A path diagram can look like a map of reality.

Boxes become observed variables. Circles become hidden constructs. Arrows imply direction. Numbers appear on paths. The diagram can be so visually persuasive that it is easy to forget what structural equation modeling actually does: it fits a specified mathematical model to patterns of means, variances, covariances and, in some extensions, categorical responses or multilevel structure.

Structural equation modeling, or SEM, is powerful because it can represent measurement error, latent variables and multiple linked relationships in one coherent system. It is also dangerous when a fitted system is mistaken for a discovered causal mechanism. The method deserves both respect and restraint.

Quick Read

• SEM combines ideas from factor analysis, regression, covariance modeling and path analysis.
• A measurement model links latent constructs to observed indicators.
• A structural model links variables or latent constructs through regression-like paths.
• Identification asks whether the available data contain enough information to estimate the parameters uniquely.
• Good global fit does not prove the model is true; many different models can fit similar covariance patterns.
• Modification indices can suggest places where the model conflicts with the data, but chasing them mechanically can overfit the sample.
• Measurement invariance matters when latent constructs are compared across groups or time.
• SEM does not create causal identification. Causal claims still require a defensible design and assumptions.

One-Sentence Answer

Structural equation modeling works by specifying a network of measurement and regression-like relationships, deriving the covariance and mean structure that network implies, and estimating parameters so the implied structure matches the observed data as closely as the chosen estimator and assumptions allow.

1. Why SEM Exists

Many important scientific quantities cannot be measured directly. “Reading comprehension”, “trust”, “mathematical anxiety”, “institutional capacity”, “household deprivation” and “service quality” are constructs. Researchers observe indicators that are intended to carry information about those constructs.

At the same time, researchers often want to model relationships among several quantities at once. Perhaps a latent learning strategy predicts a latent comprehension construct; comprehension predicts an exam outcome; and both are measured imperfectly by several observed indicators.

Running a separate regression for every arrow can ignore the measurement structure and create a fragmented analysis. SEM provides a common mathematical framework in which the measurement problem and the structural relationship problem can be estimated together.

2. Observed Variables and Latent Variables

An observed variable is recorded directly: a test response, age, income, reaction time, count or questionnaire item. A latent variable is not directly observed. It is represented through a model linking it to observed indicators.

In a simple factor model, three observed items might load on a latent factor. The factor is not a mysterious object hiding inside the data. It is a model-based construct. Its meaning depends on the indicators chosen, the measurement theory, the model constraints and the populations in which the measurement structure is supported.

This is why construct validity cannot be outsourced to software. A beautifully fitting latent variable with badly chosen indicators is still a badly defined construct.

3. The Measurement Model

The measurement model describes how observed indicators relate to latent constructs. Confirmatory factor analysis, or CFA, is one of the most common forms.

In CFA, the researcher usually specifies in advance which indicators load on which factors. That is a different scientific posture from exploratory factor analysis, where the structure is investigated with fewer a priori restrictions.

A loading tells us how strongly an indicator is related to a latent factor on the model’s scale. Residual variance represents indicator variation not explained by that factor. Cross-loadings, correlated residuals and multi-factor structures can be modeled, but every added flexibility changes the meaning and identifiability of the model.

4. The Structural Model

The structural portion specifies relationships among observed or latent variables. These relationships often resemble regression equations. A latent factor may predict another latent factor; an observed covariate may predict a latent outcome; several endogenous variables may be linked in sequence.

That makes SEM look like a causal diagram, but appearance is not identification. An arrow in a statistical model is a directional equation. Whether it has a causal interpretation depends on design, temporal ordering, confounding assumptions, measurement quality and the broader causal argument.

A cross-sectional SEM with excellent fit does not, by itself, establish that one latent construct caused another.

5. Covariance Is the Core Currency

At its heart, much of classical SEM asks whether a proposed model can reproduce the observed covariance matrix. The model implies relationships among variances and covariances. Estimation chooses parameters so the model-implied matrix is close to the observed one according to a fitting criterion.

Mean structures can also be modeled. This allows analyses of intercepts, latent means, growth curves and other questions where average levels matter in addition to covariance patterns.

This gives SEM its unifying character: factor models, path models, growth models and multiple-group models can all be expressed through constraints on moments and relationships.

6. Identification: Can the Model Be Solved?

Before asking whether parameter estimates are good, we must ask whether the model is identified. Identification concerns whether the observable information is sufficient to determine a unique set of model parameters, at least locally under the model.

A latent variable has no natural measurement unit. Its scale must be set, commonly by fixing one loading or fixing the latent variance. More complicated models can have underidentification because too many parameters are free relative to the information available, or because different parameter combinations imply the same observable covariance structure.

Software warnings about non-positive definite matrices, singular information, non-convergence or unidentified parameters are not cosmetic annoyances. They are signs that the model-data relationship may not support the intended estimation.

7. Degrees of Freedom and Model Constraint

The observed covariance matrix contains a finite number of unique elements. A model uses parameters to reproduce those elements. Roughly speaking, degrees of freedom reflect how much observed information remains after accounting for the number of free model parameters.

A saturated model reproduces the observed covariance structure exactly and therefore has no testable restrictions. A more constrained model makes stronger claims because it insists that certain relationships are zero, equal or otherwise structured.

That is a useful way to think about SEM: scientific content often enters through restrictions. A model that can reproduce anything may fit perfectly but say very little.

8. Estimation: How Parameters Are Chosen

For continuous data, maximum likelihood is a common estimator. Other approaches include generalized least squares, weighted least squares, diagonally weighted least squares and robust variants. The appropriate estimator depends on measurement level, distributional conditions, sample size, missingness and model structure.

The official lavaan documentation, for example, distinguishes multiple estimators and robust variants. For ordered endogenous variables, lavaan can switch to a weighted least squares approach designed for categorical indicators.

The practical lesson is that “SEM” is not one estimator. It is a modeling framework. Estimation is a second decision that must be aligned with the data.

9. What Does Model Fit Mean?

Model fit asks how closely the model-implied structure matches the observed data. A chi-square test can evaluate exact fit under its assumptions, while indices such as CFI, TLI, RMSEA and SRMR summarize different aspects of discrepancy.

No single fit index is a truth detector. Thresholds such as “CFI above X” or “RMSEA below Y” can be useful heuristics in some contexts, but they are not universal laws. Fit indices respond to sample size, model complexity, estimator, degrees of freedom and the kinds of misspecification present.

A strong analysis examines several sources of evidence: global fit, residuals, parameter plausibility, standard errors, theory, predictive or cross-sample performance where possible, and whether the model’s substantive conclusions survive reasonable alternatives.

10. Good Fit Does Not Mean the Model Is True

Different models can imply very similar covariance structures. A model can fit well because the sample is small, because the model is flexible, because the relevant misspecification is weakly detectable, or because multiple theoretical stories are observationally similar.

Conversely, a substantively useful model can fail an exact-fit test in a large sample because tiny deviations become detectable. The goal should not be to obtain a ceremonial collection of acceptable indices. The goal is to understand which aspects of the data the model captures, where it fails, and whether the remaining discrepancy matters for the intended inference.

11. Modification Indices: Useful Diagnostics, Dangerous Navigation

A modification index estimates how much model fit might improve if a fixed parameter were freed. It can identify where the proposed model conflicts with the observed covariance structure.

But if every large modification index becomes a new path, cross-loading or residual covariance, the analysis turns into sample-specific search. The model gradually learns the noise and idiosyncrasies of the current data while losing the theory that made it interpretable.

A good modification is theoretically defensible, transparently reported and ideally tested in new data. A modification made only because it improves fit is a hypothesis generated by the sample, not strong confirmation of an existing theory.

12. Confirmatory Does Not Mean Confirmed

“Confirmatory factor analysis” means the loading structure is specified in a confirmatory framework. It does not mean the resulting factor structure has been proven true.

A CFA can show that a proposed factor structure is compatible or incompatible with the observed data under the chosen model. Alternative structures may fit similarly. Indicators may carry local dependence. Factors may be poorly separated. The construct may behave differently across groups.

Confirmation in science accumulates across replications, populations, measurement studies, interventions and competing explanations. No single fit statistic can do that work alone.

13. Measurement Invariance: Can We Compare the Same Construct?

If a latent construct is compared across groups or time, the measurement model must be examined for invariance. Are indicators linked to the latent variable in comparable ways? Are intercepts or thresholds comparable? Are residual structures sufficiently stable for the intended comparison?

Without enough measurement invariance, an observed group difference in a latent mean may partly reflect a change in how the construct is measured rather than a genuine change in the underlying construct.

This is one reason SEM is so useful: it can make measurement comparability an explicit part of the model instead of quietly assuming that the same score means the same thing everywhere.

14. Mediation: A Decomposition, Not Automatic Causality

SEM is often used to estimate direct and indirect effects. A predictor X may relate to mediator M, which relates to outcome Y. The product of the relevant path coefficients can represent an indirect association under the model.

Calling that quantity a causal mediation effect requires much more: temporal ordering, assumptions about unmeasured confounding, appropriate treatment of exposure-mediator and mediator-outcome relationships, and a design that supports the causal interpretation.

A statistically significant indirect path in cross-sectional observational data is not enough to establish a causal mechanism.

15. Missing Data

SEM frequently encounters incomplete indicators. Some estimators can use likelihood-based methods under assumptions about the missing-data process; other workflows may rely on multiple imputation or estimator-specific approaches.

The critical point is conceptual: missingness assumptions are part of the inferential model. A sophisticated latent-variable model does not make missing data harmless. If missingness depends on unobserved values in ways not represented by the analysis, conclusions can still be biased.

16. Categorical and Ordinal Indicators

Questionnaire responses are often ordinal rather than continuous. Treating a four-category response as though the numerical distances 1–2, 2–3 and 3–4 are identical can be a strong assumption.

Modern SEM frameworks can model ordered indicators using threshold-based approaches and estimators designed for categorical data. The lavaan documentation, for example, describes diagonally weighted least squares and robust test statistics for ordered endogenous variables.

The measurement scale should therefore be part of the model design, not an afterthought applied once the diagram has been drawn.

17. Sample Size and Statistical Power

There is no universal “SEM needs 200 cases” rule. Required sample size depends on model complexity, effect sizes, loading strength, missingness, estimator, indicator distributions, desired precision and the particular hypothesis being tested.

A small simple model with strong indicators may be estimable with fewer observations than a large model containing weak factors, categorical outcomes and many freely estimated residual relationships. Simulation-based power analysis is often more informative than relying on a single rule of thumb.

18. Longitudinal and Growth Models

SEM can represent change through latent growth curves, autoregressive structures, cross-lagged models and other longitudinal formulations. A latent intercept can represent starting level; a latent slope can represent change over time.

Longitudinal models increase scientific leverage because they make temporal structure explicit, but time ordering alone does not eliminate confounding or measurement instability. Repeated measures also introduce dependence that must be modeled rather than ignored.

19. Multilevel SEM

Students sit inside classes, patients inside clinics, workers inside firms and repeated observations inside people. If the scientific structure is multilevel, SEM can be extended to separate within-group and between-group variation.

This matters because relationships at the individual level need not match relationships at the group level. A correlation observed among schools is not automatically the same relationship observed among students within schools.

Multilevel SEM combines the measurement flexibility of latent-variable models with the dependence structure of hierarchical data, at the cost of greater model complexity and stronger data requirements.

20. Standardised Coefficients: Helpful but Not Universal

Standardised path coefficients can help compare relationships measured on different scales, but they are not automatically more meaningful than unstandardised coefficients. Standardisation depends on sample variances, which can differ across populations and groups.

If a one-unit change has a clear substantive meaning—one year of age, one kilogram, one test item—the unstandardised coefficient may be easier to interpret. Good reporting often provides both where appropriate.

21. Common Failure Modes

• Drawing arrows first and constructing a theory afterwards.
• Assuming a latent factor is “real” merely because software estimates it.
• Treating good fit as proof of the model.
• Using fit-index cutoffs as mechanical pass/fail rules.
• Freeing paths until fit looks acceptable without independent validation.
• Interpreting observational paths causally without a causal design.
• Ignoring measurement invariance before comparing latent means.
• Using continuous-data estimators for strongly ordinal indicators without justification.
• Ignoring clustering, survey design or repeated observations.
• Reporting only standardised coefficients and hiding raw-scale meaning.

22. A Responsible SEM Workflow

• Define the constructs. Decide what each latent variable is supposed to mean.
• Defend the indicators. Measurement theory comes before the path diagram.
• Specify the model before looking for fit improvements. Separate confirmation from exploration.
• Check identification. The model must be estimable in principle.
• Choose an estimator matched to the data. Continuous, ordinal, missing and clustered data require different considerations.
• Inspect fit broadly. Use residuals, fit indices, parameter estimates and theoretical plausibility.
• Test measurement comparability. Especially for group and longitudinal comparisons.
• Perform sensitivity checks. Reasonable alternative models should not be hidden.
• Separate statistical direction from causal identification. An arrow is not a randomized intervention.
• Validate where possible. New samples, held-out data and replication protect against sample-specific modeling.

23. How This Connects Across the eduKate Library

• How Principal Component Analysis and Factor Analysis Work — the canonical route for exploratory dimensions, factors and loadings.
• How Regression Models Work — the regression foundation beneath many structural paths.
• How Causal Inference Works — the separate owner for causal identification, counterfactuals, confounding and treatment effects.
• How Measurement Invariance and Fair Comparisons Work — essential before comparing latent constructs across groups or time.
• How Construct Validity and Measurement Models Work — the broader owner for what a score or construct actually represents.
• How Missing Data Analysis Works — because incomplete indicators require explicit assumptions.
• How Statistical Power and Sample Size Planning Work — for study design before estimating a complex SEM.
• How Longitudinal and Panel Data Work — for repeated observations and change through time.

24. Authoritative Sources and Further Reading

• The lavaan project, Structural Equation Modeling tutorial overview, including CFA, SEM, growth, multiple groups, categorical data and estimation.
• The lavaan project, SEM example and CFA example.
• The lavaan project, Estimators, missing values, standard errors and bootstrapping.
• The lavaan project, Multiple-group models and measurement invariance.
• Rosseel, Y. (2012), lavaan: An R Package for Structural Equation Modeling, Journal of Statistical Software, 48(2), DOI 10.18637/jss.v048.i02.

Final Idea

SEM is most valuable when it makes hidden assumptions visible: what is being measured, which relationships are constrained, which errors are allowed, which groups are comparable, and what the data can actually identify. Its diagrams are useful maps, but maps are not the territory. The quality of the conclusion still depends on the quality of the measurement, the design and the reasoning that gave the arrows meaning.