The method of generalized estimating equations has become almost standard f
or analysing longitudinal and other correlated response data. However, we h
ave found that if binary responses have less than binomial variation over c
lusters, and are modelled using exchangeable correlations, prevailing softw
are implementations may give unreliable results. Bounding the negative corr
elation away from its theoretical minimum may not always be a satisfactory
solution. In such instances, using the independence working correlation str
ucture and robust SEs is a more trustworthy alternative. Copyright (C) 2000
John Wiley & Sons, Ltd.