Estimation of Monotone Treatment Effects in Network Experiments

Randomized experiments on social networks are a trending research topic. Such experiments pose statistical challenges due to the possibility of interference between units. We propose a new method for estimating attributable treatment effects under interference. The method does not require partial interference, but instead uses an identifying assumption that is similar to requiring nonnegative treatment effects. Pre-treatment network observations can be used to customize the test statistic, so as to increase power without making assumptions on the data generating process. The inversion of the test statistic is a combinatorial optimization problem which has a tractable relaxation, yielding conservative estimates of the attributable effect.