Evidence for a confidence–accuracy relationship in memory for same- and cross-race faces

ABSTRACT

Discrimination accuracy is usually higher for same- than for cross-race faces, a phenomenon known as the cross-race effect (CRE). According to prior research, the CRE occurs because memories for same- and cross-race faces rely on qualitatively different processes. However, according to a continuous dual-process model of recognition memory, memories that rely on qualitatively different processes do not differ in recognition accuracy when confidence is equated. Thus, although there are differences in overall same- and cross-race discrimination accuracy, confidence-specific accuracy (i.e., recognition accuracy at a particular level of confidence) may not differ. We analysed datasets from four recognition memory studies on same- and cross-race faces to test this hypothesis. Confidence ratings reliably predicted recognition accuracy when performance was above chance levels (Experiments 1, 2, and 3) but not when performance was at chance levels (Experiment 4). Furthermore, at each level of confidence, confidence-specific accuracy for same- and cross-race faces did not significantly differ when overall performance was above chance levels (Experiments 1, 2, and 3) but significantly differed when overall performance was at chance levels (Experiment 4). Thus, under certain conditions, high-confidence same-race and cross-race identifications may be equally reliable.

KEYWORDS:

• Confidence–accuracy relationship
• Cross-race effect
• Face recognition memory
• Metacognition

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. One concern about running ANOVAs on the proportion correct data is that the two independent variables of race of target face and confidence may be non-orthogonal. To account for this possibility, for each participant in each of the four studies, we ran a χ² test of independence on the number judgments made for same- and cross-race faces at each level of confidence. Based on these analyses, we removed participants using a cut-off of α = .05 on the χ² test of independence and then re-ran the ANOVAs. The number of participants removed per study was as follows: Experiment 1 = 2, Experiment 2 = 2, Experiment 3 = 1, and Experiment 4 = 2. With these participants removed, there were no differences in the statistical significance of any main effects or interactions in all four studies. Next, to decrease our chances of Type II error, we removed participants using a cut-off of α = .10 on the χ² test of independence and re-ran the ANOVAs. The number of additional participants removed per study was as follows: Experiment 1 = 7, Experiment 2 = 6, Experiment 3 = 2, and Experiment 4 = 2. With these additional participants removed, there were no differences in the statistical significance of any main effects or interactions in Experiments 2, 3, and 4. The main effect of race was significant (p = .02) in Experiment 1; therefore, we re-ran the t tests comparing the proportion correct at each confidence level after removing the 7 participants who had significant χ² test of independence using α = .10. The statistical significance of the t tests did not differ with these participants removed. Given that the results of the ANOVAs and t tests did not differ after removing these participants, in the manuscript, we report the ANOVA results on all participants.

2. The CA relationship was examined separately for the placebo and oxytocin conditions and was similar to the CA relationship for both groups combined. Before an alpha correction, there were statistically significant differences in same- and cross-race confidence-specific accuracy at confidence levels of 5 and 6 for the placebo group alone and at level 10 for the oxytocin group alone (ps < .05). There were no other significant differences. To be consistent with the other reported studies, we used a Bonferroni correction of α = .007, and these differences are no longer statistically significant. More critically, examining the mean proportions from the oxytocin group at level 10 indicated that high-confidence same-race judgments had an average accuracy of .97 whereas high-confidence cross-race judgments had an average accuracy of .95. There may be a real difference between high-confidence same- and cross-race judgments, but the difference is small.