The use of generalized estimating equations for risk assessment in developmental toxicity

This paper reviews and compares several approaches to fitting dose-response models to developmental toxicity data. The main issue of interest is how to appropriately account for litter effects. Among the approaches reviewed are Beta Binomial models, models that attempt to characterize the litter effect through the use of covariates, and models that avoid the complication of correlated offspring by modeling "affected litter" rather than fetus-specific outcomes. Finally, we discuss our recommended approach, which is to use Generalized Estimating Equations, or quasi-likelihood. We give a number of reasons for preferring the latter and illustrate its application with an example.