Labour and birth in water

Objectives Evidence for effects of environmental contaminants on pregnancy outcomes remains inconclusive. We investigated the associations between multiple, correlated exposures and related pregnancy outcomes using Bayesian, multivariate dimension reduction, and shrinkage regression approaches to account for multiple testing and interrelatedness of exposures and outcomes. Methods We evaluated a cohort of 1322 singletons, born to 547 mothers from Greenland, 197 from Warsaw, Poland, and 588 from Kharkiv, Ukraine, who were recruited in 2002–2004 during routine antenatal care visits. Three secondary metabolites of both diethylhexyl and diisononyl phthalates (DEHP, DINP), eight perfluorinated compounds (PFCs; including PFOS and PFOA), and organochlorines (p,p-DDE and PCB-153) were measured and detected in 72–100% of maternal serum samples. Outcomes were preterm birth (<37 weeks), birth weight, and small for gestational age (SGA; <10th percentile age- and gender-specific birth weight). We analysed exposures (clustered, high dimension predictors) and continuous and dichotomous outcomes with partial least squares (PLS) regression, and sparse PLS-discriminant analysis (sPLS-DA), respectively. We compared results with elastic net penalised regression, and Bayesian stochastic search variable selection with spike-and-slab priors of (nonlinear) generalised additive models. Results While applied methods had various degrees of sparseness, we observed generally consistent associations between DEHP metabolites, several PFCs and both organochlorines, and decreased birth weight and increased risk of SGA. There was no clear evidence of associations between contaminants and preterm birth. Conclusions Findings suggest that several environmental contaminants are independently associated with impaired fetal growth. Methods which account for correlations between variables and multiple testing may better discriminate robust exposure-response associations than conventional univariate linear and logistic regression models.