Integrating Mental Health and Juvenile Justice Outcomes: A Case Study of Model-Agnostic Interpretable Machine Learning

When the justice system places juveniles into correctional institutions and recidivism rates rise, society bears substantially higher costs compared with when juveniles are placed at home or in community-based alternatives. Research indicates that mental health conditions, rather than public safety concerns, drive decisions to send juveniles to institutions. However, juvenile justice and mental health data are rarely combined for analysis, depriving policymakers of crucial insights to improve juvenile justice outcomes. Identifying the predictive factors of institutional placement is crucial to developing targeted early interventions. We introduce marginal contribution weighting (MCW), a method that compares feature importance across predictive models using Shapley additive explanations (SHAP) values to inform policy decisions. MCW preserves Shapley properties and maintains factorial weighting for marginal contributions, allowing a fair comparison of feature contributions. We apply MCW to a case study in the U.S. state of Texas that integrates mental health and juvenile justice data to predict court-ordered out-of-home placements. We develop a machine learning model using juvenile justice data and then incorporate mental health treatment records to evaluate the impact of new mental health features on accuracy and explainability. By systematically quantifying the impact of mental health features using MCW and SHAP-based analysis, our approach provides a rigorous framework for assessing feature importance across models. This methodology highlights the predictive value of mental health treatment data and ensures fairness in feature comparisons. By identifying key risk factors, our approach provides justice, law enforcement, and correctional policymakers with data-driven insights to support preventive, community-based strategies and mitigate the adverse effects of institutionalization on youth development and long-term outcomes.

History: This paper has been accepted by Kaushik Dutta for the Special Issue on the Responsible AI and Data Science for Social Good.

Supplemental Material: The software that supports the findings of this study is available within the paper and its Supplemental Information (https://pubsonline.informs.org/doi/suppl/10.1287/ijoc.2024.0663) as well as from the IJOC GitHub software repository (https://github.com/INFORMSJoC/2024.0663). The complete IJOC Software and Data Repository is available at https://informsjoc.github.io/.