Primary Submission Category: Generalizability/Transportability
Estimating Causal Effects with Error-Prone Exposures Using Control Variates
Authors: Keith Barnatchez, Kevin Josey, Rachel Nethery, Giovanni Parmigiani,
Presenting Author: Keith Barnatchez*
Exposure measurement error poses a common, yet often ignored, challenge to performing causal inference in observational studies. Existing methods accounting for exposure measurement error largely rely on restrictive parametric assumptions for not only the measurement error mechanism, but also the outcome and exposure models used to estimate a causal effect. There remains a critical need for assumption-lean estimation methods that can flexibly accommodate different study designs while possessing desirable theoretical properties. In this paper, we address these needs by proposing estimators based on the control variates framework of Yang and Ding (2020). Drawing connections between the measurement error, generalizability and transportability, and missing data literatures, we show that our approach can be implemented in various two-stage study designs–where one obtains gold-standard exposure measurements for a small subset of the initial study sample–to address biases induced by measurement error for estimating general causal quantities. Under standard causal inference assumptions, our method inherits desirable double-robustness properties, including scenarios where the two-stage sampling probabilities are unknown. Through simulation studies, we show our approach performs favorably to leading methods under various two-stage sampling schemes. Finally, we test our method on observational electronic health record data on HIV outcomes from the Vanderbilt Comprehensive Care Clinic.