Improving spatial nitrogen dioxide prediction using diffusion tubes: A case study in West Central Scotland

Abstract

It has been well documented that air pollution adversely affects health, and epidemiological pollution-health studies utilise pollution data from automatic monitors. However, these automatic monitors are small in number and hence spatially sparse, which does not allow an accurate representation of the spatial variation in pollution concentrations required for these epidemiological health studies. Nitrogen dioxide (NO2) diffusion tubes are also used to measure concentrations, and due to their lower cost compared to automatic monitors are much more prevalent. However, even combining both data sets still does not provide sufficient spatial coverage of NO2 for epidemiological studies, and modelled concentrations on a regular grid from atmospheric dispersion models are also available. This paper proposes the first modelling approach to using all three sources of NO2 data to make fine scale spatial predictions for use in epidemiological health studies. We propose a geostatistical fusion model that regresses combined NO2 concentrations from both automatic monitors and diffusion tubes against modelled NO2 concentrations from an atmospheric dispersion model in order to predict fine scale NO2 concentrations across our West Central Scotland study region. Our model exhibits a 47% improvement in fine scale spatial prediction of NO2 compared to using the automatic monitors alone, and we use it to predict NO2 concentrations across West Central Scotland in 2006.

How robust are the estimated effects of air pollution on health? Accounting for model uncertainty using Bayesian model averaging

“…We empirically investigate the sensitivity of the estimated pollution-health effect to three modelling choices. The first is the estimation of spatially averaged NO 2 concentrations for each data zone, and we compare averaging the raw output from the atmospheric dispersion model used by DEFRA ( http://uk-air.defra.gov.uk/ , denoted DEFRA ) to averaging predictions from the fusion model proposed by Pannullo et al. (2015) (denoted Fusion ).…”

“…This highlights that the DEFRA pollution concentrations are more correlated with residual disease, thus explaining why it has a stronger effect size (see Table 2 ) compared to the pollution concentrations from the fusion model. However, in terms of pollution predictive performance, Pannullo et al. (2015) show that the DEFRA data are not as good at predicting measured pollution concentrations at the point level, since the root mean square prediction error (RMSPE) is 0.337 compared to 0.255 for the fusion model.…”

“…(2009) . However, the model we use here is that of Pannullo et al. (2015) , which uses a similar approach to that of Berrocal et al.…”

“…The second term in the mean model is a spatial random effect which accounts for residual spatial autocorrelation in the measured data and are modelled by a Gaussian process with a spatial exponential correlation matrix where ρ denotes the range parameter. Full details of this model are available from Pannullo et al. (2015) .…”

“…This model calibrates the modelled concentrations according to urban background and rural environments since the West Central Scotland study region has a large proportion of rural areas, while allowing the effect of the modelled concentrations to vary across space. Pannullo et al. (2015) showed that their fusion model produced improved predictions of the measured data in a cross-validation exercise compared to using the modelled concentrations in isolation, and predictions were made from the model on a 1 km × 1 km resolution for each year separately.…”