Description:

Xinran Li (University of Chicago): Randomization inference for distributions of individual treatment effects

Abstract: Understanding treatment effect heterogeneity is a central problem in causal inference. In this talk, I will present a randomization-based inference framework for distributions and quantiles of individual treatment effects. It builds upon the classical Fisher randomization test for sharp null hypotheses, and considers the worst-case randomization p-value for composite null hypotheses. In particular, we utilize distribution-free rank statistics to overcome the computational challenge, where the optimization of p-value often permits simple and intuitive solutions. By standard test inversion, we can further construct simultaneous confidence bands for the entire distribution of individual treatment effects. 

During the talk, I will first describe this finite-sample valid inference along with strategies for improving power, and eventually show the asymptotic sharpness of the inference. I will then extend the methodology to observational studies, where we perform sensitivity analysis for unmeasured confounding. I will finally consider sample attrition with informative missingness, and show that structural assumptions on missingness mechanisms can substantially improve inference power. The proposed methods will be illustrated with applications in education, public health, and social programs.