Primary Submission Category: Causal Inference in Networks
Representational Power of Exposure Mapping Functions in Network Interference
Authors: Yuchen Xiao, Cowrin Zigler,
Presenting Author: Yuchen Xiao*
We study the representational power of exposure mapping functions (EMF) used in network interference. The traditional SUTVA assumption does not hold because, under interference, the potential outcomes of a unit depend on its treatment as well as on the treatments of other units. The usual approach is to embed the neighborhood treatments into a lower dimensional representation with EMF. The most common EMFs typically assume neighborhood treatments share equal weights, limiting their ability to represent some forms of interference. For example, a person’s health status may be more heavily influenced by some contacts who are encountered more frequently than others and people with the same number of treated neighbors but different number of neighbors should be expressed differently. We use Graph Attention Networks (GAT), which employ a self-attention strategy, to compute the attention coefficients (i.e., weights) of neighboring units to each unit in a network. One of the advantages of GAT is the range of interference can be extend to k-hops away, which permits long-range causal dependence. After assigning weights to neighboring units, we compute treatment and spillover effects using a generalize propensity score approach that has been used previously with exposure mappings. We detail the GAT architecture and its relative computational advantage. Finally, we make comparisons between the results estimated with GAT and other common exposure mapping functions.