Incorporating residential mobility in chronic air pollution studies is fundamental to accurate exposure estimates. Boscoe  presents a review of environmental health studies that have incorporated residential histories to-date. In our study, only 40% of participants lived at their study entry residence for the entire 20 year exposure period; on average, 2.3 (SD = 1.6) different residences per subject were reported. Recall bias was present for self-reported residential histories prior to 1975, with cases recalling older residences more often than controls. This has important implications for environmental epidemiology using self-reported residential histories as many environmental exposures have decreased substantially over time. Consequently, exposure assessment based on a greater proportion of older residential histories in cases compared to controls will result in an upward bias, rather than non-differential bias typically assumed from exposure misclassification. Studies that incorporate self-reported residential histories, particularity long-term residential histories - in this case over twenty years, may need to account for reporting bias in epidemiological analysis.
This study also demonstrated the importance of estimating air pollution exposures from residential histories, both in terms of including different residential locations as well as the corresponding spatiotemporal air pollution concentration estimates. Exposure quintiles based on residential addresses at study entry had approximately 50% correspondence to exposure quintiles developed from residential histories and spatiotemporal air pollution surface. These results address one of the research opportunities suggested by Meliker and Sloan : "indentifying circumstances under which it is worthwhile to compile and incorporate extensive space-time data histories of mobility or environmental contaminants". Epidemiological studies of diseases with long latency periods (in this case lung cancer) and/or that examine spatially and temporally varying exposures (in this case ambient air pollution) are clearly such circumstances.
Despite the fact that the Canadian NAPS monitoring network is one of the longest-standing national air pollution monitoring programs worldwide and now covers the majority of urban centers in Canada, its limited spatiotemporal coverage necessitated the creation of national models that capture both urban and rural populations. We were able to use NAPS data within 50 km of residential postal codes to assign exposures to 63%, 70% and 54% of exposure-years for TSP, O3 and NO2. Very limited spatial and temporal PM2.5 monitoring data were available (only 40% of exposure-years between 1984 and 1994 could be assigned) and we therefore estimated historical PM2.5 using TSP and metropolitan area indicator variables. The resulting models predicted PM2.5 variability well; the ratio for modelled PM2.5/TSP (0.32, SD = 0.12) is very similar to that found in US metropolitan areas (PM2.5/TSP = 0.30, SD = 0.11) .
National spatial pollutant surfaces were compiled and calibrated with historical NAPS data to assign ambient pollutant concentrations to all study participants' residential postal codes between 1975 and 1994. The two approaches used to calibrate spatial pollutant surfaces differ in their approach to account for temporal and spatial change; IDW interpolation accounted for the heterogeneity in pollution level changes across Canada during the exposure period, while linear regression models incorporated a linear time-trend and population density as a spatial predictor. The interpolation approach better represented historical PM2.5 concentrations, potentially due to the larger spatial scale of PM2.5, while the linear regression models better represented historical NO2 and O3 concentration, which have finer spatial resolutions.
The creation of national spatiotemporal models allowed for the inclusion of all study participants, regardless of geographic location and NAPS monitor coverage. This was important as 42884 (23%) of exposure-years occurred in rural areas. The mean PM2.5, NO2 and O3 exposure estimates derived from the spatiotemporal models were 11.3 μg/m3 (SD = 2.6), 17.7 ppb (4.1), and 26.4 ppb (3.4) respectively. The magnitude of these exposures are less than those used in other studies, for example, the widely cited ACS study (PM2.5: 17.7 μg/m3 (3.0), NO2 21.4 ppb (7.1); and O3 45.5 ppb (7.3)) . This is likely due to the inclusion of rural study participants as well as lower ambient pollution levels in Canada. The ability to incorporate rural areas in the exposure assessment added to the variability in the studies exposure estimates, particularly for NO2 and O3, as the majority of historical NAPS measurements in Canada represent pollutant concentration in large urban areas.
The results of the retrospective air pollution modeling approach conducted here are comparable to other such studies; however, the majority of retrospective air pollution exposure assessments have been conducted solely for urban areas. For example, Bellander et al.  used emission data, dispersion models, and geographic information systems (GIS) to assess exposure to NO2, NOx and SO2 ambient air pollution during 1960, 1970 and 1980 in Stockholm, Sweden. Model evaluation using historical data was not possible, but the model was found to have high correlation (r = 0.96) with aggregated 1994-1997 data from 16 monitors. In terms of national models, Hart et al.  developed U.S. nationwide models of annual exposure to PM10 and NO2 from 1985 to 2000. Generalized additive models were used to predict spatial surfaces from monitoring data and GIS-derived covariates (e.g. distance to road, elevation, proportion of low-intensity residential, high-intensity residential, and industrial, commercial land use). Model performance (R2) for PM10 and NO2 was 0.49 and 0.88 respectively. Another national retrospective study was conducted as part of the Netherlands Cohort Study on Diet and Cancer . Ambient air pollution exposures were estimated using regional (IDW monitor interpolation), urban (regression modelling), and local (road proximity) components. This approach explained 84%, 44%, 59% and 56% of the variability in averaged monitor data between 1976 and 1997 for NO2, NO, BS and SO2, respectively. The density of monitors in the Netherlands and the use of aggregated monitoring data may explain the higher model performance than seen in this study.
The exposure assessment approach presented here capitalizes on study participants' lifetime residential histories and incorporates comprehensive modelling approaches to estimate exposures to ambient air pollution and to vehicle and industrial emissions. Nevertheless, there are several limitations to this approach that may lead to exposure misclassification. Due to privacy concerns, residential addresses were coded using a standard geographic reference of 6-digit postal codes. Using a set geographic reference reduced error from changing postal codes over time; however, the spatial accuracy of postal codes varies substantially between urban and rural areas of Canada. Proximity analyses for exposures to vehicle and industrial emissions will therefore be more accurate in urban areas. The ambient air pollution exposure assessment relies on the accuracy of NAPS monitoring data, and historical monitor locations, especially in rural areas, may have been sited to capture local pollution problems. Unfortunately, no historical data were available to evaluate the representativeness of NAPS monitoring data. Due to sparse temporal and spatial PM2.5 monitor coverage, we created historical models based on TSP monitoring data and CMA indicator variables. While the model had good prediction, it was created from a limited number of monitoring stations from 1984 to 2000. Nevertheless, several studies have estimated PM2.5 successfully from TSP [6, 27]. The accuracy of the final spatiotemporal PM2.5, NO2 and O3 surfaces is also determined from the initial concentration surface as well as fusion with historical NAPS monitoring data or predictions incorporating a linear time-trend and population density. Some anomalies exist in the current spatial surfaces, for example, high PM2.5 concentrations in mountainous regions and PM2.5 and NO2 in certain locations in the Prairies; however, few study participants lived in these locations and exposure misclassification is therefore limited. All historical monitors were used to adjust annual spatial pollution surfaces, which resulted in urban monitor ratios extrapolated to rural areas. Few rural monitors exist and it was not possible to restrict to rural monitors when adjusting the spatial pollution surfaces in rural areas. Exposure to vehicle emissions was based on proximity measures to a national 1996 road network and a clear limitation was the lack of historical road databases. Industrial emissions were based on a comprehensive database of industrial locations from 1970 to 1994; however, emission estimates were only available for major industries, which restricted the examination of specific industrial chemicals when minor industries were included.