Individual Confidence Intervals Do Not Inform Decision-Makers About the Accuracy of Risk Assessment Evaluations

Some recent articles have proposed that the confidence interval for the predicted outcome of a single case can be used to describe the predictive accuracy of risk assessments (Hart et al. Br J Psychiat 190:60–65, 2007b; Cooke and Michie, Law Hum Behav 2009). Given that the confidence intervals for an individual prediction are very large, Cooke and colleagues have questioned the wisdom of applying recidivism rates estimated from group data to single cases. In this article, we argue that the confidence intervals for the recidivism outcome predicted for a single case will range between zero to one (i.e., be uninformative) when the outcome is dichotomous and the predicted probability is between .05 and .95. This is true by definition and limits the utility of using individual confidence intervals to measure predictive accuracy. Consequently, other quality indicators (many of which are non-quantitative) are needed to determine the accuracy and error of risk evaluations.