Primary Submission Category: Causal Inference in Networks
Inverse probability weighting-based bias correction methods for causal effects estimated under misspecified interference sets
Authors: Ariel Chao, Donna Spiegelman, Ashley Buchanan, Laura Forasteire,
Presenting Author: Laura Forastiere*
Interference is often present in randomized or observational studies, where one participant’s exposure to the intervention may affect another’s outcome. To estimate causal effects, for each participant we must specify an interference set, that is, the set of those whose exposure may affect that participant’s outcome. Interference sets are conventionally assumed to be correctly specified; however, they are prone to misspecification. For example, under the partial interference assumption, interference is assumed to be contained within well-separated clusters, yet social interactions may extend across them but are falsely assumed away. In this paper, we show that when interference sets are misspecified, causal effects estimated by an inverse probability weighting (IPW) estimator are biased. In HIV studies where social behaviors drive disease transmission, correcting causal effects for bias is crucial for accurate evaluation of interventions. We propose IPW-based bias-correction methods when a validation study containing data on the true interference sets is available, and extend these methods to the setting where multiple surrogates of the interference sets may be observed. We assessed finite sample properties of our methods in a simulation study, and applied the methods to the Botswana Combination Prevention Project, where clusters defined by two geographical boundaries were regarded as the surrogate interference sets, and phylogenetic data were used to define the true ones.