Marginal Methods for Incomplete Longitudinal Data Arising in Clusters

Abstract

Inverse probability–weighted generalized estimating equations are commonly used to deal with incomplete longitudinal data arising from a missing-at-random mechanism when the marginal means are of primary interest. In many cases, however, the repeated measurements themselves may arise in clusters, which leads to both a cross-sectional and a longitudinal correlation structure. In some applications, the degree of these types of correlation may become of scientific interest. Here we develop inverse probability–weighted second-order estimating equations for monotone missing-data patterns which, under specified assumptions, facilitate consistent estimation of the marginal mean parameters and association parameters. Here the missing-data model accommodates cross-sectional clustering in the missing-data indicators, and the probabilities are estimated under a multivariate Plackett model. For computational reasons, we also consider using the alternating logistic regression algorithm for estimation of the association parameters for the responses. We investigate the importance of modeling the cross-sectional clustering in the missing-data process by simulation. An extension to deal with intermittently missing data is provided, and an application to a longitudinal cluster-randomized smoking prevention trial is presented.