#SmartPLS4 Series - 41 - Analyze, Interpret, and Report Higher Order Reflective Formative Model

Higher-order constructs that behave “reflective at the lower level but formative at the higher level” can be validated in SmartPLS4 using a two-stage approach—first validating all lower-order measurement models, then building the higher-order construct from the lower-order latent variable scores and checking convergent validity, collinearity, and indicator significance.

The session focuses on a reflective–formative higher-order model where three lower-order dimensions (Vision development and rewards) collectively form a higher-order construct (internal marketing). At the lower level, each dimension is treated as reflective, meaning multiple items load onto each dimension and standard reliability/validity checks apply. Only after those lower-order constructs pass reliability and validity are latent variable scores exported (via PLS-SEM path modeling and “export to CSV”) and re-imported into a second SmartPLS project to construct the higher-order measurement model.

Two versions of the two-stage method are contrasted: embedded and disjoint. In the embedded two-stage approach, the first stage models the entire higher-order construct, while the disjoint approach keeps stage one limited to lower-order components—no higher-order construct appears in the path model at that stage. In stage two, the higher-order construct is measured formatively using the saved lower-order latent variable scores as indicators. The practical difference in SmartPLS is that disjoint stage two uses latent variable scores for the lower-order constructs that belong to the higher-order construct, while other constructs without higher-order structure remain measured using their original indicators.

Once the higher-order reflective–formative model is specified in stage two (with measurement model directions inverted so the higher-order construct is formative), validation proceeds with criteria tailored to formative measurement. Convergent validity at the higher-order level requires a “global” single-item measure representing the whole construct (planned during questionnaire design). A redundancy analysis is run by linking the higher-order formative construct to this global reflective measure; the resulting path coefficient should exceed a threshold (the session cites 0.708), indicating the formative construct explains more than half of the variance in the alternative global measure.

Next comes collinearity diagnostics using VIF values for the formative indicators (should be below 5). Then the significance of outer weights is tested via bootstrapping. If outer weights are not significant, the procedure shifts to outer loadings: indicators can be retained when loadings are sufficiently high (the session uses 0.5 as the practical cutoff) and significant; if both weight and loading fail, the indicator should be removed.

After measurement model validation at both levels, the workflow moves to structural assessment: bootstrapping (with bias-corrected and accelerated options and one-tailed testing when directions are known) is used to evaluate path coefficients and p-values for hypotheses. The session also demonstrates mediation testing for cases where the higher-order construct participates in indirect effects. Mediation is reported through indirect effects (specific indirect effects), total effects, and direct effects; partial mediation is identified when the direct effect remains significant alongside a significant indirect effect. The end result is a repeatable reporting template: validate lower-order reflective constructs first, generate latent scores, validate the higher-order reflective–formative construct with redundancy analysis, VIF, outer weights/loadings, then test direct, indirect, and mediated relationships using bootstrapped significance.