Does SEM always use latent variables?

No. Path analysis with observed variables is also part of the broader framework.

Can SEM prove causation from a survey?

Model fitting alone cannot. Causal interpretation depends on design, assumptions, timing, and alternative explanations.

Is there one minimum sample size?

No universal number suits every model. Requirements depend on complexity, measurement quality, estimator, effect sizes, and the analysis objective.

AI Glossary / AI and machine learning / Intermediate

Structural equation modeling

Definition

Structural equation modeling, or SEM, is a statistical framework for representing and estimating relationships among variables through linked equations. A model can combine directly observed measurements with latent variables, which represent concepts inferred from multiple indicators rather than measured directly.

What it means

An ordinary regression describes one outcome using predictors. SEM can connect several such relationships and, when appropriate, include a measurement model for concepts such as satisfaction or perceived usability.

A path diagram is a convenient representation: boxes commonly indicate observed variables, circles indicate latent variables, and arrows express specified relationships. The diagram is a model proposal, not evidence that every arrow reflects a genuine cause.

SEM includes different estimation traditions. This entry focuses on the general framework and familiar covariance-based models. Analysts must specify which approach they use because assumptions, diagnostics, and interpretations can differ.

How it works

Begin with substantive reasoning and define the variables and proposed paths. If latent constructs are included, specify how their observed indicators relate to them. Then determine whether the model is identified: whether its parameters can, in principle, be uniquely estimated under the chosen assumptions.

Fit the model using an estimator appropriate to the data and check convergence, plausible estimates, and uncertainty. In covariance-based SEM, a central task is comparing model-implied relationships with observed relationships.

Inspect several diagnostics rather than treating one fit statistic as a verdict. A model can fit reasonably while omitting an important alternative explanation. Poor measurements, missing data, unsuitable distributions, or unmodeled dependencies can also affect conclusions.

Finally, distinguish statistical associations from causal interpretations. Causal claims require defensible assumptions and study design beyond a successful model fit.

Three practical examples

1. Customer research: Analysts model satisfaction using several survey questions and examine its association with reported intention to return.

2. Product usability: Researchers examine relationships among perceived effort, perceived usefulness, and adoption intention while accounting for their measurement indicators.

3. Training evaluation: A team models links among course participation, measured knowledge, and later task performance, while considering selection effects and alternative explanations.

Think of it this way

Think of a diagram of connected water tanks, where some levels are measured directly and others are inferred from gauges. Equations describe how the quantities are expected to relate.

The analogy is limited because a statistical path diagram does not prove a physical flow. Several different systems may explain similar observations, and a latent variable is not automatically a directly existing object simply because a model represents it.

Worked example

Consider a simple linear path model with observed, centered variables X, M, and Y. Suppose the fitted equations are M = 0.5X + error and Y = 0.2X + 0.6M + error.

The model's indirect path from X through M to Y has coefficient 0.5 × 0.6 = 0.3. The direct path has coefficient 0.2, giving a total modeled coefficient of 0.5 under this simple specification.

For a one-unit change in X, the model therefore associates a 0.5-unit change in Y through the two specified paths together.

This arithmetic does not establish mediation as a causal fact. Confounding, reverse direction, measurement error, or a different functional form can invalidate that interpretation. Uncertainty for the indirect effect also needs estimation; multiplying two point estimates does not supply a confidence interval.

Illustrative case study

Fictional case: Juniper Learning surveys employees after a new course and measures later task performance. The research team proposes that course engagement relates to confidence, which in turn relates to performance.

An initial model fits the survey relationships acceptably. Interviews reveal that prior experience could influence all three variables. The analysts revise their study plan and make the limitation explicit rather than presenting the original diagram as a proven causal chain.

They also check whether confidence questions measure a coherent construct. The fictional example illustrates model criticism and measurement work, not a demonstrated return on training investment.

Common misconceptions and mistakes

A good fit does not prove the model is true. Competing models may fit the same data, and large or small samples can complicate diagnostics.

A latent variable is not automatically valid because several questionnaire items load on it. Its meaning and measurement quality require evidence.

Adding paths solely to improve fit can overfit a sample and weaken the original research question.

Do not infer causation from arrow direction alone. Likewise, “structural” in the name does not make every fitted relationship a stable mechanism that will survive intervention.

Business use

SEM is useful when a question involves linked relationships or explicit measurement of concepts. Its value lies in making assumptions testable and measurement choices visible.

Before commissioning a model, define the decision it should inform, the constructs, the study design, and the limits of the data. Allocate time for questionnaire quality, missing-data treatment, and alternative specifications. Report uncertainty and substantive plausibility alongside fit indices. If a simpler regression answers the question adequately, added model complexity may not improve the decision.

Check your understanding

A model estimates a direct coefficient of 0.1 and an indirect path with coefficients 0.4 and 0.5. What is the total modeled coefficient in this simple linear setup?

Answer: The indirect coefficient is 0.4 × 0.5 = 0.2, so the total is 0.3. This arithmetic still does not establish a causal effect.

Related terms

Latent variable
Path analysis (statistics)
Mediation (statistics)
Covariance matrix
Measurement error
Causal inference
path analysis
confirmatory factor analysis
mediation
model identification

Learn next

Learn confirmatory factor analysis to separate measurement questions from relationships among constructs. Then study identification and compare two plausible path diagrams for the same research question. Practice explaining what additional evidence would be needed to interpret a path causally.

Factor analysis

Sources

lavaan authors, A SEM example: https://lavaan.ugent.be/tutorial/sem.html, and Mediation: https://lavaan.ugent.be/tutorial/mediation.html. Official measurement/structural model specification and indirect/total coefficient formulas independently consulted. Original arithmetic checked: 0.5 × 0.6 + 0.2 = 0.5; quiz 0.4 × 0.5 + 0.1 = 0.3. Coefficients are illustrative; causal and uncertainty limitations remain explicit. The training case is fictional.

Previous: Structural EM Next: Structural metadata