Primary Submission Category: Causal Inference in Networks
Covariate Adjustment in Randomized Trials under General Interference
Authors: Ralph Trane, Hyunseung Kang,
Presenting Author: Hyunseung Kang*
Covariate adjustment has been a popular approach to improve precision and power in randomized trials. However, in some randomized trials say vaccine trials or evaluation of new educational programs, study units often interact and affect each other’s responses, a phenomena known as general interference. While recent work have shown how to identify and estimate treatment effects in randomized trials under general interference, to the best of our knowledge, there is little work on how to properly adjust for covariates in this setting, especially to deal with the potentially complex and unknown dependencies between study units. In this paper, we propose a class of flexible, covariate-adjusted estimators of treatment effects under general interference. Under some smoothness conditions on the response model, our estimators are consistent, asymptotically Normal, and can incorporate some, but not all, types of flexible methods from machine learning. An important corollary of our result is that ANCOVA, a popular method of covariate adjustment in randomized trials under no interference, yields a consistent, asymptotically Normal, covariate-adjusted estimator of treatment effects under general interference. Our results are demonstrated through a simulation study under some popular network models and an empirical study. We conclude by providing concrete guidelines to practitioners on how to adjust for covariates in randomized trials under general interference.