You are here

Enhancing models and measurements of traffic-related air pollutants for health studies using Bayesian melding

Principal Investigator: 

University of Michigan

This study aims to improve estimates of concentrations of traffic-related air pollutants using source-oriented emission and dispersion models and Bayesian Melding, a novel data fusion technique that combines measured and modeled concentrations of traffic pollutants.

Funded under
In review

Abstract for the 2017 HEI Annual Conference

Enhancing Models and Measurements of Traffic-Related Air Pollutants for Health Studies Using Dispersion and Bayesian Fusion Models
Stuart Batterman1, Veronica Berrocal1, Owais Gilani1, Chad Milando1, Sarav Arunachalam2, Max Zhang3
1University of Michigan–Ann Arbor, USA; 2University of North Carolina–Chapel Hill, USA; Cornell University, Ithaca, NY, USA
Background An improved understanding of traffic-related air pollutants (TRAPs) is needed to estimate exposures and adverse health impacts in traffic corridors and other near-road environments where individuals can be exposed to elevated concentrations.  The overall objective of this project is to improve estimates of concentrations of traffic-related air pollutants for use in health-related studies, with specific attention to source-oriented dispersion models and data fusion methods that can provide the spatial and temporal resolution needed to determine near-road exposures.  This poster describes analyses related to two specific aims of the work.  First, we provide an operational evaluation of the RLINE dispersion modeling for the near-road environment.  Second, recognizing the need in epidemiological studies to estimate exposures at locations that are not monitored, we assess several spatio-temporal modeling strategies that can improve estimates of TRAPs in the near-road environment.
Methods For the first aim, we predict PM2.5, NOx and CO concentrations using a detailed linked-base on-road emissions inventory and the RLINE model, an updated point source inventory and the AERMOD dispersion model, and four years of meteorological and concentration data collected in the Detroit, MI area.  For the second aim, we evaluate several spatio-temporal models that leverage short-term concentration data monitored at nine transects across major Detroit roads with estimates from the RLINE dispersion model to predict hourly concentrations over the entire study region with associated prediction uncertainties.
Results Summary conclusions from the operational evaluation show that:  RLINE can capture both spatial and temporal features of NOx and CO concentrations at locations downwind of major roads for 24-hr periods; performance decreases for winds parallel to road; and the ability to discern traffic-related contributions of PM2.5 is limited, a result of background concentrations, the sparseness of the monitoring network, the omission and large uncertainty of certain sources (e.g., area, fugitive) and processes (formation of secondary aerosols).  Summary conclusions from the spatio-temporal modeling include: confirmation that the dispersion model outputs contain valuable information (including spatial dependence) that can supplement the limited spatially and temporally-sparse monitoring data; that the output displays a non-ignorable spatially-varying bias and thus requires calibration; leveraging the correlation among TRAPs can lead to improved predictions compared to single-pollutant Bayesian data fusion approaches; and that prediction estimates have considerable uncertainty due to the limited amount of monitoring data available and the inherent variability in the observed concentrations.
Conclusions The study shows the performance that is possible and potentially likely when dispersion models are used to predict exposures in epidemiological applications.  Spatial-temporal analyses combining these predictions with monitoring observations can improve predictions and also provide uncertainty estimates.