Optimizing Juvenile Assessment Performance

The variations were 1) item selection technique, 2) response weighting, 3) gender responsivity, 4) race-ethnicity neutrality, 5) outcome specificity, 6) prediction duration, and 7) jurisdiction variation. The combined effects of multiple variations were also examined as an eighth variation. The project used a 10-state sample of youth, who were assessed using the same assessment. The project developed risk assessment models using the seven development methods. Where required, boosted regression models were used to identify predictive items and provide coefficient weights. In addition, several sub-samples were developed to examine and compare approaches between gender and race/ethnic groups. Comparisons were made between the 10-site unified sample and models created to capture site differences. To measure model performance, k-fold validation was completed, and industry standard predictive performance metrics were developed. Findings identified consistent and substantial improvements with each of the eight hypothesized variations. Outcome and jurisdiction-specific models identified just over a full effect size improvement. In addition, the project identified an optimized set of models for the 10 sites. These were customized tools based on each data set. They can be implemented to improve predictive performance. Recommendations focus on how the field can make similar adjustments to off-the-shelf tools in optimizing predictive performance. 41 tables, 1 figure, and 99 references