Model selection applications in time series studies of air pollution and mortality

Abstract

A number of time series studies provide evidence that air pollution levels are associated with daily death counts. Most of these studies explain these associations by using Poisson log-linear regression models, after allowing for possible confounders. However, the findings of these time series studies suffer from several problems, of which the most significant is conflicting results from different approaches to model selection. To address this problem, this dissertation compares the statistical properties of these approaches, including those with and without model selection uncertainty. In the simulation experiments, the sensitivity of these statistical properties to the choice of bootstrap resampling schemes, including a parametric bootstrap and residual-based bootstrap, is investigated. Simulation results indicate that regardless of the type of resampling schemes used, Bayesian model averaging methods based on Akaike's Information Criterion (BMA-aic) perform well in predicting the mortality effect of particulate matter. Additionally, rather than the standard use of the Bayesian Information Criterion (BIC), this study suggests using Akaike's Information Criterion (AIC) for the data from Chicago, if prediction is the goal. Thus, when predicting seasonal mortality effects of air pollution in Chicago, this dissertation incorporates AIC to identify the pattern of unmeasured confounding effects for each season in that city. A common approach to control for unmeasured confounding effects is to adopt a natural cubic spline function of time with fixed degrees of freedom per year. This approach uses an annual adjustment for these effects. However, in reality, these effects may vary according to season; given this, seasonal adjustments for these effects are advocated. As such, this raises the issue of how to control the confounders adequately in the natural cubic spline function. To resolve this issue, a model is proposed to explain the association between particulate matter and mortality after allowing for the seasonal patterns of confounders. AIC is used to select the degrees of freedom for each season. The results indicate a high mortality level during seasons with high particulate matter concentrations. Show more