ORs are a good approximation of relative risk when prevalence rates are low (e.g., less than 10%)(15), but when prevalence rates are high ORs overestimate relative risk. Therefore, an additional advantage of the GEE logistic regression approach proposed under model (1.3) is that because it estimates associations “across oncogenic HPV types” (the weighted average of the effects for individual types) the OR is likely to be a good estimate of relative risk; i.e., since the prevalence of oncogenic HPV on a type-specific basis is usually less than 10% even in high risk populations. In contrast, “any oncogenic HPV” as defined under model (1.1) is binary (detection of one or more oncogenic HPV types) which can involve high prevalence in high risk populations. When a situation arises in which the prevalence of HPV is substantially greater than 10% direct estimation of prevalence ratios may be preferred to the use of an OR. This can be accomplished in GEE (or mixed effects models; see below) using log link instead of the logit link used in logistic regression (i.e, ORs), but log link is more likely than logit link to have problems with non-convergence(16,17).
