We then compared these GEE results with those obtained using a mixed effects model (model(1.5)). As expected, the mixed effects model gave much higher estimates of association, with wider confidence intervals than did GEE (Figure 1), but similar P-values (P<0.001). For example, the OR (i.e., eδ̄ in model (1.5)) of HPV16 detection associated with having a CD4+ count <200 cells/mm3 and HIV viral load >100,000 copies/mL (the most immunosupressed stratum) relative to being HIV-negative was 9.22 (95% CI,4.84–17.6). This estimate indicates that if an individual HIV-negative woman were to become HIV-positive and her immune status dropped to the most immunosupressed stratum, her risk of having HPV 16 infection would increase approximately 9-fold. In contrast, the OR (i.e., eβ̄ in model(1.3)) from the above GEE model would be interpreted as that on average an HIV-positive woman in the most immunosuppressed stratum has a 4-fold greater risk of HPV16 than an HIV-negative woman. A common limitation of mixed effects models was also revealed. Specifically, mixed effects models to assess HPV type-specific associations with host immune status failed to converge, consistent with the greater data requirement to run these models compared with GEE.
