Incorporation of air pollution dispersion principals into regression models - how far can we take it?

Abstract

TPS 652: Air pollution exposure modeling 2, Exhibition Hall, Ground floor, August 28, 2019, 3:00 PM - 4:30 PM Background: Regression modelling is the most commonly used method for personalised air pollution exposure assessment. These models often incorporate a large number of land-use variables for which the causal relationship to air pollution is not clear. Moreover, it has been challenging to apply them to exposure assessment at a high temporal resolution. Formulating a regression model which takes into account well-established dispersion principals may improve previous capabilities. Methods: we designed a non-linear regression model which is based on traditional Gaussian dispersion. This model has 9 regression parameters which were optimised to provide the best fit between modelled and observed nitrogen oxides concentrations at the 39 regulatory air quality monitoring locations in our study area. Results: Using a single representative wind direction for the entire study area, the new model performed better than Inverse-Distance-Weighing (IDW) and a simpler regression model (TI-ODM). In particular, the cross-validated Mean Spatial Pearson Correlation (MSPC) was 0.23 over a dataset of 3500 unique half-hourly time-points, versus 0.19 for TI-ODM and 0.14 of IDW. The MSPC is particularly interesting when considering the epidemiological applications of exposure models. It was striking that all models performed much worse in predicting total nitrogen oxides than the nitrogen dioxide predictions shown in previous studies over the same area. A heterogeneous wind field was incorporated through a puff-like regression model. This model was applied both dynamically (i.e. having memory terms that propagate from each half-hourly time-point to the next) and statically (each time-point modelled separately as a “clean slate”). The static version performed much better (Nash-Sutcliffe efficiency=0.3 in the static version vs. -537.9 in the dynamic version). However, the static version could be run only in cross-validation mode due to computational limitations. Conclusions: In future work, this model must be qualitatively validated. If successful, it can be applied to epidemiological studies.