Geraldine Jeckeln¹ (), Alice J. O'Toole¹; ¹The University of Texas at Dallas

Cross-race face identification is done routinely in applied settings (e.g., forensic, security). Despite the error-prone nature of cross-race identification, no publicly available tests exist to evaluate individuals’ identification ability for own- versus other-race faces. Our goal was to develop a test of African American (AA) and Caucasian (CA) test items that would: 1) challenge individuals of varying abilities, 2) ensure comparable identification accuracy for both AA and CA (no item-race effect), and 3) provide measures of item difficulty. Item selection was guided by the performance of a face-identification algorithm (Szegedy et al., 2017) and AA and CA observers (n = 34 per race). The resulting test includes 25 AA and 25 CA face-image triads. Each triad contains two images of the same person and one image of a different person. The task is to select the image of the “different” person. We found a classic cross-over “other-race effect” [interaction of item and participant race: F(1, 66) = 4.33, p = 0.041 , η2p = 0.06; AA participants (AA faces: M = 0.74; CA faces M = 0.71); CA participants (CA faces, M = 0.75; AA faces M = 0.72)], with no main effect of item or participant race (p > 0.05). Next, item difficulty was assessed using Item Response Theory (Lord, 1980). This provides participant-ability and item-difficulty estimates on the same scale. Item responses for each item-race set and participant-race group were modeled separately. Results showed that test items covered a wide range of item difficulty levels, with average difficulty falling slightly below participant ability. Overall, the novel face-identification test allows for accurate assessment of participants’ identification abilities for both AA and CA face identities. This work also provides a guideline for designing cross-race tests using the output of face-recognition systems and human observers.

Acknowledgements: Research funded by The National Institute of Standards and Technology, Grant 70NANB22H150 to A.OT.