Relative performance of different exposure modeling approaches for sulfur dioxide concentrations in the air in rural western Canada

BACKGROUND: The main objective of this paper is to compare different methods for predicting the levels of SO(2 )air pollution in oil and gas producing area of rural western Canada. Month-long average air quality measurements were collected over a two-year period (2001–2002) at multiple locations, with some side-by-side measurements, and repeated time-series at selected locations. METHODS: We explored how accurately location-specific mean concentrations of SO(2 )can be predicted for 2002 at 666 locations with multiple measurements. Means of repeated measurements on the 666 locations in 2002 were used as the alloyed gold standard (AGS). First, we considered two approaches: one that uses one measurement from each location of interest; and the other that uses context data on proximity of monitoring sites to putative sources of emission in 2002. Second, we imagined that all of the previous year's (2001's) data were also available to exposure assessors: 9,464 measurements and their context (month, proximity to sources). Exposure prediction approaches we explored with the 2001 data included regression modeling using either mixed or fixed effects models. Third, we used Bayesian methods to combine single measurements from locations in 2002 (not used to calculate AGS) with different priors. RESULTS: The regression method that included both fixed and random effects for prediction (Best Linear Unbiased Predictor) had the best agreement with the AGS (Pearson correlation 0.77) and the smallest mean squared error (MSE: 0.03). The second best method in terms of correlation with AGS (0.74) and MSE (0.09) was the Bayesian method that uses normal mixture prior derived from predictions of the 2001 mixed effects applied in the 2002 context. CONCLUSION: It is likely that either collecting some measurements from the desired locations and time periods or predictions of a reasonable empirical mixed effects model perhaps is sufficient in most epidemiological applications. The method to be used in any specific investigation will depend on how much uncertainty can be tolerated in exposure assessment and how closely available data matches circumstances for which estimates/predictions are required.