Randomization Inference of Heterogeneous Treatment Effects under Network Interference

TL;DR

This paper introduces a new randomization-based testing method for assessing heterogeneous treatment effects in networked populations, addressing challenges posed by network interference and non-sharp null hypotheses.

Contribution

It develops a novel testing procedure that constructs data-dependent focal sets and allows variation across focal units, improving inference under network interference.

Findings

The proposed method controls size asymptotically under general conditions.

Application to real network data demonstrates practical utility.

Monte Carlo simulations show improved power over existing methods.

Abstract

We develop randomization-based tests for heterogeneous treatment effects in the presence of network interference. Leveraging the exposure mapping framework, we study a broad class of null hypotheses that represent various forms of constant treatment effects in networked populations. These null hypotheses, unlike the classical Fisher sharp null, are not sharp due to unknown parameters and multiple potential outcomes. Existing conditional randomization procedures either fail to control size or suffer from low statistical power in this setting. We propose a testing procedure that constructs a data-dependent focal assignment set and permits variation in focal units across focal assignments. These features complicate both estimation and inference, necessitating new technical developments. We establish the asymptotic validity of the proposed procedure under general conditions on the test statistic and characterize the asymptotic size distortion in terms of observable quantities. The procedure is applied to experimental network data and evaluated via Monte Carlo simulations.