Structural Equation Modelling (SEM)

SEM combines factor analysis and regression into one system: the measurement model checks that questionnaire items represent their latent constructs, and the structural model tests the hypothesised paths between constructs.

Two families exist — covariance-based SEM (AMOS/LISREL, for theory confirmation) and variance-based PLS-SEM (SmartPLS, for prediction and complex models). Choosing between them is a design decision, not a software preference.

An SEM analysis is reported in two stages. The measurement model comes first: indicator loadings, construct reliability (Cronbach's alpha and composite reliability), convergent validity through average variance extracted, and discriminant validity through the Fornell–Larcker criterion or the HTMT ratio. Only once the measurement model is acceptable does the structural model mean anything — path coefficients, their significance, R² for the endogenous constructs and, in covariance-based SEM, global fit indices such as CFI, TLI, RMSEA and SRMR.

Sample size is the most common practical constraint. Covariance-based SEM generally needs 200 or more respondents for stable estimates in a moderately complex model, while PLS-SEM tolerates smaller samples but is not a licence to collect too little data — a power analysis based on the construct with the most incoming paths is the defensible way to justify your target. Examiners will also expect you to explain why each path in the model exists, so every arrow in your diagram should trace back to a hypothesis grounded in theory.