Primary Submission Category: Interference and Consistency Violations
Doubly Robust Estimators for Controlled Direct Effects in the Presence of Interference
Authors: Jimmy Kelliher, Nandita Mitra,
Presenting Author: Jimmy Kelliher*
Controlled direct effects are important causal estimands for public health scientists and policymakers interested in understanding the mechanisms by which a treatment causes an outcome. However, in both clinical trials and in observational settings, interference can pose a threat to effect identification. In this paper, we extend the notion of an exposure mapping to that of a generalized counterfactual mapping, in order to accommodate interference structures for nested counterfactuals. In particular, we allow for interference in exposure-outcome, exposure-mediator, and mediator-outcome relationships in a difference-in-differences setting. After establishing identification results, we further develop doubly robust, semi-parametric efficient estimators for the controlled direct effect when counterfactual mappings are correctly specified. We then assess the small-sample performance of these estimators in various simulation settings. Finally, we apply these methods to estimate the controlled direct effect of the 2017 Philadelphia beverage tax on the volume sales of sweetened beverages, which may be mediated by price changes.