Primary Submission Category: Interference and Consistency Violations
Learning Exposure Mapping Functions for Inferring Heterogeneous Peer Effects
Authors: Shishir Adhikari, Sourav Medya, Elena Zheleva,
Presenting Author: Shishir Adhikari*
In networked settings, individuals are influenced by the actions or behaviors of peers, such as smoking habits or vaccination preferences, yet how this influence aggregates into a composite measure of peer exposure is unknown. Existing methods for estimating peer effects rely on hand-crafted exposure mappings, such as the proportion of treated peers, implicitly assuming simplistic and often unrealistic influence mechanisms. These mappings can be misspecified and lead to biased causal effect estimates. Our work moves away from explicitly defining an exposure mapping function and instead introduces a framework that learns this function automatically. We propose EgoNetGNN, a graph neural network for heterogeneous peer effect estimation that learns the exposure mapping directly from data. The model captures complex peer influence mechanisms involving not only peer treatments but also attributes of the local neighborhood, including node, edge, and structural features. We demonstrate that misspecified mappings and naive learning strategies lead to substantial bias, while EgoNetGNN adapts to diverse and complex influence mechanisms. Across a wide range of synthetic and semi-synthetic experiments, our approach consistently achieves lower estimation error than state-of-the-art baselines. Our results suggest a shift from designing exposure mappings to learning them, to enable more robust causal inference under interference across public health, social science, and economics.