Generalised additive models for modelling personal exposure to airborne organic pollutants

Abstract

Generalised Additive Models (GAMs) are utilised in this dissertation to model the possible effects of indoor and outdoor activities and airborne volatile organic compounds (VOCs) on carcinogenic polycyclic aromatic hydrocarbons (CPAHs), the latter being combustion-related pollutants that may pose risks to public health. GAMs allow for modelling relationships between variables in a non-parametric manner, offering a more flexible approach beyond the parametric form of linear models. The GAMs studied in this dissertation utilise thin plate regression splines to represent smooth functions and employ a direct nested iterative method with penalised iteratively re-weighted least squares (PIRLS) and restricted maximum likelihood (REML) estimation to estimate model co-efficients and smoothing parameters, as presented by Wood (2017). In the interest of understanding the contribution of the various predictors under study on the Personal Exposure (PE) levels to CPAH, whilst tackling any multicollinearity risks, predictors are grouped into four different types. Generalised Linear Models (GLMs) and GAMs with CPAH as a response variable were fitted using each type. A comparison between all the fitted models is presented based on a number of performance measures: AIC, Proportion of Deviance Explained and RMSE of Prediction. Such a comparison reveals that the application of the GAM framework to explain and predict CPAH levels is worthwhile for 3 out of 4 variable groupings under consideration. The models recommended to understand the contribution of the variables to pollutant levels are those which model all the variables within each grouping. Meanwhile, to predict pollutant levels from new data, parsimonious models corresponding to the previously mentioned models are recommended. Variables potentially related to ventilation-related behaviour—such as Air Freshener and Summer—were found to provide the highest decreases in CPAH levels. In contrast, variables which are likely related to combustion-related activities such as Use Bus, ETS, and ETS Home were linked to increases in CPAH PE levels.