Spatio-temporal modelling of particulate matter and its application to assessing mortality effects of long-term exposure

Abstract

In Hong Kong, no studies have evaluated methodologies to estimate concentrations of particulate matter (PM) in small areas with complex urban morphology. Directly estimating long-term PM exposures from small number of monitoring stations alone provides little spatial variations and may lead to measurement errors. Therefore, traffic density and land-use types should be taken into consideration when determining individual-level exposures in a cohort study. This study proposed a novel method which incorporated remote sensing, meteorological and geographical data to estimate long-term PM exposures for assessing health effects.

Therefore, this thesis aims to cover two objectives: 1) to develop a spatio-temporal approach to estimate PM10 and PM2.5 concentrations in small areas from 2000 to 2011 in Hong Kong; 2) to apply this approach to determine the extent to which long-term exposure to PM was associated with mortality using the data from an elderly cohort.

For Objective 1, PM10 concentrations were estimated by twelve yearly generalized additive models. For each model, monthly PM10 averages from thirteen monitoring stations were regressed against surface extinction coefficient (SEC) derived from remote sensors, meteorological covariates, traffic counts, building density and distance to the nearest road. To reduce temporal fluctuations, each model used the data from a window of three consecutive years with the target prediction year in the centre of the window. To estimate PM2.5, because of small number of available stations, only one spatio-temporal model covering the whole study period was developed. This model included the estimated PM10, month of year and spatial covariates. R^2 and root-mean-square error (RMSE) were calculated to assess the predictive performance.

For Objective 2, residential-level PM exposures were estimated by the above models based on the residence address of each cohort subject. The association between long-term PM exposures and mortality was analysed by Cox proportional hazard model adjusting for individual- and area-level confounders. As additional analyses, the PM exposures estimated by inverse distance weighting (IDW) method were used to show the need for the proposed modelling approach.

The spatio-temporal models had high predicting power with adjusted R2 of 0.91 for PM10 and 0.87 for PM2.5, and high accuracy indicated by RMSE of 5.88μg/m3 and 4.98μg/m3, respectively. Among 61,586 subjects, the median follow-up time was 11.5 years (SD: 2.82) until the end of 2011, and there were 17,453 deaths (28.3% of the subjects). Exposure to a 10 μg/m3 increase was associated with 5% (95%CI: 4%-7%) for PM10, and 12% (10%-14%) for PM2.5 increase in death from all-natural causes; 7% (4%-10%) and 14% (10%-18%) from cardiovascular diseases; 9% (5%-12%) and 14% (10%-19%) from respiratory diseases. Females, non-smokers and subjects with high BMI were found at higher susceptibility of exposure. In the additional analyses, health effect estimates using IDW method yielded high excess risks for most mortality outcomes, including accidental mortality.

This proposed modelling approach provided a reliable and robust estimation of PM concentrations and captured both temporal and spatial variations well in small areas. The magnitudes of the mortality effects associated with long-term PM exposures were comparable with previous cohort studies.