Primary Submission Category: Causal Inference in Networks
Simulating Potential Outcomes in the Presence of Interference
Authors: Gabrielle Lemire, Ashley Buchanan, Natallia Katenka, Tingfang Lee,
Presenting Author: Gabrielle Lemire*
Simulation studies aid in understanding the behavior of statistical methods, such as measuring the finite sample performance of estimators. This is possible because we know the ground truth as we are specifying the data generating mechanism. When evaluating causal estimators, simulating potential outcomes allows the researcher to overcome both challenges due to the impossibility of perfect data collection and knowledge of the data generation mechanism, but also to have full knowledge of the counterfactual outcomes. Additional challenges arise for simulating potential outcomes when observations’ outcomes are dependent due to interference (i.e., when one individual’s exposure affects another’s outcome). We provide 1) a paradigm that considers the assumptions about the nature of the interference structure, the identification assumptions required for causal effects, the estimators being evaluated, and the computing environment; and 2) efficient tools (i.e., R functions/package available on Git Hub) for simulating potential outcomes in the presence of interference which is designed to handle both network structures with interference and cluster structures satisfying the partial interference assumption (i.e., interference only occurs within groups). Our work aids in bridging the gap between the burgeoning literature offering new causal estimators in the presence of interference and the need for reproducible evaluation of their performance in practice.