Summary
Aim
To assess whether Youth Level of Service/Case Management Inventory Australian Adaptation (YLS/CMI-AA) risk/ needs data improve recidivism prediction for young offenders under community supervision, compared to static risk data from the Bureau’s Reoffending Database (ROD).
Method
The analysis included all 1,050 young offenders who commenced a supervised community order (other than bail or parole) in 2014 with a valid YLS/CMI-AA and ROD record. Recidivism was defined as a new proven offence within 12 months of order commencement. Logistic regression assessed the individual and collective relationships of static risk factors and YLS/CMI-AA scores to recidivism. Area Under the Curve (AUC), model fit indices and multiple cross-validation methods were used to evaluate the models.
Results
Interactions between variables in models built with the full sample necessitated that separate models be built for Indigenous and non-Indigenous offenders. For non-Indigenous offenders, the AUC for the combined (ROD with YLS/CMI-AA) model (.767, 95% CI (.728, .807)) was within the acceptable range (0.7-0.8) but did not significantly outperform the ROD-only model (.740, (95% CI .698, .781)). For Indigenous offenders, AUCs were significantly lower than for nonIndigenous offenders, below the acceptable range, and also showed no significant benefit from combining YLS/CMI-AA and ROD data. Compared with AUCs for the combined model, cross-validated AUCs were lower, and corresponding AUCs for the 2013 cohort were inconsistent.
Conclusion
YLS/CMI-AA data did not significantly improve the predictive accuracy of static risk-based models of recidivism for Indigenous or non-Indigenous offenders. Validation methods suggested that the results may not generalise beyond the current cohort.