Prediction Approaches of Personal Exposure Using Spatial Analysis in Evaluating Air Pollution Effects on Health Outcome

Abstract

Many epidemiologic studies have assigned exposures to individuals using measurements from a few fixed site air monitoring stations within each community. However, this approach does not capture the individual spatial heterogeneity in air pollution exposure and may lead to the inaccuracy such as misclassification error that likely underestimates the effects of air pollution. Recently many studies increasingly estimate individual level exposures from monitor based measurements by using exposure prediction methods. However, relatively few studies have evaluated how different approaches affect health risk estimates. This study was undertaken to apply the 4 different common methods including average across all monitors, nearest monitor, inverse distance weighting, and kriging and explore how these prediction methods might influence health effect estimates for lung function in a cohort in Ulsan, Korea. Data on 2,102 subjects were extracted from the Ulsan cohort for the period 2003-2007. Daily pollution exposures at individual level were interpolated from 13 fixed site monitors based on the location of residence. We assessed associations in multiple linear regression models. These results suggest that kriging prediction method estimated more accurate results than other methods based on cross-validation results. After controlling for age, sex, and BMI, we found statistically negative significant associations between all air pollutants and FVC regardless of the exposure estimation method. However, we found no significant negative effect on FEV1 except for ozone. For health effect estimates, some exposure methods resulted in the largest magnitude of the central effect where others had the most certain effect. Health effect estimates were generally higher when the average across all monitors or kriged exposure estimates was used. With respect to the uncertainty of health effect estimates, results based on the nearest monitor were the most certain among all four exposure methods. The effects were strongest in the over 65 older or female group. This study suggests that spatial interpolation methods may be able to provide better estimates than monitoring values alone by reflecting spatial characteristics of air pollutants and the spatial variability of individual level exposures by generating estimates for locations without monitors. These results can be used to guide future work on more advanced exposure methods and epidemiological studies.