Linear Mixed Models and Missing Data

Abstract

In virtually all longitudinal studies the issues of unbalancedness and missing data arise. Some studies, such as the Baltimore Longitudinal Study of Aging (Section 3.2) and the Variceal Pressures Study (Section 4.1) are designed such that the number of measurements per subject is variable or even random. The measurement times themselves can vary across subjects and can be random as well. We term these studies unbalanced. In such unbalanced studies it is usually not possible to identify non-response, unless measurement times have been recorded, even for occasions at which no measurement was actually taken. In contrast, in a balanced study the number of measurements per subject is fixed and the measurements are usually taken at an approximately common set of occasions. In this situation, missing observations can be identified without ambiguity. For this reason, we will focus attention on missing data in the balanced case. The specific case of dropout (i. e., a subject is completely observed until a certain point in time, where after no more measurements are taken) can be handled in the unbalanced case as well. The treatment of dropout in both balanced and unbalanced cases is very similar and therefore we will suffice with a balanced example of dropout (Section 5.11).