The generalized estimating equation approach when data are not missing completely at random

We propose two methods for handling missing data in generalized estimating equation (GEE) analyses: mean imputation and multiple imputation. Each provides valid GEE estimates when data are missing at random. Missing outcomes are imputed sequentially starting from the outcome nearest in time to the observed outcome. The estimators from the two kinds of imputation are compared with the weighting method of Robins et al. We show that multiple imputation with an infinite number of replications is asymptotically equivalent to mean imputation. The methods are applied to a stroke study in which neurological outcomes are measured over time after stroke but some outcomes are missing due to death or loss to follow up.