Grouping methods for informative missing data in longitudinal studies

Missing data are common in longitudinal studies and the probability that an observation is missing may depend on the unobserved outcomes. A special case is that the missingness may depend on the outcomes through random effects representing underlying individual characteristics such as disease process or health awareness. To account for such informative missing data, some special methods have to be applied instead of standard analysis. An existing method referred to ACM method incorporates some summary measures of missing patterns as additional covariates. This dissertation investigates and generalizes another approach by grouping data according to summary measures of missingness patterns. We show how such grouping methods can provide desirable estimators and clarify the differences between the ACM and the grouping methods. Detailed steps for carrying out the grouping methods under various missingness mechanisms are shown and a new imputation method based on grouping is proposed. Simulation studies are conducted to evaluate and compare the performances of new and existing methods. These methods are also applied to an example.