Causal Inference with Interference

In observation studies, our target population may have a network structure. Interference may occur if the potential outcome of one unit depends on both individual treatment and neighborhood treatment. The present of interference will break down the traditional assumptions and framework. In this project, we review the paper by Forastiere et al.(2016), which provides a extended framework and solution for the existence of interference. The basic idea for the framework and assumption extension is to take neighbors into consideration, which means to extend potential outcomes, assignment mechanism and average does-response function as a function of both individual and neighborhood treatment. Accordingly, propensity score is generalized to a joint propensity score then decomposed to individual and neighborhood propensity score, from which a estimating procedure is proposed. Here we focus only on causal inference on main effect (individual effect) when interference presents. We reproduced the simulations by Forastiere et al. (2016) based on Facebook data obtained from Stanford Network Analysis Project (SNAP) (Leskovec and Mcaule, 2012).