Arsenic in Private Well Water and Birth Outcomes in the United States

Abstract

Background/Aim: Prenatal exposure to drinking water with arsenic concentrations >50 μg/L has been associated with adverse birth outcomes. However, evidence is sparse at concentrations ≤50 μg/L. As a collaborative effort by public health experts and geologists, we used machine learning techniques to characterize arsenic concentrations in private wells, which are unregulated for contamination, and evaluated associations with birth outcomes throughout the conterminous U.S. Methods: Groundwater arsenic concentrations from ~20,000 private wells were used to develop several machine learning models, including random forest classification (RFC). Probabilistic model predictions, along with private well usage data, were linked by county to all certificates of live birth from 2016 (n=3.4 million). Mixed-effects models were fit to term birthweight and gestational age, adjusting for potential confounders and incorporating random intercepts for spatial clustering. As a sensitivity analysis to account for uncertainty in machine learning predictions, mother-infant pairs were randomly assigned to discrete private well arsenic categories (≤5, >5 to ≤10, or >10 µg/L) based on residential county. This process was repeated 10 times with estimates combined using Rubin’s rules. Results: We generally observed non-significant inverse associations with term birthweight. For instance, in sensitivity analyses of RFC predictions, we found that relative to mothers expected to have private well arsenic concentrations ≤5 μg/L, mothers assigned concentrations >5-≤10 μg/L gave birth to infants that weighed 1.6 (95% CI: -9.9, 6.6) grams less whereas those assigned to >10 μg/L gave birth to infants that weighed 2.4 (95% CI: -9.9, 5.1) grams less. Associations with gestational age were null. Conclusion: In this large nation-wide study, we did not detect a significant association of the modeled spatial distribution of arsenic in private wells with adverse birth outcomes. Measurement error stemming from a lack of individual-level information on primary water source and consumption levels likely obscured the true associations.