Recognition memory — judging whether an item was previously studied — is naturally modeled as a signal detection task. Studied ("old") items have higher familiarity than unstudied ("new") items, and the observer sets a criterion on the familiarity axis. SDT analysis separates the ability to discriminate old from new items (d′) from the tendency to say "old" (criterion c).
Key Findings
σ_old > σ_new (old items more variable)
d_a recommended over d′ when slope ≠ 1
The most robust finding in recognition memory SDT is that zROC curves are linear with slopes around 0.80, indicating that old-item distributions have about 25% greater variance than new-item distributions. This asymmetry is attributed to encoding variability: some studied items are well-encoded and highly familiar, while others receive poor encoding.
Dual-Process Models
The debate between single-process (pure SDT) and dual-process (familiarity + recollection) models has been a central issue. Yonelinas' dual-process model proposes that recognition combines a continuous familiarity process (modeled by SDT) and a threshold recollection process. The relative contributions of these processes are estimated from ROC curve shape, with recollection producing a threshold-like component and familiarity producing the curvilinear component.