Evaluating signal detection models for eyewitness identification

Abstract

Eyewitness identification researchers have only recently employed signal detection theory (SDT) to understand witness performance on the police lineup task, in which a witness to a crime must either select one member from a (typically) six-person array who matches their memory of the perpetrator, or indicate that the perpetrator is not present. In addition to calculating empirical SDT measures from lineup data using Receiver Operating Curve analysis, researchers have fit a model called SDT-compound detection (SDT-CD) in attempt to discover the underlying theoretical parameters. However, SDT-CD has been selected for use without quantitative comparison against other potential SDT models. Of particular relevance to model selection is the contentious proposition that the sequential, rather than simultaneous, presentation of lineup members leads to superior witness decision performance, as the sequential lineup task challenges the plausibility of many SDT models that assume simultaneous presentation of items. This work compares the performance of three competing models; SDT-CD, a “maximum familiarity” model (MAX) and a novel sequential model (SDT-SEQ), in characterising sequential lineup data by using the Parametric Bootstrap Cross-fitting Method (PBCM). We tested both general model types (i.e. landscaping) and specific instances of the models as fit to 26 datasets in order to examine issues of model mimicry. Preliminary results highlight the challenges of using this approach to select between highly similar models and indicate that competing models strongly mimic each other.