Incomplete data in repeated measures analysis

Complete (or balanced) repeated measures data arise when all subjects in a study are measured at the same set of time points. Data are often incomplete, because measurements are missed, or the design of the study results in subjects being measured at different sets of time points. This article reviews methods of analysis for incomplete repeated-measures data of this form, from an applied statistician's perspective. Limitations of approaches that (a) ignore between-subject variation, or (b) impute for missing values are discussed. Two methods are advocated that are relatively easy to implement using existing software, namely between-subject analysis of within-subject summary measures, and maximum likelihood based on a model for the data. Methods are applied and compared on four real-data examples with varied analytical objectives.