Conceptualize, Analyze, and Interpret Discriminant Validity using #SmartPLS4

Discriminant validity is the quality check that confirms overlapping constructs in social science research are truly distinct. In SmartPLS 4, it’s assessed after running the PLS-SEM algorithm, and the goal is straightforward: each construct should correlate less with other constructs than it does with itself. When discriminant validity holds, researchers can trust that measurement items are capturing different underlying concepts rather than repeating the same one under different labels.

The session walks through three common approaches used in SmartPLS 4: HTMT, the Fornell–Larcker criterion, and cross-loadings. For HTMT (heterotrait–monotrait ratio), the key decision rule is a threshold of 0.85 for a conservative test, with 0.90 sometimes used as a more liberal alternative. In the example model, all HTMT ratios appear below 0.85, so the constructs are treated as distinct. A special note addresses a blank entry for “development vs development op”: those two constructs are actually the same variable, so their correlation is effectively 1, leaving no meaningful ratio to report.

Next, the Fornell–Larcker criterion is used as a more traditional benchmark. The method compares the square root of each construct’s AVE (average variance extracted) against that construct’s correlations with all other constructs. In the walkthrough, the square root of AVE for “development” is computed from 0.694, giving approximately 0.833, and this value exceeds correlations between development and the other constructs. The same logic is applied to “op,” “rewards,” and “vision”: each construct’s square root of AVE is larger than its correlations with the other constructs, including the relevant pairwise comparisons (for example, development vs op, and rewards vs op). With those inequalities satisfied across the set, discriminant validity is considered established under Fornell–Larcker.

Finally, cross-loadings provide an item-level view of discriminant validity. Each construct is measured with multiple items (development: 7 items, op: 5, rewards: 4, vision: 3). The rule is that an item should load highest on its own theoretical construct compared with loadings on other constructs. The example shows that items for development load strongly on development (e.g., dev1 at 0.841) and drop when forced to load on op or rewards or vision. The same pattern holds for op items (e.g., op1 at 0.783 on op, decreasing to 0.530 when compared against development), rewards items, and vision items. Even when some items show secondary loadings, the analysis prioritizes the earlier HTMT and Fornell–Larcker results, using cross-loadings as supporting evidence.

Taken together, the workflow demonstrates how SmartPLS 4 can confirm construct distinctiveness using HTMT thresholds, AVE-based comparisons, and item-level loading patterns—so measurement models don’t accidentally conflate constructs that should remain separate.