Assessment of the temporal stability of land use regression models for traffic-related air pollution

Background: Land-use regression (LUR) modeling is a cost-effective approach for assessing intra-urban air pollution contrasts. It has been widely used to estimate long-term exposure to traffic-related air pollution in epidemiologic studies. The application was based on the assumption that spatial patterns of pollution are stable over time so that a model developed for a particular time point could be applied to other time points. However, this assumption has not been adequately examined. This has specific relevance to cohort studies where models are developed in one particular year and then retrospectively or prospectively applied over periods of ~10 other years. Methods: Metro Vancouver LUR models for annual average NO and NO₂ were developed in 2003, based on 116 measurements. In 2010, we repeated these measurements; 73 were made at the same location as in 2003, while the remaining 43 sites were within ~50 m. We then developed new models using updated data for the same predictor variables, and also explored additional variables. The temporal stability of LUR models over a 7-year period was evaluated by comparing model predictions and measured spatial contrasts between 2003 and 2010. Results: Annual average NO and NO₂ concentrations decreased from 2003 to 2010. From the 73 sites that were identical between 2003 and 2010, the correlation between NO 2003 and 2010 measurements was r = 0.87 with a mean (sd) decrease of 11.3 (9.9) ppb, and between NO₂ measurements was r = 0.74 with a mean (sd) decrease of 2.4 (3.2) ppb. 2003 and 2010 LUR models explained similar amounts of spatial variation (R² difference of 0.01 to 0.11). The 2003 models explained more variability in 2010 measurements (R²= 0.52 – 0.65) than 2010 models did for 2003 measurements (R²= 0.38 – 0.55). Conclusions: Forecasting will be more appropriate than back-casting in the case of Metro Vancouver where concentrations and their variability decreased over time. Back-casting explains nearly the same amount of variability (R²= 0.38 – 0.55) in measured concentrations as did the original model (R² = 0.52 – 0.58). These results support the validity of applying LUR models to cohort studies over periods as long as 7 years. Medicine, Faculty ofPopulation and Public Health (SPPH), School ofGraduat