When criterion control in face matching induces correlation: Commentary on Baker et al. (2026)

The task of face matching requires deciding whether two images depict the same person (a match) or two different people (a mismatch). Perhaps curiously, previous research has shown that the relationship between match and mismatch accuracies is typically weak. Baker et al. (2026) recently argued that this pattern reflected individual differences in response criterion, and by controlling for criterion, the authors revealed a strong positive association consistent with a shared underlying ability across these trial types. Here, I examine a methodological issue that complicates this interpretation. Specifically, criterion was computed from the same hit and false-alarm rates that defined match and mismatch accuracies, meaning that the residualisation procedure conditioned on a function of the outcomes themselves. Using Monte Carlo simulations, I showed that this procedure could induce large positive correlations even when match and mismatch performance were generated independently. Further, the magnitude of the residual correlation depended on the degree of outcome-covariate coupling: it is maximal when criterion is derived from the same task, reduced when averaged across tasks, and negligible when estimated independently. Finally, reanalyses of the original datasets showed this same pattern. These findings suggest that residual correlations are not, by themselves, diagnostic of a shared ability and should be interpreted with caution. More generally, the results highlight the need to distinguish genuine shared variance from artefacts introduced by statistical conditioning in signal detection analyses.