A comparison of linear regression, regularization, and machine learning algorithms to develop Europe-wide spatial models of fine particles and nitrogen dioxide.

Empirical spatial air pollution models have been applied extensively to assess exposure in epidemiological studies with increasingly sophisticated and complex statistical algorithms beyond ordinary linear regression. However, different algorithms have rarely been compared in terms of their predictive ability. This study compared 16 algorithms to predict annual average fine particle (PM2.5) and nitrogen dioxide (NO2) concentrations across Europe. The evaluated algorithms included linear stepwise regression, regularization techniques and machine learning methods. Air pollution models were developed based on the 2010 routine monitoring data from the AIRBASE dataset maintained by the European Environmental Agency (543 sites for PM2.5 and 2399 sites for NO2), using satellite observations, dispersion model estimates and land use variables as predictors. We compared the models by performing five-fold cross-validation (CV) and by external validation (EV) using annual average concentrations measured at 416 (PM2.5) and 1396 sites (NO2) from the ESCAPE study. We further assessed the correlations between predictions by each pair of algorithms at the ESCAPE sites. For PM2.5, the models performed similarly across algorithms with a mean CV R2 of 0.59 and a mean EV R2 of 0.53. Generalized boosted machine, random forest and bagging performed best (CV R2~0.63; EV R2 0.58-0.61), while backward stepwise linear regression, support vector regression and artificial neural network performed less well (CV R2 0.48-0.57; EV R2 0.39-0.46). Most of the PM2.5 model predictions at ESCAPE sites were highly correlated (R2 > 0.85, with the exception of predictions from the artificial neural network). For NO2, the models performed even more similarly across different algorithms, with CV R2s ranging from 0.57 to 0.62, and EV R2s ranging from 0.49 to 0.51. The predicted concentrations from all algorithms at ESCAPE sites were highly correlated (R2 > 0.9). For both pollutants, biases were low for all models except the artificial neural network. Dispersion model estimates and satellite observations were two of the most important predictors for PM2.5 models whilst dispersion model estimates and traffic variables were most important for NO2 models in all algorithms that allow assessment of the importance of variables. Different statistical algorithms performed similarly when modelling spatial variation in annual average air pollution concentrations using a large number of training sites.

Spatial PM2.5, NO2, O3 and BC models for Western Europe - Evaluation of spatiotemporal stability.

BACKGROUND: In order to investigate associations between air pollution and adverse health effects consistent fine spatial air pollution surfaces are needed across large areas to provide cohorts with comparable exposures. The aim of this paper is to develop and evaluate fine spatial scale land use regression models for four major health relevant air pollutants (PM2.5, NO2, BC, O3) across Europe. METHODS: We developed West-European land use regression models (LUR) for 2010 estimating annual mean PM2.5, NO2, BC and O3 concentrations (including cold and warm season estimates for O3). The models were based on AirBase routine monitoring data (PM2.5, NO2 and O3) and ESCAPE monitoring data (BC), and incorporated satellite observations, dispersion model estimates, land use and traffic data. Kriging was performed on the residual spatial variation from the LUR models and added to the exposure estimates. One model was developed using all sites (100%). Robustness of the models was evaluated by performing a five-fold hold-out validation and for PM2.5 and NO2 additionally with independent comparison at ESCAPE measurements. To evaluate the stability of each model's spatial structure over time, separate models were developed for different years (NO2 and O3: 2000 and 2005; PM2.5: 2013). RESULTS: The PM2.5, BC, NO2, O3 annual, O3 warm season and O3 cold season models explained respectively 72%, 54%, 59%, 65%, 69% and 83% of spatial variation in the measured concentrations. Kriging proved an efficient technique to explain a part of residual spatial variation for the pollutants with a strong regional component explaining respectively 10%, 24% and 16% of the R2 in the PM2.5, O3 warm and O3 cold models. Explained variance at fully independent sites vs the internal hold-out validation was slightly lower for PM2.5 (65% vs 66%) and lower for NO2 (49% vs 57%). Predictions from the 2010 model correlated highly with models developed in other years at the overall European scale. CONCLUSIONS: We developed robust PM2.5, NO2, O3 and BC hybrid LUR models. At the West-European scale models were robust in time, becoming less robust at smaller spatial scales. Models were applied to 100 × 100 m surfaces across Western Europe to allow for exposure assignment for 35 million participants from 18 European cohorts participating in the ELAPSE study.