Abstract
In the framework of a marginally specified odds ratio model for multivariate, clustered binary data, full maximum likelihood estimation can become prohibitive, especially with large within-unit replication. We describe two alternative estimation procedures pseudolikelihood and generalized estimating equations, which are easier to fit. The second one is generally well known. Their relative merits are discussed and illustrated using data from developmental toxicity studies. Our attention in this paper is limited to exchangeable association structures. The extension to longitudinal data, which typically call for more complicated association designs, needs further investigation. A limited simulation study is performed.