Breaking the 'glass ceiling' of risk prediction in recidivism: an application of connectionist modelling to offender data

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