East Coast Ballroom
Causal Clustering: A new approach to analysis of treatment effect heterogeneity (306644)
*Kwangho Kim, Carnegie Mellon UniversityEdward Kennedy, Carnegie Mellon University
Jisu Kim, Inria Saclay
Larry Wasserman, Carnegie Mellon University
Keywords: causal inference, heterogeneity of treatment effects, clustering, conditional average treatment effects, observational study, outcome-wide study, multiple treatment
We develop \textit{Causal Clustering}, a new framework for the analysis of treatment effect heterogeneity by leveraging tools in clustering analysis. We pursue an intuitive way of discovering subpopulations with similar treatment effects - viewing each of them as a separate cluster - by harnessing widely-used clustering algorithms. Therefore, it provides an efficient way to not only investigate the existence of treatment effect heterogeneity but also ascertain subpopulations for which the treatment is more effective or harmful. Our proposed framework has several benefits over other supervised learning-based approaches, and is particularly useful for multiple-treatment or outcome-wide studies. We show that three widely-used clustering methods can be successfully adopted into our framework, only at additional cost of estimating nuisance regression functions for the outcome process without harming existing theoretical properties of each clustering algorithm. Importantly, we develop a novel estimator that attains faster convergence rates in k-means causal clustering based on semi-parametric theory. We apply our methods in a study of the effect of low-dose aspirin on pregnancy outcome.