Lineup research is facilitated by cognitive models that make testable predictions.

Witnesses discount shared facial features in lineups to make an identification.

Police lineups should be fair and prevent suspects from standing out.

Police should avoid needlessly matching facial features in lineups.

Worldwide, police test eyewitness memory using a lineup, containing an innocent or guilty suspect and several fillers. How similar-looking the fillers should be to the suspect has not been sufficiently answered, possibly due to a lack of guiding cognitive models. We made predictions using a new signal-detection model assuming feature-matching logic and discounting of shared features. Participants encoded a perpetrator with a distinctive feature. In two pre-registered experiments, we tested how replicating a similar but non-identical feature across the fillers (low feature-similarity replication) influenced identification accuracy compared to replicating an identical feature (high feature-similarity replication), and an unfair condition in which the feature was not replicated. In Experiment 1 (N = 4,915), the innocent suspect’s feature matched the description of the perpetrator’s. As predicted, low compared to high feature-similarity replication increased the hit rate without affecting the false alarm rate and increased ability to discriminate innocent from guilty suspects. In Experiment 2 (N = 1,964), the innocent suspect’s feature was identical to the perpetrator’s and, as predicted, the effect on discriminability was reversed. The model provides a useful theoretical framework, and police should not needlessly match features across a lineup that could be useful cues to identity, as this may harm witness performance.

Eyewitness identification

Filler similarity

Feature matching

Ensemble decision rule

Diagnostic feature-detection

Signal-detection theory

© 2026 The Author(s). Published by Elsevier Inc.