Quantifying Selection Bias in Cross-Sectional Studies of Ovarian Hormones

Abstract

Most studies of ovarian hormones in adult women collect data from a cross-sectional sample of participants meeting various selection criteria including not having been pregnant or breastfeeding for several months. Although this approach is intended to eliminate the effects of these factors on hormonal variation, it introduces a selection bias of unknown magnitude: in a non-contracepting population, those women with the highest fecundity are more likely to be either pregnant or lactating, and so not included in a study sample. Thus a cross-sectional sample disproportionately represents women with the lowest fecundity (and potentially the lowest hormone levels). Here we present a preliminary evaluation of the magnitude of this selection bias, focusing on progesterone ($P$) levels near the luteal peak. We use data from Project REPA, a longitudinal study of reproductive functioning in rural Bolivians, recruited without regard to reproductive status (Vitzthum, Spielvogel, and Thornburg \textit{Proceedings of the U.S. National Academy of Sciences/} 101, 1443 (2004)). Drawing from 542~non-conception cycles in 144~women, we construct simulated cross-sectional samples meeting various inclusion criteria and compare their anovulation rates and progesterone levels.