Modelling air pollution for epidemiologic research--Part I: A novel approach combining land use regression and air dispersion.

A common limitation of epidemiological studies on health effects of air pollution is the quality of exposure data available for study participants. Exposure data derived from urban monitoring networks is usually not adequately representative of the spatial variation of pollutants, while personal monitoring campaigns are often not feasible, due to time and cost restrictions. Therefore, many studies now rely on empirical modelling techniques, such as land use regression (LUR), to estimate pollution exposure. However, LUR still requires a quantity of specifically measured data to develop a model, which is usually derived from a dedicated monitoring campaign. A dedicated air dispersion modelling exercise is also possible but is similarly resource and data intensive. This study adopted a novel approach to LUR, which utilised existing data from an air dispersion model rather than monitored data. There are several advantages to such an approach such as a larger number of sites to develop the LUR model compared to monitored data. Furthermore, through this approach the LUR model can be adapted to predict temporal variation as well as spatial variation. The aim of this study was to develop two LUR models for an epidemiologic study based in Greater Manchester by using modelled NO(2) and PM(10) concentrations as dependent variables, and traffic intensity, emissions, land use and physical geography as potential predictor variables. The LUR models were validated through a set aside "validation" dataset and data from monitoring stations. The final models for PM(10) and NO(2) comprised nine and eight predictor variables respectively and had determination coefficients (R²) of 0.71 (PM(10): Adj. R²=0.70, F=54.89, p<0.001, NO(2): Adj. R²=0.70, F=62.04, p<0.001). Validation of the models using the validation data and measured data showed that the R² decreases compared to the final models, except for NO(2) validation in the measured data (validation data: PM(10): R²=0.33, NO(2): R²=0.62; measured data: PM(10): R²=0.56, NO(2): R²=0.86). The validation further showed low mean prediction errors and root mean squared errors for both models.