Psychology dissertation structural equation modelling combines measurement models and directional relationships so you can test theories involving observed variables, latent constructs, measurement error, and indirect pathways. It is powerful because one model can represent how questionnaire items measure constructs and how those constructs relate. It is demanding because every arrow, covariance, constraint, estimator, and comparison needs a defensible rationale.
This guide explains when structural equation modelling, commonly abbreviated as SEM, is appropriate and how to plan, estimate, evaluate, and report it. The focus is covariance-based SEM used in psychological research. The article complements the existing guides to factor analysis, regression analysis, and mediation analysis rather than duplicating them.
What structural equation modelling adds
Ordinary regression treats measured variables as though they were observed without error. SEM can instead define an unobserved, or latent, construct through several indicators and estimate structural relations among latent and observed variables. A dissertation might model academic self-efficacy using questionnaire items, then test whether self-efficacy predicts engagement after accounting for anxiety.
An SEM has two connected parts. The measurement model specifies relations between latent constructs and their indicators. The structural model specifies regressions, covariances, indirect effects, or group differences among constructs. Separating these parts is useful because a weak measurement model can make a theoretically appealing structural path difficult to interpret.
SEM is not automatically superior to regression. If the study has a small sample, a single reliable observed outcome, and one straightforward predictor question, a simpler model may be more credible. Choose SEM because the research question needs latent variables, simultaneous pathways, longitudinal change, multiple groups, or explicit measurement assumptions, not because a path diagram looks advanced.
Match the SEM type to the research question
| Research aim | Possible SEM approach | Psychology example | Key caution |
|---|---|---|---|
| Test a measurement structure | Confirmatory factor analysis | Do six items measure one anxiety factor? | Requires an a priori factor model |
| Relate latent constructs | Latent-variable SEM | Does self-efficacy predict engagement? | Good fit does not establish causality |
| Test an indirect pathway | Latent mediation model | Does coping link stress to wellbeing? | Temporal order needs design support |
| Compare populations | Multiple-group SEM | Are paths similar across study modes? | Measurement comparability comes first |
| Model change | Latent growth model | How do symptoms change across four waves? | Time coding and attrition matter |
| Account for clustered data | Multilevel SEM | Students nested within universities | Needs enough units at each level |
Begin with the estimand: the quantity the dissertation is intended to learn about. A direct path, total effect, latent mean difference, change parameter, or model comparison answers a different question. Define the primary estimand before opening software, then make the design, measures, sample-size plan, and reporting serve that target.
Develop a theory-led path diagram
A path diagram should translate a stated theory, not replace one. Circles usually represent latent variables, rectangles observed variables, single-headed arrows directional regressions, and double-headed arrows variances or covariances. Software conventions vary, so provide a legend and explain every non-obvious element.
Write the model in words before drawing it. Identify each construct, its indicators, the predicted direction of each relation, relevant covariates, residual covariances, and the timing of measurements. Distinguish confirmatory paths from exploratory additions. If two indicators share wording or method, a residual covariance might be plausible, but it should be justified before or transparently after analysis rather than added simply to improve fit.
The official lavaan tutorial demonstrates how one syntax can contain latent-variable definitions, regression formulas, and covariance formulas. This makes the model auditable. Preserve the syntax and annotate parameter labels so the path diagram, hypotheses, and reported table can be matched.
Check identification before estimation
An identified model contains enough independent information to estimate its free parameters. An underidentified model has infinitely many solutions, while a just-identified model reproduces the observed covariance matrix exactly and therefore cannot provide a meaningful global test of fit. An overidentified model has positive degrees of freedom and can be evaluated against the data.
Each latent factor needs a scale. Common choices fix one loading to one or fix the latent variance. Indicators must supply adequate information, and the structural portion must not contain unresolved feedback or redundant parameters. Three well-chosen indicators per factor often make identification and diagnostics easier, but indicator count alone does not guarantee a strong measure.
Before data collection, count the observed variances and covariances, list every free parameter, and check the software degrees of freedom against the intended model. After fitting, examine warnings, impossible estimates, very large standard errors, and correlations near one. A model can be formally identified yet practically weak because the sample supplies little information.
Protect the measurement model
Use indicators that represent the construct
Indicators should cover the construct’s theoretical domain without becoming interchangeable repetitions. Report item wording source, response scale, scoring, reverse coding, and prior validity evidence. Do not delete items solely because a modification index or loading is inconvenient. Removal can narrow the construct and make the final latent variable different from the one described in the introduction.
Evaluate loadings and residuals together
A loading represents the relation between an indicator and its factor under the specified model. Its magnitude, uncertainty, direction, and consistency with theory matter. Residual variance is the indicator variation not represented by the factor. A low loading is not repaired merely by reporting a high coefficient elsewhere.
Inspect standardized residuals, localized misfit, factor correlations, and theoretically plausible cross-loadings. Global fit can appear acceptable while a particular item pair is poorly represented. Conversely, one modest loading may be tolerable when the item protects important content validity.
Distinguish CFA from exploratory factor analysis
Confirmatory factor analysis tests a prespecified allocation of indicators to factors, usually constraining unplanned cross-loadings. Exploratory factor analysis estimates a rotated loading pattern when the structure is uncertain. Avoid calling an extensively modified CFA confirmatory. If the structure is genuinely unclear, the factor-analysis guide explains retention, rotation, EFA, and CFA choices.
Plan sample size for the actual model
There is no universal SEM sample-size rule. Requirements change with factor loadings, path sizes, indicator reliability, model complexity, missingness, non-normality, estimator, and the parameter being tested. A ratio such as ten cases per parameter ignores these features and can be either wasteful or seriously underpowered.
Simulation research demonstrates that sample requirements vary substantially across SEMs. Open research by Buchberger and colleagues shows how construct separability and indicator reliability affect power for selecting among competing models. These findings support model-specific planning rather than a fixed minimum.
For a simulation-based plan, define plausible population loadings, variances, paths, correlations, missingness, and estimator. Generate many datasets at candidate sample sizes, fit the exact planned model, and summarize convergence, improper solutions, parameter bias, confidence-interval coverage, and power for the primary estimand. Repeat the plan with weaker loadings or greater missingness to show sensitivity. Link the rationale to the broader power-analysis guide.
Prepare data without hiding problems
Verify item direction, legal ranges, duplicated cases, impossible patterns, group codes, and the number of observations contributing to each covariance. Examine distributions and missingness at item level. A total score can conceal sparse categories, ceiling effects, or an incorrectly reversed item that destabilizes the latent model.
Do not screen cases using the outcome of the preferred SEM. Define exclusions independently and report them. For influential observations, compare results with defensible robust methods or sensitivity analyses rather than automatically deleting multivariate outliers. See the guides to outlier analysis and missing data.
Choose an estimator that matches the data
| Data condition | Common choice | Why it may fit | What to report |
|---|---|---|---|
| Approximately continuous indicators | ML or robust ML | Models covariance and mean structures efficiently | Estimator, robust correction, missing-data method |
| Ordered categorical indicators | DWLS or WLSMV-type method | Uses thresholds and ordinal information | Ordered variables, thresholds, weight-matrix method |
| Incomplete continuous data under MAR | Full-information ML | Uses available case-wise likelihood information | Auxiliary variables, missingness rationale, sensitivity |
| Severe non-normality or small samples | Robust or Bayesian approach | Can address distributional or estimation limitations | Priors or corrections, diagnostics, convergence |
| Complex indirect effect | Bootstrap or simulation-based interval | Allows asymmetric sampling uncertainty | Resamples, interval type, random seed |
The lavaan documentation states that maximum likelihood is the default for continuous data and describes robust alternatives, full-information maximum likelihood for incomplete continuous data, and bootstrap options. Its categorical-data guide explains that declaring endogenous indicators as ordered switches to a diagonally weighted least-squares approach with robust corrections.
Do not choose an estimator by software default alone. A five-category item can sometimes behave adequately under robust maximum likelihood, but heavily skewed or sparse ordinal categories may need an ordinal model. Describe the empirical distributions and explain why the estimator is appropriate for the indicators and hypothesis.
Evaluate fit as evidence, not a pass mark
The model chi-square tests exact fit, so it is sensitive to sample size and small discrepancies. RMSEA summarizes approximate misfit per degree of freedom, CFI and TLI compare the target model with a baseline model, and SRMR summarizes standardized residual discrepancies. Each index emphasizes different aspects and can behave differently with model size, degrees of freedom, estimator, and indicator quality.
A set of popular thresholds should not be treated as universal law. Report the indices selected in advance, their confidence intervals where available, the test statistic and degrees of freedom, and any robust or scaled versions. Consider parameter plausibility, residuals, convergence, theory, and predictive or comparative evidence alongside global fit.
Good global fit means the model-implied moments are not grossly inconsistent with the observed moments under the chosen diagnostics. It does not prove that the model is true, that every path is unbiased, or that competing models would fit worse. Several theoretically different models can reproduce the same covariance pattern.
| Evidence source | Question it addresses | Misuse to avoid |
|---|---|---|
| Chi-square and degrees of freedom | Is exact covariance fit rejected? | Treating non-significance as proof |
| RMSEA with interval | How much approximate misfit is estimated? | Applying one cutoff to very low-df models |
| CFI or TLI | How does fit improve over a baseline model? | Ignoring a weak or unusual baseline |
| SRMR and residuals | Where are standardized discrepancies? | Using only the average residual |
| Parameter estimates | Are effects plausible and precise? | Celebrating fit with impossible estimates |
| Competing models | Do plausible alternatives explain the data? | Comparing only after seeing results |
Use modification indices cautiously
A modification index estimates how much the chi-square statistic may decrease if a constrained parameter is freed, subject to approximation. It does not provide a theoretical reason for adding the parameter and does not guarantee replication. Searching many indices turns a confirmatory model into a data-driven one.
If modification is necessary, identify a substantive explanation, inspect the expected parameter change, add one coherent change at a time, and describe it as exploratory. Ideally, cross-validate the revised model in an independent sample or a planned split sample. Report the original model, the modification process, and the final model rather than presenting the revised version as preregistered.
Test measurement invariance before latent comparisons
When comparing groups or time points, ask whether indicators represent the same construct on a comparable scale. Configural invariance concerns the same basic factor pattern. Metric invariance constrains loadings, supporting comparisons of relations under appropriate conditions. Scalar invariance also constrains intercepts or thresholds, which is generally needed for latent mean comparisons.
Putnick and Bornstein’s review emphasizes that invariance testing evaluates psychometric equivalence across groups or time and that reporting practices should make these decisions transparent. Do not declare invariance from one index alone. Examine parameter changes, fit changes, sample size, group imbalance, partial invariance decisions, and whether the remaining noninvariance affects the target comparison.
Avoid causal claims from cross-sectional SEM
Directional arrows encode model assumptions; they do not manufacture temporal order, randomization, or control of unmeasured confounding. A cross-sectional model in which stress predicts sleep can often be re-expressed with sleep predicting stress or with a common cause. Similar fit cannot decide the causal direction.
Use language such as association, direct path, or model-implied relation unless the design supports stronger inference. Longitudinal measurements, experiments, careful confounder control, and sensitivity analyses can strengthen a causal argument, but SEM alone is not a causal design. The same caution applies to indirect effects.
Report enough information for reproduction
APA’s quantitative reporting standards for SEM call for clear descriptions of the proposed model, data preparation, estimation, fit evaluation, alternatives, and results. Clear reporting should allow readers to understand how the model was specified and evaluated rather than merely listing selected fit indices.
State the software, version, estimator, scaling method, missing-data handling, sample size, covariance or raw-data input, and convergence criteria. Give the complete measurement and structural specification, identification constraints, correlated residuals, equality constraints, indirect-effect definitions, and preregistered comparisons. Provide a readable diagram and syntax.
Report unstandardized estimates for reproducibility and standardized estimates when they aid interpretation. Include standard errors or confidence intervals, test statistics, exact p values where used, factor loadings, residual variances, latent variances and covariances, structural paths, indirect effects, fit statistics, and model-comparison results. Explain substantive magnitudes in the scale of the research question.
Common structural equation modelling mistakes
- Starting with software: define constructs, estimands, and paths before choosing commands.
- Using SEM to rescue weak measures: protect construct validity and report indicator problems.
- Applying one sample-size rule: simulate the planned model under plausible conditions.
- Treating Likert items automatically as continuous: inspect categories and justify the estimator.
- Chasing fit thresholds: combine global fit, residuals, parameters, theory, and alternatives.
- Adding every suggested covariance: require a defensible mechanism and label exploration.
- Comparing latent means without invariance: establish measurement comparability first.
- Interpreting arrows causally: match claims to design, timing, and assumptions.
Frequently asked questions
Is SEM the same as path analysis?
Path analysis usually models relations among observed variables. SEM is a broader framework that can include path analysis, latent measurement models, growth models, multiple groups, and other structures. A path model without latent variables still belongs to the SEM family.
Can I use SEM with a small dissertation sample?
Possibly, but feasibility depends on the model, loadings, effect sizes, estimator, missingness, and primary target. A simpler observed-variable model or a smaller measurement model may be more credible. Use simulation and discuss instability rather than claiming a universal minimum.
Should I report CFI, TLI, RMSEA, and SRMR?
These provide complementary information and are commonly reported, but their interpretation depends on the model and estimator. Also report chi-square, degrees of freedom, relevant intervals, parameter plausibility, residual evidence, and the criteria chosen before analysis.
What is an improper SEM solution?
Examples include negative residual variance, a standardized loading above one, a non-positive definite covariance matrix, or implausible latent correlations. Investigate coding, model specification, sample information, and estimator assumptions. Do not silently constrain the problem away.
Can SEM prove mediation?
No. SEM estimates an indirect path under the specified model, but causal mediation requires temporal order and strong assumptions about confounding and measurement. Use proportionate language and report the design limitation.
Which software can run SEM?
Common options include lavaan in R, Mplus, Amos, Stata, SAS, EQS, OpenMx, and JASP. Capabilities and defaults differ. Choose software that supports the required estimator and model, then preserve syntax and record the version.
Conclusion
Strong psychology dissertation structural equation modelling begins with a clear theory, valid measures, an identified model, and a model-specific sample-size plan. Evaluate the measurement model before emphasizing structural paths. Match the estimator to the data, interpret fit as graded evidence, disclose modifications, and keep causal language within the limits of the design.
For ethical academic support, request feedback on your path diagram, model rationale, syntax, diagnostics, or reporting. A responsible reviewer can explain weaknesses and alternatives, but should not fabricate data, conceal failed models, or produce findings that you cannot reproduce and defend.
Authoritative references
- lavaan project: structural equation model example and syntax
- lavaan project: estimators, missing values, and bootstrapping
- lavaan project: categorical indicators and WLSMV
- Buchberger and colleagues: simulation-based SEM power planning
- Putnick and Bornstein: measurement invariance conventions and reporting
- APA Style JARS: reporting standards for structural equation modelling
