Generalized Structural Equation

Modeling Using Stata

Chuck Huber
StataCorp

Italian Stata Users Group Meeting

November 14-15, 2013
 Outline
• Introduction to SEM concepts and jargon
• Continuous outcome models using SEM
• Generalized outcome models using GSEM
• Multilevel generalized models using GSEM
 What is Structural Equation Modeling?

• Structural equation modeling encompasses a

broad array of models from linear regression to
measurement models to simultaneous equations.
• Structural equation modeling is not just an
estimation method for a particular model.
• Structural equation modeling is a way of thinking,
a way of writing, and a way of estimating.

-Stata SEM Manual, pg 2

Structural Equation Models are
often drawn as Path Diagrams:
We can draw path diagrams using Stata’s SEM Builder
We can draw path diagrams using Stata’s SEM Builder

Change to generalized SEM

Select (S)
Add Observed Variable (O)
Add Generalized Response Variable (G)
Add Latent Variable (L)
Add Multilevel Latent Variable (U)
Add Path (P)
Add Covariance (C)
Add Measurement Component (M)
Add Observed Variables Set (Shift+O)
Add Latent Variables Set (Shift+L)
Add Regression Component (R)
Add Text (T)
Add Area (A)
 Jargon
• SEM and GSEM
• Observed and Latent variables
• Paths and Covariance
• Endogenous and Exogenous variables
• Recursive and Nonrecursive models
 SEM vs GSEM?
• Structural Equation Modeling (SEM)
– Continuous outcomes
– Single level data structures
– Compatible with –svy-
• Multilevel Generalized Outcomes (GSEM)
– Generalized responses (binary, ordered, count, etc)
– Multilevel data structures
– Can use factor variable notation
Structural Equation Model (SEM)
Generalized Structural Equation Model (GSEM)
 Observed and Latent Variables
• Observed variables are variables
that are included in our dataset.
They are represented by rectangles.
The variables x1, x2, x3 and x4 are
observed variables in this path
diagram.

• Latent variables are unobserved

variables that we wish we had
observed. They can be thought of
as a composite score of other
variables. They are represented by
ovals. The variable X is a latent
variable in this path diagram.
Drawing variables in Stata’s SEM Builder

Observed continuous variable (SEM and GSEM)

Observed generalized response variable (GSEM only)

Latent variable (SEM and GSEM)

Multilevel latent variable (GSEM only)

 Paths and Covariance
• Paths are direct relationships between variables. Estimated path
coefficients are analogous to regression coefficients. They are
represented by straight arrows.
• Covariance specify that two latent variables or error terms
covary. They are represented by curved arrows.
 Exogenous and Endogenous Variables
• Exogenous variables are determined outside the system of
equations. There are no paths pointing to it. The variables
price, foreign, displacement and length are exogenous.
• Endogenous variables are determined by the system of
equations. At least one path points to it. The variables weight
and mpg are endogenous.
• Observed Exogenous: a variable in a dataset
that is treated as exogenous in the model
• Latent Exogenous: an unobserved variable
that is treated as exogenous in the model.
• Observed Endogenous: a variable in a dataset
that is treated as endogenous in the model
• Latent Endogenous: an unobserved variable
that is treated as endogenous in the model.
 Recursive and Nonrecursive Systems
• Recursive models do not have any feedback loops or correlated
errors.
• Nonrecursive models have feedback loops or correlated errors.
These models have paths in both directions between one or
more pairs of endogenous variables
 Outline
• Introduction to SEM concepts and jargon
• Continuous outcome models using SEM
• Generalized outcome models using GSEM
• Multilevel generalized models using GSEM
Continuous outcome models using SEM

• Sample means
• Pearson correlation coefficient
• Student’s t-test
• Linear regression
• Multivariate linear regression
• Seemingly unrelated regression
• Three-stage least squares
Continuous outcome models using SEM
. sysuse auto
storage display value
variable name type format label variable label

make str18 %-18s Make and Model

price int %8.0gc Price
mpg int %8.0g Mileage (mpg)
rep78 int %8.0g Repair Record 1978
headroom float %6.1f Headroom (in.)
trunk int %8.0g Trunk space (cu. ft.)
weight int %8.0gc Weight (lbs.)
length int %8.0g Length (in.)
turn int %8.0g Turn Circle (ft.)
displacement int %8.0g Displacement (cu. in.)
gear_ratio float %6.2f Gear Ratio
foreign byte %8.0g origin Car type
Sample Mean Path Diagram
 Sample Mean Syntax
Syntax using means:
mean mpg

Syntax using sem:

sem mpg
 Sample Mean Results
Results using means:
Mean estimation Number of obs = 74

Mean Std. Err. [95% Conf. Interval]

mpg 21.2973 .6725511 19.9569 22.63769

Results using sem:

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

mean(mpg) 21.2973 .6679914 31.88 0.000 19.98806 22.60654

var(mpg) 33.01972 5.428409 23.92416 45.57326

Correlation Path Diagram
 Correlation Syntax
Syntax using correlate:
correlate mpg weight length

Syntax using sem:

sem mpg weight length, standardized
 Correlation Results
Results using correlate:
mpg weight length

mpg 1.0000
weight -0.8072 1.0000
length -0.7958 0.9460 1.0000

Results using sem:

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

mean(mpg) 3.706276 .3260791 11.37 0.000 3.067173 4.34538

mean(weight) 3.9116 .3419006 11.44 0.000 3.241487 4.581713
mean(length) 8.497816 .7081231 12.00 0.000 7.10992 9.885712

var(mpg) 1 . . .
var(weight) 1 . . .
var(length) 1 . . .

cov(mpg,weight) -.8071749 .0405087 -19.93 0.000 -.8865704 -.7277793

cov(mpg,length) -.7957794 .0426321 -18.67 0.000 -.8793368 -.7122221
cov(weight,
length) .9460086 .0122139 77.45 0.000 .9220699 .9699474
Student’s t-test Path Diagram
 Student’s t-test Syntax
Syntax using ttest:
ttest mpg, by(foreign)

Syntax using sem:

sem mpg <- foreign
 Student’s t-test Results
Results using ttest:
Two-sample t test with equal variances

Group Obs Mean Std. Err. Std. Dev. [95% Conf. Interval]

Domestic 52 19.82692 .657777 4.743297 18.50638 21.14747

Foreign 22 24.77273 1.40951 6.611187 21.84149 27.70396

combined 74 21.2973 .6725511 5.785503 19.9569 22.63769

diff -4.945804 1.362162 -7.661225 -2.230384

diff = mean(Domestic) - mean(Foreign) t = -3.6308

Ho: diff = 0 degrees of freedom = 72

Ha: diff < 0 Ha: diff != 0 Ha: diff > 0

Pr(T < t) = 0.0003 Pr(|T| > |t|) = 0.0005 Pr(T > t) = 0.9997

Results using sem:

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

Structural
mpg <-
foreign 4.945804 1.343628 3.68 0.000 2.312341 7.579268
_cons 19.82692 .7326131 27.06 0.000 18.39103 21.26282

var(e.mpg) 27.90954 4.5883 20.22162 38.52027

Linear Regression Path Diagram
 Linear Regression Syntax
Syntax using regress:
regress mpg weight length foreign displacement

Syntax using sem:

sem mpg <- weight length foreign displacement
 Linear Regression Results
Results using regress:
mpg Coef. Std. Err. t P>|t| [95% Conf. Interval]

weight -.0044303 .0019544 -2.27 0.027 -.0083292 -.0005315

length -.0824511 .0554128 -1.49 0.141 -.1929966 .0280944
foreign -1.692645 1.105846 -1.53 0.130 -3.898747 .5134562
displacement .0005878 .0100245 0.06 0.953 -.0194106 .0205861
_cons 50.55702 6.300024 8.02 0.000 37.98882 63.12523

Results using sem:

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

Structural
mpg <-
weight -.0044303 .0018872 -2.35 0.019 -.0081292 -.0007315
length -.0824511 .053508 -1.54 0.123 -.1873248 .0224226
foreign -1.692645 1.067833 -1.59 0.113 -3.785559 .400268
displacement .0005878 .0096799 0.06 0.952 -.0183845 .0195601
_cons 50.55702 6.083464 8.31 0.000 38.63365 62.48039

var(e.mpg) 10.78555 1.773134 7.814581 14.88603

Multivariate Regression Path Diagram
 Multivariate Regression Syntax
Syntax using mvreg:
mvreg weight length = price displacement foreign

Syntax using sem:

sem weight length <- price displacement foreign ///
, cov( e.length*e.weight)
 Multivariate Regression Results
Results using mvreg:
Coef. Std. Err. t P>|t| [95% Conf. Interval]

weight
price .0570616 .0174226 3.28 0.002 .0223132 .0918099
displacement 5.666956 .7079099 8.01 0.000 4.255074 7.078838
foreign -324.9114 122.9021 -2.64 0.010 -570.0319 -79.79076
_cons 1646.18 131.626 12.51 0.000 1383.661 1908.7

length
price .0006938 .0006547 1.06 0.293 -.0006118 .0019995
displacement .1699625 .0265999 6.39 0.000 .1169107 .2230143
foreign -6.988084 4.618077 -1.51 0.135 -16.19855 2.22238
_cons 152.1992 4.945879 30.77 0.000 142.3349 162.0634
 Multivariate Regression Results
Results using sem:
OIM
Coef. Std. Err. z P>|z| [95% Conf. Interval]

Structural
weight <-
price .0570616 .0169452 3.37 0.001 .0238496 .0902736
displacement 5.666956 .6885115 8.23 0.000 4.317498 7.016413
foreign -324.9114 119.5343 -2.72 0.007 -559.1943 -90.6284
_cons 1646.18 128.0192 12.86 0.000 1395.268 1897.093

length <-
price .0006938 .0006367 1.09 0.276 -.0005541 .0019418
displacement .1699625 .025871 6.57 0.000 .1192563 .2206687
foreign -6.988084 4.49153 -1.56 0.120 -15.79132 1.815153
_cons 152.1992 4.81035 31.64 0.000 142.771 161.6273

var(e.weight) 101330.7 16658.66 73418.28 139854.9

var(e.length) 143.0686 23.52033 103.6591 197.4608

cov(e.weight,
e.length) 3133.569 573.2375 5.47 0.000 2010.044 4257.094
Seemingly Unrelated Regression Path Diagram
 Seemingly Unrelated Regression Syntax
Syntax using sureg:
sureg (price foreign mpg displacement) ///
(weight foreign length), isure

Syntax using sem:

sem (price <- foreign mpg displacement) ///
(weight <- foreign length), ///
cov(e.price*e.weight)
 Seemingly Unrelated Regression Results
Results using sureg:
Coef. Std. Err. z P>|z| [95% Conf. Interval]

price
foreign 2940.929 691.5458 4.25 0.000 1585.525 4296.334
mpg -105.0163 57.92716 -1.81 0.070 -218.5514 8.518872
displacement 17.22083 4.244966 4.06 0.000 8.900849 25.54081
_cons 4129.866 1942.567 2.13 0.034 322.5047 7937.228

weight
foreign -153.2515 75.33472 -2.03 0.042 -300.9048 -5.598132
length 30.73507 1.528293 20.11 0.000 27.73967 33.73047
_cons -2711.096 301.6777 -8.99 0.000 -3302.374 -2119.819
 Seemingly Unrelated Regression Results
Results using sem:
OIM
Coef. Std. Err. z P>|z| [95% Conf. Interval]

Structural
price <-
foreign 2940.929 724.7311 4.06 0.000 1520.482 4361.376
mpg -105.0163 57.93461 -1.81 0.070 -218.566 8.53347
displacement 17.22083 4.5941 3.75 0.000 8.216558 26.2251
_cons 4129.866 1984.253 2.08 0.037 240.8022 8018.931

weight <-
foreign -153.2515 76.21732 -2.01 0.044 -302.6347 -3.868275
length 30.73507 1.584743 19.39 0.000 27.62903 33.84111
_cons -2711.096 312.6813 -8.67 0.000 -3323.94 -2098.252

var(e.price) 4732491 801783.1 3395302 6596312

var(e.weight) 60253.09 9933.316 43616.45 83235.44

cov(e.price,
e.weight) 209268 73909.54 2.83 0.005 64407.92 354128
3-Stage Least Squares Path Diagram
 3-Stage Least Squares Syntax
Syntax using reg3:
reg3 (mpg = weight length) ///
(weight = price foreign displacement) ///
, sure

Syntax using sem:

sem (mpg <- weight length) ///
(weight <- price foreign displacement) ///
, cov( e.mpg*e.weight)
 3-Stage Least Squares Results
Results using reg3:
Coef. Std. Err. z P>|z| [95% Conf. Interval]

mpg
weight -.0038705 .0015516 -2.49 0.013 -.0069116 -.0008295
length -.0752459 .054147 -1.39 0.165 -.181372 .0308802
_cons 47.12534 5.95489 7.91 0.000 35.45397 58.79671

weight
price .0566983 .0169217 3.35 0.001 .0235324 .0898642
foreign -331.9931 119.3554 -2.78 0.005 -565.9254 -98.0608
displacement 5.65145 .6876367 8.22 0.000 4.303707 6.999194
_cons 1653.585 127.892 12.93 0.000 1402.921 1904.248
 3-Stage Least Squares Results
Results using sem:
OIM
Coef. Std. Err. z P>|z| [95% Conf. Interval]

Structural
mpg <-
weight -.0038758 .0015516 -2.50 0.012 -.0069168 -.0008347
length -.0739612 .0550842 -1.34 0.179 -.1819243 .0340019
_cons 46.89969 6.219489 7.54 0.000 34.70971 59.08967

weight <-
price .0565862 .0169311 3.34 0.001 .0234018 .0897706
foreign -334.0496 120.3622 -2.78 0.006 -569.9552 -98.14399
displacement 5.645549 .6888778 8.20 0.000 4.295374 6.995725
_cons 1656.051 129.4058 12.80 0.000 1402.421 1909.682

var(e.mpg) 11.19224 1.842355 8.105893 15.45373

var(e.weight) 101347.1 16664.16 73426.22 139885.1

cov(e.mpg,
e.weight) -76.49352 141.2743 -0.54 0.588 -353.3861 200.3991
 Outline
• Introduction to SEM concepts and jargon
• Continuous outcome models using SEM
• Generalized outcome models using GSEM
• Multilevel generalized models using GSEM
Generalized outcome models using GSEM

• Logistic regression
• Probit regression
• Multinomial logistic regression
• Ordered logistic regression
• Poisson regression
• Negative binomial regression
Categorical outcome models using GSEM
. use "http://www.stata-press.com/data/r13/gsem_lbw", clear
. gen ptl2 = ptl>0
. label var ptl2 "Any history of premature labor"
. recode bwt (min/2500 = 1 "VeryLow") ///
(2501/3500 = 2 "Low") ///
(3501/max = 3 "Normal") ///
, gen(bwt_cat)
. label var bwt_cat "Birthweight category"
. describe

storage display value

variable name type format label variable label

id int %8.0g subject id

low byte %8.0g birth weight < 2500g
age byte %8.0g age of mother
lwt int %8.0g weight, last menstrual period
race byte %8.0g race race
smoke byte %9.0g smoke smoked during pregnancy
ptl byte %8.0g premature labor history (count)
ht byte %8.0g has history of hypertension
ui byte %8.0g presence, uterine irritability
ftv byte %8.0g # physician visits, 1st trimester
bwt int %8.0g birth weight (g)
ptl2 float %9.0g Any history of premature labor
bwt_cat int %9.0g bwt_cat Birthweight category
Logistic Regression Path Diagram
 Logistic Regression Syntax
Syntax using logit or logistic:
logit low age i.race smoke

logistic low age i.race smoke

Syntax using gsem:

gsem (low <- age 2.race 3.race smoke, ///
family(binomial) link(logit))

gsem low <- age 2.race 3.race smoke, logit

gsem low <- age i.race smoke, logit

estat eform
 Logistic Regression Results
Results using logistic:
low Odds Ratio Std. Err. z P>|z| [95% Conf. Interval]

age .9657186 .0322573 -1.04 0.296 .9045206 1.031057

race
black 2.749483 1.356659 2.05 0.040 1.045318 7.231924
other 2.876948 1.167921 2.60 0.009 1.298314 6.375062

smoke 3.00582 1.118001 2.96 0.003 1.449982 6.231081

_cons .365111 .3146026 -1.17 0.242 .0674491 1.976395

Results using gsem and estat eform:

low exp(b) Std. Err. z P>|z| [95% Conf. Interval]

age .9657186 .0322573 -1.04 0.296 .9045206 1.031057

race
white 1 (empty)
black 2.749483 1.356659 2.05 0.040 1.045318 7.231924
other 2.876948 1.167921 2.60 0.009 1.298314 6.375062

smoke 3.00582 1.118001 2.96 0.003 1.449982 6.231081

_cons .365111 .3146026 -1.17 0.242 .0674491 1.976395
Multinomial Logistic Regression Path Diagram
 Multinomial Logistic Regression Syntax
Syntax using mlogit:
mlogit bwt_cat age i.race smoke, baseoutcome(1)

Syntax using gsem:

gsem bwt_cat <- age i.race smoke, mlogit
 Multinomial Logistic Regression Results
Results using mlogit:
bwt_cat Coef. Std. Err. z P>|z| [95% Conf. Interval]

VeryLow (base outcome)

Low
age .0383795 .0349615 1.10 0.272 -.0301437 .1069028

race
black -.5139873 .5131883 -1.00 0.317 -1.519818 .4918433
other -.6468109 .4349866 -1.49 0.137 -1.499369 .2057473

smoke -.7455204 .3948038 -1.89 0.059 -1.519322 .0282809

_cons .1573422 .9150125 0.17 0.863 -1.636049 1.950734

Normal
age .0133362 .0428736 0.31 0.756 -.0706946 .097367

race
black -2.587666 .8727109 -2.97 0.003 -4.298148 -.8771839
other -2.003564 .5383297 -3.72 0.000 -3.058671 -.948457

smoke -2.014254 .5166531 -3.90 0.000 -3.026876 -1.001633

_cons 1.155972 1.13431 1.02 0.308 -1.067235 3.379179
 Multinomial Logistic Regression Results
Results using sem:
Coef. Std. Err. z P>|z| [95% Conf. Interval]

1.bwt_cat (base outcome)

2.bwt_cat <-
age .0383795 .0349615 1.10 0.272 -.0301437 .1069028

race
black -.5139873 .5131883 -1.00 0.317 -1.519818 .4918433
other -.6468109 .4349866 -1.49 0.137 -1.499369 .2057473

smoke -.7455204 .3948038 -1.89 0.059 -1.519322 .0282809

_cons .1573422 .9150125 0.17 0.863 -1.636049 1.950734

3.bwt_cat <-
age .0133362 .0428736 0.31 0.756 -.0706946 .097367

race
black -2.587666 .8727109 -2.97 0.003 -4.298148 -.8771839
other -2.003564 .5383297 -3.72 0.000 -3.058671 -.948457

smoke -2.014254 .5166531 -3.90 0.000 -3.026876 -1.001633

_cons 1.155972 1.13431 1.02 0.308 -1.067235 3.379179
Ordinal Logistic Regression Path Diagram
 Ordinal Logistic Regression Syntax
Syntax using ologit:
ologit bwt_cat age i.race smoke

Syntax using gsem:

gsem bwt_cat <- age 2.race 3.race smoke, ologit
 Ordinal Logistic Regression Results
Results using ologit:
bwt_cat Coef. Std. Err. z P>|z| [95% Conf. Interval]

age .0192665 .0268696 0.72 0.473 -.033397 .07193

race
black -1.323717 .4380843 -3.02 0.003 -2.182347 -.465088
other -1.251173 .3399064 -3.68 0.000 -1.917377 -.5849683

smoke -1.232152 .3152545 -3.91 0.000 -1.850039 -.6142645

/cut1 -1.56245 .7303115 -2.993834 -.1310659

/cut2 .6025144 .7221697 -.8129121 2.017941

Results using gsem:

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

bwt_cat <-
age .0192665 .0268696 0.72 0.473 -.033397 .07193

race
black -1.323717 .4380843 -3.02 0.003 -2.182347 -.465088
other -1.251173 .3399064 -3.68 0.000 -1.917377 -.5849683

smoke -1.232152 .3152545 -3.91 0.000 -1.850039 -.6142645

bwt_cat
/cut1 -1.56245 .7303115 -2.14 0.032 -2.993834 -.1310659
/cut2 .6025144 .7221697 0.83 0.404 -.8129121 2.017941
Poisson Regression Path Diagram
 Poisson Regression Syntax
Syntax using poisson:
poisson ftv age i.race smoke

Syntax using gsem:

gsem ftv <- age i.race smoke, poisson
 Poisson Regression Results
Results using poisson:
ftv Coef. Std. Err. z P>|z| [95% Conf. Interval]

age .0459009 .0144906 3.17 0.002 .0175 .0743019

race
black .0336452 .2488814 0.14 0.892 -.4541534 .5214438
other -.2338308 .1988733 -1.18 0.240 -.6236152 .1559537

smoke -.1147068 .1775747 -0.65 0.518 -.4627468 .2333332

_cons -1.216246 .4113805 -2.96 0.003 -2.022537 -.4099547

Results using gsem:

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

ftv <-
age .0459009 .0144906 3.17 0.002 .0175 .0743019

race
black .0336452 .2488814 0.14 0.892 -.4541534 .5214438
other -.2338308 .1988733 -1.18 0.240 -.6236152 .1559537

smoke -.1147068 .1775747 -0.65 0.518 -.4627468 .2333332

_cons -1.216246 .4113805 -2.96 0.003 -2.022537 -.4099547
 Outline
• Introduction to SEM concepts and jargon
• Continuous outcome models using SEM
• Generalized outcome models using GSEM
• Multilevel generalized models using GSEM
 Multilevel generalized outcome
models using GSEM
• Measurement component models
• Variance component models
• Latent growth curves
• Latent growth curves for generalized
outcomes
 Measurement Components Data
. use http://www.stata-press.com/data/r13/gsem_cfa, clear
. describe

storage display value

variable name type format label variable label

school byte %9.0g School id

id long %9.0g Student id
q1 byte %9.0g result q1 correct
q2 byte %9.0g result q2 correct
q3 byte %9.0g result q3 correct
q4 byte %9.0g result q4 correct
q5 byte %9.0g result q5 correct
q6 byte %9.0g result q6 correct
q7 byte %9.0g result q7 correct
q8 byte %9.0g result q8 correct
att1 float %26.0g agree Skills taught in math class will
help me get a better job.
att2 float %26.0g agree Math is important in everyday life
att3 float %26.0g agree Working math problems makes me
anxious.
att4 float %26.0g agree Math has always been my worst
subject.
att5 float %26.0g agree I am able to learn new math concepts
easily.
test1 byte %9.0g Score, math test 1
test2 byte %9.0g Score, math test 2
test3 byte %9.0g Score, math test 3
test4 byte %9.0g Score, math test 4
Measurement Component Path Diagram

We can conceptualize the eight measured variables

q1-q8 as being realizations of a person’s math ability.
We can quantify this idea using a latent variable
MathAbility.
 Measurement Components Syntax

Syntax using gsem:

gsem (MathAbility -> q1-q8 ///
, family(bernoulli) link(logit)) ///
, latent(MathAbility ) nocapslatent
 Measurement Components Results
Results using gsem:
Coef. Std. Err. z P>|z| [95% Conf. Interval]

q1 <-
MathAbility 1 (constrained)
_cons .0373365 .1252279 0.30 0.766 -.2081058 .2827787

q2 <-
MathAbility .381626 .116809 3.27 0.001 .1526845 .6105674
_cons -.4613391 .0989722 -4.66 0.000 -.655321 -.2673571

q3 <-
MathAbility .4993762 .134314 3.72 0.000 .2361255 .7626269
_cons .1533362 .1006072 1.52 0.127 -.0438503 .3505228

q4 <-
MathAbility .3299698 .1063034 3.10 0.002 .1216189 .5383207
_cons -.3230667 .0957983 -3.37 0.001 -.510828 -.1353054

q5 <-
MathAbility .8401762 .1995336 4.21 0.000 .4490975 1.231255
_cons -.0494684 .1163093 -0.43 0.671 -.2774304 .1784937

q6 <-
MathAbility .6453722 .1639865 3.94 0.000 .3239646 .9667798
_cons -.314723 .1083049 -2.91 0.004 -.5269968 -.1024493

q7 <-
MathAbility .8163613 .2045448 3.99 0.000 .4154609 1.217262
_cons .1053404 .1152979 0.91 0.361 -.1206393 .3313201

q8 <-
MathAbility .5769516 .1473524 3.92 0.000 .2881463 .865757
_cons -.026705 .1034396 -0.26 0.796 -.2294429 .1760328

var(MathAbility) 2.151059 .7298407 1.106229 4.182728

Measurement Component Path Diagram

We can also conceptualize the five measured variables

att1-att5 as being realizations of a person’s attitude
about math. We can quantify this idea using a latent
variable MathAttitude.
 Measurement Components Syntax

Syntax using gsem:

gsem (MathAttitude -> att1-att5 ///
, family(ordinal) link(logit)) ///
, latent(MathAttitude ) nocapslatent
 Measurement Components Results
Results using gsem: att1 <-
Coef. Std. Err. z P>|z| [95% Conf. Interval]

MathAttitude 1 (constrained)

att2 <-
MathAttitude .3651316 .0947737 3.85 0.000 .1793785 .5508846

att3 <-
MathAttitude -1.325592 .3281752 -4.04 0.000 -1.968803 -.6823802

att4 <-
MathAttitude -.7319336 .1476384 -4.96 0.000 -1.0213 -.4425677

att5 <-
MathAttitude .4629576 .1117098 4.14 0.000 .2440104 .6819047

att1
/cut1 -1.14403 .1407016 -8.13 0.000 -1.4198 -.8682596
/cut2 -.2571716 .1208793 -2.13 0.033 -.4940907 -.0202526
/cut3 .3113316 .121399 2.56 0.010 .0733939 .5492693
/cut4 1.38373 .1505219 9.19 0.000 1.088713 1.678748

att2
/cut1 -1.058352 .1069425 -9.90 0.000 -1.267955 -.8487485
/cut2 -.1920422 .0946322 -2.03 0.042 -.3775179 -.0065665
/cut3 .3639243 .0957805 3.80 0.000 .1761979 .5516506
/cut4 1.139819 .1090449 10.45 0.000 .9260952 1.353543

att3
/cut1 -1.003196 .1634751 -6.14 0.000 -1.323601 -.6827905
/cut2 -.0511457 .1372565 -0.37 0.709 -.3201635 .2178721
/cut3 .5278704 .1454233 3.63 0.000 .2428459 .8128949
/cut4 1.587917 .1989801 7.98 0.000 1.197923 1.977911

att4
/cut1 -1.071316 .1214149 -8.82 0.000 -1.309285 -.8333473
/cut2 -.212007 .1074834 -1.97 0.049 -.4226707 -.0013434
/cut3 .4028505 .1092331 3.69 0.000 .1887576 .6169435
/cut4 1.393148 .1312299 10.62 0.000 1.135942 1.650354

att5
/cut1 -1.242513 .1147059 -10.83 0.000 -1.467332 -1.017693
/cut2 -.339867 .0983909 -3.45 0.001 -.5327096 -.1470243
/cut3 .2076369 .0974768 2.13 0.033 .0165858 .3986879
/cut4 .9211489 .1067054 8.63 0.000 .7120101 1.130288

var(MathAttit~e) 1.835912 .5279313 1.044917 3.225683

Variance Component Model Path Diagram
 Variance Component Model Syntax
Syntax using mixed:
mixed test1 || school:

Syntax using gsem:

gsem (M1[school] -> test1, ), latent(M1) ///
nocapslatent
 Variance Component Model Results
Results using mixed:
test1 Coef. Std. Err. z P>|z| [95% Conf. Interval]

_cons 75.548 .6529386 115.70 0.000 74.26826 76.82774

Random-effects Parameters Estimate Std. Err. [95% Conf. Interval]

school: Identity
var(_cons) 7.410363 2.697303 3.630865 15.12408

var(Residual) 27.90533 1.801282 24.58909 31.66883

Results using gsem:

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

test1 <-
M1[school] 1 (constrained)

_cons 75.548 .6529386 115.70 0.000 74.26826 76.82774

var(M1[school]) 7.410362 2.697302 3.630865 15.12407

var(e.test1) 27.90533 1.801281 24.58908 31.66882

 Latent Growth Curve Data
. use http://www.stata-press.com/data/r13/sem_lcm, clear
. gen id = _n
. describe

storage display value

variable name type format label variable label

lncrime0 float %9.0g ln(crime rate) in Jan & Feb

lncrime1 float %9.0g ln(crime rate) in Mar & Apr
lncrime2 float %9.0g ln(crime rate) in May & Jun
lncrime3 float %9.0g ln(crime rate) in Jul & Aug
id float %9.0g
Latent Growth Curve Path Diagram
 Latent Growth Curve Syntax
Syntax using mixed:
gen id = _n
reshape long lncrime, i(id) j(time)
mixed lncrime time || id: time, cov(unstr)

Syntax using sem:

reshape wide
sem (lncrime0 <- Intercept@1 Slope@0 _cons@0) ///
(lncrime1 <- Intercept@1 Slope@1 _cons@0) ///
(lncrime2 <- Intercept@1 Slope@2 _cons@0) ///
(lncrime3 <- Intercept@1 Slope@3 _cons@0), ///
latent(Intercept Slope) ///
var(e.lncrime0@var e.lncrime1@var ///
e.lncrime2@var e.lncrime3@var) ///
means(Intercept Slope) ///
nocnsreport nolog
 Latent Growth Curve Results
Results using mixed:
lncrime Coef. Std. Err. z P>|z| [95% Conf. Interval]

time .1426952 .0104574 13.65 0.000 .1221992 .1631912

_cons 5.337915 .0407501 130.99 0.000 5.258047 5.417784

Random-effects Parameters Estimate Std. Err. [95% Conf. Interval]

id: Unstructured
var(time) .0196198 .0031082 .0143829 .0267635
var(_cons) .5274091 .0446436 .4467824 .6225859
cov(time,_cons) -.034316 .0088848 -.0517298 -.0169022

var(Residual) .0981956 .0051826 .0885456 .1088972

Results using sem:

mean(Intercept) 5.337915 .0407501 130.99 0.000 5.258047 5.417784
mean(Slope) .1426952 .0104574 13.65 0.000 .1221992 .1631912

var(e.lncrime0) .0981956 .0051826 .0885457 .1088972

var(e.lncrime1) .0981956 .0051826 .0885457 .1088972
var(e.lncrime2) .0981956 .0051826 .0885457 .1088972
var(e.lncrime3) .0981956 .0051826 .0885457 .1088972
var(Intercept) .527409 .0446436 .4467822 .6225858
var(Slope) .0196198 .0031082 .0143829 .0267635

cov(Intercept,
Slope) -.034316 .0088848 -3.86 0.000 -.0517298 -.0169022
 Poisson Latent Growth Curve Data
(data in long format)

storage display value

variable name type format label variable label

subject byte %9.0g Subject ID: 1-59

seizures int %9.0g No. of seizures
treat byte %9.0g 1: progabide; 0: placebo
visit float %9.0g Dr. visit; coded as (-.3, -.1, .1,
.3)
lage float %9.0g log(age), mean-centered
lbas float %9.0g log(0.25*baseline seizures),
mean-centered
lbas_trt float %9.0g lbas/treat interaction
v4 byte %8.0g Fourth visit indicator
time float %9.0g

Sorted by: subject visit

Poisson Latent Growth Curve Path Diagram
(data in long format)
 Poisson Latent Growth Curve Path Diagram
(data in long format)

Syntax using mepoisson:

mepoisson seizures time || subject: time, cov(unstr)

Syntax using gsem:

gsem (seizures <- time c.time#S[subject] I[subject]), ///
family(poisson) link(log)
 Poisson Latent Growth Curve Path Diagram
(data in long format)
Results using mepoisson:
seizures Coef. Std. Err. z P>|z| [95% Conf. Interval]

time -.0503416 .0355162 -1.42 0.156 -.119952 .0192689

_cons 1.682242 .1386123 12.14 0.000 1.410566 1.953917

subject
var(time) .0211453 .0091765 .0090326 .0495009
var(_cons) .9545034 .2064406 .6247109 1.458397

subject
cov(_cons,time) -.0362171 .0337022 -1.07 0.283 -.1022722 .0298379
 Poisson Latent Growth Curve Path Diagram
(data in long format)

Results using gsem:

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

seizures <-
time -.0503416 .0355162 -1.42 0.156 -.119952 .0192689

c.time#
S[subject] 1 (constrained)

I[subject] 1 (constrained)

_cons 1.682242 .1386123 12.14 0.000 1.410566 1.953917

var(S[subject]) .0211453 .0091765 .0090326 .0495009

var(I[subject]) .9545034 .2064406 .6247109 1.458397

cov(I[subject],
S[subject]) -.0362171 .0337022 -1.07 0.283 -.1022722 .0298379
 Poisson Latent Growth Curve Data
(data in wide format)

storage display value

variable name type format label variable label

subject byte %9.0g Subject ID: 1-59

seizures0 int %9.0g 0 seizures
lage float %9.0g 0 lage
seizures1 int %9.0g 1 seizures
seizures2 int %9.0g 2 seizures
seizures3 int %9.0g 3 seizures

Sorted by: subject

Poisson Latent Growth Curve Path Diagram
(data in wide format)
 Poisson Latent Growth Curve Path Diagram
(data in wide format)

Syntax using gsem:

gsem (seizures0 <- Intercept@1 Slope@0) ///
(seizures1 <- Intercept@1 Slope@1) ///
(seizures2 <- Intercept@1 Slope@2) ///
(seizures3 <- Intercept@1 Slope@3), ///
nocons means(Intercept Slope) ///
family(poisson) link(log)
 Poisson Latent Growth Curve Path Diagram
(data in wide format)
Results using gsem:
Coef. Std. Err. z P>|z| [95% Conf. Interval]

seizures0 <-
Intercept 1 (constrained)
Slope 0 (omitted)
_cons 0 (omitted)

seizures1 <-
Intercept 1 (constrained)
Slope 1 (constrained)
_cons 0 (omitted)

seizures2 <-
Intercept 1 (constrained)
Slope 2 (constrained)
_cons 0 (omitted)

seizures3 <-
Intercept 1 (constrained)
Slope 3 (constrained)
_cons 0 (omitted)

mean(Intercept) 1.682215 .138856 12.11 0.000 1.410062 1.954368

mean(Slope) -.0503355 .0355747 -1.41 0.157 -.1200606 .0193896

var(Intercept) .9543458 .2067282 .6241952 1.45912

var(Slope) .0211439 .0091791 .0090294 .0495122

cov(Slope,
Intercept) -.0362289 .0339178 -1.07 0.285 -.1027065 .0302486
 Multilevel GSEM Data
. use http://www.stata-press.com/data/r13/gsem_cfa, clear
. describe

storage display value

variable name type format label variable label

school byte %9.0g School id

id long %9.0g Student id
q1 byte %9.0g result q1 correct
q2 byte %9.0g result q2 correct
q3 byte %9.0g result q3 correct
q4 byte %9.0g result q4 correct
q5 byte %9.0g result q5 correct
q6 byte %9.0g result q6 correct
q7 byte %9.0g result q7 correct
q8 byte %9.0g result q8 correct
att1 float %26.0g agree Skills taught in math class will
help me get a better job.
att2 float %26.0g agree Math is important in everyday life
att3 float %26.0g agree Working math problems makes me
anxious.
att4 float %26.0g agree Math has always been my worst
subject.
att5 float %26.0g agree I am able to learn new math concepts
easily.
test1 byte %9.0g Score, math test 1
test2 byte %9.0g Score, math test 2
test3 byte %9.0g Score, math test 3
test4 byte %9.0g Score, math test 4
Multilevel GSEM Path Diagram
 Multilevel GSEM Syntax

Syntax using gsem:

gsem (MathAbility -> q1-q8, family(bernoulli) link(logit)) ///
(School[school] -> q1-q8, family(bernoulli) link(logit)) ///
, latent(MathAbility School) ///
covstruct(_lexogenous, diagonal) ///
nocapslatent
Generalized Structural Equation Models

We can combine measurement components to

fit a dizzying variety of models that can
simultaneously combine longitudinal, latent
growth curve and multilevel structures that
cannot be modeled with other Stata commands.
Multilevel GSEM Path Diagram
 Multilevel GSEM Syntax

Syntax using gsem:

gsem (MathAbility -> q1-q8, family(bernoulli) link(logit)) ///
(MathAttitude -> MathAbility, ) ///
(MathAttitude -> att1-att5, family(ordinal) link(logit)) ///
, latent(MathAbility MathAttitude ) nocapslatent
GSEM Examples
GSEM Examples

Three level GSEM

 GSEM Examples

Crossed Model: pid1 is the students’ primary school ID and sid2

is the students’ secondary school ID number. The two multilevel
latent variables account for the nesting of students within each of
the two schools.
GSEM Examples
GSEM Examples

Multilevel Mediation Model

 Conclusions
• Many regression models are a special case of
SEM/GSEM
• SEM/GSEM allow us to fit complex structural
equation models that we cannot fit with other
regression techniques
• Stata’s SEM Builder makes it easy to draw and
estimate structural equation models
 References and Further Reading
Stata 13 Structural Equation Modeling Reference Manual:
www.stata.com/manuals13/sem.pdf

Acock, A.C. (2013) Discovering Structural Equation Modeling

Using Stata, Revised Edition . College Station, TX: Stata Press.
‎
Rabe-Hesketh, S., and A. Skrondal. (2012) Multilevel and
Longitudinal Modeling Using Stata. 3rd ed. College Station, TX:
Stata Press.