Kiyofumi Miyoshi¹ (), Dobromir Rahnev², Hakwan Lau^3,4; ¹Kyoto University, ²Georgia Institute of Technology, ³Institute for Basic Science, South Korea, ⁴Sungkyunkwan University

In visual detection tasks, participants often exhibit distinct behaviors between trials with and without target presence, commonly manifested as asymmetric type-1 receiver operating characteristic (ROC) curves. Such asymmetry suggests greater signal variability in target-present trials, violating the equal-variance assumption fundamental to the classic d′ metric. Here, we examined the utility of response time (RT) data for implementing unequal-variance signal detection theory (SDT) analysis as a cost-effective alternative to the traditional confidence-based approach. Across eleven datasets, RT- and confidence-based analyses yielded highly consistent estimates of the sensitivity metric dₐ—an unequal-variance extension of classic d′. Furthermore, the d′ metric systematically overestimated detection performance relative to dₐ, underscoring the need to incorporate unequal-variance characteristics into the analysis. We also found that the degree of this overestimation varied with observers' response criteria: because unequal-variance ROCs expand more on the low–false-alarm side, adopting a conservative criterion can spuriously inflate d′. Taken together, these findings indicate that RT-based unequal-variance SDT analysis provides a reliable means of assessing visual detection performance, especially in situations where collecting confidence ratings is impractical.

Acknowledgements: This work is supported by JSPS KAKENHI Grant Number 22K13870 and 25K00896, awarded to KM. HL is supported by the Institute for Basic Science, South Korea (Grant Number IBS-R015-D2).