#SmartPLS4 Series 18 - How to Analyze Higher Order Reflective Formative Construct?

Validating a higher-order construct that behaves differently across levels—reflective at the lower level but formative at the higher level—requires more than running one set of checks. The core requirement is to validate every lower-order component first, generate their latent variable scores, and then use those scores to build and test the higher-order model under formative rules.

The session focuses on a “disjoint two-stage approach” for SmartPLS4 as an alternative to the repeated indicators method, which can run into complications when higher-order constructs are involved. In the disjoint version, stage one estimates the measurement model using only the lower-order constructs (no higher-order construct is placed in the path model). Researchers then save the latent variable scores for just the lower-order components that feed the higher-order construct. Stage two uses those saved scores as inputs to form the higher-order construct, while constructs that do not belong to the higher-order layer remain modeled at the indicator level.

A concrete example is used: “Internal Marketing” is treated as a higher-order construct that is reflective at the lower level but formative at the higher level. Three subdimensions—Vision development, and Rewards—are reflective at the lower level and together form Internal Marketing at the higher level. Meanwhile, other constructs in the model (including Internal Service Quality) are handled as reflective–reflective higher-order constructs, but the practical walkthrough centers on the reflective–formative case.

The workflow starts by validating the lower-order measurement models: reliability and validity checks are performed first, including outer loadings, reliability metrics, and discriminant validity. Only after those lower-level constructs pass do latent variable scores be exported (e.g., to CSV) and re-imported into SmartPLS4 to create the stage-two model.

For the higher-order reflective–formative construct in stage two, the evaluation follows formative measurement-model criteria. Convergent validity is assessed by linking the higher-order construct to a global measure—an overall item or scale capturing the entire construct (for Internal Marketing, an example global statement is about whether the organization provides proper rewards and development initiatives). Next comes collinearity diagnostics: VIF values for the formative lower-order components must stay below a threshold (the session uses < 5). Then the model checks formative indicator behavior through bootstrapping: outer weights must be significant; if an outer weight is insignificant, the decision pivots to outer loadings and their significance (indicators with low, non-significant loadings are candidates for removal).

The takeaway is procedural and strict: lower-order validation is not optional, latent variable scores are the bridge between stages, and formative higher-order constructs demand checks for convergent validity, multicollinearity, and the statistical contribution of each formative component. The result is a defensible higher-order measurement model that matches how the construct functions across levels—reflective below, formative above—without relying on repeated indicators.