A statistical model for predicting risk of re-imprisonment The Criminal Re-imprisonment Estimate Scale (CRES) A statistical model for predicting risk of re-imprisonment
CSNSW Vision
LSI-R
Percentage of Recidivists and Non-Recidivists Administered an LSI-R Current Practice Percentage of Recidivists and Non-Recidivists Administered an LSI-R False Positive True Positives False Positives
Risk Screening Tools LSI-R: SV OASys (ORGS) RoR-PV GRAM ROC ROI
The Criminal Re-imprisonment Estimate Scale (CRES)
Model Development 23,000 DYNAMIC STATIC EMPLOYMENT EDUCATION HOUSING SUBSTANCE MENTAL HEALTH STATIC GENDER INDIGENOUS AGE SENTENCE LENGTH TIME IN COMMUNITY ADAPTED COPAS RATE PRIOR INCARCERATION OFFENCE 23,000
Adapted Copas rate n = # full-time custodial sentences t = age at end current imprisonment – age first adult imprisonment +5 makes the distribution closer to normal and makes it comparable for those offenders who are at the beginning of their criminal career.
Final Model Adjusted Odds Ratios of Re-imprisonment
Area under the curve indicated acceptable fit for the model Model Adequacy ROC AUC Area under the curve indicated acceptable fit for the model auc = 0.79 CRES LSI-R (Watkins, 2011) Female Indigenous Australian .712 .597 Female Non-Indigenous Australian .784 .687 Male Indigenous Australian .754 .655 Male Non-Indigenous Australian .788 .694
Reimprisonment by predicted probability % not re-imprisoned % re-imprisoned
Application of the Criminal Re-imprisonment Estimate Scale (CRES) to CSNSW Offender Management
Model Thresholds Optimal Threshold True positives vs False Positives
Model Thresholds
Classification Accuracy Screening Tool Number of Released Inmates with an LSI-R % Sensitivity (True Positives) Specificity (True Negatives) Not Applied 15317 (67) 68 34 >=.15 19337 (84) 97 25 >=.25 15322 89 50 >=.35 12972 (56) 81 62 False Positive
Classification Accuracy Current Practice CRES >=.25 Total Inmate Population 23,000 Total Inmate Population 23,000 LSI-R Administration 15,317 LSI-R Administration 15,317 Recidivists 9,826 Recidivists 9,826 True Positive 68% False Positive 64% True Positive 89% False Positive 50%
Reductions in re-imprisonment Conclusions and Implications Application of the CRES model to practice Redistribution of resources to higher risk offenders Reductions in re-imprisonment
Questions?
Classification Accuracy Screening Tool Number of LSI-R Administrations % Sensitivity Specificity Positive predictive value Negative predictive value Not Applied 15317 (100) 94 24 49 85 >=.10 13975 (91) 93 36 51 89 >=.15 12901 (84) 92 38 53 86 >=.2 11793 (77) 45 55 False Positive