Measurement of memory

For over 70 years, recognition memory has been modelled using signal detection theory. An unsolved problem with this approach is that the shapes of the distributions of memory strength for studied and unstudied items are unknown. Although they are often assumed to be Gaussian, with different location and scale parameters, such models often fail to fit observed data. This has had the effect of sustaining the viability of alternative approaches such as discrete state models, mixture models, and hybrid dual process models. However, it is now possible to estimate the shapes of the proposed memory strength distributions using the monotonic linear regression algorithm developed by Dunn and Anderson (under review). We describe this algorithm, show how it can recover the relevant distribution shapes under the signal detection model, and show that it fails to do so under alternative models. We apply it to data from three item recognition experiments. Each experiment used the same set of stimuli and varied the number of study presentations (1, 2, or 4) and the nature of the study item or the study task: visual vs. auditory presentation (Experiment 1), read vs. generate task (Experiment 2), and focused vs. divided attention task (Experiment 3). While the results confirm the predictions of the signal detection model, the recovered distributions deviate from the Gaussian. Furthermore, we show that the regression weight associated with each condition can be interpreted as a measure of memory strength for that condition, replacing traditional indices such as d-prime.