Randomization Inference of Heterogeneous Treatment Effects under Network Interference

2023-07-31

AI Summary

To address the challenge of balancing Type I error control and statistical power in heterogeneous treatment effect inference under network interference, this paper embeds network interference into the potential outcomes framework and formalizes the mechanism by which units are affected by their neighbors’ treatments via an exposure mapping. We then propose a conditional randomization test based on functions of the treatment assignment, transforming non-sharp null hypotheses into testable sharp nulls—thereby extending the applicability of randomization inference to complex dependency structures. We establish rigorous theoretical guarantees for the method’s statistical validity under interference. Monte Carlo simulations and analyses of real-world network data demonstrate that the proposed approach robustly controls Type I error across varying interference strengths while substantially improving statistical power compared to existing methods.

Abstract

We design randomization tests of heterogeneous treatment effects when units interact on a single connected network. Our modeling strategy allows network interference into the potential outcomes framework using the concept of exposure mapping. We consider a general class of null hypotheses -- representing different notions of constant and no treatment effects -- that are not sharp due to unknown parameters and multiple potential outcomes. To make the nulls sharp, we propose a conditional randomization method that expands on existing procedures. Our conditioning approach permits the use of functions of treatment as a conditioning variable, widening the scope of application of the randomization method of inference. We show that the resulting testing procedures based on our conditioning approach are valid. We demonstrate the testing methods using a network data set and also present the findings of an extensive Monte Carlo study.

Problem

• Tests for heterogeneous treatment effects under network interference
• Addresses non-sharp null hypotheses in networked populations
• Proposes valid inference method with data-dependent assignment sets

Innovation

• Randomization-based tests for heterogeneous treatment effects
• Data-dependent focal assignment set construction
• Asymptotic validity under general test statistic conditions