SpatialMed-CF: A Unified Spatial Mediation Causal Forest Framework for Evaluating Heterogeneous Effects of Urban Greening Interventions

Urban greening interventions are widely implemented to mitigate urban heat islands, but rigorous causal evaluation is hindered by three challenges: interventions are not randomly assigned (selection bias), treatment effects are heterogeneous across urban contexts, and spatial spillover effects violate the Stable Unit Treatment Value Assumption (SUTVA). To address these challenges, we propose SpatialMed-CF, a spatial mediation causal forest framework that jointly estimates heterogeneous treatment effects, accounts for spatial interference through data-driven exposure mapping, decomposes total effects into direct and mediated pathways, and optimizes treatment allocation via policy learning. We validate SpatialMed-CF against seven baseline methods on synthetic panel data with known ground truth. Results show that SpatialMed-CF reduces Average Treatment Effect (ATE) estimation error by 86.1% and increases Conditional Average Treatment Effect (CATE) rank association approximately sixfold relative to standard causal forest. The full and no-spatial variants produced identical values in the original ablation, and a policy tree derived from heterogeneous treatment effects achieves a 9.5% welfare gain over random allocation, interpreted as an in-sample model-based contrast. In the added stress tests and the New York City illustration, spatial, mediation, and policy results varied with context and assumptions.