Abstract
Estimating and comparing signal detection models to account for empirical receiving operator characteristic (ROC) curves has been a pivotal method in the studies of recognition memory. The unequal-variance signal detection (UVSD) model, following the assumptions of the standard signal detection theory (SDT), assumes that recognition memory can be captured by memory strength with an additional parameter to capture the unequal variances of the signal and the noise distribution. In contrast, the dual-process signal detection (DPSD) model combines a high-threshold component for recollection and a SDT component for familiarity-driven processes to interpret the ROC curves. The comparison between the two models in previous work has produced some inconsistent results, which could originate from two sources. First, individual-level model comparison techniques could be biased toward one of the models due to their difference in model flexibility. Second, the process of recognition memory can be a population-level random effect in nature. Specifically, there could be individual differences in whether UVSD or DPSD is the true mechanism for recognition memory in the population. The present study has thus attempted to develop a framework for estimating the frequency of each model being the “true” model in the population (i.e. model frequency) by assessing and comparing candidate methods in terms of the recovery rate of the model frequency. We found that mean posterior model probability is the most effective method for the comparison between UVSD and DPSD, while the percentage of being the best model in the individual-level model comparisons produces the highest recovery rate. We have also extended the current approach to other domains of interests (e.g., detecting shifts of the mean in Gaussian distributions). In general, we propose that population-level model frequency can provide unique and important information in model comparisons and as a result should be adopted in future studies.