Primary Submission Category: Causal Inference and SUTVA/Consistencies Violations
Estimating Spillover Effects in Longitudinal Data under Unknown Interference
Authors: Ye Wang, Michael Jetsupphasuk,
Presenting Author: Ye Wang*
In longitudinal data where units are situated in a space or social network, the outcome of one unit may not only be affected by its own treatment assignment history, but also the treatment assignment histories of others. The presence of interference raises a significant challenge for researchers aiming to discern the direct and spillover effects of treatments across these dimensions, given that the interference structure—how an observation’s outcome is influenced by the treatment status of others—is often unknown to the researcher. In this paper, we put forward a design-based framework that combines marginal structural models with recent advancements in the literature on interference to address these complexities. We define estimands that enable researchers to separate direct effect of the treatment from spillover effects under unknown interference structures. We then develop estimators that are consistent for these estimands and asymptotically normal, assuming sequential ignorability and mild constraints on the extent of dependence caused by interference. Additionally, we introduce methods for constructing valid confidence intervals. Unlike existing approaches based on exposure mapping, our method circumvents the challenge of calculating exposure probabilities in longitudinal settings. Its effectiveness is demonstrated through simulations and replications of two empirical studies.