Characterizing Interference Heterogeneity and Improving Estimation for Experiments in Networks

Randomized control trials, or “A/B tests”, have been crucial to understanding the impact of an intervention. Traditional causal inference methods rely on a critical assumption called the “stable unit treatment value assumption” (SUTVA) (Fisher, 1937; Splawa-Neyman et al., 1990), which means that a unit’s treatment response only depends on their own treatment assignment. However, SUTVA is an unrealistic assumption in settings such as networks, where a unit’s response may be affected by other units. This violation of SUTVA is referred to as network interference. The current literature on network interference has two major limitations — failing to account for social theories of interference (e.g. structural diversity or social contagion) and relying on human experts to model interference patterns. To tackle these issues, we propose a two-part machine learning approach to automatically characterize network interference conditions based on both the local network structures and the treatment assignments among users in the network neighborhood. Specifically, we first construct network motifs with treatment assignment information, referred to as causal network motifs, to characterize the network interference conditions for each unit. We then develop machine learning methods based on decision trees and nearest neighbors to map these causal network motif representations to an “exposure condition” under the framework proposed by Aronow and Samii (2017). We demonstrate the validity of our approach through reanalysis of prior experimental study data. Our results show that causal network motifs are able to more accurately account for complex interference patterns and reduce bias in treatment effect estimation even with the presence of interference