Because repeated observations of HPV16 infection share a common αi they are correlated. The simple mixed effects model (1.5) above specifically assumes an exchangeable correlation structure, analogous to that described for GEE (above); that is, it assumes that any pair of repeated observations of HPV within the same woman (e.g., at visits 1 and 5 versus at visits 3 and 14) has the same correlation. However, other correlation structures can be assumed in mixed effects models, and in general mixed effects models attempt to more accurately model the correlations rather than rely on the use of robust variance as in the GEE models. If the correlation is correctly specified, a mixed effects model is more efficient than a GEE model. However, if it is not correctly specified, the results can be biased(18).
