Discriminant validity is the extent to which a measure does not correlate with measures of theoretically distinct constructs. It is the necessary complement to convergent validity within Campbell and Fiske's multitrait-multimethod (MTMM) framework: a test may converge with related measures yet fail to differentiate from unrelated ones, in which case the construct is not adequately isolated.
Campbell and Fiske argued that monotrait–heteromethod correlations (convergence) must exceed heterotrait–monomethod correlations (different traits, same method) for discriminant validity to hold. If a speaking test correlates more strongly with a writing test taken on the same platform than with another speaking test taken in another mode, method variance is dominating trait variance and discrimination has failed. The diagonal-reading procedure they proposed inspects each row and column of the matrix for this pattern.
In structural equation modelling, discriminant validity is commonly assessed through the Fornell–Larcker criterion — the square root of the average variance extracted for a construct should exceed its correlation with any other construct — or through the heterotrait-monotrait ratio (HTMT), with values below .85 or .90 typically taken as adequate evidence.
In language testing, discriminant evidence might show that a reading-comprehension measure correlates more highly with another reading test than with a vocabulary-size test, supporting separation of the comprehension construct from lexical knowledge. Failure of discrimination often signals construct underrepresentation or construct-irrelevant variance — both threats to construct validity in Messick's unified framework.