Using Tree-Based Analytic Methods to Investigate Associations of Multiple Exposures with Pubertal Development in Urban Girls

摘要

Most research on childhood environmental contributors to early puberty has examined toxins singly in a limited window, often using parametric, regression-based modeling. A realistic assessment of health effects must incorporate multiple exposures over a longer period and allow for interaction between factors. Tree-based analytic methods recursively split the complete study population into multiple subsets based on exposure(s) that most strongly predict the outcome. Tracing down the branches created by these exposure-based splits yields exposure combinations associated with the outcome of interest, allowing for the simultaneous consideration of multiple exposures while allowing for modification by demographics and/or interaction between exposures. The goal of this analysis was to apply tree-based methods to data from a study of 192 nine-year-old New York City girls to identify exposures associated with pubertal development. Using data and biological samples collected during in-person interviews, we derived over 50 exposure variables including environmental biomarkers, ambient air pollutants, dietary components, and demographics. Pediatrician-assessed pubertal development was based on Tanner staging and recorded separately for breast and pubic hair, with stages 2 through 5 indicating puberty. We constructed multiple classification trees for pubertal outcomes utilizing all of the available environmental exposure metrics. For breast development, branches of the trees identified race-specific protective effects of five urine- or dietary-based measures of phytoestrogens. For example, breast development was less likely among Black girls with higher urinary concentrations of daidzein and among white/Latina girls with higher levels of urinary concentrations of genistein. These results illustrate how data-driven approaches can uncover environmental contributors to disease that may be masked when interactions or modification by demographics are not explicitly considered.