@c1)
..., ... cov(name1@1 name2@1 name3@1 name1*name2@c1 name1*name3@c1)
..., ... means(name1@c1 name2@c1)
(name1 <- name2@c1) ..., var(name3@c1) means(name4@c1)
Symbolic names are just names from 1 to 32 characters in length. Symbolic constraints constrain
equality. For simplicity, all constraints below will have names c1, c2, . . . .
13. Linear combinations of symbolic constraints may be specified for a path, variance, covariance,
or mean by using @ (the at symbol). For example,
(name1 <- name2@c1) (name3 <- name4@(2*c1))
..., ... var(name1@c1 name2@(c1/2))
..., ... cov(name1@1 name2@1 name3@1 name1*name2@c1 name1*name2@(c1/2))
..., ... means(name1@c1 name2@(3*c1+10))
(name1 <- name2@(c1/2)) ..., var(name3@c1) means(name4@(2*c1))
14. All equations in the model are assumed to have an intercept (to include observed exogenous
variable cons) unless the noconstant option (abbreviation nocons) is specified, and then
all equations are assumed not to have an intercept (not to include cons). (There are some
exceptions to this in gsem because some generalized linear models have no intercept or even the
concept of an intercept.)
Regardless of whether noconstant is specified, you may explicitly refer to observed exogenous
variable cons.
The following path specifications are ways of writing the same model:
(name1 <- name2) (name1 <- name3)
(name1 <- name2) (name1 <- name3) (name1 <- cons)
(name1 <- name2 name3)
(name1 <- name2 name3 cons)
There is no reason to explicitly specify cons unless you have also specified the noconstant
option and want to include cons in some equations but not others, or regardless of whether
you specified the noconstant option, you want to place a constraint on its path coefficient. For
example,
(name1 <- name2 name3 cons@c1) (name4 <- name5 cons@c1)
15. The noconstant option may be specified globally or within a path specification. That is,
(name1 <- name2 name3) (name4 <- name5), nocons
suppresses the intercepts in both equations. Alternatively,
(name1 <- name2 name3, nocons) (name4 <- name5)
suppresses the intercept in the first equation but not the second, whereas
(name1 <- name2 name3) (name4 <- name5, nocons)
suppresses the intercept in the second equation but not the first.

sem and gsem path notation Command syntax for path diagrams

519

In addition, consider the equation

(name1 <- name2 name3, nocons)
This can be written equivalently as
(name1 <- name2, nocons) (name1 <- name3, nocons)
16. Initial values (starting values) may be specified for a path, variance, covariance, or mean by
using the init(#) suboption:
(name1 <- (name2, init(0)))
(name1 <- (name2, init(0)) name3)
(name1 <- (name2, init(0)) (name3, init(5)))
..., ... var((name3, init(1)))
..., ... cov((name4*name5, init(.5)))
..., ... means((name5, init(0)))
The initial values may be combined with symbolic constraints:
(name1 <- (name2@c1, init(0)))
(name1 <- (name2@c1, init(0)) name3)
(name1 <- (name2@c1, init(0)) (name3@c2, init(5)))
..., ... var((name3@c1, init(1)))
..., ... cov((name4*name5@c1, init(.5)))
..., ... means((name5@c1, init(0)))

Also see
[SEM] sem Structural equation model estimation command
[SEM] gsem Generalized structural equation model estimation command
[SEM] sem path notation extensions Command syntax for path diagrams
[SEM] gsem path notation extensions Command syntax for path diagrams
[SEM] intro 2 Learning the language: Path diagrams and command language
[SEM] intro 6 Comparing groups (sem only)

Title
sem and gsem syntax options Options affecting interpretation of syntax
Syntax

Description

Options

Remarks and examples

Also see

Syntax
sem paths . . . , . . . syntax options
gsem paths . . . , . . . syntax options
syntax options

Description

latent(names)
nocapslatent

explicitly specify latent variable names

do not treat capitalized Names as latent

where names is a space-separated list of the names of the latent variables.

Description
These options affect some minor issues of how sem and gsem interpret what you type.

Options
latent(names) specifies that names is the full set of names of the latent variables. sem and gsem
ordinarily assume that latent variables have the first letter capitalized and observed variables have the
first letter lowercased; see [SEM] sem and gsem path notation. When you specify latent(names),
sem and gsem treat the listed variables as the latent variables and all other variables, regardless
of capitalization, as observed. latent() implies nocapslatent.
nocapslatent specifies that having the first letter capitalized does not designate a latent variable.
This option can be used when fitting models with observed variables only where some observed
variables in the dataset have the first letter capitalized.

Remarks and examples

We recommend using the default naming convention. If your dataset contains variables with the
first letter capitalized, it is easy to convert the variables to have lowercase names by typing
. rename *, lower

See [D] rename group.

Also see
[SEM] sem Structural equation model estimation command
[SEM] gsem Generalized structural equation model estimation command
[SEM] sem and gsem path notation Command syntax for path diagrams
520

Title
sem estimation options Options affecting estimation
Syntax

Description

Options

Remarks and examples

Also see

Syntax
sem paths . . . , . . . estimation options
estimation options

Description

method(method)
vce(vcetype)
nm1
noxconditional
allmissing
noivstart

method may be ml, mlmv, or adf

vcetype may be oim, eim, opg, robust, cluster clustvar,
bootstrap, or jackknife
compute sample variance rather than ML variance
compute covariances, etc., of observed exogenous variables
for use with method(mlmv)
skip calculation of starting values

noestimate

do not fit the model; instead show starting values

maximize options
control maximization process for specified model; seldom used
satopts(maximize options) control maximization process for saturated model; seldom used
baseopts(maximize options) control maximization process for baseline model; seldom used

Description
These options control how results are obtained.

Options
method() and vce() specify the method used to obtain parameter estimates and the technique used
to obtain the variancecovariance matrix of the estimates. See [SEM] sem option method( ).
nm1 specifies that the variances and covariances used in the SEM equations be the sample variances
(divided by N 1) and not the asymptotic variances (divided by N ). This is a minor technical
issue of little importance unless you are trying to match results from other software that assumes
sample variances. sem assumes asymptotic variances.
noxconditional states that you wish to include the means, variances, and covariances of the
observed exogenous variables among the parameters to be estimated by sem. See [SEM] sem
option noxconditional.
allmissing specifies how missing values should be treated when method(mlmv) is also specified.
Usually, sem omits from the estimation sample observations that contain missing values of any of
the observed variables used in the model. method(mlmv), however, can deal with these missing
values, and in that case, observations containing missing are not omitted.
Even so, sem, method(mlmv) does omit observations containing .a, .b, . . . , .z from the
estimation sample. sem assumes you do not want these observations used because the missing
value is not missing at random. If you want sem to include these observations in the estimation
sample, specify the allmissing option.
521

522

sem estimation options Options affecting estimation

noivstart is an arcane option that is of most use to programmers. It specifies that sem is to
skip efforts to produce good starting values with instrumental-variable techniques, techniques that
require computer time. If you specify this option, you should specify all the starting values. Any
starting values not specified will be assumed to be 0 (means, path coefficients, and covariances)
or some simple function of the data (variances).
noestimate specifies that the model is not to be fit. Instead, starting values are to be shown and they
are to be shown using the coeflegend style of output. An important use of this is to improve
starting values when your model is having difficulty converging. You can do the following:
. sem ..., ... noestimate
. matrix b = e(b)
. ... (modify elements of b) ...
. sem ..., ... from(b)

maximize options specify the standard and rarely specified options for controlling the maximization process for sem; see [R] maximize
 relevant options for sem are difficult,
 . The
technique(algorithm spec), iterate(#), no log, trace, gradient, showstep, hessian,
tolerance(#), ltolerance(#), and nrtolerance(#).
satopts(maximize options) is a rarely specified option and is only relevant if you specify the
method(mlmv) option. sem reports a test for model versus saturated at the bottom of the output.
Thus sem needs to obtain the saturated fit. In the case of method(ml) or method(adf), sem can
make a direct calculation. In the other case of method(mlmv), sem must actually fit the saturated
model. The maximization options specified inside satopts() control that maximization process.
It is rare that you need to specify the satopts() option, even if you find it necessary to specify
the overall maximize options.
baseopts(maximize options) is a rarely specified option and an irrelevant one unless you also
specify method(mlmv) or method(adf). When fitting the model, sem records information about
the baseline model for later use by estat gof, should you use that command. Thus sem needs
to obtain the baseline fit. In the case of method(ml), sem can make a direct calculation. In
the cases of method(mlmv) and method(adf), sem must actually fit the baseline model. The
maximization options specified inside baseopts() control that maximization process. It is rare
that you need to specify the baseopts() option even if you find it necessary to specify the overall
maximize options.

Remarks and examples

The most commonly specified option among this group is vce(). See [SEM] sem option method( ),
[SEM] intro 8, and [SEM] intro 9.

Also see
[SEM] sem Structural equation model estimation command
[SEM] sem option method( ) Specifying method and calculation of VCE
[SEM] sem option noxconditional Computing means, etc., of observed exogenous variables
[SEM] intro 8 Robust and clustered standard errors
[SEM] intro 9 Standard errors, the full story

Title
sem group options Fitting models on different groups

Syntax

Description

Options

Remarks and examples

Also see

Syntax
sem paths . . . , . . . group options
group options

Description

group(varname)
ginvariant(classname)

fit model for different groups

specify parameters that are equal across groups

classname

Description

scoef
scons

structural coefficients
structural intercepts

mcoef
mcons

measurement coefficients
measurement intercepts

serrvar
merrvar
smerrcov

covariances of structural errors

covariances of measurement errors
covariances between structural and measurement errors

meanex
covex

means of exogenous variables

covariances of exogenous variables

all
none

all the above

none of the above

ginvariant(mcoef mcons) is the default if ginvariant() is not specified.

meanex, covex, and all exclude the observed exogenous variables (that is, they include only the latent
exogenous variables) unless you specify the noxconditional option or the noxconditional
option is otherwise implied; see [SEM] sem option noxconditional. This is what you would desire
in most cases.

Description
sem can fit combined models across subgroups of the data while allowing some parameters to vary
and constraining others to be equal across subgroups. These subgroups could be males and females,
age category, and the like.
sem performs such estimation when the group(varname) option is specified. The ginvariant(classname) option specifies which parameters are to be constrained to be equal across the
groups.

523

524

sem group options Fitting models on different groups

Options
group(varname) specifies that the model be fit as described above. varname specifies the name of
a numeric variable that records the group to which the observation belongs.
If you are using summary statistics data in place of raw data, varname is the name of the group
variable as reported by ssd describe; see [SEM] ssd.
ginvariant(classname) specifies which classes of parameters of the model are to be constrained to
be equal across groups. The classes are defined above. The default is ginvariant(mcoef mcons)
if the option is not specified.

Remarks and examples

See [SEM] intro 6, and see [SEM] example 20 and [SEM] example 23.

Also see
[SEM] sem Structural equation model estimation command
[SEM] intro 6 Comparing groups (sem only)
[SEM] example 20 Two-factor measurement model by group
[SEM] example 23 Specifying parameter constraints across groups

Title
sem model description options Model description options
Syntax

Description

Options

Remarks and examples

Also see

Syntax
sem paths . . . , . . . model description options
model description options

Description

covariance()

variance()

means()

path notation for treatment of covariances;

see [SEM] sem and gsem path notation
path notation for treatment of variances;
see [SEM] sem and gsem path notation
path notation for treatment of means;
see [SEM] sem and gsem path notation

covstructure()

alternative method to place restrictions on covariances;

see [SEM] sem and gsem option covstructure( )

noconstant
nomeans
noanchor
forcenoanchor

do not fit intercepts

do not fit means or intercepts
do not apply default anchoring
programmers option

reliability()

reliability of measurement variables;

see [SEM] sem and gsem option reliability( )

constraints()
from()

specify constraints; see [SEM] sem and gsem option constraints( )

specify starting values; see [SEM] sem and gsem option from( )

Option may be specified more than once.

Description
paths and the options above describe the model to be fit by sem.

Options
covariance(), variance(), and means() fully describe the model to be fit. See [SEM] sem and
gsem path notation.
covstructure() provides a convenient way to constrain covariances in your model. Alternatively
or in combination, you can place constraints by using the standard path notation. See [SEM] sem
and gsem option covstructure( ).
noconstant specifies that all intercepts be constrained to 0. See [SEM] sem and gsem path notation.
nomeans specifies that means and intercepts not be fit. The means and intercepts are concentrated out
of the function being optimized, which is typically the likelihood function. Results for all other
parameters are the same whether or not this option is specified.
525

526

sem model description options Model description options

This option is seldom specified. sem issues this option to itself when you use summary statistics
data that do not include summary statistics for the means.
noanchor specifies that sem is not to check for lack of identification and fill in anchors where needed.
sem is instead to issue an error message if anchors would be needed. You specify this option when
you believe you have specified the necessary normalization constraints and want to hear about it
if you are wrong. See Identification 2: Normalization constraints (anchoring) in [SEM] intro 4.
forcenoanchor is similar to noanchor except that rather than issue an error message, sem proceeds
to estimation. There is no reason you should specify this option. forcenoanchor is used in testing
of sem at StataCorp.
reliability() specifies the fraction of variance not due to measurement error for a variable. See
[SEM] sem and gsem option reliability( ).
constraints() specifies parameter constraints you wish to impose on your model; see [SEM] sem
and gsem option constraints( ). Constraints can also be specified as described in [SEM] sem and
gsem path notation, and they are usually more conveniently specified using the path notation.
from() specifies the starting values to be used in the optimization process; see [SEM] sem and gsem
option from( ). Starting values can also be specified by using the init() suboption as described
in [SEM] sem and gsem path notation.

Remarks and examples

To use sem successfully, you need to understand paths, covariance(), variance(), and means();
see Using path diagrams to specify standard linear SEMs in [SEM] intro 2 and [SEM] sem and gsem
path notation.
covstructure() is often convenient; see [SEM] sem and gsem option covstructure( ).

Also see
[SEM] sem Structural equation model estimation command
[SEM] intro 2 Learning the language: Path diagrams and command language
[SEM] sem and gsem path notation Command syntax for path diagrams
[SEM] sem and gsem option covstructure( ) Specifying covariance restrictions
[SEM] sem and gsem option reliability( ) Fraction of variance not due to measurement error
[SEM] sem and gsem option constraints( ) Specifying constraints
[SEM] sem and gsem option from( ) Specifying starting values

Title
sem option method( ) Specifying method and calculation of VCE

Syntax

Description

Options

Remarks and examples

Also see

Syntax
sem . . .

, . . . method(method) vce(vcetype) . . .

method

Description

ml
mlmv
adf

maximum likelihood; the default

ml with missing values
asymptotic distribution free

vcetype

Description

oim
eim
opg
robust
cluster clustvar


bootstrap , bootstrap options


jackknife , jackknife options

observed information matrix; the default

expected information matrix
outer product of gradients
Huber/White/sandwich estimator
generalized Huber/White/sandwich estimator
bootstrap estimation
jackknife estimation

The following combinations of method() and vce() are allowed:

ml
mlmv
adf

oim

eim

opg

robust

cluster

bootstrap

jackknife

x
x
x

x
x
x

x
x

x
x

x
x

x
x
x

x
x
x

Description
sem option method() specifies the method used to obtain the estimated parameters.
sem option vce() specifies the technique used to obtain the variancecovariance matrix of the
estimates (VCE), which includes the reported standard errors.

527

528

sem option method( ) Specifying method and calculation of VCE

Options
method(method) specifies the method used to obtain parameter estimates. method(ml) is the default.
vce(vcetype) specifies the technique used to obtain the VCE. vce(oim) is the default.

Remarks and examples

See [SEM] intro 4, [SEM] intro 8, and [SEM] intro 9.

Also see
[SEM] sem Structural equation model estimation command
[SEM] intro 4 Substantive concepts
[SEM] intro 8 Robust and clustered standard errors
[SEM] intro 9 Standard errors, the full story
[SEM] example 26 Fitting a model with data missing at random

Title
sem option noxconditional Computing means, etc., of observed exogenous variables

Syntax

Description

Option

Remarks and examples

Also see

Syntax
sem . . .

, . . . noxconditional . . .

Description
sem has a noxconditional option that you may rarely wish to specify. The option is described
below.

Option
noxconditional states that you wish to include the means, variances, and covariances of the observed
exogenous variables among the parameters to be estimated by sem.

Remarks and examples

Remarks are presented under the following headings:
What is x conditional?
When to specify noxconditional
Option forcexconditional (a technical note)

What is x conditional?
In many cases, sem does not include the means, variances, and covariances of observed exogenous
variables among the parameters to be estimated. When sem omits them, the estimator of the model is
said to be x conditional. Rather than estimating the values of the means, variances, and covariances,
sem uses the separately calculated observed values of those statistics. sem does this to save time and
memory.
sem does not use the x-conditional calculation when it would be inappropriate.
The noxconditional option prevents sem from using the x-conditional calculation. You specify
noxconditional on the sem command:
. sem ..., ... noxconditional

Do not confuse the x-conditional calculation with the assumption of conditional normality discussed
in [SEM] intro 4. The x-conditional calculation is appropriate even when the assumption of conditional
normality is inappropriate.

529

530

sem option noxconditional Computing means, etc., of observed exogenous variables

When to specify noxconditional

It is never inappropriate to specify the noxconditional option. Be aware, however:
1. If you are using the default method(ml), estimated point estimates and standard errors will be
the same.
2. If you are using method(adf), estimated point estimates and standard errors will be slightly
different, asymptotically equivalent, and there is no reason to prefer one set of estimates over
the other.
3. If you are using method(mlmv), the situation is the same as in statement 1.
4. Regardless of the estimation method used, calculation of results will require more computer time
and memory. The memory requirements increase quadratically with the total number of estimated
parameters in your model. If you have k1 observed exogenous variables and k2 latent exogenous
variables, the number of added parameters from noxconditional is k1 + k1 (k1 + 1)/2 + k1 k2 .
The resulting total memory requirements can be so great as to require more memory than your
computer can provide.
To make statements 14 true, there are two cases when sem specifies noxconditional for you:
1. sem defaults to noxconditional whenever you constrain a mean, variance, or covariance of an
observed exogenous variable. For example,
.
.
.
.

sem
sem
sem
sem

...,
...,
...,
...,

...
...
...
...

means(x1@m x2@m)
var(x1@v x2@v)
cov(x1*x2@c x1*x3@c)
covstruct(_OEx, diagonal)

See [SEM] sem and gsem path notation and [SEM] sem and gsem option covstructure( ).
2. sem defaults to noxconditional whenever you use method(mlmv) and there are missing values
among the observed exogenous variables.
There are only three reasons for you to specify the noxconditional option:
1. Specify noxconditional if you subsequently wish to test means, variances, or covariances of
observed exogenous variables with postestimation commands. For example,
. sem ..., ... noxconditional
. sem, coeflegend
. test _b[means(x1):_cons] == _b[means(x2)_cons]

2. Specify noxconditional if you are fitting a model with the group() option.
3. Specify noxconditional if you also specify the ginvariant() option, and you want the
ginvariant() classes meanex, covex, or all to include the observed exogenous variables.
For example,
. sem ..., ... by(agegrp) ginvariant(all) noxconditional

You may also wish to specify noxconditional when comparing results with those from other
packages. Many packages use the noxconditional approach when using an estimation method other
than maximum likelihood (ML). Correspondingly, most packages use the x-conditional calculation
when using ML.

sem option noxconditional Computing means, etc., of observed exogenous variables

531

Option forcexconditional (a technical note)

In addition to noxconditional, sem has a forcexconditional option:


sem . . . , . . . forcexconditional . . .
This option turns off sems switching away from the x-conditional calculation when that is required.
Do not specify this option unless you are exploring the behavior of x-conditional calculation in cases
where it is theoretically inappropriate.

Also see
[SEM] sem Structural equation model estimation command

Title
sem option select( ) Using sem with summary statistics data

Syntax

Description

Option

Remarks and examples

Also see

Syntax




sem . . . , . . . select(# # . . . ) . . .

Description
sem may be used with summary statistics data (SSD), data containing only summary statistics such
as the means, standard deviations or variances, and correlations and covariances of the underlying,
raw data.
You enter SSD with the ssd command; see [SEM] ssd.
To fit a model with sem, there is nothing special you have to do except specify the select()
option where you would usually specify if exp.

Option


select(# # . . . ) is allowed only when you have SSD in memory. It specifies which groups should
be used.

Remarks and examples

See [SEM] intro 11.
sem option select() is the SSD alternative for if exp if you only had the underlying, raw data
in memory. With the underlying raw data, where you would usually type
. sem ... if agegrp==1 | agegrp==3, ...

with SSD in memory, you type

. sem ..., ... select(1 3)

You may select only groups for which you have separate summary statistics recorded in your
summary statistics dataset; the ssd describe command will list the group variable, if any. See
[SEM] ssd.
By the way, select() may be combined with sem option group(). Where you might usually
type
. sem ... if agegrp==1 | agegrp==3, ... group(agegrp)

you could type

. sem ..., ... select(1 3) group(agegrp)

532

sem option select( ) Using sem with summary statistics data

533

The above restricts sem to age groups 1 and 3, so the result will be an estimation of a combined
model of age groups 1 and 3 with some coefficients allowed to vary between the groups and other
coefficients constrained to be equal across the groups. See [SEM] sem group options.

Also see
[SEM] sem Structural equation model estimation command
[SEM] intro 11 Fitting models with summary statistics data (sem only)

Title
sem path notation extensions Command syntax for path diagrams
Syntax

Description

Options

Remarks and examples

Also see

Syntax
sem paths . . .

, covariance() variance() means()

group(varname)



paths specifies the direct paths between the variables of your model.
The model to be fit is fully described by paths, covariance(), variance(), and means().
The syntax of these elements is modified (generalized) when the group() option is specified.

Description
This entry concerns sem only.
The command syntax for describing your SEMs is fully specified by paths, covariance(),
variance(), and means(). How that works is described in [SEM] sem and gsem path notation.
See that section before reading this section.
This entry concerns the path features unique to sem, and that has to do with the group() option
for comparing different groups.

Options
covariance(), variance(), and means() are described in [SEM] sem and gsem path notation.
group(varname) allows models specified with paths, covariance(), variance(), and means() to
be automatically generalized (interacted) with the groups defined by varname; see [SEM] intro 6.
The syntax of paths and the arguments of covariance(), variance(), and means() gain an
extra syntactical piece when group() is specified.

Remarks and examples

The model you wish to fit is fully described by the paths, covariance(), variance(), and
means() that you type. The group(varname) option,
. sem ..., ... group(varname)

specifies that the model be fit separately for the different values of varname. varname might be sex
and then the model would be fit separately for males and females, or varname might be something
else and perhaps take on more than two values.
Whatever varname is, group(varname) defaults to letting some of the path coefficients, covariances,
variances, and means of your model vary across the groups and constraining others to be equal.
Which parameters vary and which are constrained is described in [SEM] sem group options, but that
is a minor detail right now.
In what follows, we will assume that varname is mygrp and takes on three values. Those values
are 1, 2, and 3, but they could just as well be 2, 9, and 12.
Consider typing
. sem ..., ...

534

sem path notation extensions Command syntax for path diagrams

535

and typing
. sem ..., ... group(mygrp)

Whatever the paths, covariance(), variance(), and means() are that describe the model,
there are now three times as many parameters because each group has its own unique set. In fact,
when you give the second command, you are not merely asking for three times the parameters, you
are specifying three models, one for each group! In this case, you specified the same model three
times without knowing it.
You can vary the model specified across groups.
1. Lets write the model you wish to fit as
. sem (a) (b) (c), cov(d) cov(e) var(f)

where a, b, . . . , f stand for what you type. In this generic example, we have two cov() options
just because multiple cov() options often occur in real models. When you type
. sem (a) (b) (c), cov(d) cov(e) var(f) group(mygrp)

results are as if you typed

. sem (1: a) (2: a) (3: a)
///
(1: b) (2: b) (3: b)
///
(1: c) (2: c) (3: c),
///
cov(1: d) cov(2: d) cov(3: d)
///
cov(1: e) cov(2: e) cov(3: e)
///
var(1: f) cov(2: f) cov(3: f) group(mygrp)

The 1:, 2:, and 3: identify the groups for which paths, covariances, or variances are being
added, modified, or constrained.
If mygrp contained the unique values 5, 8, and 10 instead of 1, 2, and 3, then 5: would appear
in place of 1:; 8: would appear in place of 2:; and 10: would appear in place of 3:.
2. Consider the model
. sem (y <- x) (b) (c), cov(d) cov(e) var(f) group(mygrp)

If you wanted to constrain the path coefficient (y <- x) to be the same across all three groups,
you could type
. sem (y <- x@c1) (b) (c), cov(d) cov(e) var(f) group(mygrp)

See item 12 in [SEM] sem and gsem path notation for more examples of specifying constraints.
This works because the expansion of (y <- x@c1) is
(1: y <- x@c1) (2: y <- x@c1) (3: y <- x@c1)

3. Consider the model

. sem (y <- x) (b) (c), cov(d) cov(e) var(f) group(mygrp)

If you wanted to constrain the path coefficient (y <- x) to be the same in groups 2 and 3, you
could type
. sem (1: y <- x) (2: y <- x@c1) (3: y <- x@c1) (b) (c),
///
cov(d) cov(e) var(f) group(mygrp)

4. Instead of following item 3, you could type

. sem (y <- x) (2: y <- x@c1) (3: y <- x@c1) (b) (c),
///
cov(d) cov(e) var(f) group(mygrp)

536

sem path notation extensions Command syntax for path diagrams

The part (y <- x) (2: y <- x@c1) (3: y <- x@c1) expands to
(1: y <- x) (2: y <- x) (3: y <- x) (2: y <- x@c1) (3: y <- x@c1)

and thus the path is defined twice for group 2 and twice for group 3. When a path is defined
more than once, the definitions are combined. In this case, the second definition adds more
information, so the result is as if you typed
(1: y <- x) (2: y <- x@c1) (3: y <- x@c1)

5. Instead of following item 3 or item 4, you could type

. sem (y <- x@c1) (1: y <- x@c2) (b) (c),
///
cov(d) cov(e) var(f) group(mygrp)

The part (y <- x@c1) (1: y <- x@c2) expands to

(1: y <- x@c1)

(2: y <- x@c1)

(3: y <- x@c1)

(1: y <- x@c2)

When results are combined from repeated definitions, then definitions that appear later take
precedence. In this case, results are as if the expansion read
(1: y <- x@c2) (2: y <- x@c1)

(3: y <- x@c1)

Thus coefficients for groups 2 and 3 are constrained. The group-1 coefficient is constrained to
c2. If c2 appears nowhere else in the model specification, then results are as if the path for
group 1 were unconstrained.
6. Instead of following item 3, item 4, or item 5, you could not type
. sem (y <- x@c1) (1: y <- x) (b) (c),
///
cov(d) cov(e) var(f) group(mygrp)

The expansion of (y <- x@c1) (1: y <- x) reads

(1: y <- x@c1)

(2: y <- x@c1)

(3: y <- x@c1)

(1: y <- x)

and you might think that 1: y <- x would replace 1: y <- x@c1. Information, however, is
combined, and even though precedence is given to information appearing later, silence does not
count as information. Thus the expanded and reduced specification reads the same as if 1: y <x was never specified:
(1: y <- x@c1)

(2: y <- x@c1)

(3: y <- x@c1)

7. Items 16, stated in terms of paths, apply equally to what is typed inside the means(),
variance(), and covariance() options. For instance, if you typed
. sem (a) (b) (c), var(e.y@c1) group(mygrp)

then you are constraining the variance to be equal across all three groups.
If you wanted to constrain the variance to be equal in groups 2 and 3, you could type
. sem (a) (b) (c), var(e.y) var(2: e.y@c1) var(3: e.y@c1), group(mygrp)

You could omit typing var(e.y) because it is implied. Alternatively, you could type
. sem (a) (b) (c), var(e.y@c1) var(1: e.y@c2) group(mygrp)

sem path notation extensions Command syntax for path diagrams

537

You could not type

. sem (a) (b) (c), var(e.y@c1) var(1: e.y) group(mygrp)

because silence does not count as information when specifications are combined.
Similarly, if you typed
. sem (a) (b) (c), cov(e.y1*e.y2@c1) group(mygrp)

then you are constraining the covariance to be equal across all groups. If you wanted to constrain
the covariance to be equal in groups 2 and 3, you could type
. sem (a) (b) (c), cov(e.y1*e.y2)
///
cov(2: e.y1*e.y2@c1) cov(3: e.y1*e.y2@c1) ///
group(mygrp)

You could not omit cov(e.y1*e.y2) because it is not assumed. By default, error variables
are assumed to be uncorrelated. Omitting the option would constrain the covariance to be 0 in
group 1 and to be equal in groups 2 and 3.
Alternatively, you could type
. sem (a) (b) (c), cov(e.y1*e.y2@c1)
cov(1: e.y1*e.y2@c2)
group(mygrp)

///
///

8. In the examples above, we have referred to the groups with their numeric values, 1, 2, and 3.
Had the values been 5, 8, and 10, then we would have used those values.
If the group variable mygrp has a value label, you can use the label to refer to the group. For
instance, imagine mygrp is labeled as follows:
. label define grpvals 1 Male
. label values mygrp grpvals

2 Female

3 "Unknown sex"

We could type
. sem (y <- x) (Female: y <- x@c1) (Unknown sex: y <- x@c1) ..., ...

or we could type
. sem (y <- x) (2: y <- x@c1) (3: y <- x@c1) ..., ...

Also see
[SEM] sem Structural equation model estimation command
[SEM] sem and gsem path notation Command syntax for path diagrams
[SEM] intro 2 Learning the language: Path diagrams and command language
[SEM] intro 6 Comparing groups (sem only)

Title
sem postestimation Postestimation tools for sem

Description

Remarks and examples

Also see

Description
The following are the postestimation commands that you can use after estimation by sem:
Command

Description

sem, coeflegend
estat framework

display b[ ] notation
display results in modeling framework (matrix form)

estat
estat
estat
estat
estat

overall goodness of fit

group-level goodness of fit
equation-level goodness of fit
matrices of residuals
Akaikes and Schwarzs Bayesian information criteria (AIC and BIC)

gof
ggof
eqgof
residuals
ic

estat mindices
estat scoretests
estat ginvariant

modification indices (score tests)

score tests
test of invariance of parameters across groups

estat eqtest
lrtest
test
lincom
nlcom
testnl
estat stdize:

equation-level Wald tests

likelihood-ratio tests
Wald tests
linear combination of parameters
nonlinear combination of parameters
Wald tests of nonlinear hypotheses
test standardized parameters

estat
estat
estat
estat

decomposition of effects
assess stability of nonrecursive systems
summary statistics for the estimation sample
variancecovariance matrix of the estimators (VCE)

teffects
stable
summarize
vce

predict

factor scores, predicted values, etc.

margins

marginal means, predictive margins, and marginal effects

estimates

cataloging estimation results

For a summary of postestimation features, see [SEM] intro 7.

538

sem postestimation Postestimation tools for sem

539

Remarks and examples

This manual entry concerns sem. For information on postestimation features available after gsem,
see [SEM] gsem postestimation.

Also see
[SEM] sem reporting options Options affecting reporting of results

Title
sem reporting options Options affecting reporting of results
Syntax
Reference

Description
Also see

Options

Remarks and examples

Syntax
sem paths . . . , . . . reporting options
sem, reporting options
reporting options

Description

level(#)
standardized
coeflegend
nocnsreport
nodescribe
noheader
nofootnote
notable
nofvlabel
fvwrap(#)
fvwrapon(style)

set confidence level; default is level(95)

display standardized coefficients and values
display coefficient legend
do not display constraints
do not display variable classification table
do not display header above parameter table
do not display footnotes below parameter table
do not display parameter table
display group values rather than value labels
allow # lines when wrapping long value labels
apply style for wrapping long value labels;
style may be word or width
report all estimated parameters

showginvariant

Description
These options control how sem displays estimation results.

Options
level(#); see [R] estimation options.
standardized displays standardized values, that is, beta values for coefficients, correlations for
covariances, and 1s for variances. Standardized values are obtained using model-fitted variances
(Bollen 1989, 124125). We recommend caution in the interpretation of standardized values,
especially with multiple groups.
coeflegend displays the legend that reveals how to specify estimated coefficients in b[ ] notation,
which you are sometimes required to use when specifying postestimation commands.
nocnsreport suppresses the display of the constraints. Fixed-to-zero constraints that are automatically
set by sem are not shown in the report to keep the output manageable.
nodescribe suppresses display of the variable classification table.
540

sem reporting options Options affecting reporting of results

541

noheader suppresses the header above the parameter table, the display that reports the final loglikelihood value, number of observations, etc.
nofootnote suppresses the footnotes displayed below the parameter table.
notable suppresses the parameter table.
nofvlabel displays group values rather than value labels.
fvwrap(#) specifies how many lines to allow when long value labels must be wrapped. Labels
requiring more than # lines are truncated. This option overrides the fvwrap setting; see [R] set
showbaselevels.
fvwrapon(style) specifies whether value labels that wrap will break at word boundaries or break
based on available space.
fvwrapon(word), the default, specifies that value labels break at word boundaries.
fvwrapon(width) specifies that value labels break based on available space.
This option overrides the fvwrapon setting; see [R] set showbaselevels.
showginvariant specifies that each estimated parameter be reported in the parameter table. The
default is to report each invariant parameter only once.

Remarks and examples

Any of the above options may be specified when you fit the model or when you redisplay results,
which you do by specifying nothing but options after the sem command:
. sem (...) (...), ...
(original output displayed)
. sem
(output redisplayed)
. sem, standardized
(standardized output displayed)
. sem, coeflegend
(coefficient-name table displayed)
. sem
(output redisplayed)

Reference
Bollen, K. A. 1989. Structural Equations with Latent Variables. New York: Wiley.

Also see
[SEM] sem Structural equation model estimation command
[SEM] example 8 Testing that coefficients are equal, and constraining them
[SEM] example 16 Correlation

Title
sem ssd options Options for use with summary statistics data
Syntax

Description

Options

Remarks and examples

Also see

Syntax
sem paths . . . , . . . ssd options
ssd options

Description

select()
forcecorrelations

alternative to if exp for SSD

allow groups and pooling of SSD correlations

Description
Data are sometimes available in summary statistics form only. These summary statistics include
means, standard deviations or variances, and correlations or covariances. These summary statistics
can be used by sem in place of the underlying raw data.

Options
select() is an alternative to if exp when you are using summary statistics data (SSD). Where you
might usually type
. sem ... if agegrp==1 | agegrp==3 | agegrp==5, ...
with SSD in memory, you type
. sem ..., ... select(1 3 5)
See [SEM] sem option select( ) and [SEM] intro 11.
forcecorrelations tells sem that it may make calculations that would usually be considered
suspicious with SSD that contain only a subset of means, variances (standard deviations), and
covariances (correlations). Do not specify this option unless you appreciate the statistical issues
that we are about to discuss. There are two cases where forcecorrelations is relevant.
In the first case, sem is unwilling to produce group() estimates if one or more (usually all) of the
groups have correlations only defined. You can override that by specifying forcecorrelations,
and sem will assume unit variances for the group or groups that have correlations only. Doing
this is suspect unless you make ginvariant() all parameters that are dependent on covariances
or unless you truly know that the variances are indeed 1.
In the second case, sem is unwilling to pool across groups unless you have provided means and
covariances (or means and correlations and standard deviations or variances). Without that information, should the need for pooling arise, sem issues an error message. The forcecorrelations
option specifies that sem ignore its rule and pool correlation matrices, treating correlations as
if they were covariances when variances are not defined and treating means as if they were 0
when means are not defined. The only justification for making the calculation in this way is that
variances truly are 1 and means truly are 0.
542

sem ssd options Options for use with summary statistics data

543

Understand that there is nothing wrong with using pure correlation data, or covariance data without
the means, so long as you fit models for individual groups. Doing anything across groups basically
requires that sem have the covariance and mean information.

Remarks and examples

See [SEM] intro 11.

Also see
[SEM] sem Structural equation model estimation command
[SEM] intro 11 Fitting models with summary statistics data (sem only)
[SEM] ssd Making summary statistics data (sem only)

Title
ssd Making summary statistics data (sem only)
Syntax

Description

Options

Remarks and examples

Stored results

Also see

Syntax
To enter summary statistics data (SSD), the commands are
ssd init varlist
 
ssd set # observations #
 
ssd set # means vector
 
ssd set # { variances | sd } vector
 
ssd set # { covariances | correlations } matrix
(to add second group)

ssd addgroup varname

(to add subsequent groups)

ssd addgroup

(to remove last group)

ssd unaddgroup #
ssd status

 
#

(to review status)

To build SSD from raw data, the command is



ssd build varlist , group(varname) clear
To review the contents of SSD, the commands are
 
ssd status #
ssd describe
 
ssd list #
In an emergency (ssd will tell you when), you may use
ssd repair
In the above,
A vector can be any of the following:
1. A space-separated list of numbers, for example,
. ssd set means 1 2 3

544

ssd Making summary statistics data (sem only)

545

2. (stata) matname
where matname is the name of a Stata 1 k or k 1 matrix, for example,
. ssd set variances (stata) mymeans

3. (mata) matname
where matname is the name of a Mata 1 k or k 1 matrix, for example,
. ssd set sd (mata) mymeans

A matrix can be any of the following:

1. A space-separated list of numbers corresponding to the rows of the matrix, with backslashes
(\) between rows. The numbers are either the lower triangle and diagonal or the diagonal
and upper triangle of the matrix, for example,
. ssd set correlations 1 \ .2 1 \ .3 .5 1

or
. ssd set correlations 1 .2 .3 \ 1 .5 \ 1

2. (ltd) # # . . .
which is to say, a space-separated list of numbers corresponding to the lower triangle and
diagonal of the matrix, without backslashes between rows, for example,
. ssd set correlations (ltd) 1

.2 1

.3 .5 1

3. (dut) # # . . .
which is to say, a space-separated list of numbers corresponding to the diagonal and upper
triangle of the matrix, without backslashes between rows, for example,
. ssd set correlations (dut) 1 .2 .3

1 .5

4. (stata) matname
where matname is the name of a Stata k k symmetric matrix, for example,
. ssd set correlations (stata) mymat

5. (mata) matname
where matname is the name of a Mata k k symmetric matrix, for example,
. ssd set correlations (mata) mymat

Description
ssd allows you to enter SSD to fit SEMs and allows you to create SSD from original, raw data to
publish or to send to others (and thus preserve participant confidentiality). Data created by ssd may
be used with sem but not gsem.

Options
group(varname) is for use with ssd build. It specifies that separate groups of summary statistics
be produced for each value of varname.
clear is for use with ssd build. It specifies that it is okay to replace the data in memory with SSD
even if the original dataset has not been saved since it was last changed.

546

ssd Making summary statistics data (sem only)

Remarks and examples

See
[SEM] intro 11

Fitting models with summary statistics data (sem only)

for an introduction, and see

[SEM] example 2
[SEM] example 19
[SEM] example 25

Creating a dataset from published covariances

Creating multiple-group summary statistics data
Creating summary statistics data from raw data

A summary statistics dataset is different from a regular, raw Stata dataset. Be careful not to use
standard Stata data-manipulation commands with SSD in memory. The commands include
generate
replace
merge
append
drop
set obs
to mention a few. You may, however, use rename to change the names of the variables.
The other data-manipulation commands can damage your summary statistics dataset. If you make
a mistake and do use one of these commands, do not attempt to repair the data yourself. Let ssd
repair your data by typing
. ssd repair

ssd is usually successful as long as variables or observations have not been dropped.
Every time you use ssd, even for something as trivial as describing or listing the data, ssd verifies
that the data are not corrupted. If ssd finds that they are, it suggests you type ssd repair:
. ssd describe
SSD corrupt
The summary statistics data should [ssd
describes the problem]. The data may be fixable;
type ssd repair.
. ssd repair
(data repaired)
. ssd describe
(usual output appears)

In critical applications, we also recommend you digitally sign your summary statistics dataset:
. datasignature set
5:5(65336):3718404259:2275399871

(data signature set)

Then at any future time, you can verify the data are still just as they should be:
. datasignature confirm
(data unchanged since 30jun2012 15:32)

The data signature is a function of the variable names. If you rename a variablesomething that is
allowedthen the data signature will change:
. rename varname newname
. datasignature confirm
data have changed since 30jun2012 15:32
r(9);

ssd Making summary statistics data (sem only)

547

In that case, you can re-sign the data:

. datasignature set, reset
5:5(71728):3718404259:2275399871

(data signature set)

Before re-signing, however, if you want to convince yourself that the data are unchanged except
for the variable name, type datasignature report. It is the part of the signature in parentheses
that has to do with the variable names. datasignature report will tell you what the new signature
would be, and you can verify that the other components of the signature match.
See [D] datasignature.

Stored results
ssd describe stores the following in r():
Scalars
r(N)
r(k)
r(G)
r(complete)
r(complete means)
r(complete covariances)
Macros
r(v#)
r(groupvar)

number of observations (total across groups)

number of variables (excluding group variable)
number of groups
1 if complete, 0 otherwise
1 if complete means, 0 otherwise
1 if complete covariances, 0 otherwise
variable names (excluding group variable)
name of group variable (if there is one)

Also see
[SEM] intro 11 Fitting models with summary statistics data (sem only)
[D] datasignature Determine whether data have changed
[SEM] example 2 Creating a dataset from published covariances
[SEM] example 19 Creating multiple-group summary statistics data
[SEM] example 25 Creating summary statistics data from raw data

Title
test Wald test of linear hypotheses
Syntax
Remarks and examples

Menu
Stored results

Description
Also see

Options

Syntax
test coeflist
test exp = exp =
test [eqno]

: coeflist

test [eqno = eqno

test (spec)

= ...

 

= . . . ] : coeflist

(spec) . . .

 

, test options

Menu
Statistics

>

SEM (structural equation modeling)

>

Testing and CIs

>

Wald tests of linear hypotheses

Description
test is a postestimation command for use after sem, gsem, and other Stata estimation commands.
test performs the Wald test of the hypothesis or hypotheses that you specify. In the case of sem
and gsem, you must use the b[ ] coefficient notation.
See [R] test. Also documented there is testparm. testparm cannot be used after sem or gsem
because its syntax hinges on use of shortcuts for referring to coefficients.

Options
See Options for test in [R] test.

Remarks and examples

See [SEM] example 8 and [SEM] example 16.
test works in the metric of SEM, which is to say path coefficients, variances, and covariances. If
you want to frame your tests in terms of standardized coefficients and correlations and you fit your
model with sem, not gsem, then prefix test with estat stdize:; see [SEM] estat stdize.

548

test Wald test of linear hypotheses

Stored results
See Stored results in [R] test.

Also see
[SEM] example 8 Testing that coefficients are equal, and constraining them
[SEM] example 16 Correlation

[R] test Test linear hypotheses after estimation

[SEM] estat stdize Test standardized parameters
[SEM] estat eqtest Equation-level test that all coefficients are zero
[SEM] lrtest Likelihood-ratio test of linear hypothesis
[SEM] lincom Linear combinations of parameters

549

Title
testnl Wald test of nonlinear hypotheses
Syntax
Remarks and examples

Menu
Stored results

Description
Also see

Options

Syntax
testnl exp = exp

testnl (exp = exp

= ...

 

, options

 


 

= . . . ) (exp = exp = . . . ) . . .
, options

Menu
Statistics

>

SEM (structural equation modeling)

>

Testing and CIs

>

Wald tests of nonlinear hypotheses

Description
testnl is a postestimation command for use after sem, gsem, and other Stata estimation commands.
testnl performs the Wald test of the nonlinear hypothesis or hypotheses. In the case of sem and
gsem, you must use the b[ ] coefficient notation.

Options
See Options in [R] testnl.

Remarks and examples

testnl works in the metric of SEM, which is to say path coefficients, variances, and covariances.
If you want to frame your tests in terms of standardized coefficients and correlations and you fit the
model with sem, not gsem, then prefix testnl with estat stdize:; see [SEM] estat stdize.

Technical note
estat stdize: is unnecessary because, with testnl, everywhere you wanted a standardized
coefficient or correlation, you could just type the formula. If you did that, you would get the
same answer except for numerical precision. In this case, the answer produced with the estat
stdize: prefix will be a little more accurate because estat stdize: is able to substitute an analytic
derivative in one part of the calculation where testnl, doing the whole thing itself, would be forced
to use a numeric derivative.

550

testnl Wald test of nonlinear hypotheses

Stored results
See Stored results in [R] testnl.

Also see
[R] testnl Test nonlinear hypotheses after estimation
[SEM] test Wald test of linear hypotheses
[SEM] lrtest Likelihood-ratio test of linear hypothesis
[SEM] estat stdize Test standardized parameters
[SEM] estat eqtest Equation-level test that all coefficients are zero
[SEM] nlcom Nonlinear combinations of parameters

551

Glossary
ADF, method(adf). ADF stands for asymptotic distribution free and is a method used to obtain fitted
parameters for standard linear SEMs. ADF is used by sem when option method(adf) is specified.
Other available methods are ML, QML, and MLMV.
anchoring, anchor variable. A variable is said to be the anchor of a latent variable if the path
coefficient between the latent variable and the anchor variable is constrained to be 1. sem and
gsem use anchoring as a way of normalizing latent variables and thus identifying the model.
baseline model. A baseline model is a covariance modela model of fitted means and covariances
of observed variables without any other pathswith most of the covariances constrained to be
0. That is, a baseline model is a model of fitted means and variances but typically not all the
covariances. Also see saturated model. Baseline models apply only to standard linear SEMs.
BentlerWeeks formulation. The Bentler and Weeks (1980) formulation of standard linear SEMs places
the results in a series of matrices organized around how results are calculated. See [SEM] estat
framework.
bootstrap, vce(bootstrap). The bootstrap is a replication method for obtaining variance estimates.
Consider an estimation method E for estimating . Let b be the result of applying E to dataset
D containing N observations. The bootstrap is a way of obtaining variance estimates for b from
repeated estimates b1 , b2 , . . . , where each bi is the result of applying E to a dataset of size N
drawn with replacement from D. See [SEM] sem option method( ) and [R] bootstrap.

vce(bootstrap) is allowed with sem but not gsem. You can obtain bootstrap results by prefixing
the gsem command with bootstrap:, but remember to specify bootstraps cluster() and
idcluster() options if you are fitting a multilevel model. See [SEM] intro 9.
Builder. The Builder is Statas graphical interface for building sem and gsem models. The Builder is
also known as the SEM Builder. See [SEM] intro 2, [SEM] Builder, and [SEM] Builder, generalized.
CFA, CFA models. CFA stands for confirmatory factor analysis. It is a way of analyzing measurement
models. CFA models is a synonym for measurement models.
CI. CI is an abbreviation for confidence interval.
clustered, vce(cluster clustvar). Clustered is the name we use for the generalized Huber/White/sandwich estimator of the VCE, which is the robust technique generalized to relax the
assumption that errors are independent across observations to be that they are independent across
clusters of observations. Within cluster, errors may be correlated.

Clustered standard errors are reported when sem or gsem option vce(cluster clustvar) is
specified. The other available techniques are OIM, OPG, robust, bootstrap, and jackknife. Also
available for sem only is EIM.
coefficient of determination. The coefficient of determination is the fraction (or percentage) of
variation (variance) explained by an equation of a model. The coefficient of determination is thus
like R2 in linear regression.
command language. Statas sem and gsem commands provide a way to specify SEMs. The alternative
is to use the Builder to draw path diagrams; see [SEM] intro 2, [SEM] Builder, and [SEM] Builder,
generalized.
552

Glossary

553

complementary log-log regression. Complementary log-log regression is a term for generalized linear
response functions that are family Bernoulli, link cloglog. It is used for binary outcome data.
Complementary log-log regression is also known in Stata circles as cloglog regression or just
cloglog. See generalized linear response functions.
conditional normality assumption. See normality assumption, joint and conditional.
constraints. See parameter constraints.
correlated uniqueness model. A correlated uniqueness model is a kind of measurement model in
which the errors of the measurements have a structured correlation. See [SEM] intro 5.
crossed-effects models. See multilevel models.
curved path. See path.
degree-of-freedom adjustment. In estimates of variances and covariances, a finite-population degreeof-freedom adjustment is sometimes applied to make the estimates unbiased.

Lets write an estimated variance as

bii and write the standard formula for the variance as

bii = Sii /N . If
bii is the variance of observable variable xi , it can readily be proven that Sii /N
is a biased estimate of the variances in samples of size N and that Sii /(N 1) is an unbiased
estimate. It is usual to calculate variances with Sii /(N 1), which is to say the standard formula
has a multiplicative degree-of-freedom adjustment of N/(N 1) applied to it.
If
bii is the variance of estimated parameter i , a similar finite-population degree-of-freedom
adjustment can sometimes be derived that will make the estimate unbiased. For instance, if i is a
coefficient from a linear regression, an unbiased estimate of the variance of regression coefficient
i is Sii /(N p 1), where p is the total number of regression coefficients estimated excluding
the intercept. In other cases, no such adjustment can be derived. Such estimators have no derivable
finite-sample properties, and one is left only with the assurances provided by its provable asymptotic
properties. In such cases, the variance of coefficient i is calculated as Sii /N , which can be
derived on theoretical grounds. SEM is an example of such an estimator.
SEM is a remarkably flexible estimator and can reproduce results that can sometimes be obtained by
other estimators. SEM might produce asymptotically equivalent results, or it might produce identical

results depending on the estimator. Linear regression is an example in which sem and gsem produce
the same results as regress. The reported standard errors, however, will not look identical because
the linear-regression estimates have the finite-population degree-of-freedom adjustment applied to
them and the SEM estimates do not. To see the equivalence, you must
p undo the adjustment on the
reported linear regression standard errors by multiplying them by {(N p 1)/N }.
direct, indirect, and total effects. Consider the following system of equations:

y1 = b10 + b11 y2 + b12 x1 + b13 x3 + e1

y2 = b20 + b21 y3 + b22 x1 + b23 x4 + e2
y3 = b30 +

b32 x1 + b33 x5 + e3

The total effect of x1 on y1 is b12 + b11 b22 + b11 b21 b32 . It measures the full change in y1 based
on allowing x1 to vary throughout the system.
The direct effect of x1 on y1 is b12 . It measures the change in y1 caused by a change in x1
holding other endogenous variablesnamely, y2 and y3 constant.
The indirect effect of x1 on y1 is obtained by subtracting the total and direct effect and is thus
b11 b22 + b11 b21 b32 .

554

Glossary

EIM, vce(eim). EIM stands for expected information matrix, defined as the inverse of the negative
of the expected value of the matrix of second derivatives, usually of the log-likelihood function.
The EIM is an estimate of the VCE. EIM standard errors are reported when sem option vce(eim)
is specified. EIM is available only with sem. The other available techniques for sem are OIM, OPG,
robust, clustered, bootstrap, and jackknife.
endogenous variable. A variable, observed or latent, is endogenous (determined by the system) if
any path points to it. Also see exogenous variable.
error, error variable. The error is random disturbance e in a linear equation:

y = b0 + b1 x1 + b2 x2 + + e
An error variable is an unobserved exogenous variable in path diagrams corresponding to e.
Mathematically, error variables are just another example of latent exogenous variables, but in sem
and gsem, error variables are considered to be in a class by themselves. All (Gaussian) endogenous
variablesobserved and latenthave a corresponding error variable. Error variables automatically
and inalterably have their path coefficients fixed to be 1. Error variables have a fixed naming
convention in the software. If a variable is the error for (observed or latent) endogenous variable
y, then the residual variables name is e.y.
In sem and gsem, error variables are uncorrelated with each other unless explicitly indicated
otherwise. That indication is made in path diagrams by drawing a curved path between the error
variables and is indicated in command notation by including cov(e.name1*e.name2) among the
options specified on the sem command. In gsem, errors for family Gaussian, link log responses
are not allowed to be correlated.
estimation method. There are a variety of ways that one can solve for the parameters of an SEM.
Different methods make different assumptions about the data-generation process, so it is important
that you choose a method appropriate for your model and data; see [SEM] intro 4.
exogenous variable. A variable, observed or latent, is exogenous (determined outside the system)
if paths only originate from it or, equivalently, no path points to it. In this manual, we do not
distinguish whether exogenous variables are strictly exogenousthat is, uncorrelated with the
errors. Also see endogenous variable.
family distribution. See generalized linear response functions.
fictional data. Fictional data are data that have no basis in reality even though they might look real;
they are data that are made up for use in examples.
first-, second-, and higher-level (latent) variables. Consider a multilevel model of patients within
doctors within hospitals. First-level variables are variables that vary at the observational (patient)
level. Second-level variables vary across doctors but are constant within doctors. Third-level
variables vary across hospitals but are constant within hospitals. This jargon is used whether
variables are latent or not.
first- and second-order latent variables. If a latent variable is measured by other latent variables
only, the latent variable that does the measuring is called first-order latent variable, and the latent
variable being measured is called the second-order latent variable.
full joint and conditional normality assumption. See normality assumption, joint and conditional.
gamma regression. Gamma regression is a term for generalized linear response functions that are
family gamma, link log. It is used for continuous, nonnegative, positively skewed data. Gamma
regression is also known as log-gamma regression. See generalized linear response functions.
Gaussian regression. Gaussian regression is another term for linear regression. It is most often used
when referring to generalized linear response functions. In that framework, Gaussian regression is
family Gaussian, link identity. See generalized linear response functions.

Glossary

555

generalized linear response functions. Generalized linear response functions include linear functions
and include functions such as probit, logit, multinomial logit, ordered probit, ordered logit, Poisson,
and more.

These generalized linear functions are described by a link function g() and statistical distribution
F . The link function g() specifies how the response variable yi is related to a linear equation of
the explanatory variables, xi , and the family F specifies the distribution of yi :

g{E(yi )} = xi ,

yi F

If we specify that g() is the identity function and F is the Gaussian (normal) distribution, then we
have linear regression. If we specify that g() is the logit function and F the Bernoulli distribution,
then we have logit (logistic) regression.
In this generalized linear structure, the family may be Gaussian, gamma, Bernoulli, binomial,
Poisson, negative binomial, ordinal, or multinomial. The link function may be the identity, log,
logit, probit, or complementary log-log.
gsem fits models with generalized linear response functions.
generalized method of moments. Generalized method of moments (GMM) is a method used to obtain
fitted parameters. In this documentation, GMM is referred to as ADF, which stands for asymptotic
distribution free and is available for use with sem. Other available methods for use with sem are
ML, QML, ADF, and MLMV.

The SEM moment conditions are cast in terms of second moments, not the first moments used in
many other applications associated with GMM.
generalized SEM. Generalized SEM is a term we have coined to mean SEM optionally allowing
generalized linear response functions or multilevel models. gsem fits generalized SEMs.
GMM. See generalized method of moments.
goodness-of-fit statistic. A goodness-of-fit statistic is a value designed to measure how well the
model reproduces some aspect of the data the model is intended to fit. SEM reproduces the firstand second-order moments of the data, with an emphasis on the second-order moments, and thus
goodness-of-fit statistics appropriate for use after sem compare the predicted covariance matrix
(and mean vector) with the matrix (and vector) observed in the data.
gsem. gsem is the Stata command that fits generalized SEMs. Also see sem.
GUI. See Builder.
identification. Identification refers to the conceptual constraints on parameters of a model that are
required for the models remaining parameters to have a unique solution. A model is said to be
unidentified if these constraints are not supplied. These constraints are of two types: substantive
constraints and normalization constraints.

Normalization constraints deal with the problem that one scale works as well as another for each
latent variable in the model. One can think, for instance, of propensity to write software as being
measured on a scale of 0 to 1, 1 to 100, or any other scale. The normalization constraints are the
constraints necessary to choose one particular scale. The normalization constraints are provided
automatically by sem and gsem by anchoring with unit loadings.
Substantive constraints are the constraints you specify about your model so that it has substantive
content. Usually, these constraints are 0 constraints implied by the paths omitted, but they can
include explicit parameter constraints as well. It is easy to write a model that is not identified for
substantive reasons; see [SEM] intro 4.

556

Glossary

indicator variables, indicators. The term indicator variable has two meanings. An indicator variable
is a 0/1 variable that contains whether something is true. The other usage is as a synonym for
measurement variables.
indirect effects. See direct, indirect, and total effects.
initial values. See starting values.
intercept. An intercept for the equation of endogenous variable y , observed or latent, is the path
coefficient from cons to y . cons is Stata-speak for the built-in variable containing 1 in all
observations. In SEM-speak, cons is an observed exogenous variable.
jackknife, vce(jackknife). The jackknife is a replication method for obtaining variance estimates.
Consider an estimation method E for estimating . Let b be the result of applying E to dataset
D containing N observations. The jackknife is a way of obtaining variance estimates for b from
repeated estimates b1 , b2 , . . . , bN , where each bi is the result of applying E to D with observation
i removed. See [SEM] sem option method( ) and [R] jackknife.

vce(jackknife) is allowed with sem but not gsem. You can obtain jackknife results by prefixing
the gsem command with jackknife:, but remember to specify jackknifes cluster() and
idcluster() options if you are fitting a multilevel model. See [SEM] intro 9.
joint normality assumption. See normality assumption, joint and conditional.
Lagrange multiplier test. Synonym for score test.
latent growth model. A latent growth model is a kind of measurement model in which the observed
values are collected over time and are allowed to follow a trend. See [SEM] intro 5.
latent variable. A variable is latent if it is not observed. A variable is latent if it is not in your dataset
but you wish it were. You wish you had a variable recording the propensity to commit violent
crime, or socioeconomic status, or happiness, or true ability, or even income accurately recorded.
Latent variables are sometimes described as imagined variables.

In the software, latent variables are usually indicated by having at least their first letter capitalized.
Also see first- and second-order latent variables, first-, second-, and higher-level (latent) variables,
and observed variables.
linear regression. Linear regression is a kind of SEM in which there is a single equation. See
[SEM] intro 5.
link function. See generalized linear response functions.
logit regression. Logit regression is a term for generalized linear response functions that are family
Bernoulli, link logit. It is used for binary outcome data. Logit regression is also known as logistic
regression and also simply as logit. See generalized linear response functions.
manifest variables. Synonym for observed variables.
measure, measurement, x a measurement of X, x measures X. See measurement variables.
measurement models, measurement component. A measurement model is a particular kind of
model that deals with the problem of translating observed values to values suitable for modeling.
Measurement models are often combined with structural models and then the measurement model
part is referred to as the measurement component. See [SEM] intro 5.
measurement variables, measure, measurement, x a measurement of X, x measures X. Observed
variable x is a measurement of latent variable X if there is a path connecting x X . Measurement
variables are modeled by measurement models. Measurement variables are also called indicator
variables.

Glossary

557

method. Method is just an English word and should be read in context. Nonetheless, method is used
here usually to refer to the method used to solve for the fitted parameters of an SEM. Those
methods are ML, QML, MLMV, and ADF. Also see technique.
MIMIC. See multiple indicators and multiple causes.
mixed-effects models. See multilevel models.
ML, method(ml). ML stands for maximum likelihood. It is a method to obtain fitted parameters. ML
is the default method used by sem and gsem. Other available methods for sem are QML, MLMV,
and ADF. Also available for gsem is QML.
MLMV, method(mlmv). MLMV stands for maximum likelihood with missing values. It is an ML
method used to obtain fitted parameters in the presence of missing values. MLMV is the method
used by sem when the method(mlmv) option is specified; method(mlmv) is not available with
gsem. Other available methods for use with sem are ML, QML, and ADF. These other methods
omit from the calculation observations that contain missing values.
modification indices. Modification indices are score tests for adding paths where none appear. The
paths can be for either coefficients or covariances.
moments (of a distribution). The moments of a distribution are the expected values of powers of a
random variable or centralized (demeaned) powers of a random variable. The first moments are
the expected or observed means, and the second moments are the expected or observed variances
and covariances.
multilevel models. Multilevel models are models that include unobserved effects (latent variables)
for different groups in the data. For instance, in a dataset of students, groups of students might
share the same teacher. If the teachers identity is recorded in the data, then one can introduce
a latent variable that is constant within teacher and that varies across teachers. This is called a
two-level model.

If teachers could in turn be grouped into schools, and school identities were recorded in the data,
then one can introduce another latent variable that is constant within school and varies across
schools. This is called a three-level (nested-effects) model.
In the above example, observations (students) are said to be nested within teacher nested within
school. Sometimes there is no such subsequent nesting structure. Consider workers nested within
occupation and industry. The same occupations appear in various industries and the same industries
appear within various occupations. We can still introduce latent variables at the occupation and
industry level. In such cases, the model is called a crossed-effects model.
The latent variables that we have discussed are also known as random effects. Any coefficients on
observed variables in the model are known as the fixed portion of the model. Models that contain
fixed and random portions are known as mixed-effects models.
multinomial logit regression. Multinomial logit regression is a term for generalized linear response
functions that are family multinomial, link logit. It is used for categorical-outcome data when the
outcomes cannot be ordered. Multinomial logit regression is also known as multinomial logistic
regression and as mlogit in Stata circles. See generalized linear response functions.
multiple correlation. The multiple correlation is the correlation between endogenous variable y and
its linear prediction.
multiple indicators and multiple causes. Multiple indicators and multiple causes is a kind of structural
model in which observed causes determine a latent variable, which in turn determines multiple
indicators. See [SEM] intro 5.

558

Glossary

multivariate regression. Multivariate regression is a kind of structural model in which each member
of a set of observed endogenous variables is a function of the same set of observed exogenous
variables and a unique random disturbance term. The disturbances are correlated. Multivariate
regression is a special case of seemingly unrelated regression.
negative binomial regression. Negative binomial regression is a term for generalized linear response
functions that are family negative binomial, link log. It is used for count data that are overdispersed
relative to Poisson. Negative binomial regression is also known as nbreg in Stata circles. See
generalized linear response functions.
nested-effects models. See multilevel models.
nonrecursive (structural) model (system), recursive (structural) model (system). A structural model
(system) is said to be nonrecursive if there are paths in both directions between one or more pairs
of endogenous variables. A system is recursive if it is a systemit has endogenous variables that
appear with paths from themand it is not nonrecursive.

A nonrecursive model may be unstable. Consider, for instance,

y1 = 2y2 + 1x1 + e1
y2 = 3y1 2x2 + e2
This model is unstable. To see this, without loss of generality, treat x1 + e1 and 2x2 + e2 as if
they were both 0. Consider y1 = 1 and y2 = 1. Those values result in new values y1 = 2 and
y2 = 3, and those result in new values y1 = 6 and y2 = 6, and those result in new values . . . .
Continue in this manner, and you reach infinity for both endogenous variables. In the jargon of the
mathematics used to check for this property, the eigenvalues of the coefficient matrix lie outside
the unit circle.
On the other hand, consider these values:

y1 =0.5y2 + 1x1 + e1
y2 =1.0y1 2x2 + e2
These results are stable in that the resulting values converge to y1 = 0 and y2 = 0. In the jargon
of the mathematics used to check for this property, the eigenvalues of the coefficient matrix lie
inside the unit circle.
Finally, consider the values

y1 =0.5y2 + 1x1 + e1
y2 =2.0y1 2x2 + e2
Start with y1 = 1 and y2 = 1. That yields new values y1 = 0.5, and y2 = 2 and that yields new
values y1 = 1 and y2 = 1, and that yields new values y1 = 0.5 and y2 = 2, and it will oscillate
forever. In the jargon of the mathematics used to check for this property, the eigenvalues of the
coefficient matrix lie on the unit circle. These coefficients are also considered to be unstable.
normalization constraints. See identification.
normality assumption, joint and conditional. The derivation of the standard, linear SEM estimator
usually assumes the full joint normality of the observed and latent variables. However, full joint
normality can replace the assumption of normality conditional on the values of the exogenous
variables, and all that is lost is one goodness-of-fit test (the test reported by sem on the output) and

Glossary

559

the justification for use of optional method MLMV for dealing with missing values. This substitution
of assumptions is important for researchers who cannot reasonably assume normality of the observed
variables. This includes any researcher including, say, variables age and age-squared in his or her
model.
Meanwhile, the generalized SEM makes only the conditional normality assumption.
Be aware that even though the full joint normality assumption is not required for the standard
linear SEM, sem calculates the log-likelihood value under that assumption. This is irrelevant except
that log-likelihood values reported by sem cannot be compared with log-likelihood values reported
by gsem, which makes the lesser assumption.
See [SEM] intro 4.
normalized residuals. See standardized residuals.
observed variables. A variable is observed if it is a variable in your dataset. In this documentation,
we often refer to observed variables by using x1, x2, . . . , y1, y2, and so on; in reality, observed
variables have names such as mpg, weight, testscore, and so on.

In the software, observed variables are usually indicated by having names that are all lowercase.
Also see latent variable.
OIM, vce(oim). OIM stands for observed information matrix, defined as the inverse of the negative
of the matrix of second derivatives, usually of the log likelihood function. The OIM is an estimate
of the VCE. OIM is the default VCE that sem and gsem report. The other available techniques are
EIM, OPG, robust, clustered, bootstrap, and jackknife.
OPG, vce(opg). OPG stands for outer product of the gradients, defined as the cross product of the
observation-level first derivatives, usually of the log likelihood function. The OPG is an estimate of
the VCE. The other available techniques are OIM, EIM, robust, clustered, bootstrap, and jackknife.
ordered complementary log-log regression. Ordered complementary log-log regression is a term
for generalized linear response functions that are family ordinal, link cloglog. It is used for
ordinal-outcome data. Ordered complementary log-log regression is also known as ocloglog in
Stata circles. See generalized linear response functions.
ordered logit regression. Ordered logit regression is a term for generalized linear response functions
that are family ordinal, link logit. It is used for ordinal outcome data. Ordered logit regression is
also known as ordered logistic regression, as just ordered logit, and as ologit in Stata circles. See
generalized linear response functions.
ordered probit regression. Ordered probit regression is a term for generalized linear response functions
that are family ordinal, link probit. It is used for ordinal-outcome data. Ordered probit regression
is also known as just ordered probit and known as oprobit in Stata circles. See generalized linear
response functions.
parameter constraints. Parameter constraints are restrictions placed on the parameters of the model.
These constraints are typically in the form of 0 constraints and equality constraints. A 0 constraint
is implied, for instance, when no path is drawn connecting x with y . An equality constraint is
specified when one path coefficient is forced to be equal to another or one covariance is forced
to be equal to another.

Also see identification.

parameters, ancillary parameters. The parameters are the to-be-estimated coefficients of a model.
These include all path coefficients, means, variances, and covariances. In mathematical notation,
the theoretical parameters are often written as = (, , , ), where is the vector of intercepts,

560

Glossary

is the vector of path coefficients, is the vector of means, and is the matrix of variances
and covariances. The resulting parameter estimates are written as b
.
Ancillary parameters are extra parameters beyond the ones just described that concern the distribution.
These include the scale parameter of gamma regression, the dispersion parameter for negative
binomial regression, and the cutpoints for ordered probit, logit, and cloglog regression, and the
like. These parameters are also included in .
path. A path, typically shown as an arrow drawn from one variable to another, states that the first
variable determines the second variable, at least partially. If x y , or equivalently y x, then
yj = + + xj + + e.yj , where is said to be the x y path coefficient. The ellipses
are included to account for paths to y from other variables. is said to be the intercept and is
automatically added when the first path to y is specified.

A curved path is a curved line connecting two variables, and it specifies that the two variables are
allowed to be correlated. If there is no curved path between variables, the variables are usually
assumed to be uncorrelated. We say usually because correlation is assumed among observed
exogenous variables and, in the command language, assumed among latent exogenous variables,
and if some of the correlations are not desired, they must be suppressed. Many authors refer to
covariances rather than correlations. Strictly speaking, the curved path denotes a nonzero covariance.
A correlation is often called a standardized covariance.
A curved path can connect a variable to itself, and in that case, it indicates a variance. In path
diagrams in this manual, we typically do not show such variance paths even though variances are
assumed.
path coefficient. The path coefficient is associated with a path; see path. Also see intercept.
path diagram. A path diagram is a graphical representation that shows the relationships among a set
of variables using paths. See [SEM] intro 2 for a description of path diagrams.
path notation. Path notation is a syntax defined by the authors of Statas sem and gsem commands
for entering path diagrams in a command language. Models to be fit may be specified in path
notation or they may be drawn using path diagrams into the Builder.
p-value. P -value is another term for the reported significance level associated with a test. Small
p-values indicate rejection of the null hypothesis.
Poisson regression. Poisson regression is a term for generalized linear response functions that are
family Poisson, link log. It is used for count data. See generalized linear response functions.
probit regression. Probit regression is a term for generalized linear response functions that are family
Bernoulli, link probit. It is used for binary outcome data. Probit regression is also known simply
as probit. See generalized linear response functions.
QML, method(ml) vce(robust). QML stands for quasimaximum likelihood. It is a method used to
obtain fitted parameters and a technique used to obtain the corresponding VCE. QML is used by sem
and gsem when options method(ml) and vce(robust) are specified. Other available methods
are ML, MLMV, and ADF. Other available techniques are OIM, EIM, OPG, clustered, bootstrap, and
jackknife.
quadrature. Quadrature is generic method for performing numerical integration. gsem uses quadrature
in any model including latent variables (excluding error variables). sem, being limited to linear
models, does not need to perform quadrature.
random-effects models. See multilevel models.
regression. A regression is a model in which an endogenous variable is written as a function of other
variables, parameters to be estimated, and a random disturbance.

Glossary

561

reliability. Reliability is the proportion of the variance of a variable not due to measurement error.
A variable without measure error has reliability 1.
residual. In this manual, we reserve the word residual for the difference between the observed
and fitted moments of an SEM. We use the word error for the disturbance associated with a
(Gaussian) linear equation; see error. Also see standardized residuals.
robust, vce(robust). Robust is the name we use here for the Huber/White/sandwich estimator of
the VCE. This technique requires fewer assumptions than most other techniques. In particular, it
merely assumes that the errors are independently distributed across observations and thus allows
the errors to be heteroskedastic. Robust standard errors are reported when the sem (gsem) option
vce(robust) is specified. The other available techniques are OIM, EIM, OPG, clustered, bootstrap,
and jackknife.
saturated model. A saturated model is a full covariance modela model of fitted means and
covariances of observed variables without any restrictions on the values. Also see baseline model.
Saturated models apply only to standard linear SEMs.
score test, Lagrange multiplier test. A score test is a test based on first derivatives of a likelihood
function. Score tests are especially convenient for testing whether constraints on parameters should
be relaxed or parameters should be added to a model. Also see Wald test.
scores. Scores has two unrelated meanings. First, scores are the observation-by-observation firstderivatives of the (quasi) log-likelihood function. When we use the word scores, this is what
we mean. Second, in the factor-analysis literature, scores (usually in the context of factor scores)
refers to the expected value of a latent variable conditional on all the observed variables. We refer
to this simply as the predicted value of the latent variable.
second-level latent variable. See first-, second-, and higher-order latent variables.
second-order latent variable. See first- and second-order latent variables.
seemingly unrelated regression. Seemingly unrelated regression is a kind of structural model in
which each member of a set of observed endogenous variables is a function of a set of observed
exogenous variables and a unique random disturbance term. The disturbances are correlated and
the sets of exogenous variables may overlap. If the sets of exogenous variables are identical, this
is referred to as multivariate regression.
SEM. SEM stands for structural equation modeling and for structural equation model. We use SEM
in capital letters when writing about theoretical or conceptual issues as opposed to issues of the
particular implementation of SEM in Stata with the sem or gsem commands.
sem. sem is the Stata command that fits standard linear SEMs. Also see gsem.
SSD, ssd. See summary statistics data.
standard linear SEM. An SEM without multilevel effects in which all response variables are given
by a linear equation. Standard linear SEM is what most people mean when they refer to just SEM.
Standard linear SEMs are fit by sem, although they can also be fit by gsem; see generalized SEM.
standardized coefficient. In a linear equation y = . . . bx + . . . , the standardized coefficient is
(b
y /b
x )b. Standardized coefficients are scaled to units of standard deviation change in y for a
standard deviation change in x.
standardized covariance. A standardized covariance between y and x is equal to the correlation of y
and x, that is, it is equal to xy /x y . The covariance is equal to the correlation when variables
are standardized to have variance 1.

562

Glossary

standardized residuals, normalized residuals. Standardized residuals are residuals adjusted so that
they follow a standard normal distribution. The difficulty is that the adjustment is not always
possible. Normalized residuals are residuals adjusted according to a different formula that roughly
follow a standard normal distribution. Normalized residuals can always be calculated.
starting values. The estimation methods provided by sem and gsem are iterative. The starting values
are values for each of the parameters to be estimated that are used to initialize the estimation
process. sem and gsem provide starting values automatically, but in some cases, these are not
good enough and you must both diagnose the problem and provide better starting values. See
[SEM] intro 12.
structural equation model. Different authors use the term structural equation model in different
ways, but all would agree that an SEM sometimes carries the connotation of being a structural
model with a measurement component, that is, combined with a measurement model.
structural model. A structural model is a model in which the parameters are not merely a description
but are believed to be of a causal nature. Obviously, SEM can fit structural models and thus so can
sem and gsem. Neither SEM, sem, nor gsem are limited to fitting structural models, however.

Structural models often have multiple equations and dependencies between endogenous variables,
although that is not a requirement.
See [SEM] intro 5. Also see structural equation model.
structured (correlation or covariance). See unstructured and structured (correlation or covariance).
substantive constraints. See identification.
summary statistics data. Data are sometimes available only in summary statistics form, as means
and covariances; means, standard deviations or variances, and correlations; covariances; standard
deviations or variances and correlations; or correlations. SEM can be used to fit models with such
data in place of the underlying raw data. The ssd command creates datasets containing summary
statistics.
technique. Technique is just an English word and should be read in context. Nonetheless, technique
is usually used here to refer to the technique used to calculate the estimated VCE. Those techniques
are OIM, EIM, OPG, robust, clustered, bootstrap, and jackknife.

Technique is also used to refer to the available techniques used with ml, Statas optimizer and
likelihood maximizer, to find the solution.
total effects. See direct, indirect, and total effects.
unstandardized coefficient. A coefficient that is not standardized. If mpg = 0.006 weight +
39.44028, then 0.006 is an unstandardized coefficient and, as a matter of fact, is measured in
mpg-per-pound units.
unstructured and structured (correlation or covariance). A set of variables, typically error variables,
is said to have an unstructured correlation or covariance if the covariance matrix has no particular
pattern imposed by theory. If a pattern is imposed, the correlation or covariance is said to be
structured.
variancecovariance matrix of the estimator. The estimator is the formula used to solve for the fitted
parameters, sometimes called the fitted coefficients. The VCE is the estimated variancecovariance
matrix of the parameters. The diagonal elements of the VCE are the variances of the parameters or
equivalent; the square roots of those elements are the reported standard errors of the parameters.
VCE. See variancecovariance matrix of the estimator.

Glossary

563

Wald test. A Wald test is a statistical test based on the estimated variancecovariance matrix of the
parameters. Wald tests are especially convenient for testing possible constraints to be placed on
the estimated parameters of a model. Also see score test.
weighted least squares. Weighted least squares (WLS) is a method used to obtain fitted parameters.
In this documentation, WLS is referred to as ADF, which stands for asymptotic distribution free.
Other available methods are ML, QML, and MLMV. ADF is, in fact, a specific kind of the more
generic WLS.
WLS. See weighted least squares.

Subject and author index

This is the subject and author index for the
Structural Equation Modeling Reference Manual.
Readers interested in topics other than structural equation
modeling should see the combined subject index (and
the combined author index) in the Glossary and Index.

A
Acock, A. C., [SEM] intro 4, [SEM] intro 5,
[SEM] intro 6, [SEM] intro 11,
[SEM] example 1, [SEM] example 3,
[SEM] example 7, [SEM] example 9,
[SEM] example 18, [SEM] example 20
adaptopt() option, see gsem option adaptopts()
addgroup, ssd subcommand, [SEM] ssd
ADF, see asymptotic distribution free
adf, see sem option method()
AIC, see Akaike information criterion
Akaike information criterion, [SEM] estat gof,
[SEM] example 4, [SEM] methods and
formulas for sem
Akaike, H., [SEM] estat gof, [SEM] methods and
formulas for sem
allmissing option, see sem option allmissing
Alwin, D. F., [SEM] example 9
anchoring, see constraints, normalization
Andrich, D., [SEM] example 28g
asymptotic distribution free, [SEM] intro 4,
[SEM] methods and formulas for sem,
[SEM] Glossary

B
Baron, R. M., [SEM] example 42g
baseline comparisons, [SEM] estat gof,
[SEM] example 4
baseline model, [SEM] estat gof, [SEM] example 4,
[SEM] methods and formulas for sem,
[SEM] Glossary
baseopts option, see sem option baseopts()
Bauldry, S., [SEM] intro 5
Bayesian information criterion, [SEM] estat gof,
[SEM] example 4, [SEM] methods and
formulas for sem
Bentham, G., [SEM] example 39g
Bentler, P. M., [SEM] estat eqgof, [SEM] estat
framework, [SEM] estat gof, [SEM] estat
stable, [SEM] example 3, [SEM] methods and
formulas for sem
BentlerRaykov squared multiple-correlation coefficient,
[SEM] estat eqgof
BentlerWeeks matrices, [SEM] intro 7, [SEM] estat
framework, [SEM] example 11, [SEM] Glossary
BIC, see Bayesian information criterion

565

binary outcome model, [SEM] intro 5,

[SEM] example 27g, [SEM] example 28g,
[SEM] example 29g, [SEM] example 30g,
[SEM] example 31g, [SEM] example 32g,
[SEM] example 33g, [SEM] example 34g
Bollen, K. A., [SEM] intro 4, [SEM] intro 5,
[SEM] estat residuals, [SEM] estat teffects,
[SEM] example 10, [SEM] example 15,
[SEM] methods and formulas for sem,
[SEM] predict after sem, [SEM] sem reporting
options
Bond, T. G., [SEM] example 28g
bootstrap, [SEM] Glossary
Boyle, P., [SEM] example 39g
Brown, T. A., [SEM] intro 4
Browne, M. W., [SEM] estat gof, [SEM] methods and
formulas for sem
build, ssd subcommand, [SEM] ssd
Builder (GUI), [SEM] Glossary

C
Campbell, D. T., [SEM] example 17
CD, see coefficient of determination
Center for Human Resource Research,
[SEM] example 38g, [SEM] example 46g
CFA, see confirmatory factor analysis
CFI, see comparative fit index
chi-squared test, [SEM] methods and formulas for sem
CI, see confidence interval
cloglog option, see gsem option cloglog
cluster, see gsem option vce(), see sem option
vce()
cluster estimator of variance, structural equation
modeling, [SEM] intro 8, [SEM] sem option
method( )
clustered, [SEM] Glossary
coefficient of determination, [SEM] estat eqgof,
[SEM] estat ggof, [SEM] estat gof,
[SEM] example 4, [SEM] example 21,
[SEM] methods and formulas for sem,
[SEM] Glossary
coeflegend option, see gsem option coeflegend, see
sem option coeflegend
collinear option, see gsem option collinear
command language, [SEM] Glossary
comparative fit index, [SEM] estat gof, [SEM] methods
and formulas for sem
complementary log-log regression, [SEM] Glossary
conditional normality, see normality, conditional
confidence interval, [SEM] Glossary
confirmatory factor analysis, [SEM] intro 5,
[SEM] example 15, [SEM] example 30g,
[SEM] Glossary
constraints, [SEM] sem and gsem option constraints( ),
[SEM] Glossary
across groups, [SEM] intro 6
normalization, [SEM] intro 4, [SEM] gsem,
[SEM] sem, [SEM] Glossary

566 Subject and author index

constraints, continued
relaxing, [SEM] intro 6, [SEM] sem and gsem path
notation, [SEM] sem path notation extensions
specifying, [SEM] intro 4, [SEM] intro 6,
[SEM] sem and gsem option constraints( ),
[SEM] sem and gsem option covstructure( ),
[SEM] sem and gsem path notation, [SEM] sem
path notation extensions
constraints() option, see gsem option
constraints(), see sem option
constraints()
contrast command, [SEM] intro 7
convergence, [SEM] intro 12, [SEM] sem, [SEM] sem
and gsem option from( )
correlated uniqueness model, [SEM] intro 5,
[SEM] example 17, [SEM] Glossary
correlation model, [SEM] intro 5, [SEM] Glossary
correlation, tests of, [SEM] estat stdize,
[SEM] example 16
count model, [SEM] intro 5, [SEM] example 34g,
[SEM] example 39g
covariance, [SEM] intro 4, [SEM] Glossary
assumptions, [SEM] gsem, [SEM] sem
covariance() option, see gsem option
covariance(), see sem option covariance()
covariances, creating dataset from, see summary
statistics data
covstructure() option, see gsem option
covstructure(), see sem option
covstructure()
Cox, C., [SEM] example 2
crossed-effects model, [SEM] example 40g,
[SEM] Glossary
Cudeck, R., [SEM] estat gof, [SEM] methods and
formulas for sem
curved path, [SEM] Glossary

D
datasignature command, [SEM] example 25,
[SEM] ssd
degree-of-freedom adjustment, [SEM] Glossary
delta method, [SEM] estat residuals, [SEM] estat
teffects
describe, ssd subcommand, [SEM] ssd
digitally signing data, see datasignature command
Duncan, O. D., [SEM] example 7

E
Eaves, R. C., [SEM] example 2
effects,
direct, [SEM] estat teffects, [SEM] example 7,
[SEM] example 42g, [SEM] methods and
formulas for sem, [SEM] Glossary

effects, continued
indirect, [SEM] estat teffects, [SEM] example 7,
[SEM] example 42g, [SEM] methods and
formulas for sem, [SEM] Glossary
total, [SEM] estat teffects, [SEM] example 7,
[SEM] example 42g, [SEM] methods and
formulas for sem, [SEM] Glossary
eform, estat subcommand, [SEM] estat eform
eigenvalue stability index, [SEM] estat stable
EIM, see expected information matrix
eim, see sem option vce()
Embretson, S. E., [SEM] example 28g,
[SEM] example 29g
empirical Bayes predictions, [SEM] intro 7,
[SEM] methods and formulas for gsem,
[SEM] predict after gsem
endogenous treatment-effects model,
[SEM] example 46g
endogenous variable, [SEM] intro 4, [SEM] Glossary
eqgof, estat subcommand, [SEM] estat eqgof
eqtest, estat subcommand, [SEM] estat eqtest
error, [SEM] Glossary
variable, [SEM] intro 4, [SEM] Glossary
estat
eform command, [SEM] intro 7, [SEM] estat
eform, [SEM] example 33g, [SEM] example 34g
eqgof command, [SEM] intro 7, [SEM] estat
eqgof, [SEM] example 3
eqtest command, [SEM] intro 7, [SEM] estat
eqtest, [SEM] example 13
framework command, [SEM] intro 7, [SEM] estat
framework, [SEM] example 11
ggof command, [SEM] intro 7, [SEM] estat ggof,
[SEM] example 21
ginvariant command, [SEM] intro 7, [SEM] estat
ginvariant, [SEM] example 22
gof command, [SEM] estat gof, [SEM] example 4
mindices command, [SEM] intro 7, [SEM] estat
mindices, [SEM] example 5, [SEM] example 9
residuals command, [SEM] intro 7, [SEM] estat
residuals, [SEM] example 10
scoretests command, [SEM] intro 7, [SEM] estat
scoretests, [SEM] example 8
stable command, [SEM] intro 7, [SEM] estat
stable, [SEM] example 7
stdize: prefix command, [SEM] estat stdize,
[SEM] example 16
summarize command, [SEM] estat summarize
teffects command, [SEM] estat teffects,
[SEM] example 7, [SEM] example 42g
estimation method, [SEM] Glossary
estimation options, [SEM] gsem estimation options,
[SEM] sem estimation options
Ex, [SEM] sem and gsem option covstructure( )
exogenous variable, [SEM] intro 4, [SEM] Glossary
expected information matrix, [SEM] Glossary

Subject and author index 567

exponentiated coefficients, [SEM] estat eform
exposure() option, see gsem option exposure()

F
factor analysis, see confirmatory factor analysis
factor scores, [SEM] intro 7, [SEM] example 14,
[SEM] methods and formulas for sem,
[SEM] predict after sem
factor-variable notation, [SEM] intro 3
family
Bernoulli, [SEM] methods and formulas for gsem
binomial, [SEM] methods and formulas for gsem
distribution, [SEM] Glossary
gamma, [SEM] methods and formulas for gsem
Gaussian, [SEM] methods and formulas for gsem
multinomial, [SEM] methods and formulas for
gsem
negative binomial, [SEM] methods and formulas
for gsem
ordinal, [SEM] methods and formulas for gsem
Poisson, [SEM] methods and formulas for gsem
family() option, see gsem option family()
feasible generalized least squares, [SEM] intro 4
feedback loops, [SEM] estat stable, [SEM] estat
teffects
fictional data, [SEM] Glossary
first-order latent variables, [SEM] Glossary
Fischer, G. H., [SEM] example 28g
Fiske, D. W., [SEM] example 17
forcecorrelations option, see sem option
forcecorrelations
forcenoanchor option, see gsem option
forcenoanchor, see sem option
forcenoanchor
forcexconditional option, see sem option
forcexconditional
Fox, C. M., [SEM] example 28g
framework, estat subcommand, [SEM] estat
framework
Freeman, E. H., [SEM] estat stable
from() option, see gsem option from(), see sem
option from()
fvstandard option, see gsem option fvstandard
fvwrap() option, see sem option fvwrap()
fvwrapon() option, see sem option fvwrapon()

G
gamma option, see gsem option gamma
gamma regression, [SEM] intro 5, [SEM] Glossary
Gaussian regression, [SEM] Glossary
generalized
least squares, feasible, see feasible generalized least
squares
linear response functions, [SEM] Glossary
method of moments, [SEM] Glossary

generalized, continued
response variables, [SEM] intro 2, [SEM] intro 5,
[SEM] gsem family-and-link options
responses, combined, [SEM] example 34g
SEM, [SEM] Glossary
ggof, estat subcommand, [SEM] estat ggof
ginvariant, estat subcommand, [SEM] estat
ginvariant
ginvariant() option, see sem option ginvariant()
GMM, see generalized method of moments
gof, estat subcommand, [SEM] estat gof
goodness of fit, [SEM] intro 7, [SEM] estat
eqgof, [SEM] estat ggof, [SEM] estat
gof, [SEM] example 3, [SEM] example 4,
[SEM] Glossary
graphical user interface, [SEM] Builder,
[SEM] Builder, generalized, [SEM] Glossary
Greenacre, M. J., [SEM] example 35g,
[SEM] example 36g
Greenfield, S., [SEM] example 37g
Gronau, R., [SEM] example 45g
group invariance test, [SEM] methods and formulas
for sem
group() option, see sem option group()
gsem command, [SEM] Builder, generalized,
[SEM] example 1, [SEM] example 27g,
[SEM] example 28g, [SEM] example 29g,
[SEM] example 30g, [SEM] example 31g,
[SEM] example 32g, [SEM] example 33g,
[SEM] example 34g, [SEM] example 35g,
[SEM] example 36g, [SEM] example 37g,
[SEM] example 38g, [SEM] example 39g,
[SEM] example 40g, [SEM] example 41g,
[SEM] example 42g, [SEM] example 43g,
[SEM] example 44g, [SEM] example 45g,
[SEM] example 46g, [SEM] gsem, [SEM] gsem
family-and-link options, [SEM] gsem model
description options, [SEM] gsem path notation
extensions, [SEM] gsem postestimation,
[SEM] methods and formulas for gsem,
[SEM] sem and gsem path notation
gsem option
adaptopts(), [SEM] gsem estimation options
cloglog, [SEM] gsem family-and-link options
coeflegend, [SEM] example 29g, [SEM] gsem
reporting options
collinear, [SEM] gsem model description
options
constraints(), [SEM] gsem model description
options, [SEM] sem and gsem option
constraints( )
covariance(), [SEM] gsem model description
options
covstructure(), [SEM] gsem model description
options, [SEM] sem and gsem option
covstructure( )
exposure(), [SEM] gsem family-and-link options
family(), [SEM] gsem family-and-link options,
[SEM] gsem model description options,
[SEM] gsem path notation extensions

568 Subject and author index

gsem option, continued
forcenoanchor, [SEM] gsem model description
options
from(), [SEM] intro 12, [SEM] gsem estimation
options, [SEM] gsem model description options,
[SEM] sem and gsem option from( )
fvstandard, [SEM] intro 3, [SEM] gsem model
description options
gamma, [SEM] gsem family-and-link options
intmethod(), [SEM] intro 12, [SEM] gsem
estimation options
intpoints(), [SEM] gsem estimation options
latent(), [SEM] sem and gsem syntax options
level(), [SEM] gsem reporting options
link(), [SEM] gsem family-and-link options,
[SEM] gsem model description options,
[SEM] gsem path notation extensions
listwise, [SEM] gsem estimation options
logit, [SEM] gsem family-and-link options
maximize options, [SEM] intro 12, [SEM] gsem
estimation options
means(), [SEM] gsem model description options
method(), [SEM] intro 8, [SEM] intro 9,
[SEM] gsem estimation options
mlogit, [SEM] gsem family-and-link options
nbreg, [SEM] gsem family-and-link options
noanchor, [SEM] gsem model description options
noasis, [SEM] gsem model description options
nocapslatent, [SEM] sem and gsem syntax
options
nocnsreport, [SEM] gsem reporting options
noconstant, [SEM] gsem model description
options
noestimate, [SEM] gsem estimation options
noheader, [SEM] gsem reporting options
notable, [SEM] gsem reporting options
ocloglog, [SEM] gsem family-and-link options
offset(), [SEM] gsem family-and-link options
ologit, [SEM] gsem family-and-link options
oprobit, [SEM] gsem family-and-link options
poisson, [SEM] gsem family-and-link options
probit, [SEM] gsem family-and-link options
regress, [SEM] gsem family-and-link options
reliability(), [SEM] intro 12, [SEM] gsem
model description options, [SEM] sem and
gsem option reliability( )
startgrid(), [SEM] intro 12, [SEM] gsem
estimation options
startvalues(), [SEM] intro 12, [SEM] gsem
estimation options
variance(), [SEM] gsem model description
options
vce(), [SEM] intro 8, [SEM] intro 9, [SEM] gsem
estimation options
gsem postestimation commands, [SEM] intro 7
GUI, see graphical user interface

H
Haller, A. O., [SEM] example 7
Hambleton, R. K., [SEM] example 28g,
[SEM] example 29g
Hancock, G. R., [SEM] estat gof, [SEM] methods and
formulas for sem
Hausman, J. A., [SEM] estat residuals, [SEM] methods
and formulas for sem
Heckman selection model, [SEM] example 43g
Heckman, J., [SEM] example 45g
higher-order models, see confirmatory factor analysis
Hocevar, D., [SEM] example 19
Hosmer, D. W., Jr., [SEM] example 33g,
[SEM] example 34g
Huber, C., [SEM] Builder, [SEM] Builder, generalized
Huber/White/sandwich estimator of variance, see robust,
Huber/White/sandwich estimator of variance
hypothesis test, [SEM] test, [SEM] testnl

I
identification, see model identification
indicator variables, [SEM] Glossary
information criteria, see Akaike information criterion,
see Bayesian information criterion
init, ssd subcommand, [SEM] ssd
initial values, [SEM] Glossary, see starting values
intercept, [SEM] intro 4, [SEM] Glossary, also see
constraints, specifying
interval regression model, [SEM] example 44g
intmethod() option, see gsem option intmethod()
intpoints() option, see gsem option intpoints()
IRT, see item response theory
item response theory, [SEM] intro 5,
[SEM] example 28g, [SEM] example 29g
iterate() option, see gsem option maximize options,
see sem option maximize options

J
jackknife, [SEM] Glossary
joint normality, see normality, joint
Joreskog, K. G., [SEM] estat residuals

K
Kenny, D. A., [SEM] intro 4, [SEM] example 42g
Kline, R. B., [SEM] intro 4, [SEM] example 3,
[SEM] example 4, [SEM] example 5
Krull, J. L., [SEM] example 42g

L
Lagrange multiplier test, [SEM] estat ginvariant,
[SEM] estat mindices, [SEM] estat scoretests,
[SEM] Glossary
Langford, I. H., [SEM] example 39g

Subject and author index 569

Laplacian approximation, [SEM] methods and
formulas for gsem
latent growth model, [SEM] intro 5,
[SEM] example 18, [SEM] Glossary
latent() option, see gsem option latent(), see sem
option latent()
latent variable, [SEM] intro 4, [SEM] Glossary
Lemeshow, S. A., [SEM] example 33g,
[SEM] example 34g
level() option, see gsem option level(), see sem
option level()
Lewis, H. G., [SEM] example 45g
LEx, [SEM] sem and gsem option covstructure( )
Li, C., [SEM] intro 4
likelihood-ratio test, [SEM] lrtest, [SEM] methods and
formulas for sem
lincom command, [SEM] intro 7, [SEM] estat stdize,
[SEM] lincom
linear regression, [SEM] intro 5, [SEM] example 6,
[SEM] Glossary
link
complementary log-log, [SEM] methods and
formulas for gsem
function, [SEM] Glossary
identity, [SEM] methods and formulas for gsem
log, [SEM] methods and formulas for gsem
logit, [SEM] methods and formulas for gsem
probit, [SEM] methods and formulas for gsem
link() option, see gsem option link()
list, ssd subcommand, [SEM] ssd
listwise option, see gsem option listwise
log likelihood, [SEM] methods and formulas for gsem,
[SEM] methods and formulas for sem
logistic and logit regression, [SEM] intro 5,
[SEM] example 33g, [SEM] example 34g,
[SEM] Glossary
logit option, see gsem option logit
lrtest command, [SEM] example 10,
[SEM] example 39g, [SEM] lrtest

M
MacKinnon, D. P., [SEM] example 42g
Mair, C. S., [SEM] example 39g
manifest variables, [SEM] Glossary
MAR, see missing values
margins command, [SEM] intro 7
Marsh, H. W., [SEM] example 19
maximize options, see gsem option maximize options,
see sem option maximize options
maximum likelihood, [SEM] intro 4, [SEM] methods
and formulas for gsem, [SEM] methods and
formulas for sem, [SEM] Glossary
with missing values, [SEM] example 26,
[SEM] Glossary
McDonald, A., [SEM] example 39g
means() option, see gsem option means(), see sem
option means()

measurement
component, [SEM] Glossary
error, [SEM] intro 5, [SEM] example 1,
[SEM] example 27g
model, [SEM] intro 5, [SEM] example 1,
[SEM] example 3, [SEM] example 20,
[SEM] example 27g, [SEM] example 30g,
[SEM] example 31g, [SEM] Glossary
variables, [SEM] Glossary
mediation model, [SEM] intro 5, [SEM] example 42g
Mehta, P. D., [SEM] example 30g
method, [SEM] Glossary
method() option, see gsem option method(), see sem
option method()
MIMIC models, see multiple indicators and multiple
causes model
mindices, estat subcommand, [SEM] estat mindices
missing values, [SEM] example 26
mixed-effects model, see multilevel model
ML, see maximum likelihood
ml, see gsem option method(), see sem option
method()
MLMV, see maximum likelihood with missing values
mlmv, see sem option method()
mlogit option, see gsem option mlogit
model identification, [SEM] intro 4, [SEM] intro 12,
[SEM] Glossary
model simplification test, [SEM] example 8,
[SEM] example 10
model-implied covariances and correlations,
[SEM] example 11
modification indices, [SEM] estat mindices,
[SEM] example 5, [SEM] methods and
formulas for sem, [SEM] Glossary
Molenaar, I. W., [SEM] example 28g
moments (of a distribution), [SEM] Glossary
MTMM, see multitraitmultimethod data and matrices
Mueller, R. O., [SEM] estat gof, [SEM] methods and
formulas for sem
multilevel latent variable, [SEM] intro 2, [SEM] gsem
path notation extensions
multilevel mixed-effects model, see multilevel model
multilevel model, [SEM] intro 5, [SEM] example 30g,
[SEM] example 38g, [SEM] example 39g,
[SEM] example 40g, [SEM] example 41g,
[SEM] example 42g, [SEM] Glossary
multinomial logistic regression, [SEM] intro 2,
[SEM] intro 5, [SEM] example 37g,
[SEM] example 41g, [SEM] Glossary
multiple correlation, [SEM] Glossary
multiple indicators and multiple causes model,
[SEM] intro 5, [SEM] example 10,
[SEM] Glossary
multiple indicators multiple causes model,
[SEM] example 36g
multitraitmultimethod data and matrices,
[SEM] intro 5, [SEM] example 17

570 Subject and author index

multivariate regression, [SEM] example 12,
[SEM] Glossary, also see seemingly unrelated
regression, see multivariate regression
Muthen, B., [SEM] example 9

N
nbreg option, see gsem option nbreg
Neale, M. C., [SEM] example 30g
negative binomial, [SEM] example 39g
negative binomial regression, [SEM] Glossary
Nelson, E. C., [SEM] example 37g
nested-effects model, [SEM] Glossary
nlcom command, [SEM] intro 7, [SEM] estat stdize,
[SEM] example 42g, [SEM] nlcom
nm1 option, see sem option nm1
noanchor option, see gsem option noanchor, see sem
option noanchor
noasis option, see gsem option noasis
nocapslatent option, see gsem option
nocapslatent, see sem option nocapslatent
nocnsreport option, see gsem option nocnsreport,
see sem option nocnsreport
noconstant option, see gsem option noconstant, see
sem option noconstant
nodescribe option, see sem option nodescribe
noestimate option, see gsem option noestimate, see
sem option noestimate
nofootnote option, see sem option nofootnote
nofvlabel option, see sem option nofvlabel
noheader option, see gsem option noheader, see sem
option noheader
noivstart option, see sem option noivstart
nomeans option, see sem option nomeans
noncursive model, see nonrecursive model
nonnormed fit index, see TuckerLewis index
nonrecursive model, [SEM] Glossary
stability of, [SEM] estat stable, [SEM] example 7
normality,
conditional, [SEM] intro 4, [SEM] Glossary
joint, [SEM] intro 4, [SEM] Glossary
normalization constraints, see constraints, normalization
normalized residuals, [SEM] estat residuals,
[SEM] methods and formulas for sem,
[SEM] Glossary
notable option, see gsem option notable, see sem
option notable
noxconditional option, see sem option
noxconditional

O
observed information matrix, [SEM] Glossary
observed variables, [SEM] intro 4, [SEM] Glossary
ocloglog option, see gsem option ocloglog
OEx, [SEM] sem and gsem option covstructure( )
offset() option, see gsem option offset()

OIM, see observed information matrix

oim, see gsem option vce(), see sem option vce()
ologit option, see gsem option ologit
OPG, see outer product of the gradient
oprobit option, see gsem option oprobit
ordered complementary log-log regression,
[SEM] Glossary
ordered logistic regression, [SEM] Glossary
ordered logit, [SEM] example 35g
ordered probit, [SEM] example 35g,
[SEM] example 36g
ordered probit regression, [SEM] Glossary
ordinal model, [SEM] intro 5, [SEM] example 31g,
[SEM] example 32g, [SEM] example 35g,
[SEM] example 36g
outer product of the gradient, [SEM] Glossary

P
p-value, [SEM] Glossary

parameter constraints, [SEM] estat ginvariant,

[SEM] Glossary
parameter values, obtaining symbolic names, see gsem
option coeflegend, see sem option coeflegend
parameters, [SEM] Glossary
combinations of, [SEM] lincom, [SEM] nlcom
path, [SEM] Glossary
adding, [SEM] intro 6
coefficients, [SEM] Glossary
constraining, [SEM] intro 4
diagrams, [SEM] intro 2, [SEM] intro 3,
[SEM] Glossary
model, [SEM] intro 5
notation, [SEM] intro 2, [SEM] intro 3,
[SEM] gsem path notation extensions,
[SEM] sem and gsem path notation, [SEM] sem
path notation extensions, [SEM] Glossary
pclose, [SEM] example 4
Perrin, E., [SEM] example 37g
Pickles, A., [SEM] Acknowledgments, [SEM] intro 2,
[SEM] example 29g, [SEM] methods and
formulas for gsem
Poisson, [SEM] example 34g, [SEM] example 39g
poisson option, see gsem option poisson
Poisson regression, [SEM] Glossary
population error, [SEM] estat gof, [SEM] example 4
Portes, A., [SEM] example 7
postestimation command, [SEM] intro 7, [SEM] gsem
postestimation, [SEM] sem postestimation
postestimation, predicted values, [SEM] intro 7,
[SEM] example 14, [SEM] example 28g,
[SEM] predict after gsem, [SEM] predict after
sem
Preacher, K. J., [SEM] example 42g
predict command, [SEM] intro 7, [SEM] example 14,
[SEM] example 28g, [SEM] predict after gsem,
[SEM] predict after sem
predicted values, see postestimation, predicted values
probit option, see gsem option probit

Subject and author index 571

probit regression, [SEM] Glossary
pwcompare command, [SEM] intro 7

Q
QML, see quasimaximum likelihood
quadrature, [SEM] Glossary
GaussHermite, [SEM] methods and formulas for
gsem
mean variance adaptive, [SEM] methods and
formulas for gsem
mode curvature adaptive, [SEM] methods and
formulas for gsem
quasimaximum likelihood, [SEM] Glossary

R
Rabe-Hesketh, S., [SEM] Acknowledgments,
[SEM] intro 2, [SEM] intro 4,
[SEM] example 28g, [SEM] example 29g,
[SEM] example 30g, [SEM] example 39g,
[SEM] example 40g, [SEM] example 41g,
[SEM] example 45g, [SEM] example 46g,
[SEM] methods and formulas for gsem,
[SEM] predict after gsem
Raftery, A. E., [SEM] estat gof
random intercept, [SEM] example 38g
random slope, [SEM] example 38g
random-effects model, [SEM] example 38g,
[SEM] Glossary
Rasch models, see item response theory
Rasch, G., [SEM] example 28g
raw residuals, [SEM] methods and formulas for sem
Raykov, T., [SEM] estat eqgof, [SEM] example 3,
[SEM] methods and formulas for sem
recursive model, [SEM] Glossary
regress option, see gsem option regress
regression, [SEM] Glossary
Reise, S. P., [SEM] example 28g, [SEM] example 29g
reliability, [SEM] intro 5, [SEM] intro 12,
[SEM] example 24, [SEM] gsem model
description options, [SEM] sem and gsem
option reliability( ), [SEM] sem model
description options, [SEM] Glossary
reliability option, see gsem option
reliability(), see sem option
reliability()
r. En, [SEM] sem and gsem option covstructure( )
repair, ssd subcommand, [SEM] ssd
replaying models, [SEM] intro 7
reporting options, [SEM] gsem reporting options,
[SEM] sem reporting options
residuals, [SEM] estat gof, [SEM] estat residuals,
[SEM] example 4, [SEM] Glossary
residuals, estat subcommand, [SEM] estat
residuals
RMSEA, see root mean squared error of approximation
robust, [SEM] Glossary
robust, see gsem option vce(), see sem option vce()

robust, Huber/White/sandwich estimator of variance,

structural equation modeling, [SEM] intro 8,
[SEM] sem option method( )
root mean squared error of approximation, [SEM] estat
gof, [SEM] example 4, [SEM] methods and
formulas for sem
R2 , [SEM] estat eqgof

S
sandwich/Huber/White estimator of variance, see robust,
Huber/White/sandwich estimator of variance
satopts() option, see sem option satopts()
saturated model, [SEM] estat gof, [SEM] example 4,
[SEM] methods and formulas for sem,
[SEM] Glossary
Schwarz, G., [SEM] estat gof, [SEM] methods and
formulas for sem
Schwarz information criterion, see Bayesian information
criterion
score test, [SEM] intro 7, [SEM] estat ginvariant,
[SEM] estat mindices, [SEM] estat scoretests,
[SEM] methods and formulas for sem,
[SEM] Glossary
scores, [SEM] Glossary
scoretests, estat subcommand, [SEM] estat
scoretests
second-order latent variables, [SEM] Glossary
seemingly unrelated regression, [SEM] intro 5,
[SEM] example 12, [SEM] Glossary
select() option, see sem option select()
SEM, see structural equation modeling
sem command, [SEM] Builder, [SEM] example 1,
[SEM] example 3, [SEM] example 6,
[SEM] example 7, [SEM] example 8,
[SEM] example 9, [SEM] example 10,
[SEM] example 12, [SEM] example 15,
[SEM] example 16, [SEM] example 17,
[SEM] example 18, [SEM] example 20,
[SEM] example 23, [SEM] example 24,
[SEM] example 26, [SEM] example 42g,
[SEM] methods and formulas for sem,
[SEM] sem, [SEM] sem and gsem path
notation, [SEM] sem model description
options, [SEM] sem path notation extensions,
[SEM] sem postestimation, [SEM] Glossary
missing values, [SEM] example 26
with constraints, [SEM] example 8
sem option
allmissing, [SEM] sem estimation options
baseopts(), [SEM] sem estimation options
coeflegend, [SEM] example 8,
[SEM] example 16, [SEM] sem reporting
options
constraints(), [SEM] sem and gsem option
constraints( ), [SEM] sem model description
options
covariance(), [SEM] sem and gsem path
notation, [SEM] sem model description options,
[SEM] sem path notation extensions

572 Subject and author index

sem option, continued
covstructure(), [SEM] intro 5,
[SEM] example 17, [SEM] sem and gsem
option covstructure( ), [SEM] sem model
description options
forcecorrelations, [SEM] sem ssd options
forcenoanchor, [SEM] sem model description
options
forcexconditional, [SEM] sem option
noxconditional
from(), [SEM] intro 12, [SEM] sem and gsem
option from( ), [SEM] sem model description
options
fvwrap(), [SEM] sem reporting options
fvwrapon(), [SEM] sem reporting options
ginvariant(), [SEM] intro 6, [SEM] example 23,
[SEM] sem group options
group(), [SEM] intro 6, [SEM] example 20,
[SEM] example 23, [SEM] sem group options,
[SEM] sem option select( ), [SEM] sem path
notation extensions
latent(), [SEM] sem and gsem syntax options
level(), [SEM] sem reporting options
maximize options, [SEM] intro 12, [SEM] sem
estimation options
means(), [SEM] intro 5, [SEM] example 18,
[SEM] sem and gsem path notation, [SEM] sem
model description options, [SEM] sem path
notation extensions
method(), [SEM] intro 4, [SEM] intro 8,
[SEM] intro 9, [SEM] example 26, [SEM] sem
estimation options, [SEM] sem option
method( ), [SEM] Glossary
nm1, [SEM] sem estimation options
noanchor, [SEM] sem model description options
nocapslatent, [SEM] sem and gsem syntax
options
nocnsreport, [SEM] sem reporting options
noconstant, [SEM] sem model description
options
nodescribe, [SEM] sem reporting options
noestimate, [SEM] sem estimation options
nofootnote, [SEM] sem reporting options
nofvlabel, [SEM] sem reporting options
noheader, [SEM] sem reporting options
noivstart, [SEM] sem estimation options
nomeans, [SEM] sem model description options
notable, [SEM] sem reporting options
noxconditional, [SEM] sem estimation options,
[SEM] sem option noxconditional
reliability(), [SEM] intro 12,
[SEM] example 24, [SEM] sem and gsem
option reliability( ), [SEM] sem model
description options
satopts(), [SEM] sem estimation options
select(), [SEM] sem option select( ), [SEM] sem
ssd options
showginvariant, [SEM] sem reporting options
standardized, [SEM] sem reporting options

sem option, continued

variance(), [SEM] sem and gsem path notation,
[SEM] sem model description options,
[SEM] sem path notation extensions
vce(), [SEM] intro 4, [SEM] intro 8,
[SEM] intro 9, [SEM] sem estimation options,
[SEM] sem option method( ), [SEM] Glossary
sem postestimation commands, [SEM] intro 7
set, ssd subcommand, [SEM] ssd
showginvariant option, see sem option
showginvariant
signing digitally data, see datasignature command
Skrondal, A., [SEM] Acknowledgments,
[SEM] intro 2, [SEM] intro 4,
[SEM] example 28g, [SEM] example 29g,
[SEM] example 30g, [SEM] example 39g,
[SEM] example 40g, [SEM] example 41g,
[SEM] example 45g, [SEM] example 46g,
[SEM] methods and formulas for gsem,
[SEM] predict after gsem
Smans, M., [SEM] example 39g
Sobel, M. E., [SEM] estat teffects
Sorbom, D., [SEM] estat ginvariant, [SEM] estat
mindices, [SEM] estat residuals, [SEM] estat
scoretests
Sparks, A. T., [SEM] example 41g
squared multiple correlation, [SEM] methods and
formulas for sem
SRMR, see standardized, root mean squared residual
SSD, see summary statistics data
ssd
addgroup command, [SEM] ssd
build command, [SEM] ssd
describe command, [SEM] ssd
init command, [SEM] ssd
list command, [SEM] ssd
repair command, [SEM] ssd
set command, [SEM] ssd
status command, [SEM] ssd
unaddgroup command, [SEM] ssd
stability of nonrecursive models, see nonrecursive
model, stability of
stable, estat subcommand, [SEM] estat stable
standard errors, see gsem option vce(), see sem option
vce()
robust, see robust, Huber/White/sandwich estimator
of variance
standard linear SEM, [SEM] Glossary
standardized
coefficients, [SEM] example 3, [SEM] example 6,
[SEM] Glossary, also see standardized
parameters
covariance, [SEM] Glossary
covariance residual, [SEM] methods and formulas
for sem
mean residual, [SEM] methods and formulas for
sem
option, [SEM] example 11

Subject and author index 573

standardized, continued
parameters, [SEM] estat stdize, [SEM] methods and
formulas for sem
residuals, [SEM] estat residuals, [SEM] methods
and formulas for sem, [SEM] Glossary
root mean squared residual, [SEM] estat
ggof, [SEM] estat gof, [SEM] example 4,
[SEM] example 21, [SEM] methods and
formulas for sem
standardized option, see sem option standardized
startgrid() option, see gsem option startgrid()
starting values, [SEM] intro 12, [SEM] sem and gsem
option from( ), [SEM] sem and gsem path
notation, [SEM] sem path notation extensions,
[SEM] Glossary
startvalues() option, see gsem option
startvalues()
status, ssd subcommand, [SEM] ssd
stdize, estat subcommand, [SEM] estat stdize
stored results, [SEM] intro 7
structural equation modeling, [SEM] methods and
formulas for gsem, [SEM] methods and
formulas for sem, [SEM] Glossary
structural model, [SEM] intro 5, [SEM] example 7,
[SEM] example 9, [SEM] example 32g,
[SEM] Glossary
structured, [SEM] Glossary
structured (correlation or covariance), [SEM] Glossary
Sturdivant, R. X., [SEM] example 33g,
[SEM] example 34g
substantive constraints, see constraints
summarize, estat subcommand, [SEM] estat
summarize
summary statistics, [SEM] estat summarize
summary statistics data, [SEM] intro 11,
[SEM] example 2, [SEM] example 19,
[SEM] example 25, [SEM] sem option
select( ), [SEM] sem ssd options, [SEM] ssd,
[SEM] Glossary
Summers, G. F., [SEM] example 9
sureg command, [SEM] intro 5, [SEM] example 12
survey data, [SEM] intro 10

T
Tarlov, A. R., [SEM] example 37g
technique, [SEM] Glossary
teffects, estat subcommand, [SEM] estat teffects
test,
chi-squared, see chi-squared test
goodness-of-fit, see goodness of fit
group invariance, see group invariance test
hypothesis, see hypothesis test
Lagrange multiplier, see Lagrange multiplier test
likelihood-ratio, see likelihood-ratio test
model simplification, see model simplification test
modification indices, see modification indices
score, see score test
Wald, see Wald test

test command, [SEM] estat stdize, [SEM] example 8,

[SEM] example 9, [SEM] example 16,
[SEM] test
testnl command, [SEM] estat stdize, [SEM] testnl
testparm command, [SEM] test
TLI, see TuckerLewis index
tobit regression model, [SEM] example 43g
treatment effects, [SEM] example 46g
TuckerLewis index, [SEM] estat gof, [SEM] methods
and formulas for sem

U
unaddgroup, ssd subcommand, [SEM] ssd
unit loading, [SEM] intro 4
unstandardized coefficient, [SEM] Glossary
unstructured, [SEM] Glossary

V
van der Linden, W. J., [SEM] example 28g,
[SEM] example 29g
variable types, [SEM] intro 4
variance,
analysis of, [SEM] intro 4
Huber/White/sandwich estimator, see robust,
Huber/White/sandwich estimator of variance
variancecovariance matrix of estimators,
[SEM] Glossary, also see gsem option vce(),
also see sem option vce()
variance() option, see gsem option variance(), see
sem option variance()
VCE, see variancecovariance matrix of estimators
vce() option, see gsem option vce(), see sem option
vce()

W
Wald test, [SEM] intro 7, [SEM] estat eqtest,
[SEM] estat ginvariant, [SEM] example 13,
[SEM] example 22, [SEM] methods and
formulas for sem, [SEM] test, [SEM] testnl,
[SEM] Glossary
Ware, J. E., Jr., [SEM] example 37g
Weeks, D. G., [SEM] estat framework
Weesie, J., [SEM] Acknowledgments
weighted least squares, [SEM] methods and formulas
for sem, [SEM] Glossary
Wheaton, B., [SEM] example 9
White/Huber/sandwich estimator of variance, see robust,
Huber/White/sandwich estimator of variance
Wiggins, V. L., [SEM] sem
Williams, T. O., Jr., [SEM] example 2
WLS, see weighted least squares
Wooldridge, J. M., [SEM] estat ginvariant,
[SEM] estat mindices, [SEM] estat scoretests,
[SEM] methods and formulas for sem
Wright, D. B., [SEM] example 41g

574 Subject and author index

Z
Zhang, Z., [SEM] example 42g
Zubkoff, M., [SEM] example 37g
Zyphur, M. J., [SEM] example 42g