Primary Submission Category: Causal Inference in Networks
Causal Inference With Contagion and Latent Homophily Under Full Interference
Authors: Yufeng Wu, Rohit Bhattacharya,
Presenting Author: Yufeng Wu*
Dependent data poses a serious challenge for valid causal inference. In some cases, we observe a single realization of a network of individuals, all of whom may depend on each other – a setting termed “full interference.” Tchetgen Tchetgen et al (2021) developed the auto-g computation method for computing network causal effects in this setting. Their method assumes that all relevant confounders are observed and that there is no dependence induced by unmeasured common causes between individuals. In other words, the method allows for contagion and interference effects, but assumes the absence of latent homophily. In this work, we propose a nonparametric test that can be used to distinguish between dependence due to contagion and latent homophily. This test acts as a verification tool for the auto-g computation method, providing a way to accept or reject its model assumptions. In cases where there is dependence due to latent homophily, the auto-g method produces biased estimates of network causal effects, so we propose a modified identification and estimation strategy for settings where dependence could be induced due to either latent homophily or contagion. We evaluate the effectiveness of our method through simulation studies and a real-world data application on social networks.
Eric J Tchetgen Tchetgen, Isabel R Fulcher, and Ilya Shpitser. Auto-g-computation of causal effects on a network. Journal of the American Statistical Association, 116(534): 833–844, 2021