How to Deal with Missing Covariate Data in Survival Analysis

• Activity Number: 457
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Torben Martinussen*+
• Keywords: missing data ; cox regression ; survival data ; non-parametric

Abstract

Missing covariate data often occurs in practice when analyzing survival data. Multiple imputation has been suggested as one way of dealing with this problem. When doing multiple imputation it is important to also use the response variable in the imputation algorithm, but for survival data it is not clear what the response really is when the survival time is censored. In this talk I will take another route based on the observed hazard function derived from assuming the Cox model for the full data setting. This leads to some recursive estimating functions for the unknown parameters that can be used for estimation. In the case where the observed covariates are discrete one can derive an estimator of the target parameters without having to model the covariate distribution, which is otherwise needed for existing methods. Large sample results are given and the method is applied to a real data example.