Introduction to GSEM in Stata

Christopher F Baum

ECON 8823: Applied Econometrics

Boston College, Spring 2016

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 1 / 39
 Generalized Structural Equation Modeling in Stata

Generalized Structural Equation Modeling in Stata

We now present an introduction to Stata’s gsem command, which

extends the facilities of the sem command to implement a broader set
of applications of structural equation modeling: thus, generalized
structural equation modeling. As gsem has many capabilities, we can
only discuss a limited subset of its features and give some illustrations
of its use.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 2 / 39
 Generalized Structural Equation Modeling in Stata Generalized Linear Model

Generalized Linear Model

To understand Stata’s extension of the SEM framework, we must

introduce the concept of the Generalized Linear Model: something that
has been a component of Stata for many years as the glm command.

The generalized linear model (GLM) framework of McCullaugh and

Nelder (1989) is common in applied work in biostatistics, but has not
been widely applied in econometrics. It offers many advantages, and
should be more widely known.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 3 / 39
 Generalized Structural Equation Modeling in Stata Generalized Linear Model

GLM estimators are maximum likelihood estimators that are based on

a density in the linear exponential family (LEF). These include the
normal (Gaussian) and inverse Gaussian for continuous data, Poisson
and negative binomial for count data, Bernoulli for binary data
(including logit and probit) and Gamma for duration data.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 4 / 39
 Generalized Structural Equation Modeling in Stata Generalized Linear Model

GLM estimators are essentially generalizations of nonlinear least

squares, and as such are optimal for a nonlinear regression model with
homoskedastic additive errors. They are also appropriate for other
types of data which exhibit intrinsic heteroskedasticity where there is a
rationale for modeling the heteroskedasticity.

The GLM estimator θ̂ maximizes the log-likelihood

N
X
Q(θ) = [a (m(xi , β)) + b(yi ) + c (m(xi , β))]
i=1

where m(x, β) = E(y |x) is the conditional mean of y , a(·) and c(·)
correspond to different members of the LEF, and b(·) is a normalizing
constant.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 5 / 39
 Generalized Structural Equation Modeling in Stata Generalized Linear Model

For instance, for the Poisson, where the mean equals the variance,
a(µ) = −µ and c(µ) = log(µ). Given definitions of these two functions,
the mean and variance are E(y ) = µ = −a0 (µ)/c 0 (µ) and
Var (y ) = 1/c 0 (µ). For the Poisson, a0 (µ) = 1, c 0 (µ) = 1/µ, so
E(y ) = Var (y ) = µ.

GLM estimators are consistent provided that the conditional mean

function is correctly specified: that E(yi |xi ) = m(xi , β). If the variance
function is not correctly specified, a robust estimate of the VCE should
be used.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 6 / 39
 Generalized Structural Equation Modeling in Stata Generalized Linear Model

To use the GLM estimator, you must specify two options: the
family(), which defines the member of the LEF to be employed, and
the link(), which is the inverse of the conditional mean function. The
family option may be chosen as gaussian, igaussian,
binomial, poisson, nbinomial, gamma.

The link function essentially expresses the transformation to be applied

to the dependent variable. Each family has a canonical link, which is
chosen if not specified: for instance, family(gaussian) has default
link(identity), so that a GLM with those two options would
essentially be linear regression via maximum likelihood.

The binomial family has a default link(logit), while the

poisson and nbinomial families share link(log). However, a
number of other combinations of family and link are valid: for
instance, link(power n) is valid for all distributional families.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 7 / 39
 Generalized Structural Equation Modeling in Stata The GLM and the GSEM

The GLM and the GSEM

What, then, is Stata’s Generalized Structural Equation Model, or

gsem? Essentially, the combination of the sem modeling capabilities
we have discussed thus far with the broader glm estimation
framework, allowing us to build models that include latent variables as
well as response variables that are not continuous measures.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 8 / 39
 Generalized Structural Equation Modeling in Stata The GLM and the GSEM

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.

gsem also has the ability to fit multilevel mixed SEMs. 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, ... This extends Stata’s
mixed framework.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 9 / 39
 Models supported by GSEM The one-factor measurement model, generalized response

Models supported by GSEM

We now consider a number of models that are supported by the SEM

methodology. The first is the single-factor measurement model, in
which we consider several observed variables as influencing a single
latent factor, as we considered earlier. The difference is that we now
allow for a generalized response, rather than assuming that the
response is continuous, driven by Gaussian errors. This can be
graphically represented:

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 10 / 39
 Models supported by GSEM The one-factor measurement model, generalized response

Bernoulli Bernoulli Bernoulli Bernoulli

x1 x2 x3 x4
probit probit probit probit

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 11 / 39
 Models supported by GSEM The one-factor measurement model, generalized response

In this model, we have four observed factors, each of which is a binary

(pass/fail) outcome. The latent factor, being related to only binary
measurements, will have different properties than a model based on
continuous measurements. Thus, the errors are presumed to follow a
Bernoulli distribution, and the GLM link function is the probit. Notice
that those specifications show up in the graphical diagram. We may
implement this model using gsem as:
gsem (x1 x2 x3 x4 <-X), probit

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 12 / 39
 Models supported by GSEM The one-factor measurement model, generalized response

If one or more of these measurements was continuous, we could use a

different family and link for that part of the model. Say that
measurement 4 was not only a pass/fail mark, but the score on a test.
Then that equation would be fit with the gsem default of Gaussian
errors and the Identity link.
gsem (x1 x2 x3 <-X, probit) (s4<-X)

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 13 / 39
 Models supported by GSEM Logistic regression

Logistic regression

We could use gsem to fit a standard logistic regression, which is

equivalent to the logit model in the GLM framework. The model here
considers the probability of low birth weight as related to a number of
observed factors about the mother’s medical condition, weight, race,
and smoking status.
We may implement this model using gsem as:
gsem (low <- age lwt i.race smoke ptl ht ui), logit

where i.race is the standard factor variable notation, indicating that

one race should be omitted and indicator variables created for each of
the other race categories. Graphically:

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 14 / 39
 Models supported by GSEM Logistic regression

age

lwt

1b.race

2.race
Bernoulli

low
3.race
logit

smoke

ptl

ht

ui

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 15 / 39
 Models supported by GSEM Ordered probit and ordered logit

Ordered probit and ordered logit

We can also use ordered probit or ordered logit models in the GSEM
framework to deal with variables, such as responses on a Likert scale,
where there is assumed to be an underlying factor, with ranges of that
latent variable ‘binned’ into observed discrete categories.
We may implement this model for a latent factor, relating attitudes
toward science in a pure measurement framework to four Likert sale
variables, as:
gsem (y1 y2 y3 y4 <- SciAtt), oprobit

Ordered logit could also be used, yielding almost identical results,

while making use of the logistic distribution rather than the Gaussian.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 16 / 39
 Models supported by GSEM Ordered probit and ordered logit

SciAtt

ordinal ordinal ordinal ordinal

y1 y2 y3 y4
probit probit probit probit

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 17 / 39
 Models supported by GSEM Tobit model

Tobit model

The Tobit regression model combines a binary outcome, which

indicates censoring, and a continuous outcome for uncensored
observations. Censoring may be from below, above or both. For
instance, we may have a response to “how much did you spend on a
new car last year?”, where responses of 0 indicate non-purchase. This
may be implemented as:
gsem mpg <- wgt, family(gaussian, lcensored(17))

where the lcensored option indicates that left-censoring at the value

17 is applied. Graphically:

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 18 / 39
 Models supported by GSEM Tobit model

Gaussian

wgt mpg ε1
identity

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 19 / 39
 Models supported by GSEM Interval regression

Interval regression

Interval regression (as implemented by Stata’s intreg) fits a model

where the response lies in an interval, as described by two dependent
variables. For instance, from a survey we may have the information
that a worker’s wage is between $10.00 and $11.99, or between
$12.00 and $13.99. Those values would appear as the lower and
upper limits in the interval regression. The GSEM implementation of
this model can be represented as:
gsem wage1 <- age c.age#c.age nev_mar rural school tenure, ///
family(gaussian, udepvar(wage2))

where wage1 would be the lower-limit values, and the udepvar option
specifies the variable containing the upper-limit values. Graphically:

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 20 / 39
 Models supported by GSEM Interval regression

age

c.age#c.age

Gaussian
nev_mar
wage1 ε1

identity
rural

school

tenure

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 21 / 39
 Models supported by GSEM Heckman selection model

Heckman selection model

The Heckman regression with selection, as implemented in heckman,

can also be considered in the GSEM framework. This model 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.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 22 / 39
 Models supported by GSEM Heckman selection model

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 L
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.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 23 / 39
 Models supported by GSEM Heckman selection model

This may be implemented as:

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)

where the variable wage is only observed when the notselected

variable is 0. The selected and notselected variables are
complements. The variables married and children are assumed to
only affect the probability of labor force participation, while educ and
age are presumed to affect both LFP and the level of the wage for
working women.

Like Roodman’s cmp, Stata considers missing values on an

equation-by-equation basis, so the fact that wage is missing for
non-working women is not a problem. Graphically:

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 24 / 39
 Models supported by GSEM Heckman selection model

ε1 a

married Gaussian

selected
1
children identity

L
1
educ
wage

age
ε2 a

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 25 / 39
 Models supported by GSEM Endogenous treatment-effects model

Endogenous treatment-effects model

The treatment-effects model attempts to measure the effect of a

“treatment” on a continuous outcome. For instance, we might
hypothesize that belonging to a labor union has an effect on wages,
and we want to measure the effect. This differs from the Heckman
selection model in that here we observe the outcome—the wage—for
all observations.

The econometric problem is that those persons with certain

characteristics, for instance, higher education, might be more or less
likely to be ‘treated’. Thus, we must take account of the
non-experimental nature of the data at hand, rather than merely
regressing wage on union with controls.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 26 / 39
 Models supported by GSEM Endogenous treatment-effects model

We may implement this, similar to the Heckman model, with a latent

variable related to the probability of being treated. Variables llunion
and ulunion are complements, reflecting the observed union
indicator.
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)

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 27 / 39
 Models supported by GSEM Endogenous treatment-effects model

1.south
ε1 a

1.black Gaussian

llunion
tenure identity
1

age L
1

grade wage

1.smsa ε2 a

1.union

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 28 / 39
 Models supported by GSEM One-parameter IRT (Rasch) model

One-parameter IRT (Rasch) model

The GSEM framework may be used to implement Item Response

Theory (IRT) models such as the Rasch model. In this example, we
have eight binary measurements from a math test, and we want to
generate a single latent factor, Math Ability, and evaluate how difficult
each of the questions were. This can be done by constraining the
effects of each observed variable on the latent factor to 1; the
estimated intercepts then gauge difficulty. A logit link is used, as all of
the observed variables follow a Bernoulli distribution.
gsem (MathAb -> (q1-q8)@1), logit

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 29 / 39
 Models supported by GSEM One-parameter IRT (Rasch) model

MathAb
1

b b b b b b b b

Bernoulli Bernoulli Bernoulli Bernoulli Bernoulli Bernoulli Bernoulli Bernoulli

q1 q2 q3 q4 q5 q6 q7 q8
logit logit logit logit logit logit logit logit

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 30 / 39
 Models supported by GSEM Two-level measurement model (multilevel, generalized response)

Two-level measurement model (multilevel,

generalized response)

We consider the Math Ability problem, noting that students are nested
within schools. We include a latent variable at the school level to
account for possible school-by-school effects. This makes the
estimation problem into a multilevel model. In the graphical
representation, school shows up as a latent variable at the school
level.
gsem (MathAb M1[school] -> q1-q8), logit

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 31 / 39
 Models supported by GSEM Two-level measurement model (multilevel, generalized response)

MathAb

1 c2 c3 c4 c5 c6 c7 c8

Bernoulli Bernoulli Bernoulli Bernoulli Bernoulli Bernoulli Bernoulli Bernoulli

q1 q2 q3 q4 q5 q6 q7 q8
logit logit logit logit logit logit logit logit

c3 c4 c5 c6
1 c2 c7 c8

school1

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 32 / 39
 Models supported by GSEM Two-factor measurement model (generalized response)

Two-factor measurement model (generalized

response)

This two-factor measurement model contains two latent variables: one

measuring Math Ability and a second measuring Math Attitude. While
the observed variables related to Ability are pass/fail grades on an
eight-question test, those underlying Math Attitude are Likert-scale
variables, necessitating the use of an ordinal estimator. The two latent
factors are assumed to be correlated. The code:
gsem (MathAb -> q1-q8, logit) ///
(MathAtt -> att1-att5, ologit)

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 33 / 39
 Models supported by GSEM Two-factor measurement model (generalized response)

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

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 34 / 39
 Models supported by GSEM Full structural equation model (generalized response)

Full structural equation model (generalized

response)

In the prior model, we allowed for (and estimated) a covariance

between the two latent factors, Math Ability and Math Attitude. We now
add a structural component to the model by assuming that there is a
causal relationship between Math Attitude and Math Ability. This allows
us to test a hypothesis regarding the way in which attitudes toward
math may affect ability, as evidenced by the observed test scores.
gsem (MathAb -> q1-q8, logit) ///
(MathAtt -> att1-att5, ologit) ///
(MathAtt -> MathAb)

As MathAb is now endogenous, an error term appears in the

specification.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 35 / 39
 Models supported by GSEM Full structural equation model (generalized response)

ordinal ordinal ordinal ordinal ordinal

att1 att2 att3 att4 att5

logit logit logit logit logit

MathAtt

MathAb ε1

Bernoulli Bernoulli Bernoulli Bernoulli Bernoulli Bernoulli Bernoulli Bernoulli

q1 q2 q3 q4 q5 q6 q7 q8
logit logit logit logit logit logit logit logit

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 36 / 39
 Models supported by GSEM Combined models (generalized responses)

We now return to a framework similar to that implemented by

Roodman’s cmp. In this application, we combine a logit equation and a
Poisson regression equation. The logit models the probability of low
birthweight, while the Poisson counts the number of premature
episodes of labor encountered by the mother. We posit a causal
relationship between premature labor and low birth weight.
gsem (low <- ptl age smoke ht lwt i.race ui, logit) ///
(ptl <- age smoke ht, poisson)

Age, smoking status and an indicator of hypertension are assumed to

affect both outcomes, where other controls only enter the logit
equation.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 37 / 39
 Models supported by GSEM Combined models (generalized responses)

age

smoke
Poisson

ptl
ht
log

lwt

1b.race
Bernoulli

2.race low
logit

3.race

ui

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 38 / 39
 Models supported by GSEM Additional models implemented in GSEM

Although we will not illustrate these models, the GSEM framework can
also be used to implement:
MIMIC model (generalized response)
Multinomial logistic regression
Random-intercept and random-slope models (multilevel)
Crossed models (multilevel)
and a number of others. See Stata’s [SEM] manual for details.

Christopher F Baum (BC / DIW) Introduction to GSEM in Stata Boston College, Spring 2016 39 / 39