Reproduce “Causal K-Means Clustering” on OpenResearch

About this paper

Kwangho Kim, Jisu Kim, Edward H. Kennedy

Causal K-Means reframes treatment-effect heterogeneity as clustering in estimated counterfactual outcome space, grouping individuals by similarity in their expected responses across treatments rather than by covariate similarity alone. Alongside a consistent plug-in method, the paper develops a cross-fitted influence-function estimator that can attain root-n convergence and support inference under suitable boundary and nuisance-estimation conditions.

Key insights

• Causal subgroups need not correspond to nearby points in covariate space; they can instead be defined by similar potential-outcome profiles.
• The influence-function formulation makes the clustering objective doubly robust in its leading behavior, with nuisance errors entering mainly through second-order products.
• The response vector is part of the scientific target: clustering on potential outcomes, treatment contrasts, or other counterfactual summaries emphasizes different notions of heterogeneity.

Summary by alphaXiv.