Marginal models for the analysis of longitudinal measurements with nonignorable non-monotone missing data

Abstract

We propose methods for the analysis of continuous responses subject to nonignorable non-monotone missing data. We form a pseudolikelihood by naively assuming independence over time and using a product of marginal likelihoods at each time point, and we obtain consistent and asymptotically normal estimators of the mean and missingness parameters. Our primary interest is in estimating the parameters of the marginal model at each time point, and we make no assumption about the correlation structure. An approximate variance is found for the parameter estimates using a sandwich estimate. We provide an example, and present a simulation study that assesses the performance of the model.