Research Article Methods for Stratified Cluster Sampling with Informative Stratification

We look at fitting regression models using data from stratified cluster samples when the strata may depend in some way on the observed responses within clusters. One important subclass of examples is that of family studies in genetic epidemiology, where the probability of selecting a family into the study depends on the incidence of disease within the family. We develop the survey-weighted estimating equation approach for this problem, with particular emphasis on the estimation of superpopulation parameters. Full maximum likelihood for this class of problems involves modelling the population distribution of the covariates which is simply not feasible when there are a large number of potential covariates. We discuss efficient semiparametric maximum likelihood methods in which the covariate distribution is left completely unspecified. We further discuss the relative efficiencies of these two approaches.