Disconcordance in statistical models of bisphenol A and chronic disease outcomes in NHANES 2003-08.

BACKGROUND

Bisphenol A (BPA), a high production chemical commonly found in plastics, has drawn great attention from researchers due to the substance's potential toxicity. Using data from three National Health and Nutrition Examination Survey (NHANES) cycles, we explored the consistency and robustness of BPA's reported effects on coronary heart disease and diabetes.

METHODS AND FINDINGS

We report the use of three different statistical models in the analysis of BPA: (1) logistic regression, (2) log-linear regression, and (3) dose-response logistic regression. In each variation, confounders were added in six blocks to account for demographics, urinary creatinine, source of BPA exposure, healthy behaviours, and phthalate exposure. Results were sensitive to the variations in functional form of our statistical models, but no single model yielded consistent results across NHANES cycles. Reported ORs were also found to be sensitive to inclusion/exclusion criteria. Further, observed effects, which were most pronounced in NHANES 2003-04, could not be explained away by confounding.

CONCLUSIONS

Limitations in the NHANES data and a poor understanding of the mode of action of BPA have made it difficult to develop informative statistical models. Given the sensitivity of effect estimates to functional form, researchers should report results using multiple specifications with different assumptions about BPA measurement, thus allowing for the identification of potential discrepancies in the data.