Linear regression models with incomplete categorical covariates (Q1855635)

This paper presents methods dealing with incomplete binary variables. The authors consider the linear model \(y=X\beta+\varepsilon\), where \(y\) is a continuous and completely observed response vector, and the \(n\times K\) design matrix \(X\) contains \(m\) missing values in one binary regressor. Single imputation, multiple imputation, the so-called pi imputation and single imputation based on a modified first order regression are compared with standard methods (complete case analysis, mean imputation and imputation of the mode). In simulation experiments, two different models are used. The first model consists of three covariables -- one constant and two binary variables (one of them incomplete). The second model contains an additional covariable which is continuous. MSE-ratio, variance and bias are used to illustrate differences within and between the approaches.