Exposure Mapping Function Learning for Peer Effect Estimation
Keywords: peer effects, causal inference, network effects, interference, heterogenous influence, exposure mapping
TL;DR: We propose learning exposure mapping function that summarizes neighborhood exposure for peer effect estimation
Abstract: In causal inference involving interacting units (e.g., individuals in a contact network), peer effects quantify how the actions or behaviors of peers (e.g., wearing a mask) affect an individual's outcome (e.g., viral infection). Measuring peer effects involves defining exposure mapping function that outputs peer exposure, a high-level causal variable summarizing peer treatments (or interventions), and estimating the difference in counterfactual outcomes for different peer exposures. Most of the existing approaches for defining exposure mapping functions consider homogeneous influence from peers and use peer exposure based on the fraction of treated peers. There is a growing interest in work that acknowledges heterogeneous influence among units (e.g., due to local neighborhood structure) and captures those influence mechanisms by automatically learning exposure mapping function. Recently, graph neural networks (GNNs) have been extensively used for causal effect estimation in networks, but their use has been mostly limited to automatic feature aggregation and addressing confounding. This work explores the capabilities of GNNs to automatically capture peer influence based on local neighborhood structure. We show GNNs using homogeneous peer exposure or GNNs learning peer exposure naively face difficulty capturing such influence mechanisms. To address this issue, we propose EgoNetGNN to learn exposure mapping function by capturing peer influence mechanisms based on local neighborhood structure. We show that our approach reduces the error in estimating peer effects using synthetic network models.
Submission Number: 30
Loading