5. Complex Higher Order Construct/Second Order Analysis with 3 Hierarchical Models (See Description)

A complex Smart PLS model can be built and tested when constructs mix reflective–formative measurement, higher-order constructs, and moderation—by using a disjoint two-stage workflow that first validates lower-order components and then validates the higher-order constructs using latent variable scores. The key move is treating each lower-order dimension as an “indicator” for its higher-order construct in stage two, while keeping the measurement type consistent (reflective–reflective for some higher-order constructs, reflective–formative for the mediator).

The session starts with a fictitious but detailed setup: an independent variable (IV) measured as reflective–reflective higher order, a mediator measured as reflective–formative higher order, and a moderating variable measured as reflective–reflective higher order. The dependent side is represented by multiple lower-order constructs (DV1–DV4), which receive paths from the mediator and from the interaction created by moderation. The moderating variable is designed to moderate the mediator-to-dependent relationships (e.g., between the mediating variable and DV1, DV2, and so on), creating moderation effects that Smart PLS must estimate in the structural model.

To handle higher-order constructs, the workflow uses the disjoint two-stage approach. In stage one, only the lower-order constructs are modeled—without the higher-order construct nodes in the path model. Each lower-order construct is linked directly to the other theoretically related constructs. This stage focuses on reliability and validity: the presenter runs the PLS algorithm, checks discriminant validity using HTMT, and flags a problem between DV3 and DV4 when HTMT exceeds 0.90. The fix is to investigate cross-loadings and/or outer loadings (even if outer loadings look acceptable at first glance), because discriminant validity issues often come from overlap between indicators.

Stage two begins after latent variable scores are generated in stage one. Those scores—one per lower-order dimension—are exported (e.g., via Excel) and imported back into the dataset. In stage two, the higher-order constructs are reintroduced, but now their indicators are the latent variable scores from stage one. Measurement validation then differs by construct type. For reflective–reflective higher-order constructs, the checks emphasize composite reliability and convergent validity; the session notes a low loading for one IV subdimension (e.g., “complexity”), suggesting potential deletion if reliability/validity thresholds are not met. Discriminant validity is rechecked for the higher-order measurement model, again focusing on the DV3–DV4 HTMT issue.

The mediator is reflective–formative at the higher-order level, so validation shifts to collinearity and indicator contribution. The procedure includes checking VIF values (must be < 5), testing outer weights via bootstrapping, and using outer loadings as a secondary check (outer loadings ideally > 0.5; if not significant, consider removing indicators). In the example results, some outer weights are insignificant, but outer loadings are significant and VIF diagnostics show no multicollinearity, so indicators are retained.

Finally, the structural model is tested with bootstrapping (noting that complex models may take time). The results show a significant effect from IV to the mediator, and the mediator significantly affects the dependent constructs. Moderation is mixed: two moderation paths are significant while the rest are insignificant. Mediation is assessed by inspecting the indirect effects in the structural results. The overall takeaway is a practical blueprint: validate lower-order constructs first, generate latent variable scores, rebuild higher-order measurement models in stage two with the correct reflective/formative settings, then run bootstrapped structural tests including moderation.