Analyses Do Not Support CYP17 Genotype-Estradiol Associations

To the Editor: Jasienska et al. (1) examined whether estradiol levels vary with CYP17 genotypes by assaying 18 mid-cycle daily saliva samples from each of 22 urban and 38 rural Polish women. Their study design is a significant advance over others in which <5 days were sampled. However, their analyses are flawed, undermining their conclusion that the A2A2 genotype is associated with significantly higher estradiol levels than the A1A1 or A1A2 genotypes.

The authors make no mention of excluding anovulatory cycles (which typically have lower estradiol levels than ovulatory cycles) from their analyses. Anovulatory cycles bias the results, an effect exacerbated by different sample sizes across the genotypes (A2A2 = 7, A1A2 = 30, and A1A1 = 21). The anovulation rate for women comparable in age to their participants (24-36 years old) is ≥10% (2). Assuming that each participant has a 10% probability of contributing an anovulatory cycle, there is a 48% probability that all 7 participants with A2A2 genotypes contributed ovulatory cycles, but this probability is only 11% for the 21 participants with an A1A1 genotype, and only 4% for the 30 participants with an A1A2 genotype. This bias is eliminated if the samples happen to have identical proportions of anovulatory cycles, but dissimilar proportions are not improbable. For example, there is a 38% chance that the samples from 7 participants with A2A2 genotypes and from 30 participants with A1A2 genotypes differ by >10% in their proportion of anovulatory cycles. If the true anovulation rate is 15%, the chance that these samples differ by >10% rises to 49%. Furthermore, even if the anovulation proportion is identical, the inclusion of anovulatory cycles introduces unnecessary error into the analyses.

Additional problems with the statistical analyses include the violation of assumptions (3). For example, univariate repeated-measures ANOVA requires sphericity, and the Wilcoxon signed-rank test requires independent observations (repeated estradiol measurements over multiple days are not independent). Therefore, the reported significance levels for some analyses are not meaningful. The reported interaction between genotype and estradiol levels pertains to the shape (not the magnitude) of the profiles and does not test the main hypothesis. Also, rural/urban differences in estradiol levels and anovulation rates are not unlikely, but were not considered.

In sum, the reported analyses cannot support the authors' conclusions. The shortcomings can be easily corrected with additional analyses.