UC Irvine

Recidivism, or the subsequent commission of a criminal offense after receiving punishment in the justice system, is a primary concern for public safety decision-makers. The quantification of an individual offender's likelihood of recidivism has been a focus of criminology for over 100 years and methods of analysis continue to develop alongside technological innovations in statistical computing. At present, machine learning approaches are being applied to criminal justice data sets in order to generate algorithms that offer predictions of recidivism by specific, individual offenders. Such algorithms underlie the risk assessment instruments that are presently used by decision-makers during the sentencing phase. Although these machine learning approaches are cutting-edge techniques of exploring patterns in data, certain strategies (especially so-called "black box algorithms") may not be appropriate for use in criminal justice applications due to their enigmatic nature conflicting with the principle of open justice. Furthermore, the underlying data used may be such a flawed reflection of recidivism that the algorithms produced by this process inevitably generate predictions which reflect and perpetuate systemic bias in the justice system. This dissertation examines the predictive validity of the Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) recidivism tool, which is widely-used in the United States at the present time. A parsimonious actuarial model (or PAM) is generated by the author, entertaining a number of state-of-the-art approaches to machine learning for its development, and employed as an alternative white-box algorithm for comparison of predictive validity. Results indicate a trivial difference in predictive validity between existing tools and open approaches. As such, a simplified and transparent model is both philosophically and practically preferable for the purpose of criminal justice applications.