Implementing new econometric tools in Stata

Christopher F Baum

Boston College and DIW Berlin

Mexican Stata Users Group Meeting, May 2013

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 1 / 73
 Introduction

Introduction

Focus of the talk: the implementation of two state-of-the-art

econometric estimators in Stata and Mata.

1 estimating a binary response with one or more limited

endogenous variables: sspecialreg.
2 estimating an equation with instrumental variables techniques
where sufficient instruments may not be available: ivreg2h.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 2 / 73
Modeling a binary outcome with an endogenous binary regressor

Topic I: Modeling a binary outcome with an

endogenous binary regressor

Acknowledgement
This presentation is based on the work of Lewbel, Dong & Yang,
“Viewpoint: Comparing features of Convenient Estimators for Binary
Choice Models With Endogenous Regressors”, Canadian Journal of
Economics, 45:3, 2012. Baum’s contribution is the review and
enhancement of the software developed in this research project.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 3 / 73
 Motivation

Motivation

Researchers often want to estimate a binomial response, or

binary choice, model where one or more explanatory variables are
endogenous or mismeasured.
For instance, in policy analysis, we want to estimate treatment
effects when treatment is not randomly assigned, or based solely
on the observables.
A linear 2SLS model, equivalent to a linear probability model with
instrumental variables, is often employed, ignoring the binary
outcome.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 4 / 73
 Motivation

Several alternative approaches exist:

linear probability model (LPM) with instruments
maximum likelihood estimation
control function based estimation
‘special regressor’ methods

Each of these estimators has advantages and disadvantages, and

some of these disadvantages are rarely acknowledged.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 5 / 73
 Motivation

In what follows, we focus on a particular disadvantage of the linear

probability model, and propose a straightforward alternative based on
‘special regressor’ methods (Lewbel, J. Econometrics, 2000; Lewbel,
Dong and Yang, Can. J. Econ., 2012).

We also propose the average index function (AIF), an alternative to the

average structural function (ASF; Blundell and Powell, Rev. Ec. Stud.,
2004), for calculating marginal effects. It is easy to construct and
estimate, as we will illustrate.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 6 / 73
 Binary choice models

Binary choice models

We define D as an observed binary variable: the outcome to be

explained. Let X be a vector of observed regressors, and β a
corresponding coefficient vector, with ε an unobserved error. In a
treatment model, X would include a binary treatment indicator T . In
general, X could be divided into X e , possibly correlated with ε, and X 0 ,
which are exogenous.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 7 / 73
 Binary choice models

A binary choice or ‘threshold crossing’ model estimated by maximum

likelihood is
D = I(X β + ε ≥ 0)
where I(·) is the indicator function. This latent variable approach is that
employed in a binomial probit or logit model, with Normal or logistic
errors, respectively. Although estimation provides point and interval
estimates of β, the choice probabilities and marginal effects are of
interest: that is, Pr[D = 1|X ] and ∂Pr[D = 1|X ]/∂X .

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 8 / 73
 Linear probability models

Linear probability models

In contrast to the threshold crossing latent variable approach, a linear

probability model (LPM) assumes that

D = Xβ + ε

so that the estimated coefficients β̂ are themselves the marginal

effects. With all exogenous regressors, E(D|X ) = Pr[D = 1|X ] = X β.

If some elements of X (possibly including treatment indicators) are

endogenous or mismeasured, they will be correlated with ε. In that
case, an instrumental variables approach is called for, and we can
estimate the LPM with 2SLS or IV-GMM, given an appropriate set of
instruments Z .

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 9 / 73
 Linear probability models

As the LPM with exogenous explanatory variables is based on

standard regression, the zero conditional mean assumption
E(ε|X ) = 0 applies. In the presence of endogeneity or measurement
error, the corresponding assumption E(ε|Z ) = 0 applies, with Z the set
of instruments, including the exogenous elements of X .

An obvious flaw in the LPM: the error ε cannot be independent of any

regressors, even exogenous regressors, unless X consists of a single
binary regressor. This arises because for any given X , ε must equal
either 1 − X β or −X β, which are functions of all elements of X .

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 10 / 73
 Linear probability models

The other, well recognized, flaw in the LPM is that its fitted values are
not constrained to lie in the unit interval, so that predicted probabilities
below zero or above one are commonly encountered. Any regressor
that can take on a large range of values will inevitably cause the LPM’s
predictions to breach these bounds.

A common rejoinder to these critiques is that the LPM is only intended

to approximate the true probability for a limited range of X values, and
that its constant marginal effects are preferable to those of the binary
probit or logit model, which are functions of the values of all elements
of X .

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 11 / 73
 Linear probability models

Consider, however, the LPM with a single continuous regressor. The

linear prediction is an approximation to the S-shape of any cumulative
distribution function: for instance, that of the Normal for the probit
model. The linear prediction departs greatly from the S-shaped CDF
long before it nears the (0,1) limits. Thus, the LPM will produce
predicted probabilities that are too extreme (closer to zero or one) even
for moderate values of X β̂ that stay ‘in bounds’.

Some researchers claim that although predicted probabilities derived

from the LPM are flawed, their main interest lies in the models’
marginal effects, and argue that it makes little substantive difference to
use a LPM, with its constant marginal effects, rather than the more
complex marginal effects derived from a proper estimated CDF, such
as that of the probit model.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 12 / 73
 Linear probability models Examples of support for the LPM approach

E XAMPLE 1

Jeffrey Wooldridge’s widely used undergraduate text, Introductory

Econometrics: A Modern Approach devotes a section of the chapter
on regression with qualitative variables to the LPM. He points out two
flaws: computation of the predicted probability and marginal
effects—and goes on to state
“Even with these problems, the linear probability model is
useful and often applied in economics. It usually works well
for values of the independent variables that are near the
averages in the sample.” (2009, p. 249)

Wooldridge also discusses the heteroskedastic nature of the LPM’s

error, which is binomial by construction, but does not address the issue
of the lack of independence that this implies.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 13 / 73
 Linear probability models Examples of support for the LPM approach

E XAMPLE 2

Joshua Angrist and Steve Pischke’s popular Mostly Harmless

Econometrics invokes the principle of Occam’s razor, arguing that
“...extra complexity comes into the inference step as well,
since we need standard errors for marginal effects.” (2009, p.
107)

This is surely a red herring for Stata users, as the margins command
in Stata 11 or 12 computes those standard errors via the delta method.
They also discuss the difficulty of computing marginal effects for a
binary regressor: again, not an issue for Stata 12 users, with the new
contrast command.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 14 / 73
 Maximum Likelihood approach

Maximum likelihood estimators

A maximum likelihood estimator of a binary outcome with possibly

endogenous regressors can be implemented for the model

D = I(X e βe + X 0 β0 + ε ≥ 0)
X e = G(Z , θ, e)

which, for a single binary endogenous regressor, G(·) probit, and ε and
e jointly Normal, is the model estimated by Stata’s biprobit
command.

Like the LPM, maximum likelihood allows endogenous regressors in

X e to be continuous, discrete, limited, etc. as long as a model for G(·)
can be fully specified, along with the fully parameterized joint
distribution of (ε, e).

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 15 / 73
 Control Function approach

Control function estimators

Control function estimators first estimate the model of endogenous

regressors as a function of instruments, like the ‘first stage’ of 2SLS,
then use the errors from this model as an additional regressor in the
main model.

This approach is more general than maximum likelihood as the first

stage function can be semiparametric or nonparametric, and the joint
distribution of (ε, e) need not be fully parameterized.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 16 / 73
 Control Function approach

To formalize the approach, consider a model D = M(X , β, ε), and

assume there are functions G, h and a well-behaved error U such that
X e = G(Z , e), ε = h(e, U), and U ⊥ (X , e).

We first estimate G(·): the endogenous regressors as functions of

instruments Z , and derive fitted values of the errors e. Then we have

D = M(X , β, h(e, u)) = M(X

e , e, β, U)

where the error term of the Me model is U, which is suitably

independent of (X , e). This model no longer has an endogeneity
problem, and can be estimated via straightforward methods.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 17 / 73
 Control Function approach

Given the threshold crossing model

D = I(X e βe + X 0 β0 + ε ≥ 0)
Xe = Zα + e

with (ε, e) jointly normal, we can first linearly regress X e on Z , with

residuals being estimates of e.

This then yields an ordinary probit model

D = I(X e βe + X 0 β0 + λe + U ≥ 0)

which is the model estimated by Stata’s ivprobit command.

Despite its name, ivprobit is a control function estimator, not an IV
estimator.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 18 / 73
 Control Function approach

A substantial limitation of control function methods in this context is

that they generally require the endogenous regressors X e to be
continuous, rather than binary, discrete, or censored. For instance, a
binary endogenous regressor will violate the assumptions necessary to
derive estimates of the ‘first stage’ error term e. The errors in the ‘first
stage’ regression cannot be normally distributed and independent of
the regressors. Thus, the ivprobit command should not be applied
to binary endogenous regressors, as its documentation clearly states.

In this context, control function estimators—like maximum likelihood

estimators—of binary outcome models require that the first stage
model be correctly specified. This is an important limitation of these
approaches. A 2SLS approach will lose efficiency if an appropriate
instrument is not included, but a ML or control function estimator will
generally become inconsistent.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 19 / 73
 Special Regressor approach

Special regressor estimators

Special regressor estimators were first proposed by Lewbel (J.

Econometrics, 2000). Their implementation is fully described in Dong
and Lewbel (2012, BC WP 604; forthcoming, Econometric Reviews).

They assume that the model includes a particular regressor, V , with

certain properties. It is exogenous (that is, E(ε|V ) = 0) and appears
as an additive term in the model. It is continuously distributed and has
a large support.

A third condition, important for the success of this approach, is that V

have a thick-tailed distribution. A regressor with greater kurtosis will be
more useful as a special regressor, and one that is strictly normally
distributed would lack the tail thickness to perform well.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 20 / 73
 Special Regressor approach

The binary choice special regressor proposed by Lewbel (2000) has

the ‘threshold crossing’ form

D = I(X e βe + X 0 β0 + V + ε ≥ 0)

or, equivalently,
D = I(X β + V + ε ≥ 0)

This is the same basic form for D as in the ML or control function (CF)
approach. Note, however, that the special regressor V has been
separated from the other exogenous regressors, and its coefficient
normalized to unity: a harmless normalization.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 21 / 73
 Special Regressor approach

Given a special regressor V , the only other requirements are those

applicable to linear 2SLS: to handle endogeneity, the set of instruments
Z must satisfy E(ε|Z ) = 0, and E(Z 0 X ) must have full rank.

The main drawback of this method is that the special regressor V must
be conditionally independent of ε. Even if it is exogenous, it could fail
to satisfy this assumption because of the way in which V might affect
other endogenous regressors. Also, V must be continuously
distributed after conditioning on the other regressors, so that a term
like V 2 could not be included as an additional regressor.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 22 / 73
 Special Regressor approach

Apart from these restrictions on V , the special regressor (SR) method

has none of the drawbacks of the three models discussed earlier:
Unlike the LPM, the SR predictions stay ‘in bounds’ and is
consistent with other threshold crossing models.
Unlike ML and CF methods, the SR model does not require
correct specification of the ‘first stage’ model: any valid set of
instruments may be used, with only efficiency at stake.
Unlike ML, the SR method has a linear form, not requiring iterative
search.
Unlike CF, the SR method can be used when endogenous
regressors X e are discrete or limited; unlike ML, there is a single
estimation method, regardless of the characteristics of X e .
Unlike ML, the SR method permits unknown heteroskedasticity in
the model errors.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 23 / 73
 Special Regressor approach

The special regressor method imposes far fewer assumptions on the

distribution of errors—particularly the errors e in the ‘first stage’
equations for X e —than do CF or ML estimation methods. Therefore,
SR estimators will be less efficient than these alternatives when the
alternatives are consistent.

SR estimators may be expected to have larger standard errors and

lower precision than other methods, when those methods are valid.
However, if a special regressor V can be found, the SR method will be
valid under much more general conditions than the ML and CF
methods.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 24 / 73
 The average index function (AIF)

The average index function (AIF)

Consider the original estimation problem

D = I(X β + ε ≥ 0)

where with generality one of the elements of X may be a special

regressor V , with coefficient one. If ε is independent of X , the
propensity score or choice probability is
Pr[D = 1|X ] = E(D|X ) = E(D|X β) = F−ε (X β) = Pr(−ε ≤ X β), with
F−ε (·) the probability distribution function of −ε. In the case of
independent errors, these measures are identical.

When some regressors are endogenous, or generally when the

assumption X ⊥ ε is violated (e.g., by heteroskedasticity), these
expressions may differ from one another.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 25 / 73
 The average index function (AIF)

Blundell and Powell (Rev. Ec. Stud., 2004) propose using the average
structural function (ASF) to summarize choice probabilities: F−ε (X β),
even though ε is no longer independent of X . In this case,
F−ε|X (X β|X ) should be computed: a formidable task.

Lewbel, Dong and Yang (Can. J. Econ., 2012) propose using the
measure E(D|X β), which they call the average index function (AIF), to
summarize choice probabilities.

Like the ASF, the AIF is based on the estimated index, and equals the
propensity score when ε ⊥ X . However, when this assumption is
violated (by endogeneity or heteroskedasticity), the AIF is usually
easier to estimate, via a unidimensional nonparametric regression of D
on X β.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 26 / 73
 The average index function (AIF)

The AIF can be considered a middle ground between the propensity

score and the ASF, as the former conditions on all covariates using
F−ε|X ; the ASF conditions on no covariates using F−ε ; and the AIF
conditions on the index of covariates, F−ε|X β .

Define the function M(X β) = E(D|X β), with derivatives m. The

marginal effects of the regressors on the choice probabilities, as
measured by the AIF, are ∂E(D|X β)/∂X = m(X β)β, so the average
marginal effects just equal the average derivatives, E(m(X β + V ))β.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 27 / 73
 The average index function (AIF)

For the LPM, the ASF and AIF both equal the fitted values of the linear
2SLS regression of D on X. For the other methods, the AIF choice
probabilities can be estimated using a standard unidimensional kernel
regression of D on X β̂: for instance, using the lpoly command in
Stata, with the at() option specifying the observed data points. This
will produce the AIF for each observation i, M bi.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 28 / 73
 The average index function (AIF)

Employing the derivatives of the kernel function, the individual-level

marginal effects m
b i may be calculated, and averaged to produce
average marginal effects:
n
1X
mβ̂ = m
b i β̂
n
i=1

Estimates of the precision of these average marginal effects may be

derived by bootstrapping.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 29 / 73
 The Stata implementation

The Stata implementation

My Stata command sspecialreg estimates the Lewbel and Dong

simple special regression estimator of a binary outcome with one or
more binary endogenous variables. It is an optimized version of earlier
code developed for this estimator, and provides significant (8–10x)
speed improvements over that code.

Two forms of the special regressor estimator are defined, depending

on assumptions made about the distribution of the special regressor V .
In the first form of the model, only the mean of V is assumed to be
related to the other covariates. In the second, ‘heteroskedastic’ form,
higher moments of V can also depend in arbitrary, unknown ways on
the other covariates. In practice, the latter form may include squares
and cross products of some of the covariates in the estimation
process, similar to the auxiliary regression used in White’s general test
for heteroskedasticity.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 30 / 73
 The Stata implementation

The sspecialreg Stata command also allows for two specifications

of the density estimator used in the model: one based on a standard
kernel density approach such as that implemented by density or Ben
Jann’s kdens, as well as the alternative ‘sorted data density’ approach
proposed by Lewbel and Schennach (J. Econometrics, 2007).
Implementation of the latter approach also benefited greatly, in terms
of speed, by being rewritten in Mata, with Ben Jann’s help gratefully
acknowledged.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 31 / 73
 The Stata implementation

Just as in a probit or ivprobit model, the quantities of interest are

not the estimated coefficients derived in the special regressor method,
but rather the marginal effects. In the work of Lewbel et al., those are
derived from the average index function (AIF) as described earlier.
Point estimates of the AIF can be derived in a manner similar to that of
average marginal effects in standard limited dependent variable
models. For interval estimates, bootstrapped standard errors for the
marginal effects are computed.

A bootstrap option was also added to sspecialreg so that the

estimator can produce point and interval estimates of the relevant
marginal effects in a single step, with the user’s choice of the number
of bootstrap samples to be drawn.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 32 / 73
 Empirical illustration 1

Empirical illustration 1

In this example of the special regressor method, taken from Dong and
Lewbel (BC WP 604), the binary dependent variable is an indicator
that individual i migrates from one US state to another. The objective
is to estimate the probability of interstate migration.

The special regressor Vi in this context is age. Human capital theory

suggests that it should appear linearly (or at least monotonically) in a
threshold crossing model. Migration is in part driven by maximizing
expected lifetime income, and the potential gain in lifetime earnings
from a permanent change in labor income declines linearly with age.
Evidence of empirical support for this relationship is provided by Dong
(Ec. Letters, 2010). Vi is defined as the negative of age, demeaned, so
that it should have a positive coefficient and a zero mean.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 33 / 73
 Empirical illustration 1

Pre-migration family income and home ownership are expected to be

significant determinants of migration, and both should be considered
endogenous. A maximum likelihood approach would require an
elaborate dynamic specification in order to model the homeownership
decision. Control function methods such as ivprobit are not
appropriate as homeowner is a discrete variable.
The sample used includes male heads of household, 23–59 years of
age, from the 1990 wave of the PSID who have completed education
and are not retired, so as to exclude those moving to retirement
communities. The observed D = 1 indicates migration during
1991–1993. In the sample of 4689 individuals, 807 were interstate
migrants.
Exogenous regressors in the model include years of education,
number of children, and indicators for white, disabled, and married
individuals. The instruments Z also include the level of government
benefits received in 1989–1990 and state median residential tax rates.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 34 / 73
 Empirical illustration 1

In the following table, we present four sets of estimates of the marginal

effects computed by ssimplereg, utilizing the sorted data density
estimator in columns 2 and 4 and allowing for heteroskedastic errors in
columns 3 and 4.

For contrast, we present the results from an IV LPM (ivregress

2sls) in column 5, a standard probit (ignoring endogeneity) in
column 6, and an ivprobit in the last column, ignoring its lack of
applicability to the binary endogenous regressor homeowner.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 35 / 73
 Empirical illustration 1

Table: Marginal effects: binary outcome, binary endogenous regressor

kdens sortdens kdens_hetero sortdens_hetero IV-LPM probit ivprobit

age 0.0146 0.0112 0.0071 0.0104 -0.0010 0.0019 -0.0005
(0.003)∗∗∗ (0.003)∗∗∗ (0.003)∗ (0.003)∗∗∗ (0.002) (0.001)∗∗ (0.007)

log income -0.0079 0.0024 0.0382 0.0176 0.0550 -0.0089 0.1406

(0.028) (0.027) (0.024) (0.026) (0.080) (0.007) (0.286)

homeowner 0.0485 -0.0104 -0.0627 -0.0111 -0.3506 -0.0855 -1.0647

(0.072) (0.065) (0.059) (0.061) (0.204) (0.013)∗∗∗ (0.708)

white 0.0095 0.0021 0.0021 0.0011 0.0086 -0.0099 0.0134

(0.008) (0.010) (0.007) (0.008) (0.018) (0.012) (0.065)

disabled 0.1106 0.0730 0.0908 0.0916 0.0114 -0.0122 0.0104

(0.036)∗∗ (0.042) (0.026)∗∗∗ (0.037)∗ (0.055) (0.033) (0.203)

education -0.0043 -0.0023 -0.0038 -0.0036 0.0015 0.0004 0.0047

(0.002)∗ (0.003) (0.002)∗ (0.002) (0.004) (0.002) (0.015)

married 0.0628 0.0437 0.0258 0.0303 0.0322 -0.0064 0.0749

(0.020)∗∗ (0.028) (0.013) (0.020) (0.031) (0.017) (0.114)

nr. children -0.0169 -0.0117 0.0006 -0.0021 0.0137 0.0097 0.0502

(0.005)∗∗∗ (0.005)∗ (0.002) (0.003) (0.006)∗ (0.005)∗ (0.023)∗
Note: bootstrapped standard errors in parentheses (100 replications)

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 36 / 73
 Empirical illustration 1

The standard errors of these estimated marginal effects are computed

from 100 bootstrap replications. The marginal effect of the ‘special
regressor’ age of head is estimated as positive by the special
regressor methods, but both the two-stage linear probability model and
the ivprobit model yield negative (but insignificant) point estimates.

Household income and homeownership status do not seem to play

significant roles in the migration decision. Among the special
regression methods, the kernel data density estimator appears to yield
the most significant results, with age of head, disabled status, years of
education, marital status and number of children all playing a role in
predicting the migration decision.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 37 / 73
 Empirical illustration 2

Empirical illustration 2

In this example of the special regressor method for firm-level data, we

use a subsample of COMPUSTAT firm-level data on US publicly traded
firms. The binary outcome variable is an indicator of whether the firm
repurchased its common or preferred stock in a given year. The
endogenous binary regressor is an indicator of whether the firm issued
long-term debt in that year.

The special regressor Vit in this context is the firm’s return on assets,
or ROA. We also include the lagged value of income over total assets
and a set of year dummies as exogenous factors. The instruments Z
also include the lagged values of two ratios: capital expenditures to
total assets and acquisitions to total assets.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 38 / 73
 Empirical illustration 2

The sample includes firm-level data from 1996–2006: a total of 30,852

firm-years. Over this period, the probability that a firm would
repurchase its own stock in a given year is 0.479, while the probability
that a firm would issue long-term debt in a given year is 0.583.

In the following table, we present four sets of estimates of the marginal

effects computed by sspecialreg, utilizing the sorted data density
estimator in columns 2 and 4 and allowing for heteroskedastic errors in
columns 3 and 4.

For contrast, we present the results from an IV LPM (ivregress

2sls) in column 5, a standard probit (ignoring endogeneity) in
column 6, and an ivprobit in the last column, ignoring its lack of
applicability to the binary endogenous regressor ltdiss.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 39 / 73
 Empirical illustration 2

Modeling firms’ stock repurchase decision

Table: Marginal effects: binary outcome, binary endogenous regressor

kdens sortdens kdens_H sortdens_H IV-LPM probit ivprobit
roa 0.4352 0.3918 0.3784∗ 0.3569 0.0122∗∗∗ 0.0065 0.0243∗∗
(0.231) (0.307) (0.189) (0.284) (0.002) (0.004) (0.008)

ltdiss -0.1944∗∗∗ -0.0847 -0.0983∗∗ -0.0892 -0.4600∗∗∗ 0.0063 -1.5950∗∗∗

(0.025) (0.092) (0.034) (0.082) (0.058) (0.007) (0.067)

linca -0.0246 -0.0098 -0.0345∗ -0.0377 0.0794∗∗∗ 0.1644∗∗∗ 0.3687∗∗∗

(0.013) (0.016) (0.014) (0.033) (0.007) (0.012) (0.028)
Standard errors in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 40 / 73
 Empirical illustration 2

The standard errors of these estimated marginal effects are computed

from 20 bootstrap replications. The marginal effect of the ‘special
regressor’, ROA, is estimated as positive and significant by the special
regressor methods.

The endogenous factor, long-term debt issuance, has a negative and

significant coefficient (at 95%) in the estimates computed by the
special regressor methods when the kernel density estimator is
employed. The estimate of that coefficient in the IV-LPM model is
almost twice as large, and its estimate from ivprobit implausibly
huge.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 41 / 73
 Empirical illustration 2

The lagged ratio of income to total assets has a negative point

estimate, varying in significance. There are sizable time effects (not
shown) in the estimates.

Among the special regression methods, the kernel data density

estimator allowing for heteroskedasticity appears to yield the most
significant results, with all three variables displaying plausible
estimates.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 42 / 73
 Summary remarks on sspecialreg

Summary remarks on sspecialreg

We have discussed an alternative to the linear probability model for

estimation of a binary outcome with one or more binary endogenous
regressors. This alternative, Lewbel and Dong’s ‘simple special
regressor’ method, circumvents the drawbacks of the IV-LPM
approach, and yields consistent estimates in this context in which
ivprobit does not.

Computation of marginal effects via the proposed average index

function approach is straightforward, requiring only a single kernel
density estimation and no iterative techniques. Bootstrapping is
employed to derive interval estimates.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 43 / 73
 IV methods with generated instruments

Topic II: IV methods with generated instruments

Acknowledgement
This presentation is based on the work of Arthur Lewbel, “Using
Heteroskedasticity to Identify and Estimate Mismeasured and
Endogenous Regressor Models,” Journal of Business & Economic
Statistics, 2012. The contributions of Baum and Mark E Schaffer are
the development of Stata software to implement Lewbel’s methodology.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 44 / 73
 Motivation

Motivation

Instrumental variables (IV) methods are employed in linear regression

models, e.g., y = Xβ + u, where violations of the zero conditional
mean assumption E[u|X] = 0 are encountered.

Reliance on IV methods usually requires that appropriate instruments

are available to identify the model: often via exclusion restrictions.

Those instruments, Z, must satisfy three conditions: (i) they must

themselves satisfy orthogonality conditions (E[uZ] = 0); (ii) they must
exhibit meaningful correlations with X; and (iii) they must be properly
excluded from the model, so that their effect on the response variable
is only indirect.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 45 / 73
 Motivation Challenges in employing IV methods

Finding appropriate instruments which simultaneously satisfy all three

of these conditions is often problematic, and the major obstacle to the
use of IV techniques in many applied research projects.

Although textbook treatments of IV methods stress their usefulness in

dealing with endogenous regressors, they are also employed to deal
with omitted variables, or with measurement error of the regressors
(‘errors in variables’) which if ignored will cause bias and inconsistency
in OLS estimates.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 46 / 73
 Motivation Lewbel’s approach

Lewbel’s approach

The method proposed in Lewbel (JBES, 2012) serves to identify

structural parameters in regression models with endogenous or
mismeasured regressors in the absence of traditional identifying
information, such as external instruments or repeated measurements.

Identification is achieved in this context by having regressors that are

uncorrelated with the product of heteroskedastic errors, which is a
feature of many models where error correlations are due to an
unobserved common factor.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 47 / 73
 Motivation Lewbel’s approach

In this presentation, we describe a method for constructing instruments

as simple functions of the model’s data. This approach may be applied
when no external instruments are available, or, alternatively, used to
supplement external instruments to improve the efficiency of the IV
estimator.

Supplementing external instruments can also allow ‘Sargan–Hansen’

tests of the orthogonality conditions to be performed which would not
be available in the case of exact identification by external instruments.

In that context, the approach is similar to the dynamic panel data

estimators of Arellano and Bond (Review of Economic Studies, 1991)
et al., as those estimators customarily make use of appropriate lagged
values of endogenous regressors to identify the model. In contrast, the
approach we describe here may be applied in a purely cross-sectional
context, as well as that of time series or panel data.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 48 / 73
 The basic framework

The basic framework

Consider Y1 , Y2 as observed endogenous variables, X a vector of

observed exogenous regressors, and ε = (ε1 , ε2 ) as unobserved error
processes. Consider a structural model of the form:

Y1 = X 0 β1 + Y2 γ1 + ε1
Y2 = X 0 β2 + Y1 γ2 + ε2

This system is triangular when γ2 = 0 (or, with renumbering, when

γ1 = 0). Otherwise, it is fully simultaneous. The errors ε1 , ε2 may be
correlated with each other.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 49 / 73
 The basic framework

If the exogeneity assumption, E(εX ) = 0 holds, the reduced form is

identified, but in the absence of identifying restrictions, the structural
parameters are not identified. These restrictions often involve setting
certain elements of β1 or β2 to zero, which makes instruments
available.

In many applied contexts, the third assumption made for the validity of
an instrument—that it only indirectly affects the response variable—is
difficult to establish. The zero restriction on its coefficient may not be
plausible. The assumption is readily testable, but if it does not hold, IV
estimates will be inconsistent.

Identification in Lewbel’s approach is achieved by restricting

correlations of εε0 with X . This relies upon higher moments, and is
likely to be less reliable than identification based on coefficient zero
restrictions. However, in the absence of plausible identifying
restrictions, this approach may be the only reasonable strategy.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 50 / 73
 The basic framework

The parameters of the structural model will remain unidentified under

the standard homoskedasticity assumption: that E(εε0 |X ) is a matrix of
constants. However, in the presence of heteroskedasticity related to at
least some elements of X , identification can be achieved.

In a fully simultaneous system, assuming that cov(X , ε2j ) 6= 0, j = 1, 2

and cov(Z , ε1 ε2 ) = 0 for observed Z will identify the structural
parameters. Note that Z may be a subset of X , so no information
outside the model specified above is required.

The key assumption that cov(Z , ε1 ε2 ) = 0 will automatically be satisfied

if the mean zero error processes are conditionally independent:
ε1 ⊥ ε2 |Z = 0. However, this independence is not strictly necessary.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 51 / 73
 Single-equation estimation

Single-equation estimation

In the most straightforward context, we want to apply the instrumental

variables approach to a single equation, but lack appropriate
instruments or identifying restrictions. The auxiliary equation or
‘first-stage’ regression may be used to provide the necessary
components for Lewbel’s method.

In the simplest version of this approach, generated instruments can be

constructed from the auxiliary equations’ residuals, multiplied by each
of the included exogenous variables in mean-centered form:

Zj = (Xj − X ) · 
where  is the vector of residuals from the ‘first-stage regression’ of
each endogenous regressor on all exogenous regressors, including a
constant vector.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 52 / 73
 Single-equation estimation

These auxiliary regression residuals have zero covariance with each of

the regressors used to construct them, implying that the means of the
generated instruments will be zero by construction. However, their
element-wise products with the centered regressors will not be zero,
and will contain sizable elements if there is clear evidence of ‘scale
heteroskedasticity’ with respect to the regressors. Scale-related
heteroskedasticity may be analyzed with a Breusch–Pagan type test:
estat hettest in an OLS context, or ivhettest (Schaffer, SSC;
Baum et al., Stata Journal, 2007) in an IV context.

The greater the degree of scale heteroskedasticity in the error process,

the higher will be the correlation of the generated instruments with the
included endogenous variables which are the regressands in the
auxiliary regressions.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 53 / 73
 Stata implementation ivreg2h

Stata implementation

An implementation of this simplest version of Lewbel’s method,

ivreg2h, has been constructed from Baum, Schaffer, Stillman’s
ivreg2 and Schaffer’s xtivreg2, both available from the SSC
Archive. The panel-data features of xtivreg2 are not used in this
implementation: only the nature of xtivreg2 as a ‘wrapper’ for
ivreg2.

In its current version, ivreg2h can be invoked to estimate

a traditionally identified single equation, or
a single equation that fails the order condition for identification:
either (i) by having no excluded instruments, or
(ii) by having fewer excluded instruments than needed for
traditional identification.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 54 / 73
 Stata implementation ivreg2h

In the former case, of external instruments augmented by generated

instruments, the program provides three sets of estimates: the
traditional IV estimates, estimates using only generated instruments,
and estimates using both generated and excluded instruments.

In the latter case, of an underidentified equation, only the estimates

using generated instruments are displayed. Unlike ivreg2 or
ivregress, ivreg2h allows the syntax
ivreg2h depvar exogvar (endogvar=)
disregarding the failure of the order condition for identification.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 55 / 73
 Empirical illustration 1

Empirical illustration 1

In Lewbel’s 2012 JBES paper, he illustrates the use of his method with
an Engel curve for food expenditures. An Engel curve describes how
household expenditure on a particular good or service varies with
household income (Ernst Engel, 1857, 1895).1 Engel’s research gave
rise to Engel’s Law: while food expenditures are an increasing function
of income and family size, food budget shares decrease with income
(Lewbel, New Palgrave Dictionary of Economics, 2d ed. 2007).

In this application, we are considering a key explanatory variable, total

expenditures, to be subject to potentially large measurement errors, as
is often found in applied research: due in part to infrequently
purchased items (Meghir and Robin, Journal of Econometrics, 1992).

1
Not to be confused with Friedrich Engels, Karl Marx’s coauthor.
CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 56 / 73
 Empirical illustration 1

The data are 854 households, all married couples without children,
from the UK Family Expenditure Survey, 1980–1982, as studied by
Banks, Blundell and Lewbel (Review of Economics and Statistics,
1997). The dependent variable is the food budget share, with a sample
mean of 0.285. The key explanatory variable is log real total
expenditures, with a sample mean of 0.599. A number of additional
regressors (age, spouse’s age, ages2 , and a number of indicators) are
available as controls. The coefficients of interest in this model are
those of log real total expenditures and the constant term.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 57 / 73
 Empirical illustration 1

Food Budget Share .6

.4

.2

0
-.5 0 .5 1 1.5 2
Log Real Total Expenditure

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 58 / 73
 Empirical illustration 1

We first estimate the model with OLS regression, ignoring any issue of
mismeasurement. We then reestimate the model with log total income
as an instrument using two-stage least squares: an exactly identified
model. As such, this is also the IV-GMM estimate of the model.

In the following table, these estimates are labeled as OLS and TSLS1.
A Durbin–Wu–Hausman test for the endogeneity of log real total
expenditures in the TSLS1 model rejects with p-value=0.0203,
indicating that application of OLS is inappropriate.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 59 / 73
 Empirical illustration 1

Table: OLS and conventional TSLS

(1) (2)
OLS TSLS,ExactID
lrtotexp -0.127 -0.0859
(0.00838) (0.0198)

Constant 0.361 0.336

(0.00564) (0.0122)
Standard errors in parentheses

These OLS and TSLS results can be estimated with standard

regress and ivregress 2sls commands. We now turn to
estimates produced from generated instruments via Lewbel’s method.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 60 / 73
 Empirical illustration 1

We produce generated instruments from each of the exogenous

regressors in this equation. The equation may be estimated by TSLS
or by IV-GMM, in each case producing robust standard errors. For
IV-GMM, we report Hansen’s J.

Table: Generated instruments only

(1) (2)
TSLS,GenInst GMM,GenInst
lrtotexp -0.0554 -0.0521
(0.0589) (0.0546)

Constant 0.318 0.317

(0.0352) (0.0328)
Jval 12.91
Jdf 11
Jpval 0.299
Standard errors in parentheses

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 61 / 73
 Empirical illustration 1

The greater efficiency available with IV-GMM is evident in the precision

of these estimates. However, reliance on generated instruments yields
much larger standard errors than identified TSLS.2

As an alternative, we augment the available instrument, log total

income, with the generated instruments, which overidentifies the
equation, estimated with both TSLS and IV-GMM methods.

2
The GMM results do not agree with those labeled GMM2 in the JBES
article. However, it appears that the published GMM2 results are not the true
optimum.
CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 62 / 73
 Empirical illustration 1

Table: Augmented by generated instruments

(1) (2)
TSLS,AugInst GMM,AugInst
lrtotexp -0.0862 -0.0867
(0.0186) (0.0182)

Constant 0.336 0.337

(0.0114) (0.0112)
Jval 16.44
Jdf 12
Jpval 0.172
Standard errors in parentheses

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 63 / 73
 Empirical illustration 1

Relative to the original, exactly-identified TSLS/IV-GMM specification,

the use of generated instruments to augment the model has provided
an increase in efficiency, and allowed overidentifying restrictions to be
tested. As a comparison:

Table: With and without generated instruments

(1) (2)
GMM,ExactID GMM,AugInst
lrtotexp -0.0859 -0.0867
(0.0198) (0.0182)

Constant 0.336 0.337

(0.0122) (0.0112)
Jval 16.44
Jdf 12
Jpval 0.172
Standard errors in parentheses

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 64 / 73
 Empirical illustration 2

Empirical illustration 2

We illustrate the use of this method with an estimated equation on

firm-level panel data from US Industrial Annual COMPUSTAT. The
model, a variant on that presented in a working paper by Baum,
Chakraborty and Liu, is based on Faulkender and Wang
(J. Finance, 2006).

We seek to explain the firm-level unexpected change in cash holdings

as a function of the level of cash holdings (C), the change in earnings
(dE), the change in non-cash assets (dNA) and market leverage (Lev ).
The full sample contains about 13,000 firms for up to 35 years.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 65 / 73
 Empirical illustration 2

For purposes of illustration, we first fit the model treating the level of
cash holdings as endogenous, but maintaining that we have no
available external instruments. In this context, ivreg2h produces
three generated instruments: one from each included exogenous
regressor. We employ IV-GMM with a cluster-robust VCE, clustered by
firm.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 66 / 73
 Empirical illustration 2

Table: Modeling ∆C1

GenInst
C -0.152∗∗∗
(-5.04)

dE 0.0301∗∗∗
(7.24)

dNA -0.0115∗∗∗
(-6.00)

Lev -0.0447∗∗∗
(-18.45)
N 117036
jdf 2
jp 0.245
t statistics in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 67 / 73
 Empirical illustration 2

The resulting model is overidentified by two degrees of freedom (jdf).

The jp value of 0.245 is the p-value of the Hansen J statistic.

We reestimate the model using the lagged value of cash holdings as

an instrument. This causes the model to be exactly identified, and
estimable with standard techniques. ivreg2h thus produces three
sets of estimates: those for standard IV, those using only generated
instruments, and those using both external and generated instruments.
The generated-instrument results differ from those shown previously
as the sample is now smaller.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 68 / 73
 Empirical illustration 2

Table: Modeling ∆C1

StdIV GenInst GenExtInst
C -0.0999∗∗∗ -0.127∗∗∗ -0.100∗∗∗
(-15.25) (-3.83) (-15.37)

dE 0.0287∗∗∗ 0.0324∗∗∗ 0.0304∗∗∗

(6.43) (7.09) (7.99)

dNA -0.0121∗∗∗ -0.0133∗∗∗ -0.0124∗∗∗

(-6.16) (-6.43) (-7.09)

Lev -0.0447∗∗∗ -0.0468∗∗∗ -0.0460∗∗∗

(-15.99) (-17.79) (-19.51)
N 102870 102870 102870
jdf 0 2 3
jp 0.691 0.697
t statistics in parentheses
∗
p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 69 / 73
 Empirical illustration 2

The results show that there are minor differences in the point
estimates produced by standard IV and those from the augmented
equation. However, the latter are more efficient, with smaller standard
errors for each coefficient. The model is now overidentified by three
degrees of freedom, allowing us to conduct a test of over identifying
restrictions. The p-value of that test,jp, indicates no problem.

This example illustrates what may be the most useful aspect of

Lewbel’s method: the ability to augment an exactly-identified equation
to both allow a test of over identifying restrictions and gain efficiency.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 70 / 73
 Empirical illustration 2

We have illustrated this method with one endogenous regressor, but it

generalizes to multiple endogenous (or mismeasured) regressors. It
may be employed as long as there is at least one included exogenous
regressor for each endogenous regressor. If there is only one, the
resulting equation will be exactly identified.

As this estimator has been implemented within the ivreg2 framework,

all of the diagnostics and options available in that program (Baum,
Schaffer, Stillman, Stata Journal, 2003, 2007) are available in this
context.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 71 / 73
 Summary remarks on ivreg2h

Summary remarks on ivreg2h

The extension of this method to the panel fixed-effects context is

relatively straightforward, and we are finalizing a version of Mark
Schaffer’s xtivreg2 which implements Lewbel’s method in this
context.

We have illustrated how this method might be used to augment the

available instruments to facilitate the use of tests of overidentification.
Lewbel argues that the method might also be employed in a fully
saturated model, such as a difference-in-difference specification with
all feasible fixed effects included, in order to test whether OLS
methods will yield consistent results.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 72 / 73
 Concluding remarks

Concluding remarks

I hope that this illustration of how cutting-edge econometric techniques

may be made available to Stata users has been enlightening. The
underlying code for both sspecialreg and ivreg2h is accessible,
as is true of nearly all user-written contributions to the Stata
community. This openness greatly enhances users’ ability to both
extend Stata’s capabilities and assure themselves that user-written
routines are of high quality and generally reliable.

CF Baum (BC / DIW) Implementing new econometric tools in Stata MXSUG, May 2013 73 / 73