Back-extrapolation of estimates of exposure from current land-use regression models

Land use regression has been used in epidemiologic studies to estimate long-term exposure to air pollution within cities. The models are often developed toward the end of the study using recent air pollution data. Given that there may be spatially-dependent temporal trends in urban air pollution and that there is interest for epidemiologists in assessing period-specific exposures, especially early-life exposure, methods are required to extrapolate these models back in time. We present herein three new methods to back-extrapolate land use regression models. During three two-week periods in 2005–2006, we monitored nitrogen dioxide (NO2) at about 130 locations in Montreal, Quebec, and then developed a land-use regression (LUR) model. Our three extrapolation methods entailed multiplying the predicted concentrations of NO2 by the ratio of past estimates of concentrations from fixed-site monitors, such that they reflected the change in the spatial structure of NO2 from measurements at fixed-site monitors. The specific methods depended on the availability of land use and traffic-related data, and we back-extrapolated the LUR model to 10 and 20 years into the past. We then applied these estimates to residential information from subjects enrolled in a case–control study of postmenopausal breast cancer that was conducted in 1996.Observed and predicted concentrations of NO2 in Montreal decreased and were correlated in time. The estimated concentrations using the three extrapolation methods had similar distributions, except that one method yielded slightly lower values. The spatial distributions varied slightly between methods. In the analysis of the breast cancer study, the odds ratios were insensitive to the method but varied with time: for a 5 ppb increase in NO2 using the 2006 LUR the odds ratio (OR) was about 1.4 and the ORs in predicted past concentrations of NO2 varied (OR∼1.2 for 1985 and OR∼1.3–1.5 for 1996). Thus, the ORs per unit exposure increased with time as the range and variance of the spatial distributions decreased, and this is due partly to the regression coefficient being approximately inversely proportional to the variance of exposure. Changing spatial variability complicates interpretation and this may have important implications for the management of risk. Further studies are needed to estimate the accuracy of the different methods.