A0631
Title: Estimating the prevalence of peer effects in network experiments
Authors: David Choi - Carnegie Mellon University (United States) [presenting]
Abstract: In randomized experiments with arbitrary and unknown interference, such as social network settings where the treated and control units may affect each other, recent work has proposed hypothesis tests for the null hypothesis of no interference, i.e., that unit is unaffected by the treatment of others. However, without further assumptions, rejection of this null only implies that at least one individual was affected by a treatment other than their own. It is shown that these tests can be inverted with no assumptions on the nature of the interference, producing one-sided interval estimates (or lower bounds) not for the peer effect itself but rather for the number of units affected by the treatment of others. This does not fully identify a peer effect but may be used to show that it exists and estimate whether it is widely prevalent.
