20. SEMinR Series. Evaluating Structural Model | Step 1 | Collinearity Diagnostics

After finishing reliability/validity checks for the measurement (outer) models, the next priority in PLS-SEM is to evaluate the structural (inner) model—specifically whether relationships between constructs are trustworthy and meaningful. The core risk at this stage is multicollinearity: structural path coefficients come from ordinary least squares regressions of each endogenous construct on its predictor constructs, so strong correlations among predictors can bias those path estimates. That’s why collinearity diagnostics come first, before judging whether any hypothesized links are significant.

The structural model assessment is laid out as a step-by-step workflow. Step one focuses on multicollinearity among predictor constructs. If predictors are highly correlated, point estimates and standard error estimates for path coefficients can become distorted, making results unreliable. The standard diagnostic is the Variance Inflation Factor (VIF). As a rule of thumb, VIF values between 3 and 5 indicate no critical multicollinearity concerns, while values above 5 suggest a collinearity problem. When collinearity is problematic, the guidance is to consider creating higher-order constructs to reduce redundancy among predictors.

In practice, the collinearity check is performed by running a summary object for the estimated model and then extracting VIF values for the antecedents (predictor constructs). In the example shown, all VIF values fall below the threshold—“less than three” across the antecedents—so multicollinearity is unlikely to be a critical issue. With that hurdle cleared, the analysis can move forward to the next stage.

Step two then evaluates the structural relationships themselves: whether the path coefficients linking constructs are statistically significant and substantively relevant. Only after establishing that the measurement model is reliable and valid—and after confirming that multicollinearity won’t bias the structural estimates—does it make sense to interpret which construct relationships hold.

The workflow also includes later steps for assessing explanatory and predictive power and for model comparisons, but those are deferred until after moderation and mediation are discussed in subsequent sessions. For now, the takeaway is clear: structural model interpretation in PLS-SEM starts with collinearity diagnostics (via VIF), because biased path coefficients can undermine every downstream conclusion about significance, relevance, and model performance.