Causal Inference for Randomized Experiments in Social Networks

• Activity Number: 347 - Recent Advances in Clustering and Mixture Models Analysis
• Type: Topic-Contributed
• Date/Time: Thursday, August 12, 2021 : 10:00 AM to 11:50 AM
• Sponsor: Section for Statistical Programmers and Analysts
• Author(s): David Choi*
• Companies: Carnegie Mellon University
• Keywords: causal inference; interference between units; social networks; randomized experiments

Abstract

In an randomized experiment where the assumption of "no interference between units" is violated, the treatment effect may spillover and indirectly affect other individuals in the study. For such experiments, such as those studying peer effects or herd immunity, it may be of interest to estimate the difference between the direct and indirect effects of treatment. To address this, we propose to estimate the difference in attributable effects for units that received different exposures to the treatment, with generalization to regression-based contrasts. For our approach, interval estimation requires no assumptions beyond randomization of treatment, even in the presence of unknown and unbounded interference between units. This is in contrast to other approaches that consider average or expected treatment effects rather than attributable ones, and require additional assumptions for interval estimation. As a result, our estimand may be of interest not only to describe variation in the treatment effect, but also as a baseline for experiments in which the underlying social mechanisms are poorly understood.