Comments on: Missing data methods in longitudinal studies: a review

摘要

We congratulate the authors on a far-reaching and thorough account of modelling for incomplete longitudinal data. Their focus is understandably on direct likelihood methods and their comprehensive treatment makes this paper a very useful reference. We would like to complement their picture by drawing attention to the potential value of multiple imputation (MI) in such settings. First, though we would like to make a distinction which we believe is important, but does not really appear in the paper. This is the difference between longitudinal settings in which intended times of measurement are common to all subjects (at least relative to the start of the study) and those in which they are not. The former is the norm with clinical trials, while both types occur commonly in observational studies. Interestingly, both the authors' examples are of the first type. In many ways, the problem of missing data (and other aspects of the analysis) are greatly simplified when times of measurement are common to all subjects.