Statistical methodologies for dealing with incomplete longitudinal outcomes due to dropout missing at random

Abstract

Longitudinal studies are based on repeatedly measuring the outcome of interest and covariates over a sequences of time points. These studies play a vital role in many disciplines of science, such as medicine, epidemiology, ecology and public health. However, data arising from such studies often show inevitable incompleteness due to dropouts or even intermittent missingness that can potentially cause serious bias problems in the analysis of longitudinal data. In this chapter we confine our considerations to the dropout missingness pattern. Given the problems that can arise when there are dropouts in longitudinal studies, the following question is forced upon researchers: What methods can be utilized to handle these potential pitfalls? The goal is to use approaches that better avoid the generation of biased results. This chapter considers some of the key modelling techniques and basic issues in statistical data analysis to address dropout problems in longitudinal studies. The main objective is to provide an overview of issues and different methodologies in the case of subjects dropping out in longitudinal data for both the case of continuous and discrete outcomes. The chapter focusses on methods that are valid under the missing at random (MAR) mechanism and the missingness patterns of interest will be monotone; these are referred to as dropout in the context of longitudinal data. The fundamental concepts of the patterns and mechanisms of dropout are discussed. The techniques that are investigated for handling dropout are: (1) Multiple imputation (MI); (2) Likelihood-based methods, in particular Generalized linear mixed models (GLMMs) ; (3) Multiple imputation based generalized estimating equations (MI-GEE) ; and (4) Weighted estimating equations (WGEE) . For each method, useful and important assumptions regarding its applications are presented. The existing literature in which we examine the effectiveness of these methods in the analysis of incomplete longitudinal data is discussed in detail. Two application examples are presented to study the potential strengths and weaknesses of the methods under an MAR dropout mechanism.