Primary Submission Category: Sensitivity Analysis
Robust Causal Inferences from the Sensitivity Analysis of Multiple Estimands
Authors: Nathan Cheng, Jose Zubizarreta,
Presenting Author: Nathan Cheng*
When the assumption of unconfoundedness is suspect, one may reason about the robustness of a causal finding via a sensitivity analaysis. Traditional techniques in sensitivity analysis are typically designed to reason about a single estimand at a time. However, when the investigator is interested in a suite of hypotheses—answered by way of multiple estimands—and wishes to control some joint error rate, dependencies among the estimands can be leveraged to yield more precise assessments of sensitivity. In this work, we introduce an approach for the sensitivity analysis of multiple estimands in a general setting where weighting estimators are used to estimate causal quantities such as the average treatment effect (ATE). We show that useful sensitivity regions—high-dimensional sets that contain the unidentifiable ATE with high probability—can be tractably computed, and useful inferences can be extracted from them. Using simulations, we demonstrate the benefits of leveraging estimand dependence compared to naively combining one-at-a-time sensitivity analyses. Lastly, we describe some creative applications of our approach, and apply our method to real data for illustration.