3. Reflective-Formative Higher Order Construct/Second Order Analysis and Reporting in SmartPLS

Q: What makes a reflective–formative higher order construct “type two,” and how does that shape the analysis steps?

In a type two reflective–formative higher order construct, the lower-order constructs are reflective while the higher-order construct is formative. That split determines the workflow: reflective lower-order blocks are validated first (stage one), and their latent variable scores become the formative indicators used to estimate and validate the higher-order construct in stage two.

Q: How does the disjoint two-stage approach differ from the embedded two-stage approach in SmartPLS?

In the embedded two-stage approach, the higher-order construct is present in stage one (the entire higher-order construct is modeled early). In the disjoint two-stage approach, stage one includes only the lower-order components; the higher-order construct is created in stage two using the saved latent variable scores from stage one. The transcript notes that both approaches can yield similar results, but disjoint is widely used for reflective–formative models.

Q: What measurement validation checks apply to stage one (lower-order reflective constructs)?

Stage one uses standard PLS-SEM measurement model evaluation for reflective constructs: outer loadings for indicator reliability, reliability/validity metrics, convergent validity via AVE (commonly AVE > 0.50), and discriminant validity via Fornell–Larcker (the square root of AVE for a construct should exceed correlations with other constructs). Collinearity is also checked using VI/VIF values, with VI values below 5 treated as acceptable.

Q: Why are latent variable scores saved after stage one, and how are they used in stage two?

Latent variable scores from the lower-order constructs are needed because, in the disjoint approach, the higher-order construct is not estimated in stage one. In stage two, those saved scores become the indicators used to form the higher-order formative construct (e.g., CSR formed from the latent scores of ECC, LC, EC, and DC). The transcript emphasizes importing these scores into SmartPLS and re-specifying CSR as formative in the second-stage model.

Q: How does the transcript suggest reporting and interpreting structural and mediation results?

After validating both measurement layers, structural paths are tested using bootstrapping (the transcript mentions 5,000 resamples). Path coefficients are reported with t-statistics and p-values. For mediation, the transcript describes interpreting direct and indirect effects (e.g., TI mediating CSR’s effect on OP) and notes that mediation may appear under a more lenient significance threshold (like 10%) even if it does not meet stricter 5% criteria.

TL;DR

Use the disjoint two-stage approach for reflective–formative higher order constructs: validate lower-order reflective blocks first, then build and validate the higher-order formative construct using saved latent variable scores.

Briefing Cornell Notes

Briefing

Reflective–formative higher order constructs in SmartPLS are best handled with a disjoint two-stage approach: first validate the lower-order reflective components, then use their latent variable scores to build and validate the higher-order formative construct before running the structural model. The key practical payoff is that the higher-order construct (formative) is not included in the first-stage PLS run, which keeps the initial measurement assessment focused on the reflective blocks that generate the scores used later.

In this model type, multiple lower-order constructs are reflective, while the higher-order construct is formative—specifically a “type two” reflective–formative higher order construct. The disjoint two-stage method differs from the embedded version in how the higher-order construct is treated. In the disjoint approach, stage one estimates only the lower-order constructs and saves their latent variable scores. Stage two then uses those saved scores as indicators to estimate the higher-order construct along with the rest of the study constructs.

Validation happens twice. For stage one, the lower-order reflective constructs follow standard PLS-SEM measurement checks: indicator reliability via outer loadings, internal consistency/reliability, convergent validity using AVE (with a common benchmark of AVE > 0.50), and discriminant validity using the Fornell–Larcker criterion (the square root of AVE for each construct should exceed the off-diagonal correlations). Collinearity is also monitored through VIF/VI values, with VI values below 5 indicating no serious collinearity concerns.

Stage two adds a second layer of measurement validation because the higher-order formative construct must be assessed as a formative measurement model. The process starts with collinearity checks again (VI values for the higher-order construct’s formative indicators). Next comes the formative indicator evaluation: outer weights must be significant; if an outer weight is not significant, the decision shifts to outer loadings. The transcript lays out a decision rule: if outer loadings are above 0.5 and significant, the indicator can remain even when the weight is insignificant; if outer loadings are below 0.5 and not significant, the formative indicator is typically deleted—unless removal would harm the model’s results.

Once the higher-order construct (formative) is validated, the structural model is straightforward: bootstrapping (often 5,000 resamples) is used to test path coefficients and mediation. The example structural relationships test the impact of the higher-order construct CSR on TI and OP, including TI’s mediating role. Results are interpreted using significance thresholds (with the transcript noting that mediation may appear at a more lenient level such as 10% but not at 5% if strict criteria are applied). Reporting guidance is also emphasized: higher-order construct validation tables should include outer weights, outer loadings, and VI values for the formative indicators, while structural results should be reported as path coefficients with t-statistics and p-values.

A practical walkthrough ties it together using a second-order higher-order construct CSR formed by four first-order reflective dimensions (ECC, LC, EC, and DC). The workflow is: run stage one without CSR, validate the four reflective dimensions, save their latent variable scores, import them into SmartPLS, then re-specify CSR in stage two as formative using those saved scores, validate CSR’s formative measurement properties, and finally estimate the structural paths and mediation.

Cornell Notes

Reflective–formative higher order constructs require two measurement validations in SmartPLS when using the disjoint two-stage approach. Stage one estimates and validates the lower-order reflective constructs, then saves their latent variable scores. Stage two treats the higher-order construct as formative and uses those saved scores as formative indicators, checking collinearity (VI < 5), outer weights (significant), and—when needed—outer loadings (keep indicators with loadings > 0.5 and significant; drop those < 0.5 and not significant). After the higher-order construct is validated, bootstrapping tests the structural paths and any mediation (e.g., TI mediating CSR’s effect on OP). This workflow clarifies how to report both reflective and formative measurement results.

What makes a reflective–formative higher order construct “type two,” and how does that shape the analysis steps?

In a type two reflective–formative higher order construct, the lower-order constructs are reflective while the higher-order construct is formative. That split determines the workflow: reflective lower-order blocks are validated first (stage one), and their latent variable scores become the formative indicators used to estimate and validate the higher-order construct in stage two.

How does the disjoint two-stage approach differ from the embedded two-stage approach in SmartPLS?

In the embedded two-stage approach, the higher-order construct is present in stage one (the entire higher-order construct is modeled early). In the disjoint two-stage approach, stage one includes only the lower-order components; the higher-order construct is created in stage two using the saved latent variable scores from stage one. The transcript notes that both approaches can yield similar results, but disjoint is widely used for reflective–formative models.

What measurement validation checks apply to stage one (lower-order reflective constructs)?

Stage one uses standard PLS-SEM measurement model evaluation for reflective constructs: outer loadings for indicator reliability, reliability/validity metrics, convergent validity via AVE (commonly AVE > 0.50), and discriminant validity via Fornell–Larcker (the square root of AVE for a construct should exceed correlations with other constructs). Collinearity is also checked using VI/VIF values, with VI values below 5 treated as acceptable.

How is the higher-order formative construct validated in stage two?

Stage two begins with collinearity checks for the formative indicators (VI values; the transcript uses the threshold VI < 5). Then it checks outer weights for significance. If an outer weight is not significant, the decision moves to outer loadings: indicators with outer loadings > 0.5 and significant are typically retained; indicators with outer loadings < 0.5 and not significant are candidates for deletion, unless removing them would materially change the model’s results.

Why are latent variable scores saved after stage one, and how are they used in stage two?

Latent variable scores from the lower-order constructs are needed because, in the disjoint approach, the higher-order construct is not estimated in stage one. In stage two, those saved scores become the indicators used to form the higher-order formative construct (e.g., CSR formed from the latent scores of ECC, LC, EC, and DC). The transcript emphasizes importing these scores into SmartPLS and re-specifying CSR as formative in the second-stage model.

How does the transcript suggest reporting and interpreting structural and mediation results?