Affiliation: Gillings School of Global Public Health, Department of Environmental Sciences and Engineering
To support the Women’s Health Initiative (WHI) Memory Study (WHIMS), a nationwide cohort study, accurate ozone exposure estimates for ambient concentrations needed to be generated at a national scale for years 1993-2010. For this large spatial and temporal coverage we investigated different geo-statistical approaches to generate estimates that integrate routine monitoring from surface ozone observations and episodic chemical transport model (CTM) outputs. The goal is to take advantage of the accuracy of the observational data and the continuous spatial/temporal coverage of CTM model outputs. In this work, we demonstrate a Bayesian Maximum Entropy (BME) data integration geo-statistical approach for making national scale ozone estimates that models the non-linear and non-homoscedastic relation between air pollution observations and CTM predictions. This is the first application of BME that fully accounts for variability in CTM model performance through our novel Regionalized Air Quality Model Performance (RAMP) approach. A validation analysis was completed using only non-collocated data outside of a validation radius r_v and the error statistics between observations and re-estimated values were obtained. We show that by accounting for the spatial and temporal variability in model performance there is 3-12 fold increase in R2 (the squared Pearson correlation coefficient) percentage change for the daily ozone concentrations compared to estimates that assume model performance does not change across space and time. Our second project is to investigate the differences of the predictive capacity for two upscaling methods: USM1 (data aggregation from hourly to daily followed by BME approach estimation) and USM2 (perform BME approach estimation on hourly ozone followed by data aggregation). We found that the less computationally intensive method USM1 outperforms the method USM2. This highlights the capability of the RAMP approach that was able to capture the spatial temporal variability in CTM model performance at time scale of interest. Thus, we recommend to use upscaling method USM1 to integrate CTM model predictions through RAMP approach because USM1 can achieve higher estimation accuracy and also associated with much lower computational cost.