Let Yij represent the HPV16 infection status (0/1) for the ith person at jth semiannual visit, we model Pij = P(Yij = 1) by (1.2)logitPij=β0+βZij+γWij where Zijand Wijare defined similarly as in (1.1). The difference with standard logistic regression is in how the parameters are estimated. Specifically, GEE uses an estimating equation to estimate β in which a particular correlation structure (referred to as the working correlation) is assumed between repeated measures of Y. Several choices of working correlations are available in most statistical packages, including “independent” (suitable when a small correlation between data at different visits is assumed), and “exchangeable correlation” (suitable when the correlation between all pairs of observations can be assumed to be similar; e.g., visits 1 and 2 have the same correlation as visits 3 and 14), as well as other working correlations. Often we do not know how the repeated HPV outcomes are correlated with each other, but a useful feature of the GEE approach is that even if the working correlation is mis-specified, the estimate of β is still unbiased. Further, because a robust
