A Bayesian spatio-temporal model of panel design data: airborne particle number concentration in Brisbane, Australia

TL;DR

This paper develops a semi-parametric Bayesian spatio-temporal model for dense time but sparse space panel data on airborne particle concentrations, enabling detailed analysis of variability and exposure in Brisbane, Australia.

Contribution

It introduces a novel hierarchical Bayesian model using R-INLA for efficient inference on spatio-temporal particle data from a split panel design.

Findings

Model captures daily and weekly PNC cycles at schools.

Identifies peaks related to traffic and particle formation.

Describes spatial variation in PNC across schools.

Abstract

This paper outlines a methodology for semi-parametric spatio-temporal modelling of data which is dense in time but sparse in space, obtained from a split panel design, the most feasible approach to covering space and time with limited equipment. The data are hourly averaged particle number concentration (PNC) and were collected, as part of the Ultrafine Particles from Transport Emissions and Child Health (UPTECH) project. Two weeks of continuous measurements were taken at each of a number of government primary schools in the Brisbane Metropolitan Area. The monitoring equipment was taken to each school sequentially. The school data are augmented by data from long term monitoring stations at three locations in Brisbane, Australia. Fitting the model helps describe the spatial and temporal variability at a subset of the UPTECH schools and the long-term monitoring sites. The temporal variation is modelled hierarchically with penalised random walk terms, one common to all sites and a term accounting for the remaining temporal trend at each site. Parameter estimates and their uncertainty are computed in a computationally efficient approximate Bayesian inference environment, R-INLA. The temporal part of the model explains daily and weekly cycles in PNC at the schools, which can be used to estimate the exposure of school children to ultrafine particles (UFPs) emitted by vehicles. At each school and long-term monitoring site, peaks in PNC can be attributed to the morning and afternoon rush hour traffic and new particle formation events. The spatial component of the model describes the school to school variation in mean PNC at each school and within each school ground. It is shown how the spatial model can be expanded to identify spatial patterns at the city scale with the inclusion of more spatial locations.