|
Abstract:
|
Missing data are encountered often in longitudinal studies despite every attempt to collect full data. When the missingness is informative, it is generally difficult to model and analyze incomplete data because the distributional assumptions about missing data are not easily verifiable. Pattern-mixture models are used commonly in practice, in which patterns are determined by time to occurrence of missing data. Marginalization over patterns can be tricky, as outcomes are discrete. We propose a joint-modeling approach within a Bayesian framework in which patterns as random effects are marginalized within a generalized linear mixed model framework. Advantages of this approach include its being capable for both continuous and discrete responses, it is avoidant of the problem that not all parameters are estimable, it is computationally efficient, and it provides marginal estimates.
|
