A1617
Title: Hypothesis testing for detecting network interference in randomized experiments
Authors: Christopher Harshaw - Columbia University (United States) [presenting]
Chao Gao - University of Chicago (United States)
Fredrik Savje - Yale University (United States)
Yitan Wang - Yale University (United States)
Abstract: Across disciplines, the randomized experiment is a widely used method for estimating causal effects. In order to meaningfully estimate causal effects, experimenters have to assume structure on the potential outcome functions. The simplest assumption is that of no-interference, although a growing body of work has provided various network interference models under which causal effects can be estimated. Still, this raises the question: to what extent can these assumptions be tested? For example, can interference be detected when it is present in the experiment? Can one test between stronger and weaker models of network interference? To this end, previous work has focused on controlling Type I error, but Type II error of proposed methods remains less well understood. Hypothesis testing frameworks for detecting interference are further investigated. The main contribution is a series of negative results. Roughly speaking, it is shown that there is no test for detecting interference which has small Type I and Type II error, even as the sample size increases. This finding has serious consequences for the epistemological nature of the no-interference and various network interference assumptions: namely, that they are, in a rigorous sense, untestable. On the positive side, it is shown that there exists a uniformly consistent test of no-interference when the alternative hypotheses are restricted to a linear-in-means Type model and the graph is sufficiently sparse.
