Discriminant validity
Also known as: Divergent validity
A measure should not correlate with things it's supposed to be distinct from.
What it means
Discriminant validity is the degree to which a measure is empirically distinguishable from measures of conceptually different constructs, correlating only weakly with them. It guards against a measure being redundant with, or contaminated by, a neighboring concept it should be separate from. Assessed jointly with convergent validity in the multitrait-multimethod framework, it requires that correlations among different constructs be lower than correlations among different measures of the same construct. Failures often arise from shared method variance, when two unrelated traits correlate merely because they were measured the same way.
How it is assessed
Campbell and Fiske originally read discriminant validity off the multitrait-multimethod matrix by eye: a measure passed if correlations between different traits stayed below correlations between different methods measuring the same trait. Modern practice replaced this with formal rules. The Fornell-Larcker criterion asks whether a construct's average variance extracted, the share of variance its indicators hold in common, exceeds the squared correlation it shares with any other construct. A related check inspects cross-loadings, requiring each item to load more strongly on its own factor than on rivals. All three ask the same question in different arithmetic: does the measure carve out territory the neighboring constructs do not already occupy?
Where the standard tests fail
The workhorse tests are weaker than their popularity suggests. Henseler, Ringle and Sarstedt ran simulations in which two constructs were known to be nearly identical, then checked how often each method caught it. The Fornell-Larcker criterion flagged the problem in only about a fifth of cases; the cross-loadings check did even worse, catching essentially none. Their proposed alternative, the heterotrait-monotrait ratio (HTMT), estimates the true correlation between constructs after correcting for how tightly each is measured, and detected the same failures far more reliably. A later refinement, HTMT2, uses a geometric mean to relax an assumption that indicators are interchangeable. The lesson is that passing Fornell-Larcker is weak evidence; two constructs can clear it while being empirically the same thing.
Where it matters
Discriminant validity is a gatekeeping concern wherever survey constructs proliferate. Marketing and information-systems research, built heavily on structural equation models, lean on it to justify that perceived usefulness, satisfaction and loyalty are distinct drivers rather than one halo measured thrice. Clinical psychometrics uses it to separate depression from anxiety, or attention deficit from oppositional traits, though genuine comorbidity complicates the reading. Organizational research invokes it to defend that engagement, commitment and satisfaction are not the same survey relabeled. In each field the risk is the same: a literature that appears to test relationships among several constructs may in fact be regressing a variable on lightly disguised copies of itself, inflating explained variance and manufacturing findings.
Limits and caveats
Discriminant validity is a matter of degree, not a threshold to be cleared and forgotten. Rönkkö and Cho argue for defining it as the disattenuated correlation between measures, the association that remains once measurement error is removed, and warn against treating any single cutoff as decisive. Two related-but-distinct constructs may correlate genuinely and strongly; a high correlation is a prompt to investigate, not automatic proof of redundancy. Failures also frequently trace to shared method rather than the traits themselves, so a diagnosis of poor distinctness should ask whether design, not concept, produced it. The honest conclusion is often that two constructs are partly separable, which the binary language of pass and fail obscures.
Examples
A measure of self-esteem shows discriminant validity if it correlates far less with extraversion than with other self-esteem measures.
A new employee-engagement survey correlates almost perfectly with an existing job-satisfaction scale. It may simply be measuring the old construct under a fresher name rather than a distinct one.
Two supposedly unrelated traits look linked when both are scored from the same manager's ratings. The shared method, not the traits, produced the correlation — a classic false pass on distinctness.
A new depression questionnaire correlates as strongly with an anxiety inventory as with rival depression scales. Either the two disorders genuinely overlap, or the instrument fails to isolate depression from general distress.
In a brand-equity model, perceived quality and loyalty each pass the Fornell-Larcker check, yet their HTMT ratio sits near 0.95, hinting the two survey blocks measure one underlying attitude, not two.
First described in Campbell & Fiske (1959).
Key references
- Rönkkö, M., & Cho, E. (2022). An updated guideline for assessing discriminant validity. Organizational Research Methods, 25(1), 6-14. doi.org/10.1177/1094428120968614
- Roemer, E., Schuberth, F., & Henseler, J. (2021). HTMT2 — an improved criterion for assessing discriminant validity in structural equation modeling. Industrial Management & Data Systems, 121(12), 2637-2650. doi.org/10.1108/IMDS-02-2021-0082
- Henseler, J., Ringle, C. M., & Sarstedt, M. (2015). A new criterion for assessing discriminant validity in variance-based structural equation modeling. Journal of the Academy of Marketing Science, 43(1), 115-135. doi.org/10.1007/s11747-014-0403-8
- Fornell, C., & Larcker, D. F. (1981). Evaluating structural equation models with unobservable variables and measurement error. Journal of Marketing Research, 18(1), 39-50. doi.org/10.1177/002224378101800104
- Campbell, D. T., & Fiske, D. W. (1959). Convergent and discriminant validation by the multitrait-multimethod matrix. Psychological Bulletin, 56(2), 81-105. doi.org/10.1037/h0046016