Background

Prenatal exposure to drinking water with arsenic concentrations >50 μg/L is associated with adverse birth outcomes, with inconclusive evidence for concentrations ≤50 μg/L. In a collaborative effort by public health experts, hydrologists, and geologists, we used published machine learning model estimates to characterize arsenic concentrations in private wells-federally unregulated for drinking water contaminants-and evaluated associations with birth outcomes throughout the conterminous U.S.

Methods

Using several machine learning models, including boosted regression trees (BRT) and random forest classification (RFC), developed from measured groundwater arsenic concentrations of ∼20,000 private wells, we characterized the probability that arsenic concentrations occurred within specific ranges in groundwater. Probabilistic model estimates and private well usage data were linked by county to all live birth certificates from 2016 (n = 3.6 million). We evaluated associations with gestational age and term birth weight using mixed-effects models, adjusted for potential confounders and incorporated random intercepts for spatial clustering.

Results

We generally observed inverse associations with term birth weight. For instance, when using BRT estimates, a 10-percentage point increase in the probability that private well arsenic concentrations exceeded 5 μg/L was associated with a -1.83 g (95% CI: -3.30, -0.38) lower term birth weight after adjusting for covariates. Similarly, a 10-percentage point increase in the probability that private well arsenic concentrations exceeded 10 μg/L was associated with a -2.79 g (95% CI: -4.99, -0.58) lower term birth weight. Associations with gestational age were null.

Conclusion

In this largest epidemiologic study of arsenic and birth outcomes to date, we did not observe associations of modeled arsenic estimates in private wells with gestational age and found modest inverse associations with term birth weight. Study limitations may have obscured true associations, including measurement error stemming from a lack of individual-level information on primary water sources, water arsenic concentrations, and water consumption patterns.