Modeling heterogeneity in air pollution mixture effects on birth weight: A spatially varying coefficient approach

Abstract

Purpose:

To extend the existing quantile g-computation framework for studying environmental exposure mixtures to estimate local effects of ambient air pollution mixtures on birth weight. This framework has traditionally been applied to estimate global mixture effects without accounting for spatial heterogeneity.

Methods:

First, pregnancy-wide maternal exposure to five common air pollutants is estimated for nearly 1.5 million births occurring in Georgia, USA between 2005 and 2016. Then, a recently developed spatially varying coefficient model based on Bayesian additive regression trees (BART) is applied to estimate spatially heterogeneous mixture effects using quantile g-computation. Results are compared with those obtained from traditional conditional autoregressive models, as well as spatially agnostic modeling approaches.

Results:

We find evidence of county-level spatially varying mixture associations, where for 21 of 159 counties in Georgia, elevated concentrations of a mixture of PM_2.5, nitrogen dioxide, sulfur dioxide, ozone, and carbon monoxide were associated with a reduction in birthweight by as much as -14.77 grams (95% credible interval: -21.24, -9.78) per decile increase in all five air pollutants.

Conclusions:

Spatially varying coefficient models based on BART outperform alternative approaches when modeling the relationships between air pollution mixtures and birth weight for the majority of counties in Georgia.