A note on the analysis of censored regression data by multiple imputation

Wei and Tanner (1991, Biometrics 47, 1297-1309) considered two approximations to the Data Augmentation algorithm for the analysis of semiparametric linear regression models with censored response and unspecified residual distribution. On the basis of a simulation study, they concluded that the approximations have smaller mean squared errors than the Buckley-James estimator over a range of settings. We show that these conclusions result from the particular choice of censoring mechanism, starting value, and stopping rule for the iterations, and that they do not appear to hold in general. Even in the cases considered by Wei and Tanner, one appears to do at least as well with the same starting values and stopping rule using the Buckley-James iterations.