| dc.description.abstract | Air pollution is the biggest environmental risk to global health and it is estimated that, globally, 7 million deaths can be attributed to air pollution each year \citep{WHO2018}. The World Bank estimates that, in 2016, the overall cost of ambient air pollution to the global economy was an estimated US \$5.7 trillion or 4.4 per cent of global GDP \citep{worldbank}.
A number of different air pollutants have been associated with adverse health effects, including fine particulate matter (PM$_{2.5}$), nitrogen dioxide and ozone. In studies of the effects of air pollution, exposure information is often
obtained from a fixed number of monitoring sites within the region of interest. However, an increasing number of models of air pollution are being used that provide estimates of concentrations. These are used to represent exposures at every location in a health study area, rather than just at a number of fixed measurement locations. Another use of modelling of air pollution is to provide short-term forecasts that can be used to inform the behaviour of vulnerable people.
In this thesis, we develop statistical approaches to modelling, and forecasting, daily concentrations of $\mbox{PM}_{2.5}$ in urban areas. We consider two different approaches, both in terms of model formulation and performing inference. The first approach is Dynamic Space-Time Models (DSTM). Under this framework, a \textit{data} model relates observations (measurements) to a \textit{process} model that specifies the dynamic evolution of the "true" underlying process. This approach is implemented using two different methods for estimation: methods of moments and expectation-maximisation. We also develop an approach using Bayesian Hierarchical Spatio-Temporal modelling (BHSTM). The inference is done using computational efficient methods for Bayesian inference (integrated nested Laplace approximations). This model allows predictions of daily PM$_{2.5}$ over both space and time, which can be used to interpolate both past measurements and future predictions. Both approaches were implemented using data from Greater London, with their performance evaluated in terms of their ability to predict daily concentrations of PM$_{2.5}$ over time at different measuring sites. Both methods were able to accurately predict future values of daily PM$_{2.5}$ at different locations, with one-day ahead predictions being more accurate than those used for longer periods, as might be expected. One of the major advantages of the BHSTM approach is that it provides a straightforward method for producing estimates of the uncertainty that is associated with predictions. | en_GB |
