Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

Intercity transferability of land use regression models for estimating ambient concentrations of nitrogen dioxide

Abstract

Land use regression (LUR) is a method for predicting the spatial distribution of traffic-related air pollution. To facilitate risk and exposure assessment, and the design of future monitoring networks and sampling campaigns, we sought to determine the extent to which LUR can be used to predict spatial patterns in air pollution in the absence of dedicated measurements. We evaluate the transferability of one LUR model to two other geographically comparable areas with similar climates and pollution types. The source model, developed in 2003 to estimate ambient nitrogen dioxide (NO2) concentrations in Vancouver (BC, Canada) was applied to Victoria (BC, Canada) and Seattle (WA, USA). Model estimates were compared with measurements made with Ogawa® passive samplers in both cities. As part of this study, 42 locations were sampled in Victoria for a 2-week period in June 2006. Data obtained for Seattle were collected for a different project at 26 locations in March 2005. We used simple linear regression to evaluate the fit of the source model under three scenarios: (1) using the same variables and coefficients as the source model; (2) using the same variables as the source model, but calculating new coefficients for local calibration; and (3) developing site-specific equations with new variables and coefficients. In Scenario 1, we found that the source model had a better fit in Victoria (R2=0.51) than in Seattle (R2=0.33). Scenario 2 produced improved R2-values in both cities (Victoria=0.58, Seattle=0.65), with further improvement achieved under Scenario 3 (Victoria=0.61, Seattle=0.72). Although it is possible to transfer LUR models between geographically similar cities, success may depend on the between-city consistency of the input data. Modest field sampling campaigns for location-specific model calibration can help to produce transfer models that are equally as predictive as their sources.

Introduction

The study of ambient air pollution from traffic-related sources continues to be a priority for air quality management and public health research. The literature examining the effects of motor vehicle emissions on respiratory and cardiovascular health is extensive (e.g., Künzli et al., 2000; Brauer et al., 2002, 2007; Brunekreef and Holgate, 2002; Hoek et al., 2002a, 2002b; Gauderman et al., 2005; Krzyzanowski et al., 2005; Wilhelm and Ritz, 2005). Geographic estimates of ambient concentration distributions enable assessment of within-city variability of traffic-related pollutants for epidemiological studies of their health impacts. Field monitoring and computer modeling are commonly used to derive these estimates for use in exposure assessment.

Land use regression (LUR) is an emerging method used to predict the spatial distribution of ambient air pollutants for use in risk assessment and epidemiological studies, with specific application to traffic-generated air pollution. Regression methods are used to model pollutant concentrations measured at given sites on the basis of variables that characterize their surrounding land use, population density and traffic patterns, as described elsewhere (Jerrett et al., 2005; Sahsuvaroglu et al., 2006; Henderson et al., 2007).

First used in Europe by Briggs et al. (1997), LUR has been applied in a variety of European (Briggs et al., 2000; Lebret et al., 2000; Brauer et al., 2003) and North American settings (Gilbert et al., 2005; Sahsuvaroglu et al., 2006; Smith et al., 2006; Henderson et al., 2007; Jerrett et al., 2007). The method is cost effective (Jerrett et al., 2005), captures greater spatial variability than air quality monitoring networks (Brauer et al., 2003; Sahsuvaroglu et al., 2006; Henderson et al., 2007) and performs well in comparison to other methods, such as dispersion modeling (Cyrys et al., 2005; Jerrett et al., 2005) and spatial interpolation (Briggs et al., 2000). In addition, LUR offers practical advantages over these other approaches. Dispersion models require highly specific data inputs and considerable expertise to initialize and parameterize (Jerrett et al., 2005). Spatial pollutant interpolation relies exclusively on data from monitoring networks, which may be sparsely distributed and/or may fail to capture variability of pollution concentrations over short distances (Larssen et al., 1993; Briggs et al., 1997, 2000; Wu et al., 2005). In contrast, LUR typically utilizes existing land use data in combination with either readily available air quality measurements (e.g., from regulatory monitoring networks) (Ross et al., 2007) and/or measurements from dedicated sampling campaigns (Brauer et al., 2003; Gilbert et al., 2005; Sahsuvaroglu et al., 2006; Smith et al., 2006; Henderson et al., 2007).

The largest expense when implementing LUR is that associated with a dedicated field monitoring campaign. Costs can be reduced by representing all traffic pollution with a single proxy pollutant like NO2 (Hochadel et al., 2006; Ross et al., 2006; Sahsuvaroglu et al., 2006). This indicator is particularly useful because it can be measured using relatively inexpensive passive samplers. The response variable for LUR modeling is typically obtained by deploying several such monitors throughout the study area for a period of 1–2 weeks. To minimize bias introduced by seasonal trends in background concentrations, sampling sessions may either be conducted during a period of the year that represents the annual average (Henderson et al., 2007) or throughout multiple seasons (Brauer et al., 2007; Jerrett et al., 2007).

Placement of monitors throughout the study area is an important consideration when designing a sampling plan. Location-allocation, a technique used in many of the major LUR studies (Gilbert et al., 2005; Kanaroglou et al., 2005; Sahsuvaroglu et al., 2006; Henderson et al., 2007; Jerrett et al., 2007) makes use of a geographic information system and existing monitoring data to determine optimal locations for temporary monitoring sites to maximize their ability to represent spatial variations in pollutant concentrations. No methodology currently exists for determining the minimum number of sampling locations to perform LUR. In previous studies, large samples of 100+ locations have been used (Jerrett et al., 2005; Sahsuvaroglu et al., 2006; Henderson et al., 2007), though smaller sample sizes ranging from 20 to 60 sites have been used in others (Brauer et al., 2003; Gilbert et al., 2005; Gonzales et al., 2005; Ross et al., 2006; Ryan et al., 2007; Hoek et al., 2002a). It should be noted that other LUR studies have relied solely on regulatory network data and have developed models without new field monitoring (Ross et al., 2007). The representation of pollution levels of such models is heavily dependent on the specific regulatory site locations, many of which are located to represent average urban concentrations. Using such data may fail to capture the actual range in values that exist throughout the area, as the placement of these monitors may exclude hotspots or high concentration areas.

The ability to apply one LUR model to locations outside its source region may significantly reduce the time and expense necessary to construct new models. It may be possible to transfer a source model to a new area directly, or to use a limited number of sampling locations to locally calibrate the original model for a new setting. So far, the transferability of LUR models has received little attention. Briggs et al. (2000) applied one LUR model developed for Huddersfield, UK, to four other cities in Europe. The source model was found to both over- and under-predict pollution concentrations when applied in the new study areas. However, local calibration using a small number of supplementary samples (10) provided estimates within 1.5 of the actual mean in 70–90% of cases. More recently, Jerrett et al. (2005) applied the coefficients from an Amsterdam model (Briggs et al., 2000) to Hamilton, Ontario. As shown by their example, model transfer resulted in a surface that lacked spatial variability and was over-predicted in most areas. One likely explanation lies in the land use, vehicle fleet, and population density differences between European and North American cities (Gilbert et al., 2005). We hypothesize that LUR transfer may be more successful between geographically comparable areas with similar characteristics, but that local calibration will improve transfer results by accounting for difference in meteorology and topography. We test this hypothesis by transferring an LUR model developed for Vancouver, BC to Victoria, BC and Seattle, WA.

Study Area

The Georgia Basin-Puget Sound (GBPS) airshed spans the western coasts of Canada and the United States, as shown in Figure 1. This region is currently home to more than seven million people, with growth expected to increase the population to nine million by 2020. Vancouver, Victoria and Seattle are the three major urban centers in this area. The Greater Vancouver Regional District is the largest metropolitan area in western Canada, with a 2001 population of nearly 2.2 million (Statistics Canada). Victoria, which lies at the southern tip of Vancouver Island, is the second largest city in BC with a 2001 population of >300,000 (Statistics Canada). Seattle, south of both Vancouver and Victoria, is the largest city in the airshed, with a metropolitan population of approximately 3.2 million in 2005 (US Census Bureau). Although air quality in the GBPS airshed has historically met both Canadian and US standards, future population growth and increasing motor vehicle use may require new management strategies to maintain compliance with relevant standards.

Figure 1
figure1

Location of the three study cities within the Georgia Basin Puget Sound airshed.

Methods

Source LUR Model

The source LUR model was a modified version of that developed for Vancouver by Henderson et al. (2007):

Where:

Transfer Scenarios

We used three scenarios to evaluate the utility of the Vancouver source model for predicting ambient concentrations of NO2 in Victoria and Seattle. In Scenario 1, the source model was applied directly to both cities. Field measurements of NO2 were compared to the concentrations predicted by the source model at the same locations with linear regression. This represents an evaluation of how the source model would perform in new locations with no additional field measurements. In Scenario 2, the variables in the source model were retained, but a new regression model was developed to produce coefficients specific to the field data from both new cities. This represents an evaluation of how well local calibration of an existing model performs, given the limited field monitoring. In Scenario 3, additional spatial variables were generated and linear regression analysis was used to determine if an improved R2 value could be achieved with limited field data.

Data Sources and Spatial Variable Generation

Predicted NO2 surfaces for Vancouver and Victoria were generated using geographic data sets from the same data provider. These included road length, land use, population density and elevation data from DMTI Spatial (Markham, Ontario). The same data sets are not available for Seattle, WA, and the following alternatives were used: (1) roads from the US Bureau of Transportation and Statistics (2000); (2) land use from the US Environmental Protection Agency (1998); (3) population from the US Census Bureau (2000) and (4) elevation from the US Geological Survey (2000). These data sets are analogous to those used in Vancouver and Victoria, although they are from different data providers, and there may be discrepancies between how they are collected and classified. The same variable categories, subcategories and buffer radii distances used by Henderson et al. (2007) were created for Victoria and Seattle, with the exception of land use subcategories for Seattle. While the Canadian land use data were classified into 5 categories, the American land use data were classified into 19 categories. Seattle data were therefore reclassified into categories similar to those used for Vancouver and Victoria (Table 1).

Table 1 Reclassification of Seattle land use data

For Scenario 3, a full complement of potentially predictive variables was generated for both cities, including: (1) the length of highways (RD1) and major roads (RD2) at 100, 200, 300, 500, 750 and 1000 meter buffers; (2) the area of different land use types at 300, 400, 500 and 750 meter buffers; (3) population density (POP) at 750, 1000, 1250, 1500, 2000 and 2500 meter buffers; (4) elevation (ELEV), latitude (X) and longitude (Y) and (5) distance to the nearest highway (HWY) and coastline (COAST). We had a total of 43 variables for Victoria and 47 for Seattle due to the slightly different land use categories. The same model-building approach described by Henderson et al. (2007) was used to identify best-fitting regression models for Victoria and Seattle. Briefly, all variables within each group (e.g., RD1, POP) were ranked according to the absolute strength of their correlation with NO2. Variables that were highly correlated (Pearson's r>0.60) with the group's highest-ranking member were dropped from further consideration. The remaining variables were entered into stepwise regression, and those (1) with insignificant t-statistics or (2) that violated a priori assumptions (e.g., road length variables are expected to have a positive coefficient, pollution concentrations are expected to decrease with increasing distance from roads) were removed. Variables that contributed less than 1% to the R2 were also removed to identify a final parsimonious model.

Field Monitoring

The Vancouver model was developed using integrated NO2 measurements taken with Ogawa® passive diffusion samplers (Ogawa and Co., USA) during two periods in 2003 (Henderson et al., 2007). Ogawa samplers work by trapping the NO2 on filters coated with thriethanolamine, which are then extracted into water so that the nitrite concentration can be analyzed by ion chromatography. More detailed descriptions of sampler design and performance can be found elsewhere (Yamada et al., 1999; Mukerjee et al., 2004; Sather et al., 2007).

The same type of Ogawa samplers were used to measure NO2 at 40 sites in Victoria (Figure 2). Samplers were deployed on June 22 or June 23, 2006 and collected in the same order on July 6 or 7. Site selection was determined by using the Vancouver model to generate a preliminary map of concentration estimates for Victoria, which we classified into five categories by natural breaks. Seven locations were randomly selected along road networks within each category at a minimum between-site distance of 500 m using the ArcView (ESRI, Redlands, CA, USA) extension Random Point Generator v.1.3 (Jenness, 2005). In addition to these 35 locations, five samplers were placed at sites of special community interest. Two further samplers were collocated with the existing BC Ministry of Environment (MOE) monitoring stations in Victoria. To ensure sample quality (1) duplicates were taken at three locations, (2) five field blanks were momentarily exposed at a random selection of sites, and (3) five laboratory blanks were included. All samplers were exposed for a period of 336±3 h, at a height of 2.5 m, and were analyzed by ion chromatography at the UBC School of Occupational and Environmental Hygiene. The limit of detection was 0.7 p.p.b. of NO2 in air. During this sampling period no precipitation was observed and the average temperature was 17.6°C.

Figure 2
figure2

Locations of Victoria sample sites.

Field samples and blanks were analyzed between 13 and 18 July 2006. Two outliers were identified and removed from the data prior to analysis. First, one site selected to represent low concentrations was found to be located near a works yard serviced by high volume diesel traffic. Second, air flow at another site selected to represent high concentrations was restricted by access barriers.

Field monitoring of community-level NOX concentrations at 26 locations in Seattle was performed as part of the MESA Air pilot study (Cohen et al., 2007). We used this data set in a transferability exercise, although the sample was not designed for this purpose. This sample was primarily undertaken to refine a field sampling protocol for the MESA Air study, with the majority of sampling locations being clustered near major highways and streets to capture concentration gradients. Most samplers were placed between 5 and 350 m from three target roadways (one highway and two major arterial roads) oriented in different directions. Additional samplers were deployed to complete a grid-like pattern throughout the north end of Seattle (Figure 3). One sampler was collocated with the Washington State Department of Ecology's (WDOE) continuous NOx analyzer on Beacon Hill, and three duplicates were collected. Samplers were located approximately 2.5 m above the ground on telephone or utility poles. All samplers were deployed on 3 March 2005 and retrieved on 17 March, resulting in exposure durations between 334 and 338 h. Analysis by ion chromatography was conducted at the University of Washington's Exposure Assessment Laboratory.

Figure 3
figure3

Locations of Seattle sample sites.

To remove bias introduced by seasonal trends in background concentrations, the NO2 data from Victoria and Seattle were converted to effective annual averages. First, the ratios of the campaign averages to the annual average were calculated using data from continuous NO2 monitors in both cities. Second, the concentrations measured by Ogawa samplers were divided by those ratios to obtain an effective annual average for each site (Hoek et al., 2002a, 2002b). Spatial variation is preserved by assuming that the same ratio applies to all NO2 measurement sites within a city. For Victoria, data from the continuous sampler at Royal Roads University were used. The ratio for the June 2006 sampling period over the 2005 annual mean was 5.5 p.p.b./11.3 p.p.b., or 0.49. For Seattle, data from the continuous sampler on Beacon Hill were used. Owing to a large number of invalid records during the second half of 2005, the period from July 2004 to June 2005 was taken as the annual average. The ratio for the March 2005 sampling period over this annual mean was 19.3 p.p.b./17.9 p.p.b., or 1.1. Annualized values were used for the development and evaluations of all subsequent model.

Results and discussion

Field Monitoring Results

In Victoria, good agreement between duplicate samplers (<1 p.p.b. difference) was found. Laboratory testing revealed a systematic increase in concentrations (13%) between the samplers analyzed on July 13 and the field blanks analyzed on 18 July. This trend was discovered by reanalyzing a random set of samplers from the earlier batch with the field blanks to determine how the extra time influenced recorded concentrations. Field blanks were therefore adjusted down by 13% prior to subtracting their mean concentration from all samplers. The measured NO2 field concentrations at the 40 study sites ranged from 0.4 to 10.2 p.p.b. with mean, median and standard deviation values of 4.9, 4.4 and 2.6 p.p.b., respectively.

Agreement between collocated passive samplers and the two continuous monitors in Victoria was poor, with the difference in measurements between Ogawa samplers and continuous monitors ranging from 3.96 to 8.19 p.p.b. Additional monitoring was conducted during 2 weeks in August 2006 to determine whether an error in laboratory preparation, field handling or analysis might explain the discrepancies. Another two Ogawa samplers were collocated with each of the two monitors in Victoria, and three samplers were located at one site in Vancouver. In Victoria, at Site 1: Ogawa no.1=1.1, Ogawa no.2=3.6, Continuous=5.4. At Site 2: Ogawa no. 1=4.4, Ogawa no. 2=8.3, Continuous=14.1. At the Vancouver Site: Ogawa no. 1=10.6, Ogawa no. 2=12.9, Ogawa no. 3=7.9, Continuous=13.9. In both Victoria and Vancouver, collocated Ogawa samplers recorded different observations from each other as well as from the continuous monitors. In both the June and August sampling periods, the NO2 concentrations at Victoria Site no. 1 were very low (0.8, 1.1 and 3.6 p.p.b.). Victoria Site no. 2 may have been impacted by frequent truck traffic within several meters of the monitor, as a city works yard was being used as a depot for landscaping material during the sampling period. Additionally, lack of access to the continuous monitoring sites required that we attach our Ogawa monitors to the fences surrounding the sites, at locations up to 10 m away from the continuous monitor roof-top intake. NO2 values can vary significantly over short distances, especially in the vertical gradient, and this is most likely the largest factor in the poor collocation results of the duplicate samplers relative to each other as well as to the continuous readings. Despite the poor agreement between the Ogawa samplers and the continuous monitors in Victoria, the good agreement observed in Seattle and in other sampling campaigns using the same laboratory and methods (Henderson et al., 2007) and the good agreement between the measured and predicted concentrations supports the use of these data as valid measurements of ambient concentrations.

Good agreement between duplicate samples was also observed in Seattle (<1 p.p.b. difference). Results from the collocated sampler showed that the NO2 concentration measured by the passive sampler was 3.4% lower than that measured by the continuous analyzer. Measured field concentrations at the other 26 sites ranged from 11.6 to 21.4 p.p.b., with mean, median and standard deviation values of 16.0, 16.1 and 2.3 p.p.b., respectively. The average of the annualized NO2 concentrations was 10.3 p.p.b. (range: 0.8–20.9) in Victoria and 14.8 p.p.b. (range: 10.8–19.9) in Seattle.

Scenario 1: Same Variables, Same Coefficients

The R2 value for predicted vs. observed NO2 concentrations in Victoria was 0.51, which approaches that for the Vancouver source model (0.53), but in Seattle it was substantially lower (0.33). Figure 4 shows the linear relationships. Greater consistency of the input data and other similarities between Vancouver and Victoria may explain this difference in transferability. In both cities, LUR predictions produce higher means than do field measurements, though the difference is more pronounced in Victoria. This suggests that the Vancouver source model over-predicts NO2 concentrations in Victoria, a city with a much lower population density.

Figure 4
figure4

Comparison of measured vs. predicted NO2 concentrations (p.p.b.) and R2 values for Scenarios 1–3 in Victoria and Seattle.

It is a possibility that road proximity variables of the source model are responsible for the over-prediction of concentrations in Victoria. The RD1 and RD2 networks in Victoria are less utilized or at/near capacity less of the time than the corresponding network in Vancouver. Road length in a given buffer size is translated into NO2 concentration by the source model on the basis of the actual NO2 concentrations resulting from that road length in Vancouver. As vehicle emissions per nearby length of road is most likely less in Victoria, the model developed for Vancouver will attribute more pollution from the nearby roads than is actually the case in Victoria.

Also, the measured concentrations are more variable than the model-predicted concentrations in Victoria, while in Seattle the measured concentrations are less variable than the predicted concentrations. This difference is most likely due to the sampling design, in which the Seattle measurements represent less spatial variability than those in either Victoria or Vancouver.

Scenario 2: Same Variables, New Coefficients

As shown in Table 2, local calibration significantly improved model performance, with the Victoria R2 value increasing to 0.58 and the Seattle R2 value increasing to 0.65. However, some variables that were included in the source LUR equation are no longer significant in the new models. In Victoria, the variables RD2.200, RD1.1000 and POP.2500 are retained, while RD1.100, COMM.750 and ELEV are not. In Seattle, POP.2500 is the only predictor variable that remains significant. These results support the idea that Vancouver and Victoria are more similar, either in terms of characteristics or input data. The different R2 values of the Scenario 2 models (0.58–0.65) are similar to the results of Briggs et al. (2000) from the transfer of their Huddersfield model to other cities (R2 values ranging from 0.51–0.76 in new locations).

Table 2 Results of regression analysis using the same variables as source equation but new coefficients (Scenario 2)

Scenario 3: New Variables, New Coefficients

New LUR models were developed with field data from each city, resulting in improved R2 values over Scenarios 1 and 2 (Victoria=0.61, Seattle=0.72), although the improvement from Scenario 2 to 3 is not as great as that achieved by locally calibrating the Vancouver LUR model (Scenario 1 to 2). Table 3 shows the final models. All variables have significant t scores and acceptable multicollinearity, as demonstrated by the low variance inflation factors.

Table 3 LUR models generated from field measurements and best-fit variables for Victoria and Seattle (Scenario 3)

In Victoria, the new LUR model has four predictor variables, only one of which (ELEV) was present in the source LUR model and none of which were present under Scenario 2. Predictors from the RD1 and RD2 categories are significant but in different buffer radii than the original source model. Error for the Victoria model was estimated using leave-one-out cross-validation, with a mean value of zero and a standard deviation of 34%.

Similar to the Victoria model, the new Seattle LUR has four predictor variables. POP.2500 is the best population density predictor and its coefficient in Scenario 3 is very similar to the original Vancouver model. In the first iteration of developing a city-specific LUR model, residential land use within 500 m was the most predictive variable in Seattle, but its coefficient was negative. The influence of this variable was found to predict negative NO2 concentrations in outlying areas with few nearby major roads and low population density. A restriction was imposed that coefficients be positive and which yielded TRANS.750, a transportation, communications, and utilities land use designation, as the best land use predictor variable. The radius of influence is the same as the source model land use variable, commercially zoned land. RD1.100 and RD2.300 are both significant in Scenario 3, with the same radius as the Vancouver model for highways (RD1), and different for major roads (RD2) by only 100.

The high R2 value of the Scenario 3 model for Seattle may be due to the sampling design used for field data collection. As noted by Briggs et al. (2000), when selecting sample sites for local calibration, it is important to capture the range of actual values in the study area. In Victoria, we were able to use the predicted NO2 surface to help design the sampling locations, and this resulted in field data providing a good range of NO2 concentrations (mean=10.34, CV=51%) for subsequent use in developing a new model. Similarly, the European study by Briggs et al. (2000) selected sample sites stratified in terms of the Department of Environment, Transport and Regions classification and land use type in capture an adequate range. The NO2 data from Seattle, however, were from a previous pilot study and were collected predominantly to characterize gradients near major roadways. This may not have captured enough range in NO2 concentrations (mean=14.89, CV=15%) or the explanatory variables (Table 4). Also, the Seattle model is built on a smaller number (n=26) of observation points than the Victoria model (n=40). Reported R2 values for LUR models developed in other studies are higher when a smaller number of sample sites are used (Henderson et al., 2007), and this in combination with choice of sampler placement may explain the higher R2 values produced for Seattle.

Table 4 Descriptive statistics of Scenario 3 model variables for Victoria and Seattle

Conclusion

When input data are available from the same source, an LUR model may transfer to a geographically similar area reasonably well; however, given the input data from different sources or less geographic similarity, the transferability can be marginal. We are pleased with the relatively successful transferability between Vancouver and Seattle, despite their locations in different countries, where different agencies develop spatial data on the basis of different classification systems. To determine whether input data or physical dissimilarities have greater effect on transferability, an LUR model transfer could be evaluated for two cities with the same data sources that exhibit greater geographical differences than Vancouver and Victoria.

Allowing for local calibration of a source model substantially improves transferability; however, some field monitoring data are required. This study used a greater number of calibration sites (n=40 and 26) than were previously used in the European transfer (n=10) (Briggs et al., 2000) and obtained similar results. Given the field monitoring data for a new area, we recommend developing a new LUR model rather than attempting to directly transfer or locally calibrate an existing LUR model, as our results show that transferability is unpredictable and new models based on field monitoring may not contain the variables present in the source model. The utility of applying an existing LUR model comes from using it to inform a field sampling design for a relatively small number of sampling sites, although the location-allocation approach is also available. As the location-allocation technique requires a demand surface, transferring an existing LUR model may also be suitable for this purpose.

Our results generally confirm those reported in other studies of the transferability of LUR models. The immediate transfer of models between areas can result in both over- and under-prediction of NO2 concentrations. However, calibration using a small number of field samples can generate models with R2 values comparable to location-specific LUR models. The placement of monitors is a key consideration when designing a calibration sample, as monitors should be located to capture the range of pollutant levels and land use characteristics that exist throughout the study area. Existing data from monitors that provide a reasonable degree of spatial coverage, even if not representing the complete range of pollutant concentrations and land uses, may still give useful though not optimal results. The use of regulatory networks with no additional field monitoring is not recommended unless their locations reflect the full range of variability present in the study area. Generally, this is not the case in the typical North American setting, and it may be better suited in areas where regulatory monitors are not typically used to capture urban background concentrations.

References

  1. Brauer M., Hoek G., Smit H.A., de Jongste J.C., Gerritsen J., Postma D.S., Kerkhof M., and Brunekreef B. Air Pollution and the development of asthma, allergy and infections in a birth cohort. Eur Respir J 2007: 29: 879–888.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  2. Brauer M., Hoek G., van Vliet P., Meliefste K., Fischer P., Gehring U., Heinrich J., Cyrys J., Bellander T., Lewne M., and Brunekreef B. Estimating long-term average particulate air pollution concentrations: application of traffic indicators and geographic information systems. Epidemiology 2003: 14 (2): 228–239.

    Google Scholar 

  3. Brauer M., Hoek G., Van Vliet P., Meliefste K., Fischer P.H., Wijga A., Koopman L.P., Neijens H.J., Gerritsen J., Kerkhof M., Heinrich J., Bellander T., and Brunekreef B. Air pollution from traffic and the development of respiratory infections and asthmatic and allergic symptoms in children. Am J Respir Crit Care Med 2002: 166 (8): 1092–1098.

    Article  PubMed  PubMed Central  Google Scholar 

  4. Briggs D., Collins S., Elliott P., Fischer P., Kingham S., Lebret E., Pryl K., Reeuwijk H.V., and Smallbone K. Mapping urban air pollution using GIS: a regression-based approach. Int J Geogr Inf Sci 1997: 11: 669–718.

    Article  Google Scholar 

  5. Briggs D., de Hoogh C., Gulliver J., Wills J., Elliott P., Kingham S., and Smallbone K. A regression-based method for mapping traffic-related air pollution: application and testing in four contrasting urban environments. Sci Total Environ 2000: 253: 151–167.

    CAS  Article  Google Scholar 

  6. Brunekreef B., and Holgate S.T. Air pollution and health. Lancet 2002: 360 (9341): 1233–1242.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  7. Cohen M., Allen R., Gould T., Adar S., Hardie D., Kaufman J., Larson T., and Hinckley-Stukovsky K. Multi-Ethnic Study of Atherosclerosis—Air Pollution (MESA Air Pollution): Exposure Assessment Methodology. In press, spring, 2007.

    Google Scholar 

  8. Cyrys J., Hochadel M., Gehring U., Hoek G., Diegmann V., Brunekreef B., and Heinrich J. GIS-based estimation of exposure to particulate matter and NO2 in an urban area: stochastic versus dispersion modeling. Environ Health Perspect 2005: 113 (8): 987–992.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  9. Gauderman W., Avol E., Lurmann F., Kuenzli N., Gilliland F., Peters J., and McConnel R. Childhood asthma and exposure to traffic and nitrogen dioxide. Epidemiology 2005: 16 (6): 737–743.

    Article  Google Scholar 

  10. Gilbert N., Goldberg M., Beckerman B., Brook J., and Jerrett M. Assessing spatial variability of ambient nitrogen dioxide in Montréal, Canada, with a land-use regression model. J Air Waste Manag Assoc 2005: 55: 1059–1063.

    CAS  Article  Google Scholar 

  11. Gonzales M., Qualls C., Hudgens E., and Neas L. Characterization of a spatial gradient of nitrogen dioxide across a United States-Mexico border city during winter. Sci Total Environ 2005: 337 (1–3): 163–173.

    CAS  Article  Google Scholar 

  12. Henderson S.B., Beckerman B., Jerrett M., and Brauer M. Application of land use regression to estimate long-term concentrations of traffic-related nitrogen oxides and fine particulate matter. Environ Sci Technol 2007: 41: 2422–2428.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  13. Hochadel M., Heinrich J., Gehring U., Morgenstern V., Kuhlbusch T., Link E., Wichmann H., and Krämer U. Predicting long-term average concentrations of traffic-related air pollutants using GIS-based information. Atmos Environ 2006: 40: 542–553.

    CAS  Article  Google Scholar 

  14. Hoek G., Brunekreef B., Goldbohm S., Fischer P., and van den Brandt P.A. Association between mortality and indicators of traffic-related air pollution in the Netherlands: a cohort study. Lancet 2002a: 360 (9341): 1203–1209.

    Article  PubMed  PubMed Central  Google Scholar 

  15. Hoek G., Meliefste K., Cyrys J., Lewné Brauer M., Fischer P., Gehring U., van Vliet P., Heinrich J., Bellander T., and Brunekreef B. Spatial variability of fine particle concentrations in three European countries. Atmos Environ 2002b: 36: 4077–4088.

    CAS  Article  Google Scholar 

  16. Jenness J. Random point generator (randpts.avx) extension for ArcView 3.x, v 1.3. Jenness Enterprises, 2005 Available at: http://www.jennessent.com/arcview/random_points.htm.

  17. Jerrett M., Arain A., Kanaroglou P., Beckerman B., Crouse D., Gilbert N., Brook J., Finkelstein N., and Finkelstein M. Modeling the intraurban variability of ambient traffic pollution in Toronto, Canada. J Toxicol Environ Health A 2007: 70: 200–212.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  18. Jerrett M., Arain A., Kanaroglou P., Beckerman B., Potoglou D., Sahsuvaroglou T., Morrison J., and Giovis C. A review and evaluation of intraurban air pollution exposure models. J Expo Anal Environ Epidemiol 2005: 15: 185–204.

    CAS  Article  Google Scholar 

  19. Kanaroglou P., Jerrett M., Morrison J., Beckerman B., Arain M.A., Gilbert N., and Brook J. Establishing an air pollution monitoring network for intra-urban population exposure assessment; a location-allocation approach. Atmos Environ 2005: 39: 2399–2409.

    CAS  Article  Google Scholar 

  20. Krzyzanowski M., Kuna-Dibbert B., and Schneider J., (eds) Health effect of transport related air pollution. Report. World Health Organization 2005.

  21. Künzli N., Kaiser R., Medina S., Studnicka M., Chanel O., Filliger P., Herry M., Horak Jr F., Puybonnieux-Texier V., Quénel P., Schneider J., Seethaler R., Vergnaud J.-C., and Sommer H. Public-health impact of outdoor and traffic-related air pollution: a European assessment. Lancet 2000: 356 (9232): 795–801.

    Article  PubMed  PubMed Central  Google Scholar 

  22. Larssen S., Tonnesen D., Clenchaas J., Aarnes M.J., and Arnesen K. A model for car exhaust exposure calculations to investigate health-effects of air-pollution. Sci Total Environ 1993: 134: 51–60.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  23. Lebret E., Briggs D., van Reeuwijk H., Fischer P., Smallbone K., Harssema H., Bohumir K., Gorynski P., and Elliott P. Small area variations in ambient NO2 concentrations in four European areas. Atmos Environ 2000: 34: 177–185.

    CAS  Article  Google Scholar 

  24. Mukerjee S., Smith L., Norris G., Morandi M., Gonzales M., Noble C., Neas L., and Ozkaynak A. Field method comparison between passive air samplers and continuous monitors for volatile organic compounds and NO2 in El Paso, Texas, USA. J Air Waste Manag Assoc 2004: 54 (3): 307–319.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  25. Ross Z., English P., Scalf R., Gunier R., Smorodinsky S., Wall S., and Jerrett M. Nitrogen dioxide prediction in Southern California using land use regression modeling: potential for environmental health analyses. J Expo Sci Environ Epidemiol 2006: 16: 106–114.

    CAS  Article  Google Scholar 

  26. Ross Z., Jerrett M., Kazuhiko I., Tempalski B., and Thurston G. A land use regression for predicting fine particulate matter concentrations in the New York City region. Atmos Environ 2007: 41: 2255–2269.

    CAS  Article  Google Scholar 

  27. Ryan P.H., LeMasters G.K., Biswas P., Levin L., Hu S., Lindsey M., Bernstein D.I., Lockey J., Villareal M., Hershey G.K., and Grinshpun S.A. A comparison of proximity and land use regression traffic exposure models and wheezing in infants. Environ Health Perspect 2007: 115: 278–284.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  28. Sahsuvaroglu T., Arain A., Kanaroglou P., Finkelstein N., Newbold B., Jerrett M., Beckerman B., Brook J., Finkelstein M., and Gilbert N. A land use regression model for predicting ambient concentration of nitrogen dioxide in Hamilton, Ontario, Canada. J Air Waste Manage Assoc 2006: 56: 1059–1069.

    CAS  Article  Google Scholar 

  29. Sather M., Slonecker E., Johnson M., Daughtrey H., and Williams D. Evaluation of ogawa passive sampling devices as an alternative measurement method for the nitrogen dioxide annual standard in El Paso, Texas. Environ Monit Assess 2007: 124 (1–3): 211–221.

    CAS  Article  Google Scholar 

  30. Smith L., Mukerjee S., Gonzales M., Stallings C., Neas L., Norris G., and Özkaynak H. Use of GIS and ancillary variables to predict volatile organic compound and nitrogen dioxide levels at unmonitored locations. Atmos Environ 2006: 40: 3773–3787.

    CAS  Article  Google Scholar 

  31. Statistics Canada. Population and Dwelling Counts for Canada, Provinces and Territories, Census Metropolitan Areas and Census Agglomerations, 2001 and 1996 Census- 100%; available on the Statistics Canada web site. http://www12.statcan.ca/english/census01/products/standard/popdwell/tables.cfm (accessed March 2008).

  32. Wilhelm M., and Ritz B. Local variations in CO and particulate air pollution and adverse birth outcomes in Los Angeles County, California, USA. Environ Health Perspect 2005: 113 (9): 1212–1221.

    CAS  Article  PubMed  PubMed Central  Google Scholar 

  33. Wu J., Lurmann F., Winer A., Lu R., Turco R., and Funk T. Development of an individual exposure model for application to the Southern California children's health study. Atmos Environ 2005: 39: 259–273.

    CAS  Article  Google Scholar 

  34. Yamada E., Kimura M., Tomozawa K., and Fuse Y. Simple analysis of atmospheric NO2, SO2 and O3 in mountains by using passive samplers. Environ Sci Technol 1999: 33 (23): 4141–4145.

    CAS  Article  Google Scholar 

Download references

Acknowledgements

We acknowledge Steve Sakiyama and Warren McCormick from the BC Ministry of Environment for financial support as well as for providing data and access to monitoring locations; BC Hydro for the use of utility poles; and David Hardie for analyzing the Ogawa samplers collected in Seattle.

Author information

Affiliations

Authors

Corresponding author

Correspondence to Karla Poplawski.

Rights and permissions

Reprints and Permissions

About this article

Cite this article

Poplawski, K., Gould, T., Setton, E. et al. Intercity transferability of land use regression models for estimating ambient concentrations of nitrogen dioxide. J Expo Sci Environ Epidemiol 19, 107–117 (2009). https://doi.org/10.1038/jes.2008.15

Download citation

Keywords

  • air pollution
  • exposure modeling
  • land use regression
  • geographic information systems

Further reading

Search

Quick links