SEM and GSEM in Stata: Choosing Commands, Specifying Models, and Interpreting Results

1. Start with the Dependent Variable Rather Than the Diagram

A diagram with many arrows does not justify gsem. Tabulate each equation: is the dependent variable continuous, binary, ordinal, or a count; are observations clustered; which variables are latent; and on which scale should coefficients be interpreted?

sem suits linear structural equation models within a covariance framework. gsem extends to other families and links and supports many generalized and multilevel structures. These extensions change coefficient meaning and available postestimation measures (Skrondal and Rabe-Hesketh, 2004; StataCorp, 2025).

2. Creating Do-Files, Logs, and Project Structure

A do-file records all commands in order. A log saves commands and results as text. Stata documentation recommends do-files and logs to preserve an analysis trail (StataCorp, n.d.).

The version command makes Stata interpret syntax according to the specified release. Replace 19 with the version used for the final analysis. Relative paths make projects portable between computers.

version 19
clear all
set more off
capture log close _all

log using "output/analysis.log", text replace

import delimited "data/raw/survey.csv", ///
    varnames(1) encoding(utf8) clear

describe
codebook id gender age
isid id
duplicates report id

3. Checking Data Before Model Specification

Model identification cannot repair basic coding errors. Before sem or gsem, check IDs, variable types, ranges, missing data, rare groups, and outcome distributions.

The following example treats 99 as missing according to its codebook. If 99 is valid in the actual data, replace would destroy information. Check the codebook first.

misstable summarize ///
    satisfaction1-satisfaction4 ///
    trust1-trust4 outcome

foreach v of varlist satisfaction1-satisfaction4 ///
                     trust1-trust4 outcome {
    replace `v' = . if `v' == 99
}

assert inrange(satisfaction1, 1, 5) ///
    if !missing(satisfaction1)

generate satisfaction2_r = 6 - satisfaction2 ///
    if !missing(satisfaction2)

egen satisfaction_mean = rowmean( ///
    satisfaction1 satisfaction2_r ///
    satisfaction3 satisfaction4)

egen n_satisfaction = rownonmiss( ///
    satisfaction1 satisfaction2_r ///
    satisfaction3 satisfaction4)

replace satisfaction_mean = . ///
    if n_satisfaction < 3

summarize satisfaction_mean outcome, detail

4. Model Specification and Identification

Establish identification before assessing fit. Each latent variable needs a scale, commonly by fixing a loading or variance. The path structure must provide enough information to estimate free parameters.

Identification is a property of the model and data rather than a technical option. Bollen (1989) presents SEM’s foundations of specification, identification, and estimation. If Stata reports nonidentification or an abnormal solution, check equations and measurement scales before adding constraints.

5. Practicing sem with Continuous Outcomes

The first example is an observed-variable path model. satisfaction_mean predicts trust_mean and outcome; trust_mean also predicts outcome. method(mlmv) illustrates maximum likelihood estimation with missing values supported by sem. Its applicability conditions must still be assessed.

estat gof provides fit indices after sem. Do not decide model correctness using one cutoff. Examine specification, residual magnitudes, parameter estimates, and theoretical plausibility.

sem ///
    (trust_mean <- satisfaction_mean age) ///
    (outcome <- trust_mean satisfaction_mean age), ///
    method(mlmv)

estat gof, stats(all)
estat stdize, all
estat teffects

estimates store sem_path

6. Practicing Measurement Models with sem

Stata latent variable names usually begin with uppercase letters to distinguish them from observed variables. The following example specifies two reflective constructs. By default, the first indicator of each construct sets its scale.

After estimation, check loadings, error variances, construct correlations, and residuals. A good fit index does not guarantee that indicators measure the intended conceptual content.

sem ///
    (Satisfaction -> ///
        satisfaction1 satisfaction2_r ///
        satisfaction3 satisfaction4) ///
    (Trust -> trust1 trust2 trust3 trust4), ///
    method(mlmv) standardized

estat gof, stats(all)
estat residuals, standardized

estimates store cfa_two_factor

7. Practicing gsem with Ordinal Outcomes

Suppose loyalty_level contains ordered categories. Ordered logit models cumulative probabilities through the logit link. The satisfaction_mean coefficient lies on the log-odds scale and should not be interpreted as a direct increase in loyalty points.

After gsem, many familiar sem fit indices cannot be used in the same way. Focus on family and link specification, convergence, estimates, prediction, and justified model comparisons (StataCorp, 2025).

gsem ///
    (loyalty_level <- ///
        satisfaction_mean trust_mean age), ///
    ologit vce(robust)

estimates store gsem_ordinal

margins, at( ///
    satisfaction_mean = (1 2 3 4 5) ///
) predict(outcome(4))

marginsplot, ///
    ytitle("Predicted probability") ///
    xtitle("Satisfaction mean")

graph export ///
    "output/figures/ordinal_probability.png", ///
    width(1600) replace

8. Do Not Mechanically Compare sem and gsem Using One Number

Compare two models directly only when they describe the same outcome and sample and have a nesting relationship appropriate to the test. sem with a continuous outcome and gsem with an ordinal outcome specify different probability models.

If conclusions are sensitive to outcome treatment, present a sensitivity analysis. Explain why one specification is primary, what the other assesses, and which conclusions change.

9. Exporting Output and Verifying the Final Run

A complete do-file closes the log and saves derived data when needed. datasignature creates an identifier for the current data, helping detect file changes between runs.

Thesis tables must be traceable to saved models. When copying manually, record the estimates store name or model ID below the table for cross-checking.

compress
save "data/derived/survey_analysis.dta", replace

datasignature set, reset
datasignature report

estimates dir
log close

10. Reporting Template and Checklist

• Classify each outcome by support and measurement scale.
• Provide a methodological rationale for sem or gsem.
• State the family, link, estimator, and standard-error method.
• Check latent-variable scaling and identification conditions.
• Report missing data and case counts for each model.
• Interpret coefficients on the correct scale.
• Do not apply sem fit cutoffs to gsem.
• Ensure the do-file runs from a fresh session and save the log.
• Clearly label example data or hypothetical results.