Primary Submission Category: Heterogeneous Treatment Effects
A Practical Minimax Approach to Causal Inference under Limited Overlap
Authors: Yuanzhe Ma, Yian Huang, Hongseok Namkoong,
Presenting Author: Yuanzhe Ma*
A central challenge in observational analysis is that the effective sample size may be prohibitively small when there is little overlap between treated and control populations. Data sparsity becomes more pronounced in modern operational problems that involve high-dimensional covariates. Existing observational methods, which truncate data with extreme propensity scores, are only valid in the large sample limit and silently fail in practical instances with limited effective sample size. In this work, we propose a new inferential framework that provides always-valid uncertainty quantification, which provides a sensitivity analysis framework against unforeseen data sparsity. Our work builds on the theory of minimax estimation for linear functionals that can generate an always-valid confidence interval. We operationalize the above minimax approach by decomposing the data into overlap and non-overlap regions. We use standard asymptotic inference tools (e.g., AIPW) on the region of overlap, and only apply the conservative minimax approach described above on the non-overlap region while also utilizing information from the overlap region. Through simulated and real data, we demonstrate our method ensures robustness against unforeseen data sparsity by quantifying the bias induced by standard truncation techniques.