Primary Submission Category: Causal Inference in Networks
A Bayesian approach to Estimate Causal Peer Influence Using Latent Location for Unmeasured Confounding of Homophily
Authors: Seungha Um, Samrachana Adhikari,
Presenting Author: Seungha Um*
Researchers have been focused on estimating causal inferences to understand how individual’s behavior is influenced by the behaviors of their peers in observational studies on social networks. Identifying and estimating the peer influence, however, is challenging due to frequent confounding with homophily, where people tend to connect with those who share similar characteristics with them. Moreover, as the attributes driving homophily are generally not directly observed and serve as unobserved confounders, the identification and estimation of causal peer influence is not possible. In this paper, we address this challenge by leveraging latent locations inferred from the network itself to disentangle homophily from causal peer influence, and extend this approach to multiple networks by adopting a Bayesian hierarchical modeling framework. To model nonlinear response surfaces capturing the effective range of peer influence, we employ a Bayesian nonparametric model, specifically Bayesian Additive Regression Trees (BART). We propose an integrated Bayesian framework to account for the uncertainty in inferring latent locations and outcome modeling. We establish a nonparametric identification of causal peer influence in the presence of unmeasured network confounding without imposing fany parametric restrictions on the outcome model. To illustrate the applicability of our method in estimating causal peer influence, we utilize both simulated data and advice-seeking networks real data.