Abstract
Opioid use disorder (OUD) is a chronic condition affecting more than 2.1 million individuals in the United States and over 16 million worldwide. The HEALing Communities Study® aimed to reduce opioid overdose deaths through data-driven interventions. In this study, we evaluate alternative statistical frameworks for interpreting community-level data on three interrelated outcomes—opioid overdose deaths, substance use treatment, and naloxone (Narcan) administration. Given their dynamic interdependence over time, these outcomes were modeled using panel vector autoregression model estimated using both the Maximum Likelihood Estimator (MLE) and the Generalized Method of Moments (GMM). The models incorporate geospatial effects and social determinants of health to capture spatial dependencies and societal factors influencing opioid-related outcomes. Simulation studies were conducted to assess model selection and compare estimator performance under varying conditions. Additionally, out-of-sample forecasting analyses were performed to evaluate predictive accuracy across modeling approaches. Findings indicate that the MLE provides more stable parameter estimates and superior forecasting performance relative to GMM methods in this application.