A shift representable signal detection model of recognition memory.

A major theoretical question in psychology is how recognition memory can be understood in terms of latent variables reflecting underlying familiarity and response criteria. To address this, we investigate the viability of a shift representable signal detection model in which the shape of the familiarity distribution is the same for all old and new items. We present a new technique for fitting such models to recognition memory data using only ordinal properties of these data based on the monotonic linear regression algorithm described by Dunn and Anderson (Psychological Methods, 2024). We apply this methodology to analyze the results of three recognition memory experiments that vary the number of study presentations, study-test modality, focused vs. divided attention, and instructions to read study items or to generate them from cues. In each case, the data are well-described by a shift representable model with a recovered distributional form that has a negative skew that approximates the minimum extreme value, or Gumbel, distribution. We propose that if these results can be generalized across different experiments, it becomes possible to define a standard unit of measurement of familiarity that will enable direct comparisons to be made across different experiments and conditions. We discuss the implications of these results for current models of memory and for future application of our approach to other domains.