Breaking the 'Glass Ceiling' of Risk Prediction in Recidivism: An Application of Connectionist Modelling to Offender Data

The present thesis explored the capability of connectionist models to break through the ‘glass ceiling’ of accuracy currently in operation in recidivism prediction (e.g., Yang, Wong, & Coid, 2010). Regardless of the inclusion of dynamic items, all risk measures rarely exceed .75 in terms of the area under the receiver operating characteristic curve (AUC) (Hanley & McNeil, 1982). This may reflect the emphasis of multiple regression equations on main effects of a few key variables tapping long-term anti-social potential. Connectionist models, not used in criminal justice, represent a promising alternative means of combining predictors given their ability to model interactions automatically. To promote learning from other fields a systematic review of the literature on the application of connectionist models to operational data is presented. Lessons were then taken forward in the development of a connectionist model suitable for the present data which comprised fields from the Offender Assessment System (OASys) (Home Office, 2002) relating to 4,048 offenders subject to probation supervision. Included in the items for modelling was the Offender Group Reconviction Scale (OGRS) (Copas & Marshall, 1998; Taylor, 1999). Combining static and dynamic items using conventional statistical methods showed a maximum cross-validated AUC of .82. Using the connectionist model however a substantial increase in accuracy was observed, AUC=.98, and this largely maintained when variations in time to recidivism were controlled. Variation to model parameters suggested that performance linked to the resources in the middle layer, responsible for modelling rare patterns and interactions between items. Model pruning confirmed that while the connectionist model exploited a wide range of variables in its classification decisions, the linear model was affected mainly by OGRS and a limited number of other variables. Results are discussed in terms of the theoretical and practical benefits of developing the use of connectionist models for better incorporating individuals’ dynamic risk and protective factors in recidivism assessments, and reducing the costs associated with false classifications.