Because GEE and mixed effects models make use of all available data they generally produce a more efficient estimate and have greater statistical power than standard logistic regression, which involves a single study visit and a single HPV endpoint. It is, however, often difficult to determine an exact power estimate for GEE and mixed effects models. Statistical simulations are sometimes necessary to estimate power for such studies. Here we discuss a simplified approach, which may be used as a first order approximation(23,24). To obtain a “ball park” estimate of power for correlated binary outcomes, we first assume an average correlation (ρ) among repeated observations within a subject. If on average each subject contributes J observations and there are a total of n subjects, then the “effective size” (i.e., the equivalent size if each subject only contributes one observation) is nJ/(1+(J-1)ρ). For example, if ρ=0.3 and there are J=5 multiple observations per subject and 100 subjects, the power based on these 100 subjects is equivalent to the power based on 100*5/(1+(5−1)*0.3)=227 subjects with only one observation per subject. Standard software packages
