Skip to content
Back to search results

Structural Equation Modeling

Introduced by
Sewall Wright
Year
1921
Type
Method
Family
multivariate-modeling
Font size

What does it do?

Structural equation modeling (SEM) is a comprehensive statistical framework combining path analysis (Sewall Wright, 1921) and confirmatory factor analysis to test complex causal models linking observed and latent variables.

Formalized by Jöreskog (1973) with LISREL software, SEM enables simultaneous estimation of measurement relationships (how variables measure latent constructs) and structural relationships (how constructs influence outcomes), making it powerful for theory testing in psychology, epidemiology, organizational research, and health sciences where complex mediation, moderation, and latent processes require integrated analysis.

The intuition behind the method

Many research questions involve complex causal chains: does a variable X affect outcome Y directly, or indirectly through mediator M? Do multiple observed variables measure a single latent construct? SEM integrates these questions into one model. The framework includes two parts: (1) measurement model—how observed variables relate to latent constructs (factor loadings), and (2) structural model—how constructs causally relate to each other (path coefficients). Unlike regression (which ignores measurement error) or factor analysis alone (which ignores causal relationships), SEM estimates structural paths while accounting for measurement error. This dual approach yields more valid inference about underlying processes and enables tests of indirect effects (mediation), moderation, and reciprocal relationships.

History

Sewall Wright developed path analysis (1921) to decompose correlations into direct and indirect causal effects in genetics. Karl Jöreskog unified path analysis with factor analysis to create SEM, releasing LISREL (Linear Structural Relations) software in 1973, enabling complex hypothesis testing. Bentler's EQS (1985) and Arbuckle's AMOS (1995) made SEM more accessible. Modern developments include robust estimation (Yuan & Bentler, 2000), multilevel SEM (Hox & Maas, 2001), dynamic structural models (latent growth curves; Bollen & Curran, 2006), and Bayesian SEM (Lee, 2007).

When to use it

Use SEM when testing complex multivariate theories involving mediation, moderation, latent constructs, and reciprocal causation. Psychology: testing whether cognitive coping (latent) mediates the effect of stress on depression, or how personality affects behavior through motivation. Epidemiology: modeling disease risk pathways (e.g., socioeconomic status → health behaviors → cardiovascular outcomes). Organizational research: testing whether leadership (measured via multiple items) affects performance through employee engagement (latent). Education: modeling effects of school climate (latent construct from teacher and student reports) on student achievement. SEM handles latent variables naturally, making it ideal when constructs are abstract or measured via multiple indicators with error.

Limitations and considerations

  • Requires large samples and balanced design: parameter estimation unstable with small n or extreme sample characteristics; minimum n ≥ 100–200 with multiple latent variables.
  • Assumes correct model specification a priori: SEM tests specified model; incorrect specification (omitted paths, wrong causal direction) may fit well despite being wrong (equivalent models).
  • Multiple equivalent models possible: different path structures can reproduce the same covariance matrix; SEM cannot distinguish among them without additional theory or constraints.
  • Non-experimental data do not imply causality: path coefficients represent regression weights, not causal effects; confounding, reverse causality, and selection bias still threaten validity.
  • Complex interactions and nonlinear relationships are difficult to specify and estimate in SEM; typically assumes linear relationships.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Integrated measurement and structural components: simultaneously estimates how variables measure constructs (validity) and how constructs relate (theory testing).
  • Accounts for measurement error: path coefficients corrected for unreliability in measurement, yielding more valid structural estimates than regression.
  • Tests complex relationships: mediation, moderation, indirect effects, and reciprocal causation in one coherent framework.
  • Flexible for missing data: full-information maximum likelihood enables analysis with incomplete data under MCAR assumption.
  • Comprehensive fit diagnostics: multiple indices (CFI, RMSEA, residuals) enable thorough model evaluation and identification of specific misfit sources.

Applications

  • Clinical psychology: testing whether trauma exposure affects depression through negative cognitive patterns (mediation); modeling latent dimensions of psychopathology.
  • Public health: evaluating whether social-ecological interventions affect health behaviors (e.g., community engagement → knowledge → behavior change → health outcomes).
  • Educational research: assessing whether school resource allocation affects student learning through teacher quality and school climate (latent construct).
  • Organizational behavior: testing whether leadership style affects organizational performance through employee motivation and commitment (latent mediators).
  • Longitudinal developmental research: latent growth curve models tracking how individuals change over time, predicting trajectories from baseline characteristics.

How it works

How it works

1

Administer SEM by: (1) specifying the model a priori based on theory (draw or diagram the hypothesized causal structure), (2) defining latent variables (abstract constructs you observe indirectly through measured variables) and their indicators, (3) specifying structural paths (causal hypotheses linking variables/constructs), (4) collecting data on all observed variables (typically continuous; SEM can handle ordinal/categorical with appropriate estimation), (5) ensuring adequate sample size (minimum n ≥ 100, ideally n ≥ 200; rule of thumb: n > 10× number of parameters estimated). Collect data with minimal missing values; SEM's full-information maximum likelihood (FIML) handles missing data under missing completely at random (MCAR) assumption. Define exogenous variables (causes, typically uncorrelated) and endogenous variables (effects, may be caused by other variables).

2

Estimate SEM using maximum likelihood (ML) or robust estimation methods (MLR for non-normal data). Output includes: (1) Standardized path coefficients (regression weights between variables): directly comparable across paths; coefficient of 0.4 indicates moderate effect. (2) Factor loadings: correlations between latent variables and their indicators (≥0.5 acceptable, ≥0.7 strong). (3) Standard errors and z-statistics: test if coefficients differ significantly from zero. (4) Indirect effects (mediation): product of path coefficients along indirect route(s); test via bootstrapping (≥1000 iterations). (5) R² for endogenous variables: proportion of variance explained. (6) Fit indices: CFI > 0.95, TLI > 0.95, RMSEA < 0.06, SRMR < 0.08 (Hu & Bentler, 1999). (7) Residuals: standardized values should be near zero; large residuals indicate model misfit at specific relationships. (8) Modification indices: suggest alternative paths; however, data-driven modifications compromise validity—only modify based on theory.

3

Interpret standardized path coefficients as: a coefficient of 0.5 for path X→Y means a 1-SD increase in X predicts a 0.5-SD increase in Y (direct effect), holding other variables constant. If coefficient p < 0.05, the effect is statistically significant. Indirect effects: multiply coefficients along mediating path. For example, if X→M coefficient is 0.3 and M→Y is 0.4, indirect effect = 0.3×0.4 = 0.12; bootstrap 95% CI should exclude zero for significance. Total effect = direct + indirect; compare to zero to assess overall association. Model fit: meeting all fit criteria (CFI, RMSEA, etc.) indicates the model adequately reproduces observed covariance structure. Residuals near zero confirm fit at individual pathways. Modification indices > 3.84 suggest paths that might improve fit, but adding ad-hoc paths inflates Type I error; only respecify models based on theoretical justification or cross-validate on independent sample.

Common pitfalls

  • Data-driven model modification: iteratively adding paths based on modification indices capitalizes on chance and produces models that may not replicate. Modifications must be theoretically justified a priori or tested on independent sample.
  • Ignoring measurement model: focusing only on structural paths while allowing measurement error to be estimated freely; inadequate measurement undermines structural validity.
  • Reporting model fit without examining residuals: good global fit indices do not guarantee all pathways fit well. Examine standardized residuals and modification indices to identify localized misfit.
  • Treating SEM as confirmatory when used exploratorily: if model was data-derived (post-hoc specification), report as exploratory and validate on independent sample.
  • Confusing statistical significance with practical significance: significant path (p < 0.05) does not mean the effect is large or meaningful. Report effect sizes (standardized coefficients) and confidence intervals.
  • Oversaturating model with latent variables: too many latent constructs relative to indicators and sample size produces unstable estimates; use multigroup SEM or bifactor models cautiously.

Frequently asked questions

What is the difference between a direct effect and an indirect effect (mediation)?

A direct effect is the path from X to Y, quantifying how X influences Y directly. An indirect effect is the path X→M→Y, quantifying how X influences Y through mediator M. The indirect effect equals the product of the X→M path times the M→Y path. Total effect = direct + indirect. Mediation occurs when: (1) X→Y direct effect decreases after controlling for M, and (2) X→M→Y indirect effect is significant (bootstrap CI excludes zero). Full mediation: direct effect becomes nonsignificant after including mediator. Partial mediation: both direct and indirect effects significant, meaning X affects Y through multiple pathways.

How do I know if my model fits well? What do CFI and RMSEA mean?

Model fit is judged by multiple indices (Hu & Bentler, 1999 criteria): CFI (Comparative Fit Index) compares your model to a null model (all variables uncorrelated); CFI > 0.95 indicates good fit (ranges 0–1). RMSEA (Root Mean Square Error of Approximation) quantifies discrepancy between model and data per degree of freedom; RMSEA < 0.06 excellent, 0.06–0.08 good, >0.10 poor (close to zero better). TLI > 0.95 and SRMR < 0.08 also recommended. Meeting all criteria simultaneously indicates good fit. However, good overall fit does not guarantee all pathways fit; examine standardized residuals (should be < |2|) and modification indices.

What is a latent variable, and how do I measure it?

A latent variable is an unobserved construct (e.g., depression, intelligence, organizational culture) measured indirectly through multiple observed indicators (questionnaire items, test scores). Each indicator reflects the latent construct plus measurement error. In SEM, you specify which observed variables measure which latent variables via factor loadings. For example, the latent variable 'depression' might be measured by items: sadness, hopelessness, sleep disturbance (observed indicators). Each loading should be ≥0.5; if an indicator loads weakly, consider removing it or reconceptualizing the latent construct.

Can I use SEM to establish causality from observational data?

No. SEM estimates associations and tests whether a hypothesized causal model reproduces the correlation structure in data. Path coefficients represent regression weights (conditional associations), not causal effects. Causality requires: (1) temporal precedence (X measured before Y), (2) covariation (X and Y correlate), (3) no plausible alternatives (confounding, reverse causality). Observational SEM violates criterion 3: confounding variables and unmeasured causes threaten validity. SEM's statistical fit alone cannot prove causality. Use SEM with theory, qualitative evidence, and study design (randomization, instrumental variables) to strengthen causal inference. Consider sensitivity analyses testing alternative models.

Related methods

Sources

  1. Jöreskog, K. G., & Sörbom, D. (1973). LISREL: A general computer program for estimating a linear structural equation system. Research Bulletin 73-5. University of Stockholm.

    Open publication reference
    Copied.
  2. Hu, L. T., & Bentler, P. M. (1999). Cutoff criteria for fit indices in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling, 6(1), 1–55.

    Open DOI reference
    Copied.
  3. Wright, S. (1921). Correlation and causation. Journal of Agricultural Research, 20(7), 557–585.

    Open publication reference
    Copied.