Primer on Risk Assessment and the Statistics Used to Evaluate Its Accuracy

Abstract

The pervasiveness of risk assessment in correctional decision-making necessitates a better understanding of the nature of risk scales and the methods used to assess their accuracy. Risk is a continuous dimension, which means that risk assessment is a prognostic task as opposed to a diagnostic task. Risk scales can also be considered criterion-referenced as opposed to norm-referenced. Predictive accuracy can be divided into discrimination and calibration. Area under the curves (AUCs), Cox regression, Harrell’s C, Cohen’s d, and logistic regression are appropriate for analyses of discrimination. There is no consensus on calibration statistics, but both chi-square tests and the Expected/Observed (E/O) index have been used and show promise. Statistics intended for dichotomous diagnostic decisions (e.g., positive predictive accuracy and negative predictive accuracy, number needed to detain, number needed to discharge) are inappropriate for risk scales because of the prognostic nature of risk scales. In many circumstances, diagnostic statistics provide more information about the base rate of recidivism than about the risk scale.