Sequential BART for Imputation of Missing Covariates in the Presence of Auxiliary Covariates

• Activity Number: 551 - New Innovations in Handling Incomplete Biomedical Data in the Era of Data Science
• Type: Invited
• Date/Time: Wednesday, August 2, 2017 : 2:00 PM to 3:50 PM
• Sponsor: International Indian Statistical Association
• Author(s): Dandan Xu* and Michael J Daniels
• Companies: The University of Texas at Austin and University of Texas at Austin
• Keywords: Auxiliary variable MAR ; Multiple imputation ; Compatible imputation

Abstract

Missing covariate data is a common issue in regression analysis and in comparative effectiveness analysis. To make a missing at random (MAR) assumption more realistic, auxiliary covariates are often included. In the presence of auxiliary covariates, the response model conditional on the covariates and auxiliary covariates needs to be specified in the imputation model. However standard multiple imputation approaches typically do not require the response model in the imputation model to be marginalized to the inference response model on covariates, which results in incompatible imputations. We extend the sequential BART approach, a flexible Bayesian nonparametric approach to impute missing covaraites, such that the imputation model is always compatible with the inference model by introducing a subject specific intercept in the imputation model. We provide details on the computational algorithms and compare the proposed approach to the original sequential BART approach and two versions of multiple imputation by chained equations approach.