Dealing With and Understanding Endogeneity

Enrique Pinzón

StataCorp LP

October 20, 2016

Barcelona

(StataCorp LP) October 20, 2016 Barcelona 1 / 59

Importance of Endogeneity

Endogeneity occurs when a variable, observed or unobserved,

that is not included in our models, is related to a variable we
incorporated in our model.

Model building
Endogeneity contradicts:
I Unobservables have no effect or explanatory power
I The covariates cause the outcome of interest
Endogeneity prevents us from making causal claims
Endogeneity is a fundamental concern of social scientists (first to
the party)

(StataCorp LP) October 20, 2016 Barcelona 2 / 59

Importance of Endogeneity

Endogeneity occurs when a variable, observed or unobserved,

that is not included in our models, is related to a variable we
incorporated in our model.

Model building
Endogeneity contradicts:
I Unobservables have no effect or explanatory power
I The covariates cause the outcome of interest
Endogeneity prevents us from making causal claims
Endogeneity is a fundamental concern of social scientists (first to
the party)

(StataCorp LP) October 20, 2016 Barcelona 2 / 59

Outline

1 Defining concepts and building our intuition

2 Stata built in tools to solve endogeneity problems
3 Stata commands to address endogeneity in non-built-in situations

(StataCorp LP) October 20, 2016 Barcelona 3 / 59

Defining concepts and building our intuition

(StataCorp LP) October 20, 2016 Barcelona 4 / 59

Building our Intuition: A Regression Model

The regression model is given by:

yi = β0 + β1 x1i + . . . + βk xki + εi
E (εi |x1i , . . . , xki ) = 0

Once we have the information of our regressors, on average what

we did not include in our model has no importance.

E (yi |x1i , . . . , xki ) = β0 + β1 x1i + . . . + βk xki

(StataCorp LP) October 20, 2016 Barcelona 5 / 59

Building our Intuition: A Regression Model

The regression model is given by:

yi = β0 + β1 x1i + . . . + βk xki + εi
E (εi |x1i , . . . , xki ) = 0

Once we have the information of our regressors, on average what

we did not include in our model has no importance.

E (yi |x1i , . . . , xki ) = β0 + β1 x1i + . . . + βk xki

(StataCorp LP) October 20, 2016 Barcelona 5 / 59

Graphically

(StataCorp LP) October 20, 2016 Barcelona 6 / 59

Examples of Endogeneity

We want to explain wages and we use years of schooling as a

covariate. Years of schooling is correlated with unobserved ability,
and work ethic.
We want to explain to probability of divorce and use employment
status as a covariate. Employment status might be correlated to
unobserved economic shocks.
We want to explain graduation rates for different school districts
and use the fraction of the budget used in education as a
covariate. Budget decisions are correlated to unobservable
political factors.
Estimating demand for a good using prices. Demand and prices
are determined simultaneously.

(StataCorp LP) October 20, 2016 Barcelona 7 / 59

A General Framework

If the unobservables, what we did not include in our model is

correlated to our covariates then:

E (ε|X ) 6= 0

Omitted variable “bias”

Simultaneity
Functional form misspecification
Selection “bias”
A useful implication of the above condition

E X 0 ε 6= 0

(StataCorp LP) October 20, 2016 Barcelona 8 / 59

A General Framework

If the unobservables, what we did not include in our model is

correlated to our covariates then:

E (ε|X ) 6= 0

Omitted variable “bias”

Simultaneity
Functional form misspecification
Selection “bias”
A useful implication of the above condition

E X 0 ε 6= 0

(StataCorp LP) October 20, 2016 Barcelona 8 / 59

A General Framework

If the unobservables, what we did not include in our model is

correlated to our covariates then:

E (ε|X ) 6= 0

Omitted variable “bias”

Simultaneity
Functional form misspecification
Selection “bias”
A useful implication of the above condition

E X 0 ε 6= 0

(StataCorp LP) October 20, 2016 Barcelona 8 / 59

Example 1: Omitted Variable “Bias”

The true model is given by

y = β 0 + β 1 x1 + β 2 x2 + ε
E (ε|x1 , x2 ) = 0

the researcher does not incorporate x2 , i.e. they think

y = β0 + β1 x1 + ν

The objective is to estimate β1 . In our framework we get a consistent

estimate if
E (ν|x1 ) = 0

(StataCorp LP) October 20, 2016 Barcelona 9 / 59

Example 1: Omitted Variable “Bias”

The true model is given by

y = β 0 + β 1 x1 + β 2 x2 + ε
E (ε|x1 , x2 ) = 0

the researcher does not incorporate x2 , i.e. they think

y = β0 + β1 x1 + ν

The objective is to estimate β1 . In our framework we get a consistent

estimate if
E (ν|x1 ) = 0

(StataCorp LP) October 20, 2016 Barcelona 9 / 59

Example 1: Endogeneity

Using the definition of the true model

y = β 0 + β 1 x1 + β 2 x2 + ε
E (ε|x1 , x2 ) = 0

We know that
ν = β2 x2 + ε
and
E (ν|x1 ) = β2 E (x2 |x1 )
E (ν|x1 ) = 0 only if β2 = 0 or x2 and x1 are uncorrelated

(StataCorp LP) October 20, 2016 Barcelona 10 / 59

Example 1: Endogeneity

Using the definition of the true model

y = β 0 + β 1 x1 + β 2 x2 + ε
E (ε|x1 , x2 ) = 0

We know that
ν = β2 x2 + ε
and
E (ν|x1 ) = β2 E (x2 |x1 )
E (ν|x1 ) = 0 only if β2 = 0 or x2 and x1 are uncorrelated

(StataCorp LP) October 20, 2016 Barcelona 10 / 59

Example 1 Simulating Data

. clear
. set obs 10000
number of observations (_N) was 0, now 10,000
. set seed 111
. // Generating a common component for x1 and x2
. generate a = rchi2(1)
. // Generating x1 and x2
. generate x1 = rnormal() + a
. generate x2 = rchi2(2)-3 + a
. generate e = rchi2(1) - 1
. // Generating the outcome
. generate y = 1 - x1 + x2 + e

(StataCorp LP) October 20, 2016 Barcelona 11 / 59

Example 1 Estimation

. // estimating true model

. quietly regress y x1 x2
. estimates store real
. //estimating model with omitted variable
. quietly regress y x1
. estimates store omitted
. estimates table real omitted, se

Variable real omitted

x1 -.98710456 -.31950213
.00915198 .01482454
x2 .99993928
.00648263
_cons .9920283 .32968254
.01678995 .02983985
legend: b/se

(StataCorp LP) October 20, 2016 Barcelona 12 / 59

Example 2: Simultaneity in a market equilibrium

The demand and supply equations for the market are given by

Qd = βPd + εd
Qs = θPs + εs

If a researcher wants to estimate Q d and ignores that P d is

simultaneously determined, we have an endogeneity problem that fits
in our framework.

(StataCorp LP) October 20, 2016 Barcelona 13 / 59

Example 2: Assumptions and Equilibrium

We assume:
All quantities are scalars
β < 0 and θ > 0
E (εd ) = E (εs ) = E (εd εs ) = 0
E ε2d ≡ σd2

The equilibrium prices and quantities are given by:

εs − εd
P =
β−θ
βεs − θεd
Q =
β−θ

(StataCorp LP) October 20, 2016 Barcelona 14 / 59

Example 2: Endogeneity
This is a simple linear model so we can verify if

E (Pd εd ) = 0

Using our equilibrium conditions and the fact that εs and εd are
uncorrelated we get
 
εs − εd
E (Pd εd ) = E εd
β−θ
E (εs εd ) E ε2d


= −
β−θ β−θ
2


E εd
= −
β−θ
σ2
= − d
β−θ

(StataCorp LP) October 20, 2016 Barcelona 15 / 59

Example 2: Endogeneity
This is a simple linear model so we can verify if

E (Pd εd ) = 0

Using our equilibrium conditions and the fact that εs and εd are
uncorrelated we get
 
εs − εd
E (Pd εd ) = E εd
β−θ
E (εs εd ) E ε2d


= −
β−θ β−θ
2


E εd
= −
β−θ
σ2
= − d
β−θ

(StataCorp LP) October 20, 2016 Barcelona 15 / 59

Example 2: Graphically

(StataCorp LP) October 20, 2016 Barcelona 16 / 59

Example 3: Functional Form Misspecification

Suppose the true model is given by:

y = sin(x) + ε
E (ε|x) = 0

But the researcher thinks that:

y = xβ + ν

(StataCorp LP) October 20, 2016 Barcelona 17 / 59

Example 3: Functional Form Misspecification

Suppose the true model is given by:

y = sin(x) + ε
E (ε|x) = 0

But the researcher thinks that:

y = xβ + ν

(StataCorp LP) October 20, 2016 Barcelona 17 / 59

Example 3: Real vs. Estimated Predicted values

(StataCorp LP) October 20, 2016 Barcelona 18 / 59

Example 3: Endogeneity

Adding zero we have

y = xβ − xβ + sin(x) + ε
y = xβ + ν
ν ≡ sin(x) − xβ + ε

For our estimates to be consistent we need to have E (ν|X ) = 0 but

E (ν|x) = sin(x) − xβ + E (ε|x)

= sin(x) − xβ
6= 0

(StataCorp LP) October 20, 2016 Barcelona 19 / 59

Example 3: Endogeneity

Adding zero we have

y = xβ − xβ + sin(x) + ε
y = xβ + ν
ν ≡ sin(x) − xβ + ε

For our estimates to be consistent we need to have E (ν|X ) = 0 but

E (ν|x) = sin(x) − xβ + E (ε|x)

= sin(x) − xβ
6= 0

(StataCorp LP) October 20, 2016 Barcelona 19 / 59

Example 3: Endogeneity

Adding zero we have

y = xβ − xβ + sin(x) + ε
y = xβ + ν
ν ≡ sin(x) − xβ + ε

For our estimates to be consistent we need to have E (ν|X ) = 0 but

E (ν|x) = sin(x) − xβ + E (ε|x)

= sin(x) − xβ
6= 0

(StataCorp LP) October 20, 2016 Barcelona 19 / 59

Example 4: Sample Selection

We observe the outcome of interest for a subsample of the

population
The subsample we observe is based on a rule For example we
observe y if y 2 ≥ 0
In a linear framework we have that:

E (y |X1 , y2 ≥ 0) = X1 β + E (ε|X1 , y2 ≥ 0)

If E (ε|X1 , y2 ≥ 0) 6= 0 we have selection bias

In the classic framework this happens if the selection rule is
related to the unobservables

(StataCorp LP) October 20, 2016 Barcelona 20 / 59

Example 4: Endogeneity

If we define X ≡ (X1 , y2 ≥ 0) we are back in our framework

E (y |X ) = X1 β + E (ε|X )

And we can define endogeneity as happening when:

E (ε|X ) 6= 0

(StataCorp LP) October 20, 2016 Barcelona 21 / 59

Example 4: Simulating data

. clear
. set seed 111
. quietly set obs 20000
.
. // Generating Endogenous Components
.
. matrix C = (1, .8\ .8, 1)
. quietly drawnorm e v, corr (C)
.
. // Generating exogenous variables
.
. generate x1 = rbeta(2 ,3)
. generate x2 = rbeta(2 ,3)
. generate x3 = rnormal()
. generate x4 = rchi2(1)
.
. // Generating outcome variables
.
. generate y1 = x1 - x2 + e
. generate y2 = 2 + x3 - x4 + v
. quietly replace y1 = . if y2 <=0

(StataCorp LP) October 20, 2016 Barcelona 22 / 59

Example 4: Estimation

. regress y1 x1 x2, nocons

Source SS df MS Number of obs = 14,847
F(2, 14845) = 813.88
Model 1453.18513 2 726.592566 Prob > F = 0.0000
Residual 13252.8872 14,845 .892750906 R-squared = 0.0988
Adj R-squared = 0.0987
Total 14706.0723 14,847 .990508004 Root MSE = .94485

y1 Coef. Std. Err. t P>|t| [95% Conf. Interval]

x1 1.153796 .0290464 39.72 0.000 1.096862 1.210731

x2 -.7896144 .0287341 -27.48 0.000 -.8459369 -.7332919

(StataCorp LP) October 20, 2016 Barcelona 23 / 59

What have we learnt

Endogeneity manifests itself in many forms

This manifestations can be understood within a general framework
Mathematically E (ε|X ) 6= 0 which implies E (X ε) 6= 0
Considerations that were not in our model (variables, selection,
simultaneity, functional form) affect the system and the model.

(StataCorp LP) October 20, 2016 Barcelona 24 / 59

Built-in tools to solve for endogeneity

(StataCorp LP) October 20, 2016 Barcelona 25 / 59

ivregress, ivpoisson, ivtobit, ivprobit, xtivreg
etregress, etpoisson, eteffects
biprobit, reg3, sureg, xthtaylor
heckman, heckprobit, heckoprobit

(StataCorp LP) October 20, 2016 Barcelona 26 / 59

Instrumental Variables

We model Y as a function of X1 and X2

X1 is endogenous
We can model X1
X1 can be divided into two parts; an endogenous part and an
exogenous part
X1 = f (X2 , Z ) + ν
Z are variables that affect Y only through X1
Z are referred to as intrumental variables or excluded instruments

(StataCorp LP) October 20, 2016 Barcelona 27 / 59

Instrumental Variables

We model Y as a function of X1 and X2

X1 is endogenous
We can model X1
X1 can be divided into two parts; an endogenous part and an
exogenous part
X1 = f (X2 , Z ) + ν
Z are variables that affect Y only through X1
Z are referred to as intrumental variables or excluded instruments

(StataCorp LP) October 20, 2016 Barcelona 27 / 59

Instrumental Variables

We model Y as a function of X1 and X2

X1 is endogenous
We can model X1
X1 can be divided into two parts; an endogenous part and an
exogenous part
X1 = f (X2 , Z ) + ν
Z are variables that affect Y only through X1
Z are referred to as intrumental variables or excluded instruments

(StataCorp LP) October 20, 2016 Barcelona 27 / 59

What Are These Instruments Anyway?

We are modeling income as a function of education. Education is

endogenous. Quarter of birth is an instrument, albeit weak.
We are modeling the demand for fish. We need to exclude the
supply shocks and keep only the demand shocks. Rain is an
instrument.

(StataCorp LP) October 20, 2016 Barcelona 28 / 59

Solving for Endogeneity Using Instrumental Variables

The solution is the get a consistent estimate of the exogenous

part and get rid of the endogenous part
An example is two-stage least squares
In two-stage least squares both relationships are linear

(StataCorp LP) October 20, 2016 Barcelona 29 / 59

Simulating the Model

. clear
. set seed 111
. set obs 10000
number of observations (_N) was 0, now 10,000
. generate a = rchi2(2)
. generate e = rchi2(1) -3 + a
. generate v = rchi2(1) -3 + a
. generate x2 = rnormal()
. generate z = rnormal()
. generate x1 = 1 - z + x2 + v
. generate y = 1 - x1 + x2 + e

(StataCorp LP) October 20, 2016 Barcelona 30 / 59

Estimation using Regression

. reg y x1 x2
Source SS df MS Number of obs = 10,000
F(2, 9997) = 1571.70
Model 12172.8278 2 6086.41388 Prob > F = 0.0000
Residual 38713.3039 9,997 3.87249214 R-squared = 0.2392
Adj R-squared = 0.2391
Total 50886.1317 9,999 5.08912208 Root MSE = 1.9679

y Coef. Std. Err. t P>|t| [95% Conf. Interval]

x1 -.4187662 .007474 -56.03 0.000 -.4334167 -.4041156

x2 .4382175 .0209813 20.89 0.000 .39709 .479345
_cons .4425514 .0210665 21.01 0.000 .4012569 .4838459

. estimates store reg

(StataCorp LP) October 20, 2016 Barcelona 31 / 59

Manual Two-Stage Least Squares (Wrong S.E.)

. quietly regress x1 z x2
. predict double x1hat
(option xb assumed; fitted values)
. preserve
. replace x1 = x1hat
(10,000 real changes made)
. quietly regress y x1 x2
. estimates store manual
. restore

(StataCorp LP) October 20, 2016 Barcelona 32 / 59

Estimation using Two-Stage Least Squares (2SLS)

. ivregress 2sls y x2 (x1=z)

Instrumental variables (2SLS) regression Number of obs = 10,000
Wald chi2(2) = 1613.38
Prob > chi2 = 0.0000
R-squared = .
Root MSE = 2.5174

y Coef. Std. Err. z P>|z| [95% Conf. Interval]

x1 -1.015205 .0252942 -40.14 0.000 -1.064781 -.9656292

x2 1.005596 .0348808 28.83 0.000 .9372314 1.073961
_cons 1.042625 .0357962 29.13 0.000 .9724656 1.112784

Instrumented: x1
Instruments: x2 z
. estimates store tsls

(StataCorp LP) October 20, 2016 Barcelona 33 / 59

Estimation

. estimates table reg tsls manual, se

Variable reg tsls manual

x1 -.41876618 -1.0152049 -1.0152049

.007474 .02529419 .02026373
x2 .4382175 1.0055965 1.0055965
.02098126 .03488076 .02794373
_cons .44255137 1.0426249 1.0426249
.02106646 .03579622 .02867713
legend: b/se

(StataCorp LP) October 20, 2016 Barcelona 34 / 59

Other Alternatives

sem, gsem, gmm

These are tools to construct our own estimation
sem and gsem model the unobservable correlation in multiple
equations
gmm is usually used to explicitly model a system of equations
where we model the endogenous variable

(StataCorp LP) October 20, 2016 Barcelona 35 / 59

What are sem and gsem

SEM is for structural equation modeling and GSEM is for

generalized structural equation modeling
sem fits linear models for continuous responses. Models only
allow for one level.
gsem continuous, binary, ordinal, count, or multinomial, responses
and multilevel modeling.
Estimation is done using maximum likelihood
It allows unobserved components in the equations and correlation
between equations

(StataCorp LP) October 20, 2016 Barcelona 36 / 59

What are sem and gsem

SEM is for structural equation modeling and GSEM is for

generalized structural equation modeling
sem fits linear models for continuous responses. Models only
allow for one level.
gsem continuous, binary, ordinal, count, or multinomial, responses
and multilevel modeling.
Estimation is done using maximum likelihood
It allows unobserved components in the equations and correlation
between equations

(StataCorp LP) October 20, 2016 Barcelona 36 / 59

What is gmm

Generalized Method of Moments

Estimation is based on being to write objects in the form

E [g (x, θ)] = 0

θ is the parameter of interest

If you can solve directly we have a method of moments.
When we have more moments than parameters we need to give
weights to the different moments and cannot solve directly.
The weight matrix gives more weight to the more efficient
moments.

(StataCorp LP) October 20, 2016 Barcelona 37 / 59

What is gmm

Generalized Method of Moments

Estimation is based on being to write objects in the form

E [g (x, θ)] = 0

θ is the parameter of interest

If you can solve directly we have a method of moments.
When we have more moments than parameters we need to give
weights to the different moments and cannot solve directly.
The weight matrix gives more weight to the more efficient
moments.

(StataCorp LP) October 20, 2016 Barcelona 37 / 59

Estimation Using sem
. sem (y <- x2 x1) (x1 <- x2 z), cov(e.y*e.x1) nolog
Endogenous variables
Observed: y x1
Exogenous variables
Observed: x2 z
Structural equation model Number of obs = 10,000
Estimation method = ml
Log likelihood = -71917.224

OIM
Coef. Std. Err. z P>|z| [95% Conf. Interval]

Structural
y <-
x1 -1.015205 .0252942 -40.14 0.000 -1.064781 -.9656292
x2 1.005596 .0348808 28.83 0.000 .9372314 1.073961
_cons 1.042625 .0357962 29.13 0.000 .9724656 1.112784

x1 <-
x2 .9467476 .0244521 38.72 0.000 .8988225 .9946728
z -.987925 .0241963 -40.83 0.000 -1.035349 -.9405011
_cons 1.011304 .0243764 41.49 0.000 .9635269 1.059081

var(e.y) 6.337463 .2275635 5.90678 6.799549

var(e.x1) 5.941873 .0840308 5.779438 6.108874
cov(e.y,e.x1) 4.134763 .1675226 24.68 0.000 3.806424 4.463101

LR test of model vs. saturated: chi2(0) = 0.00, Prob > chi2 = .

. estimates store sem

(StataCorp LP) October 20, 2016 Barcelona 38 / 59

Estimation Using gmm
. gmm (eq1: y - {xb: x1 x2 _cons}) ///
> (eq2: x1 - {xpi: x2 z _cons}), ///
> instruments(x2 z) ///
> winitial(unadjusted, independent) nolog
Final GMM criterion Q(b) = 4.70e-33
note: model is exactly identified
GMM estimation
Number of parameters = 6
Number of moments = 6
Initial weight matrix: Unadjusted Number of obs = 10,000
GMM weight matrix: Robust

Robust
Coef. Std. Err. z P>|z| [95% Conf. Interval]
xb
x1 -1.015205 .0252261 -40.24 0.000 -1.064647 -.9657627
x2 1.005596 .0362111 27.77 0.000 .934624 1.076569
_cons 1.042625 .0363351 28.69 0.000 .9714094 1.11384

xpi
x2 .9467476 .0251266 37.68 0.000 .8975004 .9959949
z -.987925 .0233745 -42.27 0.000 -1.033738 -.9421118
_cons 1.011304 .0243761 41.49 0.000 .9635274 1.05908

Instruments for equation eq1: x2 z _cons

Instruments for equation eq2: x2 z _cons
. estimates store gmm

(StataCorp LP) October 20, 2016 Barcelona 39 / 59

 y = β0 + x1 β1 + x2 β2 + ε
x1 = π0 + x2 π1 + zπ2 + ν
Z ≡ (x2 z)
E (Z ε) = E (Z ν) = 0

Where

ε = y − (β0 + x1 β1 + x2 β2 )
ν = x1 − (π0 + x2 π1 + zπ2 )

(StataCorp LP) October 20, 2016 Barcelona 40 / 59

 y = β0 + x1 β1 + x2 β2 + ε
x1 = π0 + x2 π1 + zπ2 + ν
Z ≡ (x2 z)
E (Z ε) = E (Z ν) = 0

Where

ε = y − (β0 + x1 β1 + x2 β2 )
ν = x1 − (π0 + x2 π1 + zπ2 )

(StataCorp LP) October 20, 2016 Barcelona 40 / 59

Summarizing the results of our estimation

. estimates table reg tsls sem gmm, eq(1) se ///

> keep(#1:x1 #1:x2 #1:_cons)

Variable reg tsls sem gmm

x1 -.41876618 -1.0152049 -1.0152049 -1.0152049

.007474 .02529419 .02529419 .02522609
x2 .4382175 1.0055965 1.0055965 1.0055965
.02098126 .03488076 .03488076 .03621111
_cons .44255137 1.0426249 1.0426249 1.0426249
.02106646 .03579622 .03579622 .03633511
legend: b/se

(StataCorp LP) October 20, 2016 Barcelona 41 / 59

Control Function Type Solutions

The key element here is to model the correlation between the

unobservables between the endogenous variable equation and
the outcome equation
This is what is referred to as a control function approach
Heckman selection is similar to this approach

(StataCorp LP) October 20, 2016 Barcelona 42 / 59

Heckman Selection

. clear
. set seed 111
. quietly set obs 20000
.
. // Generating Endogenous Components
.
. matrix C = (1, .4\ .4, 1)
. quietly drawnorm e v, corr (C)
.
. // Generating exogenous variables
.
. generate x1 = rbeta(2 ,3)
. generate x2 = rbeta(2 ,3)
. generate x3 = rnormal()
. generate x4 = rchi2(1)
.
. // Generating outcome variables
.
. generate y1 = -1 - x1 - x2 + e
. generate y2 = (1 + x3 - x4)*.5 + v
. quietly replace y1 = . if y2 <=0
. generate yp = y1 !=.

(StataCorp LP) October 20, 2016 Barcelona 43 / 59

Heckman Solution

Estimate a probit model for the selected observations as a

function of a set of variables Z
Then use the probit models to estimate:

E (y |X1 , y2 ≥ 0) = X1 β + E (ε|X1 , y2 ≥ 0)
φ (Z γ)
= X1 β + βs
Φ (Z γ)
φ(Z γ)
In other words regress y on X1 and Φ(Z γ)

(StataCorp LP) October 20, 2016 Barcelona 44 / 59

Heckman Solution

Estimate a probit model for the selected observations as a

function of a set of variables Z
Then use the probit models to estimate:

E (y |X1 , y2 ≥ 0) = X1 β + E (ε|X1 , y2 ≥ 0)
φ (Z γ)
= X1 β + βs
Φ (Z γ)
φ(Z γ)
In other words regress y on X1 and Φ(Z γ)

(StataCorp LP) October 20, 2016 Barcelona 44 / 59

Heckman Estimation
. heckman y1 x1 x2, select(x3 x4)
Iteration 0: log likelihood = -25449.645
Iteration 1: log likelihood = -25449.586
Iteration 2: log likelihood = -25449.586
Heckman selection model Number of obs = 20,000
(regression model with sample selection) Censored obs = 9,583
Uncensored obs = 10,417
Wald chi2(2) = 1098.75
Log likelihood = -25449.59 Prob > chi2 = 0.0000

y1 Coef. Std. Err. z P>|z| [95% Conf. Interval]

y1
x1 -1.117284 .0464766 -24.04 0.000 -1.208377 -1.026192
x2 -1.049901 .0458861 -22.88 0.000 -1.139836 -.9599656
_cons -.9559192 .0329022 -29.05 0.000 -1.020406 -.891432

select
x3 .4990633 .0104891 47.58 0.000 .478505 .5196216
x4 -.4785327 .0101864 -46.98 0.000 -.4984976 -.4585677
_cons .4807396 .0125354 38.35 0.000 .4561707 .5053084

/athrho .4614032 .0321988 14.33 0.000 .3982946 .5245117

/lnsigma -.0047001 .0092076 -0.51 0.610 -.0227466 .0133465

rho .4312271 .0262112 .3784888 .4811747

sigma .995311 .0091644 .9775102 1.013436
lambda .4292051 .0288551 .3726501 .4857601

LR test of indep. eqns. (rho = 0): chi2(1) = 208.78 Prob > chi2 = 0.0000
. estimates store heckman

(StataCorp LP) October 20, 2016 Barcelona 45 / 59

Two Steps Heuristically

. quietly probit yp x3 x4
. matrix A = e(b)
. quietly predict double xb, xb
. quietly generate double mills = normalden(xb)/normal(xb)
. quietly regress y1 x1 x2 mills
. matrix B = A, _b[x1], _b[x2], _b[_cons], _b[mills]

(StataCorp LP) October 20, 2016 Barcelona 46 / 59

GMM Estimation
. local xb {b1}*x1 + {b2}*x2 + {b0b}
. local mills (normalden({xp:})/normal({xp:}))
. gmm (eq2: yp*(normalden({xp: x3 x4 _cons})/normal({xp:})) - ///
> (1-yp)*(normalden(-{xp:})/normal(-{xp:}))) ///
> (eq1: y1 - (`xb´) - {b3}*(`mills´)) ///
> (eq3: (y1 - (`xb´) - {b3}*(`mills´))*`mills´), ///
> instruments(eq1: x1 x2) ///
> instruments(eq2: x3 x4) ///
> winitial(unadjusted, independent) quickderivatives ///
> nocommonesample from(B)
Step 1
Iteration 0: GMM criterion Q(b) = 2.279e-19
Iteration 1: GMM criterion Q(b) = 2.802e-34
Step 2
Iteration 0: GMM criterion Q(b) = 5.387e-34
Iteration 1: GMM criterion Q(b) = 5.387e-34
note: model is exactly identified
GMM estimation
Number of parameters = 7
Number of moments = 7
Initial weight matrix: Unadjusted Number of obs = *
GMM weight matrix: Robust
Robust
Coef. Std. Err. z P>|z| [95% Conf. Interval]
x3 .4992753 .0106148 47.04 0.000 .4784706 .52008
x4 -.4779557 .0104455 -45.76 0.000 -.4984285 -.4574828
_cons .4798264 .012609 38.05 0.000 .4551132 .5045397
/b1 -1.115395 .0472637 -23.60 0.000 -1.20803 -1.02276
/b2 -1.048694 .0455168 -23.04 0.000 -1.137905 -.9594823
/b0b -.9514073 .0332245 -28.64 0.000 -1.016526 -.8862885
/b3 .4199921 .0296825 14.15 0.000 .3618155 .4781686

* Number of observations for equation eq2: 20000

Number of observations for equation eq1: 10417
Number of observations for equation eq3: 10417
Instruments for equation eq2: x3 x4 _cons
(StataCorp LP) October 20, 2016 Barcelona 47 / 59
SEM Estimation of Heckman

. gsem (y1 <- x1 x2 L@a)(yp <- x3 x4 L@a, probit), ///

> var(L@1) nolog
Generalized structural equation model Number of obs = 20,000
Response : y1 Number of obs = 10,417
Family : Gaussian
Link : identity
Response : yp Number of obs = 20,000
Family : Bernoulli
Link : probit
Log likelihood = -25449.586
( 1) - [y1]L + [yp]L = 0
( 2) [var(L)]_cons = 1

Coef. Std. Err. z P>|z| [95% Conf. Interval]

y1 <-
x1 -1.117284 .0464766 -24.04 0.000 -1.208377 -1.026192
x2 -1.049901 .0458861 -22.88 0.000 -1.139836 -.9599656
L .7287588 .0296352 24.59 0.000 .6706749 .7868426
_cons -.9559206 .0329017 -29.05 0.000 -1.020407 -.8914345
yp <-
x3 .6175268 .0142797 43.24 0.000 .589539 .6455146
x4 -.5921228 .0140871 -42.03 0.000 -.619733 -.5645125
L .7287588 .0296352 24.59 0.000 .6706749 .7868426
_cons .5948535 .017244 34.50 0.000 .561056 .6286511
var(L) 1 (constrained)
var(e.y1) .4595557 .0322516 .4004984 .5273215
. estimates store hecksem

(StataCorp LP) October 20, 2016 Barcelona 48 / 59

Comparing SEM and HECKMAN

. estimates table heckman hecksem, eq(1) se ///

> keep(#1:x1 #1:x2 #1:L #1:_cons)

Variable heckman hecksem

x1 -1.117284 -1.1172841
.04647661 .04647661
x2 -1.0499007 -1.0499007
.04588611 .04588611
L .72875877
.02963515
_cons -.95591918 -.95592061
.03290222 .03290166

legend: b/se

(StataCorp LP) October 20, 2016 Barcelona 49 / 59

Non Built-In Situations

(StataCorp LP) October 20, 2016 Barcelona 50 / 59

Control Function Approach in a Linear Model: The
Model

. clear
. set seed 111
. set obs 10000
number of observations (_N) was 0, now 10,000
. generate a = rchi2(2)
. generate e = rchi2(1) -3 + a
. generate v = rchi2(1) -3 + a
. generate x2 = rnormal()
. generate z = rnormal()
. generate x1 = 1 - z + x2 + v
. generate y = 1 - x1 + x2 + e

(StataCorp LP) October 20, 2016 Barcelona 51 / 59

Estimation Using a Control Function Approach
The underlying model is

y = X1 β1 + X2 β2 + ε
X2 = X1 Π1 + Z Π2 + ν
ε = νρ + 
E (|X1 , X2 ) = 0

This implies that:

y = X1 β1 + X2 β2 + νρ + 

We can regress y on X1 , X2 , and ν

We can test for endogeneity

(StataCorp LP) October 20, 2016 Barcelona 52 / 59

Estimation Using a Control Function Approach
The underlying model is

y = X1 β1 + X2 β2 + ε
X2 = X1 Π1 + Z Π2 + ν
ε = νρ + 
E (|X1 , X2 ) = 0

This implies that:

y = X1 β1 + X2 β2 + νρ + 

We can regress y on X1 , X2 , and ν

We can test for endogeneity

(StataCorp LP) October 20, 2016 Barcelona 52 / 59

Estimation Using a Control Function Approach
The underlying model is

y = X1 β1 + X2 β2 + ε
X2 = X1 Π1 + Z Π2 + ν
ε = νρ + 
E (|X1 , X2 ) = 0

This implies that:

y = X1 β1 + X2 β2 + νρ + 

We can regress y on X1 , X2 , and ν

We can test for endogeneity

(StataCorp LP) October 20, 2016 Barcelona 52 / 59

Estimation of Control Function Using gmm

. local xbeta {b1}*x1 + {b2}*x2 + {b3}*(x1-{xpi:}) + {b0}

. gmm (eq3: (x1 - {xpi:x2 z _cons})) ///
> (eq1: y - (`xbeta´)) ///
> (eq2: (y - (`xbeta´))*(x1-{xpi:})), ///
> instruments(eq3: x2 z) ///
> instruments(eq1: x1 x2) ///
> winitial(unadjusted, independent) nolog
Final GMM criterion Q(b) = 1.45e-32
note: model is exactly identified
GMM estimation
Number of parameters = 7
Number of moments = 7
Initial weight matrix: Unadjusted Number of obs = 10,000
GMM weight matrix: Robust
Robust
Coef. Std. Err. z P>|z| [95% Conf. Interval]

x2 .9467476 .0251266 37.68 0.000 .8975004 .9959949

z -.987925 .0233745 -42.27 0.000 -1.033738 -.9421118
_cons 1.011304 .0243761 41.49 0.000 .9635274 1.05908
/b1 -1.015205 .0252261 -40.24 0.000 -1.064647 -.9657627
/b2 1.005596 .0362111 27.77 0.000 .934624 1.076569
/b3 .6958685 .0284014 24.50 0.000 .6402028 .7515342
/b0 1.042625 .0363351 28.69 0.000 .9714094 1.11384
Instruments for equation eq3: x2 z _cons
Instruments for equation eq1: x1 x2 _cons
Instruments for equation eq2: _cons

(StataCorp LP) October 20, 2016 Barcelona 53 / 59

Ordered Probit with Endogeneity

The model is given by:

y1∗ = y2 β + xΠ + ε
y2 = xγ1 + zγ2 + ν
y1 = j if κj−1 < y1∗ < κj
κ0 = −∞ < κ1 < . . . < κk = ∞
ε ∼ N (0, 1)
cov (ν, ε) 6= 0

(StataCorp LP) October 20, 2016 Barcelona 54 / 59

gsem Representation

∗
y1gsem = y2 b + xπ + t + Lα
t ∼ N (0, 1)
L ∼ N (0, 1)
∗
Where y1gsem = My1∗ and M is a constant. Noting that
∗
y1gsem = My1∗
y2 b + xπ + t + Lα = y2 Mβ + xMΠ + Mε

Which implies that

Mε = t + Lα
2
M Var (ε) = Var (t + Lα)
M2 = 1
p+ α
2

M = 1 + α2
(StataCorp LP) October 20, 2016 Barcelona 55 / 59
gsem Representation

∗
y1gsem = y2 b + xπ + t + Lα
t ∼ N (0, 1)
L ∼ N (0, 1)
∗
Where y1gsem = My1∗ and M is a constant. Noting that
∗
y1gsem = My1∗
y2 b + xπ + t + Lα = y2 Mβ + xMΠ + Mε

Which implies that

Mε = t + Lα
2
M Var (ε) = Var (t + Lα)
M2 = 1
p+ α
2

M = 1 + α2
(StataCorp LP) October 20, 2016 Barcelona 55 / 59
gsem Representation

∗
y1gsem = y2 b + xπ + t + Lα
t ∼ N (0, 1)
L ∼ N (0, 1)
∗
Where y1gsem = My1∗ and M is a constant. Noting that
∗
y1gsem = My1∗
y2 b + xπ + t + Lα = y2 Mβ + xMΠ + Mε

Which implies that

Mε = t + Lα
2
M Var (ε) = Var (t + Lα)
M2 = 1
p+ α
2

M = 1 + α2
(StataCorp LP) October 20, 2016 Barcelona 55 / 59
Ordered Probit with Endogeneity: Simulation

. clear
. set seed 111
. set obs 10000
number of observations (_N) was 0, now 10,000
. forvalues i = 1/5 {
2. gen x`i´ = rnormal()
3. }
.
. mat C = [1,.5 \ .5, 1]
. drawnorm e1 e2, cov(C)
.
. gen y2 = 0
. forvalues i = 1/5 {
2. quietly replace y2 = y2 + x`i´
3. }
. quietly replace y2 = y2 + e2
.
. gen y1star = y2 + x1 + x2 + e1
. gen xb1 = y2 + x1 + x2
.
. gen y1 = 4
.
. quietly replace y1 = 3 if xb1 + e1 <=.8
. quietly replace y1 = 2 if xb1 + e1 <=.3
. quietly replace y1 = 1 if xb1 + e1 <=-.3
. quietly replace y1 = 0 if xb1 + e1 <=-.8

(StataCorp LP) October 20, 2016 Barcelona 56 / 59

Ordered Probit with Endogeneity: Estimation
. gsem (y1 <- y2 x1 x2 L@a, oprobit)(y2 <- x1 x2 x3 x4 x5 L@a), var(L@1) nolog
Generalized structural equation model Number of obs = 10,000
Response : y1
Family : ordinal
Link : probit
Response : y2
Family : Gaussian
Link : identity
Log likelihood = -18948.444
( 1) [y1]L - [y2]L = 0
( 2) [var(L)]_cons = 1

Coef. Std. Err. z P>|z| [95% Conf. Interval]

y1 <-
y2 1.284182 .0217063 59.16 0.000 1.241638 1.326725
x1 1.28408 .0290087 44.27 0.000 1.227224 1.340936
x2 1.293582 .0287252 45.03 0.000 1.237282 1.349883
L .7968852 .0155321 51.31 0.000 .7664428 .8273275
y2 <-
x1 .9959898 .0099305 100.30 0.000 .9765263 1.015453
x2 1.002053 .0099196 101.02 0.000 .9826106 1.021495
x3 .9938048 .0096164 103.34 0.000 .974957 1.012653
x4 .9984898 .0095031 105.07 0.000 .9798642 1.017115
x5 1.002206 .0095257 105.21 0.000 .9835358 1.020876
L .7968852 .0155321 51.31 0.000 .7664428 .8273275
_cons .0089433 .0099196 0.90 0.367 -.0104987 .0283853
y1
/cut1 -1.017707 .0291495 -34.91 0.000 -1.074839 -.9605751
/cut2 -.4071202 .0273925 -14.86 0.000 -.4608085 -.3534319
/cut3 .4094317 .0275357 14.87 0.000 .3554628 .4634006
/cut4 1.017637 .029513 34.48 0.000 .9597921 1.075481
var(L) 1 (constrained)
var(e.y2) .348641 .0231272 .3061354 .3970482

(StataCorp LP) October 20, 2016 Barcelona 57 / 59

Ordered Probit with Endogeneity: Transformation

. nlcom _b[y1:y2]/sqrt(1 + _b[y1:L]^2)

_nl_1: _b[y1:y2]/sqrt(1 + _b[y1:L]^2)

Coef. Std. Err. z P>|z| [95% Conf. Interval]

_nl_1 1.004302 .0189557 52.98 0.000 .9671491 1.041454

. nlcom _b[y1:x1]/sqrt(1 + _b[y1:L]^2)

_nl_1: _b[y1:x1]/sqrt(1 + _b[y1:L]^2)

Coef. Std. Err. z P>|z| [95% Conf. Interval]

_nl_1 1.004222 .0214961 46.72 0.000 .9620909 1.046354

. nlcom _b[y1:x2]/sqrt(1 + _b[y1:L]^2)

_nl_1: _b[y1:x2]/sqrt(1 + _b[y1:L]^2)

Coef. Std. Err. z P>|z| [95% Conf. Interval]

_nl_1 1.011654 .0213625 47.36 0.000 .9697838 1.053523

(StataCorp LP) October 20, 2016 Barcelona 58 / 59

Conclusion

We established a general framework for endogeneity where the

problem is that the unobservables are related to observables
We saw solutions using instrumental variables or modeling the
correlation between unobservables
We saw how to use gmm and gsem to estimate this models both in
the cases of existing Stata commands and situations not available
in Stata

(StataCorp LP) October 20, 2016 Barcelona 59 / 59