Increasing sample size in prospective birth cohorts: back-extrapolating prenatal levels of persistent organic pollutants in newly enrolled children.

Study sample size in prospective birth cohorts of prenatal exposure to persistent organic pollutants (POPs) is limited by costs and logistics of follow-up. Increasing sample size at the time of health assessment would be beneficial if predictive tools could reliably back-extrapolate prenatal levels in newly enrolled children. We evaluated the performance of three approaches to back-extrapolate prenatal levels of p,p'-dichlorodiphenyltrichloroethane (DDT), p,p'-dichlorodiphenyldichloroethylene (DDE) and four polybrominated diphenyl ether (PBDE) congeners from maternal and/or child levels 9 years after delivery: a pharmacokinetic model and predictive models using deletion/substitution/addition or Super Learner algorithms. Model performance was assessed using the root mean squared error (RMSE), R2, and slope and intercept of the back-extrapolated versus measured levels. Super Learner outperformed the other approaches with RMSEs of 0.10 to 0.31, R2s of 0.58 to 0.97, slopes of 0.42 to 0.93 and intercepts of 0.08 to 0.60. Typically, models performed better for p,p'-DDT/E than PBDE congeners. The pharmacokinetic model performed well when back-extrapolating prenatal levels from maternal levels for compounds with longer half-lives like p,p'-DDE and BDE-153. Results demonstrate the ability to reliably back-extrapolate prenatal POP levels from levels 9 years after delivery, with Super Learner performing best based on our fit criteria.