|Spatial modeling of PM10 and NO2 in the continental United States, 1985-2000.|
|Jump to Full Text|
|PMID: 20049118 Owner: NLM Status: MEDLINE|
|BACKGROUND: Epidemiologic studies of air pollution have demonstrated a link between long-term air pollution exposures and mortality. However, many have been limited to city-specific average pollution measures or spatial or land-use regression exposure models in small geographic areas.
OBJECTIVES: Our objective was to develop nationwide models of annual exposure to particulate matter < 10 microm in diameter (PM(10)) and nitrogen dioxide during 1985-2000.
METHODS: We used generalized additive models (GAMs) to predict annual levels of the pollutants using smooth spatial surfaces of available monitoring data and geographic information system-derived covariates. Model performance was determined using a cross-validation (CV) procedure with 10% of the data. We also compared the results of these models with a commonly used spatial interpolation, inverse distance weighting.
RESULTS: For PM(10), distance to road, elevation, proportion of low-intensity residential, high-intensity residential, and industrial, commercial, or transportation land use within 1 km were all statistically significant predictors of measured PM(10) (model R(2) = 0.49, CV R(2) = 0.55). Distance to road, population density, elevation, land use, and distance to and emissions of the nearest nitrogen oxides-emitting power plant were all statistically significant predictors of measured NO(2) (model R(2) = 0.88, CV R(2) = 0.90). The GAMs performed better overall than the inverse distance models, with higher CV R(2) and higher precision.
CONCLUSIONS: These models provide reasonably accurate and unbiased estimates of annual exposures for PM(10) and NO(2). This approach provides the spatial and temporal variability necessary to describe exposure in studies assessing the health effects of chronic air pollution.
|Jaime E Hart; Jeff D Yanosky; Robin C Puett; Louise Ryan; Douglas W Dockery; Thomas J Smith; Eric Garshick; Francine Laden|
Related Documents :
|15112818 - Fugacity-based indoor residential pesticide fate model.
12959828 - Overview of the reanalysis of the harvard six cities study and american cancer society ...
17149818 - A process-based mathematical model on methane production with emission indices for cont...
17169408 - The gis-based safeairview software for the concentration assessment of radioactive poll...
6879168 - Critical considerations in the assessment of health and environmental risks - what we h...
19477248 - "toxrtool", a new tool to assess the reliability of toxicological data.
|Type: Comparative Study; Journal Article; Research Support, N.I.H., Extramural; Validation Studies Date: 2009-06-29|
|Title: Environmental health perspectives Volume: 117 ISSN: 1552-9924 ISO Abbreviation: Environ. Health Perspect. Publication Date: 2009 Nov|
|Created Date: 2010-01-05 Completed Date: 2010-03-26 Revised Date: 2014-09-12|
Medline Journal Info:
|Nlm Unique ID: 0330411 Medline TA: Environ Health Perspect Country: United States|
|Languages: eng Pagination: 1690-6 Citation Subset: IM|
|APA/MLA Format Download EndNote Download BibTex|
Air Pollution / analysis*
Environmental Monitoring / methods
Inhalation Exposure / statistics & numerical data*
Nitrogen Dioxide / analysis
Particulate Matter / analysis
United States / epidemiology
Vehicle Emissions / analysis
|ES00002/ES/NIEHS NIH HHS; R01 CA090792/CA/NCI NIH HHS; R01 CA090792-05/CA/NCI NIH HHS; R01 CA90792/CA/NCI NIH HHS; T32 ES007069-29/ES/NIEHS NIH HHS|
|0/Air Pollutants; 0/Particulate Matter; 0/Vehicle Emissions; S7G510RUBH/Nitrogen Dioxide|
Journal ID (nlm-ta): Environ Health Perspect
Publisher: National Institute of Environmental Health Sciences
This is an Open Access article: verbatim copying and redistribution of this article are permitted in all media for any purpose, provided this notice is preserved along with the article's original DOI.
Received Day: 26 Month: 3 Year: 2009
Accepted Day: 29 Month: 6 Year: 2009
Print publication date: Month: 11 Year: 2009
Electronic publication date: Day: 29 Month: 6 Year: 2009
Volume: 117 Issue: 11
First Page: 1690 Last Page: 1696
PubMed Id: 20049118
Publisher Id: ehp-117-1690
|Spatial Modeling of PM10 and NO2 in the Continental United States, 1985?2000|
|Jaime E. Hart123|
|Jeff D. Yanosky1|
|Robin C. Puett1456|
|Douglas W. Dockery13|
|Thomas J. Smith1|
1 Exposure, Epidemiology and Risk Program, Department of Environmental Health, Harvard School of Public Health, Boston, Massachusetts, USA
2 Channing Laboratory, Department of Medicine, Brigham and Women?s Hospital and Harvard Medical School, Boston, Massachusetts, USA
3 Department of Epidemiology, Harvard School of Public Health, Harvard School of Public Health, Boston, Massachusetts, USA
4 South Carolina Cancer Prevention and Control Program, University of South Carolina, Columbia, South Carolina, USA
5 Department of Environmental Health Sciences and
6 Department of Epidemiology and Biostatistics, Arnold School of Public Health, University of South Carolina, Columbia, South Carolina, USA
7 Department of Biostatistics, Harvard School of Public Health, Boston, Massachusetts, USA
8 Pulmonary and Critical Care Medicine Section, Medical Service, VA Boston Healthcare System, Boston, Massachusetts, USA
|Correspondence: Address correspondence to J.E. Hart, 181 Longwood Ave., Boston, MA 02115 USA. Telephone: (617) 525-2289. Fax: (617) 525-2578. E-mail: Jaime.
The authors declare they have no competing financial interests.
Acute exposures to particulate and gaseous air pollutants have been associated with morbidity and mortality in a large number of time-series studies [Pope and Dockery 2006; U.S. Environmental Protection Agency (EPA) 1993, 2004]. There are fewer cohort studies where it has been possible to examine the association of long-term exposures and mortality (Dockery et al. 1993; Finkelstein et al. 2003; Hoek et al. 2002; Jerrett et al. 2005b, 2005c; Laden et al. 2006; Lipfert et al. 2006; Miller et al. 2007; Nafstad et al. 2004; Nyberg et al. 2000; Pope et al. 1995, 2004; Rosenlund et al. 2006). In most long-term studies, exposure assessment has been limited mainly to city-specific average pollution measures or spatial or geographic information system (GIS)?based exposure models in small geographic areas (Adar and Kaufman 2007; Brauer et al. 2003; Briggs et al. 2000; Jerrett et al. 2005a; Liao et al. 2006; Ryan and LeMasters 2007; Su et al. 2008; Wheeler et al. 2008; Wong et al. 2004). One recent study has described a monthly spatiotemporal exposure model for the northeastern United States using a combination of spatial and GIS-derived covariates that outperformed models with spatial smoothing alone (Yanosky et al. 2008, 2009). Another recent report has detailed the use of universal kriging to predict pollution levels for the European Union (Beelen et al. 2009). The purpose of this analysis is to develop nationwide models of annual exposure to particulate matter < 10 ?m in diameter (PM10) and nitrogen dioxide, using a combination of spatial smoothing and regression of GIS-derived covariates. To date, few countrywide models have been available for these pollutants over our time scale of interest (1985?2000). We apply the model to the addresses of the workers in the Trucking Industry Particle Study (Garshick et al. 2008; Laden et al. 2007), a retrospective cohort study of male U.S. unionized trucking company workers, to illustrate its potential use in exposure assessment for long-term epidemiologic studies with members spread over the continental United States.
Details of the Trucking Industry Particle Study (TrIPS) are provided elsewhere (Garshick et al. 2008; Laden et al. 2007). Briefly, using personnel records from four large companies we identified 54,973 males with at least 1 day of work in 1985. Information was available on demographic variables, daily job and work location, and residential home address. Using an outside vendor (TeleAtlas, Lebanon, NH), we geocoded the last known residential addresses of 53,822 members living within the continental United States to at least the ZIP code level.
We obtained information on annual average PM10 (parameter codes 81102 and 85101) and NO2 from the U.S. EPA Air Quality System (AQS). The U.S. EPA provided these annual averages on a set of DVDs compiled in 2004 for U.S. EPA Science to Achieve Results program grant 83054501-0. Data from 1985?2000 were used for this study if an annual mean was reported, regardless of the primary monitoring objective of the monitor. All monitors in the continental United States were included, because excluding monitors such as those located near point or mobile sources would prevent us from incorporating all sources of spatial variability represented in the monitoring network. Latitude and longitude of each monitor were obtained from the AQS database and used to map the monitor locations using ArcGIS (version 9.2; ESRI, Redlands, CA). All monitors were checked for latitude/longitude accuracy and precision to the county level before inclusion.
We used generalized additive models (GAMs) to predict annual outdoor levels of PM10 and NO2 using smooth spatial surfaces and GIS-derived covariates. GAMs use semiparametric methods to model nonlinear, one-dimensional, and multidimensional functions using penalized splines (Hastie and Tibshirani 1990; Wood 2003, 2004, 2006). For both pollutants, models were constructed using 90% of the available monitoring locations for each calendar year. The remaining randomly selected 10% of monitors were used to perform cross-validation as described below.
First, the average spatial surface for each pollutant, 1985?2000, was generated in a GAM containing a bivariate thin-plate spline of the projected x- and y-coordinates of the monitoring locations and indicator variables for calendar year to adjust for temporal trends (Wood 2006). To obtain information on fine-scale long-term spatial patterns, we included one-dimensional penalized splines for a priori selected GIS-derived time-invariant covariates. The covariates we considered included distance to road, population density, elevation, surrounding land use, distance to and emission from power plants, and variables for census region of the country (northeast, west, south, and midwest) to adjust for regional patterns. These variables have previously been shown to be important predictors of ambient pollution (Adar and Kaufman 2007; Jerrett et al. 2005a; Ryan and LeMasters 2007; Yanosky et al. 2008, 2009). Each characteristic was assigned to the monitoring locations using ArcGIS.
Information from the StreetMap data set (ESRI) was used to determine distance to the nearest road. Road segments were first classified by U.S. Census Feature Class Code as A1 (primary roads, typically interstates, with limited access), A2 (primary major, noninterstate roads), or A3 (smaller, secondary roads, usually with more than two lanes) (U.S. Census Bureau 1993). The distance from each location to the nearest road of each road class was then calculated in meters. Land use data were compiled from the U.S. Geological Survey (USGS) 1992 National Land Cover Dataset (USGS 2007b), which provides data on 19 categories of land use in raster image files with 1 arc-sec (about 30 m) spatial resolution (Vogelmann et al. 2001). The proportion of low-intensity residential, high-intensity residential, and industrial/commercial/transportation land uses within 1 km of each location was calculated. Population density values were assigned to each monitoring location using data from the 2000 U.S. Census at the block group level (U.S. Census Bureau 1993). Elevation data for each location were compiled from the USGS National Elevation Dataset (USGS 2007a). Information on the tons of nitrogen oxides emitted annually from all U.S. power plants in 2004 was obtained from the U.S. EPA 2006 Emissions and Generation Resource Integrated Database (U.S. EPA 2007a). The distance to and the emissions from the nearest facility were determined for each NO2 monitoring location.
Each potential covariate (or groups of covariates for distance to road, land use, and power plant distance/emissions) was first considered separately in models that included the bivariate spline for the 1985?2000 spatial surface and the indicator variables for calendar year. We constructed multivariate models including all covariates that were statistically significant (p < 0.05) and led to a higher adjusted model R2. If covariates were no longer significant when included in the multivariate model, we omitted them unless they led to better model fit as determined by Akaike?s information criterion (AIC) and cross-validation testing.
To assess annual differences from the long-term spatial patterns of pollution, we first calculated the residuals from the final long-term multivariate GAM models. Then, for each calendar year, we created a bivariate smooth of the residuals using a two-dimensional thin-plate spline. Therefore, the annual average pollution at any location was predicted using the sum of the prediction from the long-term average surface/GIS-derived covariates and the prediction from the calendar-year specific residual spatial variability surface.
To perform cross-validation, we used regression parameters from the final models and the annual spatial surfaces to predict annual pollutant levels at the 10% of monitoring locations that were held out from the original models. We assessed the potential bias of each final model by calculating the prediction error as the difference between the observed and predicted values at each cross-validation monitoring location. We also assessed bias in the models by examining the intercept and slopes from linear regression of the predicted values on the measured values. The precision of the model was estimated by taking the square root of the mean of the squared prediction errors (RMSPE). In addition, a cross-validation R2 was obtained using the squared Pearson correlation between the measured values at the held-out observations and the model predictions.
For comparison, we also predicted exposures using a simpler spatial interpolation method, inverse distance weighting (IDW), which had been frequently used in the air pollution literature. For the IDW models, the annual predictions for any given location (cross-validation monitor location or cohort member address) were calculated by taking the average of the measured value at each monitor location times the inverse of the squared distance between each location and each monitor. IDW modeling was performed in ArcGIS (Johnston et al. 2004). The bias and precision of this simpler exposure modeling method was determined using cross-validation.
After the final GAM models were determined and cross-validated, the regression parameters were used to predict annual pollutant levels at the 53,822 residential addresses of the TrIPS cohort members. For comparison, IDW was also used to predict annual pollutant levels at the residential addresses. Statistical analyses were performed in PC SAS version 9.1 (SAS Institute Inc. 2006) and Unix R 2.7.0 (R Development Core Team 2006).
The number of monitors used in the models and annual distributions of pollutant levels are shown in Table 1. The levels of both pollutants decreased over time. The median value of PM10 in 1985 was 38.2 ?g/m3, and it fell to 23.0 ?g/m3 by 2000 (a 40% decrease). The median NO2 level decreased 23% over the same period, from 19.0 ppb to 14.6 ppb. The distributions of the GIS-derived covariates at the monitor locations considered in the GAM models are shown in Table 2. The covariate distributions were quite similar for both sets of monitors. As shown in Figure 1, the cohort participants are located throughout the continental U.S., and most live close to the monitoring locations. Specifically, the cohort members lived a median distance of 10.2 km from PM10 monitoring sites and 16.6 km from NO2 sites. Seventy-five percent of the cohort was no more than 21.1 km from a PM10 monitor included in the model and 35.6 km from an NO2 monitor included in the model.
The model with only the spatial spline and calendar year indicator variables had a model R2 of 0.48. Region of the country, distance to all three census classes of road, block group population density, elevation, proportion of low-intensity residential, high-intensity residential, and industrial, commercial, or transportation land use within 1 km were all statistically significant independent predictors of measured PM10 concentrations in univariate models. In a multivariate model, all predictors except population density (p = 0.15) remained statistically significant predictors of measured PM10 annual concentrations (Table 3). Population density was removed from the final model, because it did not increase the cross- validation R2 or model fit as determined by AIC. The final model had an R2 of 0.49. Increases in the proportion of surrounding land use used for high-intensity residential or for industrial, commercial, or transportation uses were associated with increases in measured PM10 levels. Increases in all other covariates were associated with decreases in measured PM10. The cross-validation R2 of the final model was 0.55. The median [and interquartile range (IQR)] prediction error of the final model was 0.24 (7.0) ?g/m3. The intercept and slope from the regression of observed and predicted measurements were 1.49 and 0.94, respectively, and the RMSPE was 9.1 ?g/m3. A plot of the observed versus expected values from the cross-validation is presented in Supplemental Material, available online (doi:10.1289/ehp.0900840.S1 via http://dx.doi.org/).
The model with only the spatial spline and calendar year indicators had a model R2 of 0.73. Region of the country, distance to road, block group population density, elevation, surrounding land use, distance to nearest NOx-emitting power plant, and the level of emissions from that power plant were all statistically significant predictors of measured NO2 concentrations in univariate models. In a multivariate model, all predictors remained statistically significant predictors of measured NO2 annual concentrations (Table 3). The final multivariate model had an R2 of 0.88. Increases in the block group population density, NOx emissions of the nearest power plant, and the proportion of surrounding land use used for low- or high-intensity residential or for industrial, commercial, or transportation uses were associated with increases in measured NO2 levels. Increases in all other covariates were associated with decreases in measured NO2. The cross validation R2 of the final model was 0.90. The median (and IQR) prediction error of the final model was 0.10 (3.7) ppb, the intercept and slope of the regression of observed and predicted measurements were 0.00 and 1.04, and the RMSPE was 3.5 ppb. A plot of the observed versus expected values from the cross-validation is presented in Supplemental Material (doi:10.1289/ehp.0900840.S1).
A summary of the cross-validation parameters for the IDW exposure models is presented in Table 4. For both pollutants, the cross-validation R2 of the IDW model (R2 = 0.44 for PM10 and 0.67 for NO2) was lower than those from the GAMs (R2 = 0.55 for PM10 and 0.90 for NO2). For PM10, the slope from regression for the IDW model was 0.76 and the slope for the GAM was 0.94, indicating greater accuracy. The median prediction error for the IDW model was almost half that of the GAM, also indicating greater accuracy, but the RMSPE was higher, indicating lower precision. In contrast, for NO2 the IDW prediction error was 10-fold higher than the GAM, and the RMSPE was almost twice as large.
The distribution of the GIS-derived variables for the residential addresses (n = 53,822) of the TrIPS cohort is presented in Table 5. The home addresses tended to be further away, on average, from each of the census road classes and from power plants than the monitors used to develop the models. The addresses were also located in areas with a lower proportion of high-intensity residential or industrial, commercial, or transportation land use, and the addresses were located further away from power plants than monitors, with lower annual emissions of NOx from the nearest plant, on average. The distributions of the covariates tended to be tighter than those of the monitoring locations but were not significantly different.
Figure 2 shows the distribution of the pollution values for each year at the cohort addresses. The mean predicted levels of the two pollutants decreased over the follow-up period, although there was little change in the overall spread of the distributions. The spatial distributions of the predictions for both PM10 and NO2 are shown in Figure 3. At all three time points shown, PM10 values are higher in the western half of the United States than in the east. For NO2, however, the levels in all time periods are highest in major cities. To compare the two prediction methods, Figure 4 shows the cohort predictions for PM10 at base-line (1985), midpoint (1993), and last year of follow-up (2000). There is moderate correlation between the results of the GAM and IDW PM10 models, although the IDW models tend to be lower than the predictions of the GAMs (thus their lower slope of 0.76 vs. 0.94 for the GAM when both are compared with measured concentrations). The Spearman correlations between the two prediction types were 0.66 for 1985, 0.64 for 1993, and 0.77 for 2000. As shown in Figure 4, there is also moderate correlation between the GAM and IDW NO2 models. Specifically, the Spearman correlation is 0.63 for 1985, 0.53 for 1993, and 0.51 for 2000. Overall, the IDW models tend to be lower than the GAM predictions and tend to have less variance (heterogeneity).
Our results show that GAMs with a combination of spatial smoothing and GIS-derived covariates are a practical method for predicting annual outdoor air pollution values for a cohort dispersed across the continental United States. The PM10 and NO2 GAM models were reasonably accurate and precise. The final model for NO2 had a model R2 of 0.88 and a cross-validation R2 of 0.90, whereas the final model R2 for PM10 was 0.49 and the cross-validation R2 was 0.55. Overall, the GAMs for both PM10 and NO2 outperformed the simpler IDW models, although there was a greater difference in the performance of the two modeling approaches for NO2.
As expected, based on the growing literature of land-use regression models, many GIS-derived predictors were important in the pollution models. Distance to the nearest road of each road class, distance to and emissions from the nearest power plant, and land-use terms defining the surrounding area, variables previously shown to represent major sources of ambient NO2 in the United States (U.S. EPA 2007b), were all statistically significant predictors of NO2. In PM10 models, distance to the nearest road of each road class was the most important class of predictors, likely representing traffic, an important local source of particulate matter (U.S. EPA 2004). These covariates did not improve the model R2 as much for PM10 as they did for NO2. It is possible that there are other important sources of PM10 that we have not included (e.g., sea salt, crustal materials) that would improve the model R2 more.
A growing number of studies have used spatial smoothing methods or models based on GIS-derived variables to predict ambient air pollution levels for use in epidemiologic studies (Adar and Kaufman 2007; Jerrett et al. 2005a; Ryan and LeMasters 2007). Many of these studies have relied on proximity to specific pollution sources or monitoring locations to assign exposures. Others have focused on characterizing pollution from a specific source, typically on-road vehicles (Hoek et al. 2001). The most commonly used GIS-based methods have used information on traffic volume and distance to roadways as surrogates of exposure (Adar and Kaufman 2007; Bayer-Oglesby et al. 2006; Forastiere and Galassi 2005; Garshick et al. 2003; Kan et al. 2007; Nitta et al. 1993; Oosterlee et al. 1996; Venn et al. 2005). In many of these studies, distance to road is divided into categories, or individuals are classified as exposed or not exposed, based on an a priori chosen distance. This method likely leads to exposure misclassification in many of these studies and is likely also quite sensitive to the buffer or category size selected. Another popular GIS-based exposure method is land use regression (Briggs et al. 1997; Hoek et al. 2001; Ryan and LeMasters 2007; Ryan et al. 2007; Su et al. 2008). This approach is typically used in smaller areas to model local spatial variability, and roadway networks and traffic are often inputs to these models, although some also include information on surrounding land use, meteorology, and ambient air pollution monitoring locations. Other studies have used spatial smoothing techniques of the ambient measurements in single cities or counties (Jerrett et al. 2005b; Meng et al. 2007). Although direct comparisons are not appropriate, our NO2 model R2 of 0.88 is higher than those observed in many land-use regression models (0.52?0.76) (Briggs et al. 2000; Cyrys et al. 2005; Gilbert et al. 2005; Rosenlund et al. 2008) or in an EU-wide model based on ordinary kriging (Beelen et al. 2009).
On a larger spatial scale, in an exposure assessment for the Women?s Health Initiative, kriging in ArcGIS was used to generate daily PM2.5 and PM10 estimates for the entire continental United States for the year 2000 (Liao et al. 2006; Szpiro et al. 2007). For PM10, the authors report a median prediction error of 0.04 ?g/m3 and an RMSPE of 19.48 ?g/m3. In a recent exposure assessment for the Nurses? Health Study, a combination of spatial smoothing and GIS-derived covariates was used to produce monthly predictions of PM10 1988?2002 for residences in the northeastern United States (Yanosky et al. 2008). This model has a mean prediction error of ?0.4 ?g/m3 and an RMSPE of 6.4 ?g/m3 across the entire region, with no discernable differences by state or level of urbanization. Our models are similar to this modeling approach: Both include spatial smoothing and GIS-based covariates to generate predictions. The Yanosky model allows the generation of monthly estimates of PM10 through a complex spatiotemporal model and allows the inclusion of time-varying covariates and control for seasonality. In contrast, although the model presented here also uses spatial smoothing and GIS-based covariates, it is more appropriate for annual means and is less computationally intensive. Therefore, for PM10, the amount of bias [measured by average (mean or median) prediction error] and precision (measured by RMSPE) in our final model are comparable to that of other studies in the United States.
Our exposure model has several important limitations. We rely on air pollution data from existing networks that are not uniformly distributed across the continental United States. However, the measures of precision and accuracy determined by cross-validation for the held-out monitoring locations indicated good predictive performance of the models. Additionally, most of the members of the specific cohort we are using in this analysis live close to monitoring locations, so the mismatch between monitor and subject locations is unlikely to be a large source of error in exposure for our chosen application. For studies where the cohort is located much further from monitoring locations, this would likely be a larger source of error. In focusing our modeling on annual means, we are likely missing important seasonal and temporal variability occurring within each year. In years with fewer monitoring locations, it is possible that our model is underpowered to detect annual differences from the long-term spatial trends; however, in later years, only 20?40 degrees of freedom were needed to fit these surfaces, so this may not be a large issue. Our model also does not include information on time-varying covariates (such as point-source pollution or weather, especially wind direction and speed, mixing height, and precipitation) or interactions between our chosen covariates and calendar year. It is likely that information on these factors would improve the predictive ability of our model; however, it would require a different modeling approach than the one we have chosen. By treating population density, distance from road, and land use as time invariant, we are assuming that these did not vary during the study period. This is not likely to be true and will lead to increased error in areas with rapidly changing infrastructure during this time period. Finally, we are using a spatial smoothing model for the entire continental United States. It has been suggested that regional models may be more appropriate for the continental United States (Szpiro et al. 2007); however, it has been shown that for daily predictions, regional models do not substantially outperform a single countrywide model (Liao et al. 2006). Our models are adjusted for region of the country (using indicator variables), and although including region did improve the fit of the models, the regional terms themselves were not significant.
In conclusion, our air pollution exposure model combining spatial smoothing techniques and GIS-based predictors is a useful way to provide estimates of U.S.-wide annual exposures for PM10 and NO2. These models can be used to produce reasonably accurate and precise measures of pollution at the residential addresses of participants in epidemiologic studies focusing on the adverse effects of constituents of air pollution as far back as 1985.
This study was supported by grant R01 CA90792 from the National Institutes of Health/National Cancer Institute, National Institute of Environmental Health Sciences (NIEHS) Center grant ES00002, and NIEHS T32 ES007069-29 (J.E.H.).
Supplemental Material is available online (doi:10.1289/ehp.0900840.S1 via http://dx.doi.org/).
We thank C. Paciorek for helpful statistical advice and M. Jacobson Canner for programming assistance.
|Adar SD,Kaufman JD. Year: 2007Cardiovascular disease and air pollutants: evaluating and improving epidemiological data implicating traffic exposureInhal Toxicol19suppl 113514917886061|
|Bayer-Oglesby L,Schindler C,Hazenkamp-von Arx ME,Braun-Fahrlander C,Keidel D,Rapp R,et al. Year: 2006Living near main streets and respiratory symptoms in adults: the Swiss Cohort Study on Air Pollution and Lung Diseases in AdultsAm J Epidemiol164121190119817032694|
|Beelen R,Hoek G,Pebesma E,Vienneau D,de Hoogh K,Briggs DJ. Year: 2009Mapping of background air pollution at a fine spatial scale across the European UnionSci Total Environ40761852186719152957|
|Brauer M,Hoek G,van Vliet P,Meliefste K,Fischer P,Gehring U,et al. Year: 2003Estimating long-term average particulate air pollution concentrations: application of traffic indicators and geographic information systemsEpidemiology14222823912606891|
|Briggs DJ,Collins S,Elliott P,Fischer P,Kingham S,Lebret E. Year: 1997Mapping urban air pollution GIS: a regression-based approachInt J Geogr Inf Sci117699718|
|Briggs DJ,de Hoogh C,Gulliver J,Wills J,Elliott P,Kingham S,et al. Year: 2000A regression-based method for mapping traffic-related air pollution: application and testing in four contrasting urban environmentsSci Total Environ2531?315116710843339|
|Cyrys J,Hochadel M,Gehring U,Hoek G,Diegmann V,Brunekreef B,et al. Year: 2005GIS-based estimation of exposure to particulate matter and NO2 in an urban area: stochastic versus dispersion modelingEnviron Health Perspect11398799216079068|
|Dockery DW,Pope CA III,Xu X,Spengler JD,Ware JH,Fay ME,et al. Year: 1993An association between air pollution and mortality in six U.S. citiesN Engl J Med32924175317598179653|
|Finkelstein MM,Jerrett M,DeLuca P,Finkelstein N,Verma DK,Chapman K,et al. Year: 2003Relation between income, air pollution and mortality: a cohort studyCMAJ169539740212952800|
|Forastiere F,Galassi C. Year: 2005Self report and GIS based modelling as indicators of air pollution exposure: is there a gold standard?Occup Environ Med62850850916046601|
|Garshick E,Laden F,Hart JE,Caron A. Year: 2003Residence near a major road and respiratory symptoms in U.S. veteransEpidemiology14672873614569190|
|Garshick E,Laden F,Hart JE,Rosner B,Davis ME,Eisen EA,et al. Year: 2008Lung cancer and vehicle exhaust in trucking industry workersEnviron Health Perspect1161327133218941573|
|Gilbert NL,Goldberg MS,Beckerman B,Brook JR,Jerrett M. Year: 2005Assessing spatial variability of ambient nitrogen dioxide in Montreal, Canada, with a land-use regression modelJ Air Waste Manag Assoc5581059106316187576|
|Hastie TJ,Tibshirani R. Year: 1990Generalized Additive Models New YorkChapman and Hall|
|Hoek G,Brunekreef B,Goldbohm S,Fischer P,van den Brandt PA. Year: 2002Association between mortality and indicators of traffic-related air pollution in the Netherlands: a cohort studyLancet36093411203120912401246|
|Hoek G,Fischer P,Van Den Brandt P,Goldbohm S,Brunekreef B. Year: 2001Estimation of long-term average exposure to outdoor air pollution for a cohort study on mortalityJ Expo Anal Environ Epidemiol11645946911791163|
|Jerrett M,Arain A,Kanaroglou P,Beckerman B,Potoglou D,Sahsuvaroglu T,et al. Year: 2005aA review and evaluation of intraurban air pollution exposure modelsJ Expo Anal Environ Epidemiol15218520415292906|
|Jerrett M,Burnett RT,Ma R,Pope CA III,Krewski D,Newbold KB,et al. Year: 2005bSpatial analysis of air pollution and mortality in Los AngelesEpidemiology16672773616222161|
|Jerrett M,Buzzelli M,Burnett RT,DeLuca PF. Year: 2005cParticulate air pollution, social confounders, and mortality in small areas of an industrial citySoc Sci Med60122845286315820591|
|Johnston K,Van Hoef JM,Krivoruchko K,Lucas N. Year: 2004Using ArcGIS Geostatistical Analyst:ArcGIS 9Redlands, CAESRI Press|
|Kan H,Heiss G,Rose KM,Whitsel E,Lurmann F,London SJ. Year: 2007Traffic exposure and lung function in adults: the Atherosclerosis Risk in Communities studyThorax621087387917442705|
|Laden F,Hart JE,Smith TJ,Davis ME,Garshick E. Year: 2007Cause-specific mortality in the unionized U.S. trucking industryEnviron Health Perspect1151192119617687446|
|Laden F,Schwartz J,Speizer FE,Dockery DW. Year: 2006Reduction in fine particulate air pollution and mortality: extended follow-up of the Harvard Six Cities studyAm J Respir Crit Care Med173666767216424447|
|Liao D,Peuquet DJ,Duan Y,Whitsel EA,Dou J,Smith RL,et al. Year: 2006GIS approaches for the estimation of residential-level ambient PM concentrationsEnviron Health Perspect1141374138016966091|
|Lipfert FW,Baty JD,Miller JP,Wyzga RE. Year: 2006PM2.5 constituents and related air quality variables as predictors of survival in a cohort of U.S. military veteransInhal Toxicol18964565716864555|
|Meng YY,Wilhelm M,Rull RP,English P,Ritz B. Year: 2007Traffic and outdoor air pollution levels near residences and poorly controlled asthma in adultsAnn Allergy Asthma Immunol98545546317521030|
|Miller KA,Siscovick DS,Sheppard L,Shepherd K,Sullivan JH,Anderson GL,et al. Year: 2007Long-term exposure to air pollution and incidence of cardiovascular events in womenN Engl J Med356544745817267905|
|Nafstad P,Haheim LL,Wisloff T,Gram F,Oftedal B,Holme I,et al. Year: 2004Urban air pollution and mortality in a cohort of Norwegian menEnviron Health Perspect11261061515064169|
|Nitta H,Sato T,Nakai S,Maeda K,Aoki S,Ono M. Year: 1993Respiratory health associated with exposure to automobile exhaust. I. Results of cross-sectional studies in 1979, 1982, and 1983Arch Environ Health48153587680850|
|Nyberg F,Gustavsson P,Jarup L,Bellander T,Berglind N,Jakobsson R,et al. Year: 2000Urban air pollution and lung cancer in StockholmEpidemiology11548749510955399|
|Oosterlee A,Drijver M,Lebret E,Brunekreef B. Year: 1996Chronic respiratory symptoms in children and adults living along streets with high traffic densityOccup Environ Med5342412478664961|
|Pope CA III,Dockery DW. Year: 2006Health effects of fine particulate air pollution: lines that connectJ Air Waste Manag Assoc5670974216805397|
|Pope CA III,Thun MJ,Namboodiri MM,Dockery DW,Evans JS,Speizer FE,et al. Year: 1995Particulate air pollution as a predictor of mortality in a prospective study of U.S. adultsAm J Respir Crit Care Med1513 Pt 16696747881654|
|R Development Core TeamYear: 2006R: A Language and Environment for Statistical ComputingVienna, AustriaR Foundation for Statistical Computing|
|Rosenlund M,Berglind N,Pershagen G,Hallqvist J,Jonson T,Bellander T. Year: 2006Long-term exposure to urban air pollution and myocardial infarctionEpidemiology17438339016699471|
|Rosenlund M,Forastiere F,Stafoggia M,Porta D,Perucci M,Ranzi A,et al. Year: 2008Comparison of regression models with land-use and emissions data to predict the spatial distribution of traffic-related air pollution in RomeJ Expo Sci Environ Epidemiol18219219917426734|
|Ryan PH,LeMasters GK. Year: 2007A review of land-use regression models for characterizing intraurban air pollution exposureInhal Toxicol19suppl 112713317886060|
|Ryan PH,LeMasters GK,Biswas P,Levin L,Hu S,Lindsey M,et al. Year: 2007A comparison of proximity and land use regression traffic exposure models and wheezing in infantsEnviron Health Perspect11527828417384778|
|SAS Institute IncYear: 2006SAS Statistical Software 9Cary, NCSAS Institute Inc|
|Su JG,Brauer M,Ainslie B,Steyn D,Larson T,Buzzelli M. Year: 2008An innovative land use regression model incorporating meteorology for exposure analysisSci Total Environ3902?352052918048083|
|Szpiro AA,Sheppard L,Sampson PD,Kim SY. Year: 2007Validating national kriging exposure estimationEnviron Health Perspect115A33817637891|
|U.S. Census BureauYear: 1993A Guide to State and Local Census Geography Princeton, NJAssociation of Public Data Users|
|U.S. EPAYear: 1993Air Quality Criteria for Oxides of NitrogenI?III EPA/600/8-91/049aF-cF. Washington, DCU.S. Environmental Protection Agency|
|U.S. EPAYear: 2004Air Quality Criteria for Particulate Matter (October 2004) EPA 600/P-99/002aF-bF. Washington, DCU.S. Environmental Protection Agency|
|US EPA (U.S. Environmental Protection Agency)Year: 2007aEmissions and Generation Resource Integrated Database (eGRID) Available:http://www.epa.gov/cleanenergy/energy-resources/egrid/index.html. [accessed 2 October 2007].|
|U.S. EPA (U.S. Environmental Protection Agency)Year: 2007bNitrogen Dioxide Health Assessment Plan?Scope and Methods for Exposure and Risk AssessmentResearch Triangle Park, NCU.S. Environmental Protection Agency, Office of Air Quality Planning and Standards|
|USGS (U.S. Geological Survey)Year: 2007aNational Elevation Database Available: http://ned.usgs.gov/. [accessed 21 August 2007]|
|USGS (U.S. Geological Survey)Year: 2007bNational Land Cover Dataset 1992 (NLCD 1992) Available: http://landcover.usgs.gov/natllandcover.php. [accessed 18 August 2007].|
|Venn A,Yemaneberhan H,Lewis S,Parry E,Britton J. Year: 2005Proximity of the home to roads and the risk of wheeze in an Ethiopian populationOccup Environ Med62637638015901884|
|Vogelmann JE,Howard SM,Yang L,Larson CR,Wylie BK,Van Driel JN. Year: 2001Completion of the 1990?s National Land Cover Data Set for the conterminous United StatesPE&RS67650662|
|Wheeler AJ,Smith-Doiron M,Xu X,Gilbert NL,Brook JR. Year: 2008Intraurban variability of air pollution in Windsor, Ontario?measurement and modeling for human exposure assessmentEnviron Res106171617961539|
|Wong DW,Yuan L,Perlin SA. Year: 2004Comparison of spatial interpolation methods for the estimation of air quality dataJ Expo Anal Environ Epidemiol14540441515361900|
|Wood SN. Year: 2003Thin plate regression splinesJ R Stat Soc Ser B195114|
|Wood SN. Year: 2004Stable and efficient multiple smoothing parameter estimation for generalized additive modelsJ R Stat Soc Ser A99673686|
|Wood SN. Year: 2006Generalized Additive Models: An Introduction with RBoca Raton, FLChapman and Hall/CRC Press|
|Yanosky JD,Paciorek C,Schwartz J,Laden F,Puett R,Suh H. Year: 2008Spatiotemporal modeling of chronic PM10 exposure for the Nurses? Health StudyAtmos Environt421840474062|
|Yanosky J,Paciorek C,Suh H. Year: 2009Predicting chronic fine and coarse particulate exposures using spatiotemporal models for the northeastern and midwestern United StatesEnviron Health Perspect11752252919440489|
Keywords: GIS, nitrogen dioxide, outdoor air pollution, particulate matter.
Previous Document: Comparative toxicity of size-fractionated airborne particulate matter collected at different distanc...
Next Document: The short-chain fatty acid methoxyacetic acid disrupts endogenous estrogen receptor-alpha-mediated s...