INTEGRATING SPATIAL CLUSTERING AND WEIGHTED DID FOR RETAIL INTERVENTION ASSESSMENT: A FRAMEWORK FOR SPILLOVER CONTROL AND STRATEGIC RISK EVALUATION

Abstract

Quantifying the causal impact of spatial treatments such as retail pricing changes, logistics incentives, or geo-promotions is challenging due to spatial interference, spillovers, and non-random treatment assignment. This paper develops a geo-based causal inference framework consisting of unsupervised clustering, spatial buffering, and econometric estimation for such challenges. Using hybrid PCA-based K-Means and DBSCAN clustering, covariate-balanced treatment-control spatial contiguous pairs are created using Haversine distance buffers for spillover control. The framework is illustrated on a large-scale retail intervention for which a national retailer removed pickup fees for store pickup orders. Analysis integrates store metadata and transaction data across several regions, encompassing over 10,000 retail outlets. Treatment impacts are estimated based on an Inverse Propensity Weighting (IPW) weighted Difference-in-Differences (DiD) model controlling time-invariant confounders. We find a 3.79% lift in sales, corroborating significant post-intervention gains and robust parallel trends. The paper contributes a spatial causal inference framework with a scalable framework for marketing, strategy, strategic risk, and policy evaluation, with accompanying spatial spillover management and market risk management insights for localised retail decision-making.