Improved Techniques for Assessing the Accuracy of Recidivism Prediction Scales

• NCJ Number: 126668
• Date Published: 1990
• Author(s): J Cohen; S E Zimmerman
• Length: 59 pages
• Grant Number(s): 86-IJ-CX-0039
• Agencies: NIJ-Sponsored
• Publication Type: Survey

Annotation

Four prediction scales designed to predict recidivism were used to examine the Random Improvement Over Chance (RIOC) measure as an indicator of prediction accuracy that is independent of the base and selection rates of individual samples.

Abstract

Each of the four prediction scales, INSLAW, RAND, CFS81, and the CGR (Center for Governmental Research), was designed to predict different outcomes. Four datasets were selected to reflect different geographical areas, Albuquerque, Miami, New York City, and California, and a mix of case processing stages in the criminal justice system, of arrest, conviction, and incarceration. All four scales were applied to each of the four datasets. The RIOC pluses and the RIOC negatives were identical for each scale. This was the case despite the large differences in their potential ranges. The accuracy achieved for false positives and false negatives represents the same relative level of improvement within their respectives ranges. The two error rates, appearing very different in their absolute magnitudes, were actually very similar relative to chance accuracy and maximum possible accuracy in a dataset. The RIOC statistics provide a measure of accuracy that is standardized relative to the varying constraints on accuracy. As it is free of such data dependencies, the RIOC emerges as a powerful indicator of relative accuracy for both recidivist and nonrecidivist predictions. 18 references, 6 tables, 5 figures, and 1 appendix