The Geography of Impact: Endogenous Spatial Clustering for Difference-in-Differences Estimation

Standard policy evaluation methods typically assume that treatment effects are homogeneous within fixed administrative units. However, the true policy relevant boundaries are typically unknown to the researcher, as latent territorial characteristics, such as institutional quality or local economic structure, generate unobserved spatial heterogeneity that does not align with administrative borders. To address this challenge, we propose a novel unsupervised learning algorithm that endogenously identifies geographic regimes heterogeneous in terms of causal impact. Unlike existing clustering methods that group units based on geometric density or outcome similarity, our approach partitions spatial units specifically on the basis of their causal response to treatment. By explicitly maximizing treatment effect variance subject to spatial coherence, we identify where policies have differential impacts, recovering latent economic boundaries while maintaining identification requirements. We validate the estimator through Monte Carlo simulations, demonstrating its robustness in recovering latent economic structures even in high-noise environments. Finally, we apply the method to analyse the local labour market effects of the 2001 Chinese import competition shock in the United States, revealing distinct latent spatial regimes of industrial resilience that cut across state lines.