Comparison of land-use regression models for predicting spatial NOx contrasts over a three year period in Oslo, Norway

Background: Spatial modelling of traffic-related air pollution through land-use regression (LUR) is increasingly applied in epidemiological studies. These models provide highly spatially resolved data, but assume that the spatial contrasts are stable over long periods of time. It is not known to which extent these models can be used to predict concentrations in earlier or later periods. We aimed at testing the stability of measured and modelled spatial contrasts over a three year period in order to assess the relevance for future assessments of individual exposure to traffic-related air pollutants in epidemiological studies. Methods: A land-use regression model was previously developed to estimate address-level outdoor concentrations of traffic-related air pollution based on samples of nitrogen oxides (NOx) at 80 locations during the winter of 2005. In the winter of 2008, we measured NOx again at 69 of these 80 locations and developed a new LUR model. This enabled us to compare short-term measurements and model predictions with three years apart in the same area. Results: Measurements conducted in 2008 agreed well with measurements sampled in 2005 at the same locations (r = 0.91-0.95). The LUR models from 2005 and 2008 explained 66-77% and 60-74% of the variability of the measured concentrations, respectively. The 2008 LUR models explained 55-68% of the spatial variability of the 2005 measurements, while the 2005 LUR models explained 53-66% of the spatial variability of the 2008 measurements. The models performed better for NOx and NO2 compared to NO, and were shown to be equally valid when using leave-one-out cross-validation and validation of models based on independent training sets. Conclusion: We found a good agreement between short-term measured spatial contrasts in outdoor NOx over a three year period. LUR models for this area performed equally well using two different validation methods. These models predicted the spatial variation well for this area both forward and backward in time. (C) 2011 Elsevier Ltd. All rights reserved.