County-Level Surveillance and Prediction of Opioid Cost Burden Using Administrative Utilization Metrics and Social Determinants of Health

The opioid epidemic continues to impose substantial social and financial burdens across the United States. Few studies have quantitatively integrated pharmacy-benefit-management (PBM) levers with social determinants of health (SDOH) to understand geographic variation in opioid-related spending. This study develops an interpretable county-level machine-learning framework linking Medicare Part D opioid prescribing patterns (2013–2023) with socioeconomic, behavioral, and healthcare-access indicators from County Health Rankings. Using a Random Forest regression pipeline with automated feature engineering and five-fold cross-validation, the model achieved an R² of 0.97 and RMSE of $445, accurately predicting opioid cost per capita across U.S. counties. Utilization-driven variables—including cost per claim, opioid prescribing rate, and claim density—were the dominant predictors, while structural vulnerabilities such as unemployment, income inequality, and limited mental-health-provider availability contributed additional upward pressure. Counterfactual simulations quantified the fiscal effects of PBM strategies (formulary tightening, utilization guidance, and cost-per-claim reduction) and SDOH improvements (expanded primary care and behavioral-health capacity, reductions in smoking and obesity). PBM interventions produced short-term reductions in predicted opioid cost per capita, while SDOH improvements generated broader and more sustained declines. The combined PBM + SDOH scenario produced the largest modeled reduction nationally. This integrated framework demonstrates how coupling PBM program design with social and structural determinants can improve fiscal planning, target high-burden counties, and inform equitable opioid policy. This study evaluates whether publicly available PBM utilization indicators and county-level social determinants of health can jointly predict opioid cost burden and inform policy-relevant simulation scenarios.