Explaining the predictive performance of police in cases of domestic abuse

Police in England and Wales are asked to predict the likelihood of serious recidivism in domestic abuse cases, with little support beyond a flawed questionnaire. To analyze their decisions, we first develop methods to deal with a censoring problem that arises because the officer's prediction may change the outcome she was attempting to predict. Even after adjusting for censoring, we find predictive performance to be low, even lower in some cases than what one would expect by chance. We next ask whether their predictions represent mistakes, and provide several types of confirmatory evidence. We ask how officers formulate their predictions, and we find evidence consistent with representativeness bias, overreaction, and categorization and selective attention. We find that higher-skill officers use information not captured by the questionnaire to improve their predictions, whereas lower-skill officers use such information in ways that reduce accuracy.