bneR: A collaborative workflow for air pollution exposure modeling and uncertainty characterization using the Bayesian Nonparametric Ensemble.

Background: Air pollution is a major public health threat globally. Health studies, regulatory actions, and policy evaluations typically rely on air pollutant concentrations from single exposure models, assuming accurate estimations and ignoring related uncertainty. We developed a modeling framework, bneR, to apply the Bayesian Nonparametric Ensemble (BNE) prediction model that combines existing exposure models as inputs to provide air pollution estimates and their spatio-temporal uncertainty.

Methods: The bneR modeling framework (1) harmonizes air pollutant datasets to use standardized inputs for the BNE algorithm; (2) applies the BNE algorithm to obtain the posterior predictive distribution of pollutant concentrations; and (3) generates visualizations. We applied bneR to estimate NO concentrations and characterize uncertainty levels at high spatio-temporal resolution (daily, 1 km) over New York State (NYS) for 2015. We met with stakeholders and modelers to discuss bneR user-friendliness and interpretation of its estimates.

Results: Using bneR, we harmonized the spatial scale of four input NO models (using the finer resolution, 1 km for BNE estimations), applied BNE to obtain the NO daily posterior predictive distribution, and visualized the results. Over NYS, the daily average NO concentration was 6.0 (interquartile range, IQR: 4.6-6.8) pbb with daily average uncertainty (as SD) of 1.2 (IQR: 1.0-1.3) ppb. BNE performed well with cross-validated RMSE=2.84 ppb and R=0.80.

Conclusion: Meeting stakeholders and modelers allowed us to understand that efficient communication on how uncertainty is estimated and interpreted is a key feature for these communities to engage in using bneR and its data products.