Structural Equation Modeling (SEM) Explained 2026: Path Models, CFA, and Fit Indices

Your committee has asked why you chose structural equation modeling over multiple regression, and you need an answer that goes beyond “the literature used it.” Structural equation modeling is a multivariate statistical framework that tests complex theoretical relationships between observed and latent variables simultaneously — capabilities no single regression model can replicate. Yet the jargon surrounding fit indices, latent constructs, and two-stage model building trips up even experienced researchers working through their first SEM analysis.

This reference-grade guide unpacks structural equation modeling from first principles: what separates the measurement model from the structural model, how confirmatory factor analysis (CFA) establishes your latent variables, which fit statistics your reviewers will scrutinise (CFI, RMSEA, SRMR), and how to implement the full workflow in R’s lavaan package or IBM AMOS.

What Is Structural Equation Modeling?

Structural equation modeling is a statistical technique that simultaneously estimates a network of hypothesised relationships among variables. Unlike ordinary least-squares regression, which handles one dependent variable at a time and assumes all variables are directly measured without error, SEM can model latent constructs — unobservable theoretical concepts such as “academic self-efficacy,” “research anxiety,” or “organisational commitment” — as well as the directional paths connecting them.

SEM converges two older techniques into a single, more powerful framework:

• Factor analysis — which models how measured indicators (survey items, scale scores, test results) reflect underlying constructs
• Path analysis — which tests directional relationships among variables using a system of regression equations

Because SEM handles both simultaneously, it produces more realistic estimates by accounting for measurement error in each indicator rather than treating every variable as if it were perfectly observed. This error correction is the primary reason SEM path estimates are less biased than observed-variable regression when your theoretical constructs contain inherent measurement noise.

SEM is particularly common in education research, psychology, management, and health sciences — fields where the constructs of greatest theoretical interest cannot be directly observed. Researchers handling non-latent causal chains may find the companion guide on Directed Acyclic Graphs (DAGs) for causal inference equally useful for mapping the underlying logic of variable relationships before specifying a model.

The Measurement Model: CFA and Latent Variables

Every full SEM analysis begins with the measurement model, tested through confirmatory factor analysis (CFA). The purpose is to verify that your observed indicators adequately represent the latent constructs you have theorised — before examining any relationships between those constructs. Running a full SEM without validating the measurement model first is the single most common structural error in quantitative dissertations.

What Is CFA?

Confirmatory factor analysis differs from exploratory factor analysis (EFA) in a fundamental way: you specify the factor structure in advance based on theory, then test whether the data confirm it. In lavaan, factor loading relationships are specified with the =~ operator:

model <- '
  # Latent variable definitions (measurement model)
  SelfEfficacy    =~ se1 + se2 + se3 + se4
  ResearchAnxiety =~ ra1 + ra2 + ra3
'
fit <- cfa(model, data = mydata)
summary(fit, fit.measures = TRUE)

Each observed indicator (se1, se2, and so on) loads onto its assigned latent factor. CFA estimates standardised factor loadings (expected between 0.40 and 0.90 for well-constructed scales), intercepts, and residual variances for each indicator. Items with standardised loadings below 0.40 contribute little to defining the construct and may warrant revision or removal before proceeding to the structural model.

Latent Variables: The Core Advantage

Latent variables are inferred from covariance patterns among their indicators. Because the model explicitly estimates measurement error for each observed item, the resulting latent variable is a purer representation of the theoretical construct than a simple composite score or mean. This is the mechanism through which SEM produces less-biased path estimates than multiple regression when your predictors and outcomes involve psychological, social, or educational constructs.

Assessing Convergent and Discriminant Validity

After fitting a CFA model, two forms of construct validity must be established before the structural model is defensible:

Measurement consistency across raters or instruments is a related concern: the guide to inter-rater reliability covering ICC, Krippendorff’s Alpha, and Bland-Altman provides complementary grounding if your indicators involve coded or observer-rated data rather than self-report scales.

The Structural Model: Path Diagrams and Relationships

Once the measurement model passes CFA scrutiny, you add the structural model — specifying directional regression paths between the validated latent constructs. This layer is what directly tests your theoretical hypotheses.

In a path diagram, latent variables appear as ovals, observed variables as rectangles, single-headed arrows represent regression paths, and double-headed curved arrows denote correlations. A complete SEM merges both layers in one model specification:

model <- '
  # Measurement model
  SelfEfficacy    =~ se1 + se2 + se3 + se4
  ResearchAnxiety =~ ra1 + ra2 + ra3
  Performance     =~ p1  + p2  + p3

  # Structural model (regression paths among latents)
  Performance     ~ SelfEfficacy + ResearchAnxiety
  ResearchAnxiety ~ SelfEfficacy
'
fit <- sem(model, data = mydata)
summary(fit, fit.measures = TRUE, standardized = TRUE)

Here, SelfEfficacy predicts both ResearchAnxiety and Performance directly, while ResearchAnxiety also predicts Performance. SEM therefore estimates both the direct effect of SelfEfficacy on Performance and the indirect effect routed through ResearchAnxiety — simultaneously, and with measurement-error corrections that simple Baron-Kenny mediation ignores.

The two-step modelling approach (Anderson & Gerbing, 1988) is broadly recommended: establish and confirm the CFA first, then add structural paths. This sequencing prevents structural misspecification from contaminating the measurement parameter estimates that underpin your latent constructs.

How to Assess Model Fit: CFI, RMSEA, and SRMR

Model fit in SEM evaluates how closely the model-implied covariance matrix matches the observed sample covariance matrix. No single index captures all aspects of fit, which is why reporting at least three complementary indices is now standard across journals in psychology, management, and education.

The Chi-Square Problem

The model chi-square test remains standard output in every SEM software package, but it is almost never the primary criterion for evaluating fit. Chi-square is extraordinarily sensitive to sample size: with N above roughly 400, even trivially misspecified models tend to produce a statistically significant result. Always report chi-square and its degrees of freedom for transparency, then base substantive fit judgements on CFI, RMSEA, and SRMR.

Critical Caveats on Cutoff Values

The Hu-Bentler cutoffs were derived from simulations using continuous, normally distributed indicators with N ≥ 250 and small-to-moderate models. They are guidelines, not universal laws. Three adjustments are commonly required:

• Small samples (N < 200): RMSEA is artificially inflated at low N. Always report the 90% confidence interval alongside the point estimate so reviewers can assess whether apparent poor fit reflects genuine misspecification or sampling uncertainty.
• Large models (many indicators): CFI naturally drifts downward as model complexity increases. A CFI of 0.92 in a 50-indicator model warrants different interpretation than the same value in a 12-indicator model.
• Categorical or ordinal data: Request scaled versions in lavaan (cfi.scaled, rmsea.scaled) via the WLSMV estimator. Standard ML-based cutoffs do not apply directly to ordinal data. See the Cornell Statistical Consulting Unit’s SEM fit index reference for a detailed breakdown of index behaviour under varied conditions.

SEM Software: lavaan vs AMOS

Two software ecosystems dominate SEM practice in English-speaking academic research in 2026:

The UCLA Statistical Consulting Group’s lavaan tutorial is a freely available, peer-vetted resource recommended by many university graduate programmes. AMOS retains a strong foothold in management and organisational behaviour journals where SPSS familiarity predominates, but lavaan’s transparent syntax makes it the stronger choice for dissertations where methodological reproducibility is a committee or examiners’ concern.

The software and statistical approach you choose for SEM is one decision within a broader methodological strategy. If you are still working through the foundational choice between quantitative and qualitative paradigms, the guide to qualitative vs quantitative research design covers those trade-offs systematically.

Four Common SEM Mistakes to Avoid

1. Skipping the CFA stage. Running the full SEM without first confirming that the measurement model fits is the most common structural error. A misspecified measurement model corrupts all downstream structural parameter estimates. Validate CFA independently, review loadings, AVE, and CR, then add structural paths.
2. Reporting chi-square as the sole fit criterion. Reviewers in 2026 expect a minimum of CFI, RMSEA (with 90% CI), and SRMR. Any submission presenting chi-square alone will draw a request for reanalysis at first revision.
3. Using modification indices without theoretical justification. lavaan’s modindices() function will always surface parameter tweaks that improve chi-square. Adding them without theoretical warrant inflates fit statistics through capitalisation on chance — equivalent to HARKing (hypothesising after results are known) and reviewable as post-hoc model fishing.
4. Under-identified models. A model with more free parameters than known data points has negative degrees of freedom and cannot be estimated. The identification rule: df = [p(p+1)/2] − free parameters, where p is the number of observed variables. Models must be over-identified (df > 0) to permit fit evaluation. Check identification before running estimation.

Frequently Asked Questions