A1468
Title: Mixed-effects and latent space approaches for causal inference under network interference
Authors: Vanessa McNealis - University of Glasgow (United Kingdom) [presenting]
Erica Moodie - McGill University (Canada)
Nema Dean - University of Glasgow (United Kingdom)
Abstract: Causal analyses have grown popular, yet most methods rely on a no-interference assumption, in which an individual's outcome is unaffected by the exposures of others. However, some exposures exhibit spillover, affecting both the direct recipient and their social contacts, as in a sexual health promotion intervention delivered in schools. Estimating causal spillover is complicated by homophily, where individuals connect with others sharing similar characteristics, and by unmeasured cluster-level factors. The purpose is to discuss recent advances addressing unmeasured homophily, contextual confounding, and realistic evaluation. First, a Bayesian joint mixed-effects framework for clustered network data is highlighted, that combines outcome and exposure models with direct standardization to enable causal estimation under cluster-level confounding. Second, latent space approaches to construct estimators of direct and spillover effects are described, accounting for uncertainty in latent positions in the social space and allowing flexible outcome models. Finally, developed plasmode simulations are presented to evaluate causal inference methods on realistic social networks. Together, these approaches provide tools for causal inference under interference, accounting for homophily and unmeasured confounding.
