#SmartPLS4 Series 9 - How to Test Discriminant Validity?

Discriminant validity is the checkpoint that confirms constructs in a social-science measurement model are truly distinct rather than overlapping in what they measure. In SmartPLS 4, it’s defined as the empirical distinctiveness of one construct from the others—an essential safeguard because concepts in the same study often correlate and can blur together if the measurement model isn’t properly validated.

A classic way to establish discriminant validity is the Fornell–Larcker criterion (proposed by Fornell and Larcker in 1981). The method compares, for each construct, the square root of its average variance extracted (AVEs) against its correlations with other constructs. The logic is straightforward: the within-construct variance (captured by the square root of AVE) should be larger than the shared variance (captured by inter-construct correlations). In the example used, the square root of AVE for “collaborative culture” (0.786) exceeds its correlation with the other construct “op” (0.618), so discriminant validity holds. When a third construct “O” is added, the same comparison is repeated: the square root of AVE for each construct must be greater than its correlations with every other construct. The transcript emphasizes that correlations are symmetric, so the same value can be reused (e.g., correlation between CC and O matches correlation between O and CC), which simplifies the Excel comparisons.

The session then shifts to a newer, more conservative approach: the HTMT ratio (heterotrait–monotrait ratio). HTMT measures discriminant validity by looking at correlations between indicators from different constructs (heterotrait) relative to correlations among indicators measuring the same construct (monotrait). SmartPLS 4 reports HTMT values directly, and the interpretation hinges on thresholds. Values below 0.85 are treated as a conservative standard, while 0.90 is also mentioned as an acceptable recommendation. In the example, all HTMT values fall well under 0.85 (shown as green), indicating no discriminant validity problems and supporting the claim that constructs are distinct.

Finally, the transcript covers cross-loadings, a practical item-level test. Each indicator should load more strongly on its own parent construct than on other constructs. In SmartPLS 4, the loadings for items tied to “CC” are checked against loadings on “O” and “op”; the same inspection is repeated for items tied to “O” and “op.” The rule of thumb applied is that the indicator’s loading on its own construct should be substantially higher than its loadings on other constructs, and the transcript notes that no problematic cross-loading appears (including the observation that none of the competing comparisons exceed 0.70). With Fornell–Larcker, HTMT, and cross-loadings all indicating clean separation, discriminant validity is treated as established, setting up the next step: addressing any discriminant validity issues if they arise in other models.