Causal Inference in the Presence of Interference and Unmeasured Confounders

• Activity Number: 181
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Social Statistics Section
• Author(s): Cheng Zheng*+
• Companies: University of Washington
• Keywords: Causal Inference ; Interference ; Confounding ; Mediation Analysis ; Instrumental Variable

Abstract

In this work, we discussed how to make causal inference when both interference between subjects and unmeasured confounder exist. An important assumption made for Rubin's original causal framework is the requirement of no interference between subjects. This assumption is less plausible when the outcome of interests is social behaviors. VanderWeele (2013) proposed a method for mediation analysis when the spillover effects exist. However, no unmeasured confounding assumption was used. In this work, we proposed a model that allows the existence of both unmeasured confounder and interference. Using instrumental variable, we derived a consistent estimator for both direct and spillover causal effects of the post-randomization variable on the outcome under a quite general model. The identifiability assumptions and asymptotic property were discussed. Using efficacy study under two-stage randomization and mediation analysis under group randomization as examples, we discussed the implication of identifiability assumption in specific forms and performed simulation to show our proposed estimator is well behaved with finite sample size.