Associations of urinary caffeine metabolites with sex hormones: comparison of three statistical models

AimsThe association between urinary caffeine and caffeine metabolites with sex hormones remains unclear. This study used three statistical models to explore the associations between urinary caffeine and its metabolites and sex hormones among adults.MethodsWe selected the participants aged ≥18 years in the National Health and Nutrition Examination Survey (NHANES) data 2013–2014 as our study subjects. We performed principal components analysis (PCA) to investigate the underlying correlation structure of urinary caffeine and its metabolites. Then we used these principal components (PCs) as independent variables to conduct multiple linear regression analysis to explore the associations between caffeine metabolites and sex hormones (E2, TT, SHBG). We also fitted weighted quantile sum (WQS) regression, and Bayesian kernel machine regression (BKMR) methods to further assess these relationships.ResultsIn the PCA-multivariable linear regression, PC2 negatively correlates with E2: β = −0.01, p-value = 0.049 (male population). In the WQS regression model, the WQS indices were associated with SHBG and TT both in male (SHBG: WQS index = −0.11, p < 0.001; TT: WQS index = −0.10, p < 0.001) and female (SHBG: WQS index = −0.10, p < 0.001; TT: WQS index = −0.04, p < 0.001) groups. Besides, the WQS index was significantly associated with E2 in females (p < 0.05). In the BKMR model, despite no significant difference in the overall association between caffeine metabolites and the sex hormones (E2, TT, SHBG), there was nonetheless a declining trend in the male population E2 group, in the male and female population SHBG groups also observed a downward trend.ConclusionWhen considering the results of these three models, the whole-body burden of caffeine metabolites, especially the caffeine metabolites in the PC2 metabolic pathway was significantly negatively associated with E2 in males. Considering the advantages and disadvantages of the three statistical models, we recommend applying diverse statistical methods and interpreting their results together.