Assessing predictions of violence: Being accurate about accuracy.

The prediction of violence occupies a prominent and controversial place in public mental health practice. Productive debate about the validity of violence predictions has been hampered by the use of methods for quantifying accuracy that do not control for base rates or biases in favor of certain outcomes. This article describes these problems and shows how receiver-operating characteristic analysis can be used to solve them. The article also reanalyzes 58 data sets from 44 published studies of violence prediction. Taken together, these data strongly suggest that mental health professionals' violence predictions are substantially more accurate than chance. Short-term (1-7 day) clinical predictions seem no more accurate than long-term (> 1 year) predictions. Past behavior alone appears to be a better long-term predictor of future behavior than clinical judgments and may also be a better indicator than cross-validated actuarial techniques.