Title: Robust variance estimation when combining MI and IPW for missing data in EHR-based studies
Authors: Tanayott Thaweethai - Massachusetts General Hospital (United States) [presenting]
Sebastien Haneuse - Harvard TH Chan School of Public Health (United States)
Abstract: Due to the complex process by which electronic health records (EHR) are generated and collected, missing data is a huge challenge when conducting large observational studies using EHR data. Most standard methods to adjust for selection bias due to missing data fail to address the heterogeneous structure of EHR data. We consider a framework that modularizes the data provenance, or the sequence of specific decisions made by patients, health care providers, and the health system in which they interact, that leads to observing complete data in the EHR. Under this framework, one strategy is to combine inverse probability weighting and multiple imputation at different stages to address missingness. We propose an estimator based on this strategy, show that it is consistent and asymptotically Normal, and derive a consistent estimator of the asymptotic variance. Unlike Rubin's standard combining rules for multiple imputation, this variance estimator is robust to model misspecification of the selection, imputation, and analysis models. We demonstrate these methods in an EHR-based study of long-term weight change following bariatric surgery.