To Model or Not to Model in Regression with Missing Covariates

• Activity Number: 239
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Nanhua Zhang*+ and Rod Little
• Companies: University of Michigan and University of Michigan
• Keywords: Complete-case analysis ; Ignorable likelihood ; Nonignorable modeling ; Outcome dependency

Abstract

We consider regression with missing covariates in this paper. Common methods include: (1) Complete-case analysis (CC), which discards the incomplete cases; (2) Ignorable likelihood methods (IL), which base inference on the observed likelihood given a model for the variables, without modeling the missing data mechanism; (3) Nonignorable modeling (NIM), which bases inference on the joint distribution of variables and the missing data indicators. CC and IL methods do not model the missing data mechanism while NIM models the joint distribution of variables and the missing data indicators. In this paper, we study the effect of covariate missingness on the estimation of the regression and answer the question when it is necessary to model the missing data mechanism. We will study two aspects of covariate missingness on the estimation of regression: (1) nonignorability, which concerns mainly how IL methods perform under varying levels of nonignorability; (2) outcome dependency, which studies the relatedness of covariate missingness to the outcome on the estimation of regression. We compare different methods for regression with missing covariates using a series of simulation experiments.