Gustavo Diaz (University of Illinois, Urbana–Champaign)
Abstract: Spillovers or interference feature in many questions of interest in the social sciences. Current approaches to analyze spillovers in experiments assume that the researcher observes all relevant networks, implying knowledge of how units are connected. However, social science theories rarely inform these exposure mappings, and while researchers can implement experiments with spillovers in mind, direct network manipulation is not always feasible. This project introduces a variable selection protocol within the context of supervised learning to inform modeling choices around spillovers in experiments and observational studies. The protocol proceeds in three steps. First, the researcher specifies a theoretically relevant pathway through which spillovers travel. Second, an adaptive lasso tunes how far away exposure to treated units affects the outcome of control units. Third, the researcher estimates or tests for spillovers using their preferred modeling strategy. I illustrate the effectiveness of this approach by reproducing the results of an experiment that captures the spillover effect of election observers on voter registration irregularities in Ghana.