CLEAN: Confounding Latent Embedding Adjustment for Networks

Political scientists are often interested in studying the spread of behavior, attitude, or movementacross units (e.g., voters, legislators, countries), over some form of network (e.g., conversation, collaboration, alliances). To identify spread (i.e., contagion, diffusion), it is necessary to control for the fact that units that are alike, or naturally tend towards the same behaviors/attitudes, are more likely to tie in the first place (i.e., control for homophily). Separating contagion from homophily is a long-lastingchallenge in social networks research. Over the past few decades, the methodological literature on thistopic has indicated that it is very challenging to separate these two network effects from each other inobservational data. However, in a recent paper, McFowland and Shalizi (2023, Journal of the AmericanStatistical Association, 118(541), 707–718) propose a methodology to address this challenge by inferringand adjusting for latent variables that confound inferences. Through a method that draws upon common cluster-analysis approaches applied to network data, this method successfully separates social influence from homophily. We term this methodology confounding latent embedding adjustment for networks (CLEAN). We illustrate the properties of CLEAN via simulation, and further assess its statistical power. We then apply CLEAN to replication data from articles in international politics to test its substantive power. We conclude with a discussion on how this method can be effectively applied to the substantive scenarios where contagion and homophily are mixed in a complex manner.