Missing Binary Covariate Data and Imputation in Regression Models

This paper presents a simple way to handle missing values in categorical covariates, namely conditional probability imputation . Properties of this technique are given for various patterns of missing data in regression studies . An example shows its use in the proportional hazards model . The probability imputation technique is furthermore compared with multiple imputation and model-based approaches . It can be concluded that for certain patterns of missing data occuring typically in prognostic factor studies, the probability imputation technique has properties not inferior to more sophisticated but also more difficult-to-implement methods, and is outperforming standard techniques like complete case analysis or omission of covariates with missing values .