ABSTRACT
The mixed-effect model approach, on the other hand, requires the specification of a conditional mean model with an association structure induced by random effects. The generalized estimating equation approach specifies population-averaged models that require a marginal mean model as well as a correlation model that directly accounts for the association structure among observations. The random effects can be assumed to be independent, which leads to an exchangeable correlation structure among cluster members, or as a vector can be assumed to have a certain dependence structure. From the experimental design's point of view, the difference between a cross-sectional SW trial and a closed-cohort SW trial lies in whether the same patients are measured multiple times across steps, which affects r, the autocorrelation between cluster-step sample means. Researchers have assumed different models for data arising from cross-sectional SW trials.
