On Estimating the Diagnosticity of Eyewitness Nonidentifications

TLDR: In this paper, a Bayesian model of information gain is used to mathematically prove that if an eyewitness identification of a suspect increases the probability that the suspect is the criminal, then a non-identification must decrease the probability of the suspect being the criminal.

Abstract: The criminal justice system's practice of treating eyewitness lineup identifications of suspects as highly informative while treating nonidentifications (i.e., no-choice responses or choices of foi ls) as uninformative is questioned. A Bayesian model of information gain is used to mathematically prove that (a) if an eyewitness identification of a suspect increases the probability that the suspect is the criminal, then a nonidentification must decrease the probability that the suspect is the criminal; and (b) the relative diagnosticity of identifications versus nonidentifications (regarding the probability that the suspect is the criminal) is determined by the probability of obtaining an identification versus nonidentification, with nonidentifications being more diagnostic if they are relatively less frequent than identifications. An application of the Bayesian model to previously published data suggests that greater diagnosticity for nonidentifications than identifications is more than just a theoretical possibili ty; the available data show nonidentifications to be more than one and a half times as diagnostic as identifications regarding the probability that the suspect is the criminal. A breakdown of nonidentifications into two types, eyewitness choices of a lineup foil versus no-choice decisions, suggests that the latter is more informative than the former regarding the probability that the suspect is innocent. The cognitive mechanisms that may be responsible for criminal justice investigators' discounting of nonidentifications are discussed in relation to research on human judgment.