Connor Jerzak (Harvard University)
Abstract: With the rise of online social networks, there has been growing interest in modeling how experimental units influence each another---a phenomenon known as "interference'' in the causal inference literature. Current models for interference generally requires knowledge of the way in which units are connected. Yet, in most field experiments, such data is unavailable. In this paper, we propose a general method for detecting and characterizing interference when the relationship data is unobserved. We provide analytical guarantees of the method's performance and illustrate its behavior on simulated data. We apply the method to ten political science field experiments---detecting interference in about half of them---and summarize the kinds of social dynamics suggested by our results. We validate our results by comparing the network inferred by our algorithm with one observed in the context of a school experiment.