Causal Inference for Observational Studies on Networks

Abstract

Causal inference methods are powerful techniques which allow the practitioner to estimate treatment effects while simultaneously injecting cause and effect into the problem at hand. However, when causal estimation procedures are applied to observational studies on a network, the connectedness of the network often results in problems. One major problem is that Stable Unit Treatment Value Assumptions (SUTVA) often do not hold as interference manifests itself through the network. I consider the theoretical conditions necessary to properly model interference and then use a new Bayesian propensity score estimator to obtain more accurate estimates of treatment effects. Another major problem I explore is the importance the covariates of a unit's neighbor might play when trying to estimate unit-level treatment effects. I demonstrate how bias results from neglecting to incorporate neighborhood covariates, and how the theoretical setup necessary to model them. These are supported with extensive simulations which provide new analytical ways to handle such networked setups.