Using Machine Learning Analyses to Explore Relations Between Eyewitness Lineup Looking Behaviors and Suspect Guilt

Objective: We conducted 2 experiments using machine learning to better understand which lineup looking behaviors postdict suspect guilt., Hypotheses: We hypothesized that (a) lineups with guilty suspects would be subject to shorter viewing duration of all images and fewer image looks overall than lineups with innocent suspects, and (b) confidence and accuracy would be positively correlated. The question of which factors would combine to best postdict suspect guilt was exploratory. Method: Experiment 1 included 405 children (6-14 years; 43% female) who each made 2 eyewitness identifications after viewing 2 live targets. Experiment 2 included 342 adult participants (M-age = 21.00; females = 75%) who each made 2 identifications after viewing a video including 2 targets. Participants made identifications using an interactive touchscreen simultaneous lineup in which they were restricted to viewing one image at a time and their interaction with the lineup was recorded. Results: In Experiment 1, five variables (filler look time, suspect look time, number of suspect looks, number of filler looks, and winner look time) together postdicted (with a 67% accuracy score) target presence. In Experiment 2, four variables (number of suspect looks, number of filler looks, number of loser looks, and winner looks) together postdicted (with a 73% accuracy score) target presence. Conclusions: Further exploration of witness search behaviors can provide context to identification decisions. Understanding which behaviors postdict suspect guilt may assist with interpretation of identification decisions in the same way that decision confidence is currently used.