Advancing Predictive Modeling in Behavioral Health with Gradient-Boosted Bootstrap Modeling

• In Event: AERA Poster Session Thursday 7:45am
  In Poster Session: Poster Session: Statistical Theory, Quantitative Methods and Applications

Abstract

This study introduced a Gradient-Boosted Bootstrap Model (GBBM) to advance predictive modeling in behavioral health, focusing on treatment completion in substance use disorder (SUD) programs. Using data from 1,983 adults in a community-based program, the GBBM combined bootstrap resampling with gradient boosting to optimize the bias–variance tradeoff and improve model stability. SHapley Additive exPlanations (SHAP) enhanced interpretability, providing global and individualized insights into treatment retention. The GBBM achieved strong discrimination (AUC = 0.92) and calibration (Brier score = 0.09). Behavioral engagement indicators (e.g., attendance, client decision) predicted completion better than demographic or psychiatric variables, suggesting dynamic behavioral processes. The GBBM offers a transparent, data-driven approach to guide precision behavioral health and equitable strategies to improve retention.

Author

• Graham Zulu, University of Denver