Bridging prediction and causality in environmental justice: Evaluating superfund impacts with explainable AI and econometrics

Abstract

This study presents an integrated framework combining causal inference and explainable machine learning to evaluate both the determinants and effects of environmental remediation under the U.S. Superfund program. Using a nationwide panel of census tracts from 1970 to 2019, we apply a doubly robust Difference-in-Differences (DiD) model with inverse probability weighting to estimate the causal impact of Superfund designation on neighborhood socioeconomic outcomes. Concurrently, we develop a predictive model using XGBoost and SHAP to uncover the structural features that drive site designation. Our findings reveal that Superfund sites are disproportionately placed in tracts with lower income, older housing stock, higher industrial activity, and greater minority presence—patterns consistent with environmental injustice. Post-treatment analysis shows moderate but significant gains in income, housing turnover, and labor participation. By comparing predictive salience with causal responsiveness, we identify aligned, divergent, and mixed features—highlighting where policy is both well-targeted and effective, and where structural inequalities persist. This dual-inference approach offers novel insight into the design, evaluation, and equity of environmental interventions.