The first release in the Research Series on the Recidivism of Federal Offenders, this report examines in detail the predictive statistical power of the Chapter Four Criminal History guidelines. This report serves as a “performance review” of criminal history’s predictive ability. Much like performance reviews for employees, the performance review of the criminal history measure includes a discussion of areas where performance is in need of improvement, is satisfactory, or is exceeding expectations. As such, it assesses the predictive power of the criminal history measure, determining whether it predicts better than random chance, and, if so, by how much. Emanating from this performance analysis, the reports in the recidivism project series examine the recidivism contributions of current criminal history components and suggest modifications or changes to improve predictive accuracy.
The study uses pre-conviction and instant offense information for a sample of guideline federal offenders sentenced in fiscal year 1992, matched with their post-sentencing criminal behavior collected from FBI records. The current study examines in detail the predictive statistical power of the criminal history measure, responds to the Sentencing Commission’s initial intentions, fills the void of empirical evidence about the criminal history measure, and addresses current criticisms of the CHC. Both tabular and statistical models of recidivism outcomes report findings by criminal history category and point groupings, as well as by offender demographics, instant offense characteristics, and recidivating offense types. (May 2004)
Based on the Commission's analysis the following observations were made:
- Both CHCs and criminal history points predict recidivism, regardless of whether recidivism is measured with the primary definition or by a re-conviction only definition.
- The most predictive result occurs when criminal history points are the measure used to predict the primary recidivism measure. The AUC for this combination is the highest, at 0.6979. This finding confirms the expectation that the criminal history points measure represents the most accurate predictive instrument.