Structural Equation Modeling (SEM)

• Structural Equation Modeling (SEM) is a unified statistical framework that combines measurement models and structural models to analyze interrelated latent and observed variables.
• It employs techniques like regression, confirmatory factor analysis, and path analysis to handle measurement error and ensure construct validity.
• Recent advancements in SEM include Bayesian methods, nonlinear modeling, and robust fit indices, enhancing its application in complex research domains.

Structural Equation Model (SEM) is a statistical framework for the specification, estimation, and testing of models that represent complex systems of relationships among observed (manifest) and unobserved (latent) variables. SEM unifies regression analysis, path analysis, confirmatory factor analysis, and models for mediation and measurement error under a single system of (typically linear) equations. The approach integrates measurement models for latent constructs with structural models for causal or associative pathways, making it central to social and behavioral sciences, psychology, economics, political science, as well as a rapidly expanding set of computational fields (Zheng, 30 Mar 2025).

1. Mathematical Structure: Measurement and Structural Models

At its core, SEM decomposes into two submodels:

• Measurement model (e.g., CFA):

$y = \Lambda_y \, \eta + \varepsilon$

where $y$ is a $p \times 1$ vector of observed indicators, $\Lambda_y$ a $p \times m$ matrix of factor loadings, $\eta$ the $m \times 1$ latent factor vector, and $\varepsilon$ the measurement error ($\mathrm{Cov}(\varepsilon) = \Theta_\varepsilon$).

• Structural model (path analysis):

$\eta = B\,\eta + \Gamma\,\xi + \zeta$

where $B$ is an $m \times m$ matrix of coefficient paths among endogenous factors, $\Gamma$ models effects from exogenous variables $\xi \in \mathbb{R}^q$, and $\zeta$ is the structural disturbance ($\mathrm{Cov}(\zeta)=\Psi$).

The implied covariance for the vector of observed variables is matched to the sample covariance matrix $S$. Estimation proceeds by finding parameter values $\theta$ that minimize the discrepancy between $S$ and the model-implied covariance $\Sigma(\theta)$, typically under a likelihood or least-squares criterion (Zheng, 30 Mar 2025).

2. Model Specification, Identification, and Construct Typology

SEM requires rigorous model specification and identification:

• t-rule (identification): number of distinct sample moments $\left(\frac{1}{2}p(p+1)\right) \geq$ number of free parameters.
• Scaling: Each latent factor is scaled by fixing a factor loading or variance.
• Specification: Over- or under-specification can cause non-convergence or unstable/inadmissible solutions.

Construct type is critical. Recent simulation evidence highlights strong estimator bias if the true construct type (reflective latent variables, formative/causal indicators, or composites) does not match the analyst's specification, and that fit indices (RMSEA, CFI, TLI) lack specificity for this kind of misspecification (Bauer et al., 29 Jul 2025). Reflective models require indicators to covary and be interchangeable; formative constructs define the domain without a requirement for inter-correlation; composites are researcher-specified indices.

Table 1. Construct Types and Measurement Equations (from (Bauer et al., 29 Jul 2025))

Construct Type	Measurement Model	Substantive Role
Reflective latent	$x = \Lambda\,\xi + \delta$	Traits cause indicators
Formative/causal-formative	$\xi = \Gamma^\top x + \zeta$	Indicators define the construct
Composite	$y = B^\top x + \varepsilon$	Composite as a weighted sum/index

Misspecification can induce |bias| of $0.14$–$0.18$ and high type I error in standard model fit tests (Bauer et al., 29 Jul 2025).

3. Estimation Methods and Fit Assessment

The dominant estimation approach is normal-theory maximum likelihood (ML):

$$F_{\min} = \log|\Sigma(\hat\theta)| + \operatorname{tr}(S\Sigma(\hat\theta)^{-1}) - \log|S| - p$$

with test statistic

$\chi^2 = (N - 1)\,F_{\min}$

and degrees of freedom $\mathrm{df} = \frac{1}{2}p(p+1) - t$, where $t$ is the number of free parameters.

Alternative estimators:

• Structured Least Squares (GLS, WLS): weighted to account for non-normality and categorical indicators, e.g., WLSMV.
• Satorra–Bentler scaled ML (MLR): adjusts $\chi^2$ for non-normality.
• Reweighted Least Squares (RLS/RGLS): improved small-sample stability (Zheng, 30 Mar 2025).

Fit indices include RMSEA, CFI, TLI, and SRMR. However, these are sensitive to sample size, model complexity, and misspecification; index cutoffs (e.g., RMSEA < .06, CFI > .95) should not be applied rigidly (Zheng, 30 Mar 2025). Information criteria (BIC, AICc) can also be used for model comparison, with BIC generally recommended for larger models or sample sizes (Hertzog, 2018).

4. Extensions: Composites, Nonlinearity, Bayesian and Regularized SEM

Modern SEM accommodates both latent variables and composites within a unified covariance structure (Schamberger et al., 8 Aug 2025):

• Unified variance–covariance for mixed models:

$$\Sigma(\theta) = \Lambda\,\operatorname{Var}(\eta)\,\Lambda' + \Theta, \quad \operatorname{Var}(\eta) = (\mathbf{I} - \mathbf{B})^{-1}\,\Psi\,(\mathbf{I} - \mathbf{B})^{-T}$$

Emergent directions include:

• Non-normal/nonlinear SEM: Case-based methods (CLSSEM) allow estimation directly from raw data, handling nonlinear, piecewise, or non-smooth models and user-defined constraints (Oldenburg, 2021).
• Bayesian and mixture models: Mean-field variational Bayes for Gaussian-mixture SEM enables scalable inference for non-Gaussian indicators (skewness/multimodality), yielding fast, closed-form updates and model selection via variational information criteria (Dang et al., 2024). Bayesian approaches allow for modeling latent heteroscedasticity with regression structures on latent means and variances (Fazio et al., 2024).
• Computation graph and regularized SEM: Graph-based implementations (e.g., tensorsem) facilitate automatic differentiation, regularization (LASSO/ridge/spike-and-slab), and robust fit criteria (LAD) (Kesteren et al., 2019).
• Sparse and convex SEM: Convex relaxations allow high-dimensional sparse model estimation with global convergence, enabling inference in large path models and complex networks (climate, neuroimaging) (Pruttiakaravanich et al., 2018).

5. Causal Inference, Mediation, and Longitudinal Structure

SEM identifies direct, indirect (mediation), and total effects as algebraic functions of structural parameters. In mediation:

$$\text{Indirect effect} = (\Gamma_{x\to\eta})(B_{\eta\to y}), \quad \text{Total} = \text{Direct} + \text{Indirect}$$

Contemporary practice extends classic Baron-Kenny approaches with causal mediation analysis (CMA), embedding sequential ignorability and permitting nonparametric link functions (Zheng, 30 Mar 2025). For group comparisons and experimental designs, multi-group CFA (MG-CFA) is deployed to verify measurement invariance.

For longitudinal/panel data:

• CLPM/RI-CLPM: Cross-lagged panel models, with random intercept decomposition, separate between-person from within-person dynamics, addressing biases in time-series mediation.

Advances in targeted learning (TMLE) provide a doubly-robust, machine-learning-based alternative, outperforming SEM when model misspecification is likely (Ma et al., 2 Nov 2025).

6. Practical Recommendations and Empirical Applications

Best practices, as synthesized in recent reviews (Zheng, 30 Mar 2025, Bauer et al., 29 Jul 2025):

• Always report $\chi^2$, degrees of freedom, and p-values with multiple fit indices.
• Avoid "cherry-picking" fit indices; interpret in light of sample size and theoretical context.
• Perform formal identification checks (t-rule, constraint matrices).
• Select estimators based on data properties (MLR for non-normal, WLSMV for categorical, RLS for small $N$).
• Use LM and Wald tests for specification, with cross-validation or resampling of modifications.
• Implement advanced causal and invariance testing where warranted.
• Explicitly assess construct type for each block and compare rival models if theoretical ambiguity exists.

Empirical SEM applications include consumer behavior analysis (Rao, 3 Feb 2026), latent-growth and mixture modeling in opinion dynamics, and input–output analysis in health/environmental systems using non-negative matrix factorization with structural equations (NMF-SEM) (Satoh, 20 Dec 2025).

7. Limitations, Pitfalls, and Frontiers

Critical challenges and limitations remain:

• Fit indices have limited power to detect deep model misspecification, especially regarding construct type (Bauer et al., 29 Jul 2025).
• Linear Gaussian frameworks are sensitive to nonlinearity, latent heteroscedasticity, and omitted interactions.
• Misspecification can induce severe bias not rescued by traditional fit assessment (Zheng, 30 Mar 2025).
• Correct model selection, especially in high-dimensional or compositional domains, requires theory-driven specification, regularization, and comparative modeling across frameworks.

Open areas for development include formal diagnostics for construct mis-specification, extensions to hybrid and higher-order constructs, robust or Bayesian estimators for mixed construct types, and educational resources to address persistent training gaps in applied fields (Zheng, 30 Mar 2025, Bauer et al., 29 Jul 2025).