Effects of spatial sampling density and spatial extent on linear land use regression modelling of NO2 estimates in an automobile-oriented city

Residents of densely populated cities face elevated exposure risk to ambient air pollution. Epidemiological studies often estimate exposure risk with land use regression (LUR) models that predict the spatial distribution of air pollutants using surrounding land use characteristics. We examined the effects of estimating city-wide air pollution concentrations using three spatial extents of sampling (31 km2, 94 km2, and 292 km2). Passive sampling of NO2 was completed with 33 samples allocated to each of the three spatial extents. Observed concentrations varied from 6.8 ppb to 19.9 ppb. Linear land use regression models were developed using a manual stepwise approach to ensure consistency between models. The LUR model developed with the largest sampling extent (292 km2 demonstrated the highest performance (Adj. R2 = 0.78; LOOCV R2 = 0.76); however, the resulting air pollution surface showed the least variability. Model performance was not related to spatial extent, with the smallest extent (31 km2) performing better (Adj. R2 = 0.67; LOOCV R2 = 0.56) than the medium extent (Adj. R2 = 0.63; LOOCV R2 = 0.54). Variability in the air pollution surface was related to spatial extent with the highest variability generated by the small extent model. A cross-model validation was completed to examine how well models performed for predicting the other spatial extents. All models demonstrated strong performance with RMSE≤2.5 ppb for all cases, with a mean RMSE of 2.0 ppb. No relationship was apparent between spatial extent and the ability to predict another spatial extent. Our findings indicate linear LUR modelling is robust to variations in spatial extent and density of sampling.