Primary Submission Category: Randomized Designs and Analyses
Randomization Tests for Distributions of Individual Treatment Effects using Multiple Rank Statistics
Authors: David Kim, Yongchang Su, Jake Bowers, Xinran Li,
Presenting Author: Jake Bowers*
In this paper we study quantiles of individual treatment effect. Recent developments on randomization-based approaches provides finite-sample valid inference for quantiles in both completely randomized and stratified randomized experiments. However, since previous methods are using Stephenson’s Rank Statistic, where the most powerful hyperparameter choice is not fixed, there exists a burden of choices to use them. We propose combining polynomial rank sum statistic to enhance the power of existing approaches. Combining rank scores offers much more power to detect rare treatment effects or common treatment effects in long-tailed outcomes than single rank scores-based testing procedures such as the Wilcoxon Rank test and the Stephenson’s Rank Test. Moreover, using Polynomial functions of rank scores further increases power and eases application of this approach from completely randomized experiments to strata-randomized experiments.
Â