2. SPSS AMOS - Introduction to Structural Equation Modelling (SEM) and Its Concepts - Research Coach

Structural equation modeling (SEM)—often implemented in IBM SPSS AMOS—is a statistical framework for testing relationships among multiple variables at once, while explicitly modeling measurement quality and error. Instead of running separate regressions for each outcome, SEM estimates a whole system of relationships simultaneously, letting researchers specify independent variables, mediators, moderators, and dependent variables in one model. That matters because many constructs in social science (like job satisfaction or honesty) are not measured directly; SEM treats them as latent constructs measured through questionnaire items, then evaluates whether those items reliably and validly represent the constructs.

SEM is described as a family of related techniques rather than a single procedure. It combines ideas from regression and factor analysis: regression-like modeling handles directional relationships (e.g., whether an independent variable influences a dependent variable), while factor-analytic logic supports the measurement side. A key feature is the use of a covariance matrix as input, which allows SEM to assess how variables relate to each other based on observed patterns of covariation. The approach is frequently confirmatory: researchers begin with a theory-driven model, draw a schematic diagram of hypothesized links, and then test whether observed data provide evidence for the proposed directions and strengths.

The method’s core architecture separates two layers. The measurement model checks reliability and validity of the indicators for each latent construct—essentially whether the questionnaire items accurately capture the underlying concept. The structural model then tests relationships among constructs, including the significance and size of effects between constructs (e.g., from an independent construct to a dependent construct). In SEM diagrams, latent constructs are typically shown as ellipses, observed indicators as rectangles, and error terms as part of the model—making measurement error an explicit part of estimation rather than something ignored.

A common misconception addressed is causation. SEM is often called “causal modeling,” but high correlation in a covariance-based framework does not automatically prove that one variable causes another. Establishing causation generally requires experimental design—changing the putative cause and observing changes in the outcome. Still, SEM can analyze experimental data when it exists, so the method is compatible with causal inference setups, even if it doesn’t magically create causality from correlation.

The transcript also clarifies terminology: a variable is directly measured (age, income, exam score), while a construct is measured indirectly through multiple items (e.g., job satisfaction measured via several questionnaire statements). SEM is positioned as especially useful for latent constructs, which ordinary least squares regression in SPSS cannot handle in the same way.

Finally, guidance is offered on sample size, emphasizing that requirements depend on model complexity and measurement characteristics. Suggested rules of thumb include: about 100 cases for up to five latent constructs with more than three items each; about 150 for up to seven latent constructs with more than three items; around 300 for just-identified models when some constructs have fewer than three items; and roughly 500 when there are more than seven latent constructs and some have fewer than three items. The overall takeaway is that SEM—via AMOS—lets researchers test an entire theory-based model, accounting for measurement error and multiple interlocking relationships in one estimation step.