Nonstationary spatiotemporal Bayesian data fusion for pollutants in the near‐road environment

Abstract

Concentrations of near‐road air pollutants (NRAPs) have increased to very high levels in many urban centers around the world, particularly in developing countries. The adverse health effects of exposure to NRAPs are greater when the exposure occurs in the near‐road environment as compared to background levels of pollutant concentration. Therefore, there is increasing interest in monitoring pollutant concentrations in the near‐road environment. However, due to various practical limitations, monitoring pollutant concentrations near roadways and traffic sources is generally rather difficult and expensive. As an alternative, various deterministic computer models that provide predictions of pollutant concentrations in the near‐road environment, such as the research line‐source dispersion model (RLINE), have been developed. A common feature of these models is that their outputs typically display systematic biases and need to be calibrated in space and time using observed pollutant data. In this paper, we present a nonstationary Bayesian data fusion model that uses a novel data set on monitored pollutant concentrations (nitrogen oxides or NOx and fine particulate matter or PM2.5) in the near‐road environment and, combining it with the RLINE model output, provides predictions at unsampled locations. The model can also be used to evaluate whether including the RLINE model output leads to improved pollutant concentration predictions and whether the RLINE model output captures the spatial dependence structure of NRAP concentrations in the near‐road environment. A defining characteristic of the proposed model is that we model the nonstationarity in the pollutant concentrations by using a recently developed approach that includes covariates, postulated to be the driving force behind the nonstationary behavior, in the covariance function.

Accounting for survey design in Bayesian disaggregation of survey-based areal estimates of proportions: an application to the American Community Survey

“…Much of our predictive performance evaluation is based on the empirical probability that credible and/or prediction intervals cover the true value, which inherently conflates a frequentist property (empirical coverage probability) with Bayesian modeling frameworks. This type of assessment is in line with the notion of calibrated Bayes (Little, 2006) and recommended in a predictive context (Dawid, 1982); moreover, it is the authors' experience that coverage probabilities are frequently used in assessing Bayesian models, particularly in a spatial context (see Entezari, Brown and Rosenthal (2019); Gilani, Berrocal and Batterman (2019); Berrocal, Gelfand and Holland (2010) as example), where prediction is the main goal.…”

Quantifying the impact of the modifiable areal unit problem when estimating the health effects of air pollution

“…Data on air pollution concentrations are available as point‐level measurements and/or grid‐level modeled concentrations, and average concentrations for each areal unit have been estimated using simple averaging (Lee & Sarran, 2015) or spatial prediction techniques such as block Kriging (Zhu, Carlin, & Gelfand, 2003). A number of different statistical issues have been addressed in the literature in relation to modeling air pollution and its health effects, including different approaches for (i) spatiotemporal pollution prediction (Gilani, Berrocal, & Batterman, 2019; Nicolis, Díaz, Sahu, & Marín, 2019); (ii) estimating individual‐level exposures (Clifford et al, 2019); (iii) the impacts of preferential sampling of pollution (Lee, Szpiro, Kim, & Sheppard, 2015); and (iv) allowing for errors in exposure variables (Keller & Peng, 2019; Strand, Sillau, Grunwald, & Rabinovitch, 2015).…”

Data fusion with Gaussian processes for estimation of environmental hazard events

“…Boaz, Lawson, and Pearce (2019) develop a multivariate fusion framework to deal with air pollution prediction with partial missingness. Wilkie et al (2019) treat data for fusion as realisations of smooth temporal functions, Gilani, Berrocal, and Batterman (2019) accounts for nonstationarity by incorporating covariates, postulated to drive the nonstationary behavior, in the covariance function and Ma and Kang (2020) develop a stochastic expectation–maximization that facilitates the use of large spatiotemporal datasets. Forlani, Bhatt, Cameletti, Krainski, and Blangiardo (2020) demonstrate the added benefit of allowing multiple sources of model output in a framework that combines stochastic partial differential equations and the integrated nested Laplace approximation (Lindgren, Rue & Lindström, 2011).…”