A Multiply-Robust Method to Handle Missing Confounder in Observational Studies

• Activity Number: 456
• Type: Invited
• Date/Time: Wednesday, August 7, 2013 : 8:30 AM to 10:20 AM
• Sponsor: ENAR
• Author(s): Lingling Li*+ and Changyu Shen and Xiaochun Li and James Robins
• Companies: Harvard Medical School and Harvard Pilgrim Health Care Institute and Indiana University School of Medicine and Indiana University School of Medicine and HSPH
• Keywords: multiply robust ; doubly robust ; observational data ; missing covariate ; comparative effectiveness ; electronic healthcare database

Abstract

Large existing healthcare databases present not only opportunities but also challenges for comparative effectiveness research. Confounding bias and missing data are two main issues. We developed a multiply-robust method to tackle both issues in a unified manner. The new method uses the doubly-robust theory for causal inference and missing data models in a nested fashion and constructs a multiply-robust estimator in two stages. The multiply-robust estimator depends on four working parametric models for (i) the conditional mean of the outcome given exposure and confounders, (ii) the conditional mean of exposure given confounders (i.e., propensity score), (iii) the conditional probability of observing full data given the observed data, and (iv) the conditional distribution of the potentially missing confounder given the observed data. Under the assumptions of no unmeasured confounder and missing at random, the multiply-robust estimator is consistent if any of the following four conditions holds: working models (i) and (iii) are correct, working models (i) and (iv) are correct, working models (ii) and (iii) are correct, and working models (ii) and (iv) are correct.