Skip to content


Research Volume 122 | Issue 8 | August 2014

Environ Health Perspect; DOI:10.1289/ehp.1307287

Characterizing Spatial Patterns of Airborne Coarse Particulate (PM10–2.5) Mass and Chemical Components in Three Cities: The Multi-Ethnic Study of Atherosclerosis

Kai Zhang,1 Timothy V. Larson,2,3 Amanda Gassett,3 Adam A. Szpiro,4 Martha Daviglus,5 Gregory L. Burke,6 Joel D. Kaufman,3,7 and Sara D. Adar8

Author Affiliations open
1Division of Epidemiology, Human Genetics and Environmental Sciences, University of Texas School of Public Health, Houston, Texas, USA; 2Department of Civil and Environmental Engineering, 3Department of Environmental and Occupational Health Sciences, and 4Department of Biostatistics, University of Washington, Seattle, Washington, USA; 5Department of Medicine, University of Illinois at Chicago, Chicago, Illinois, USA; 6Division of Public Health Sciences, Wake Forest University School of Medicine, Winston-Salem, North Carolina, USA; 7Department of Epidemiology, University of Washington, Seattle, Washington, USA; 8Department of Epidemiology, University of Michigan, Ann Arbor, Michigan, USA

PDF icon PDF Version (3 MB)

  • Background: The long-term health effects of coarse particular matter (PM10–2.5) are challenging to assess because of a limited understanding of the spatial variation in PM10–2.5 mass and its chemical components.

    Objectives: We conducted a spatially intensive field study and developed spatial prediction models for PM10–2.5 mass and four selected species (copper, zinc, phosphorus, and silicon) in three American cities.

    Methods: PM10–2.5 snapshot campaigns were conducted in Chicago, Illinois; St. Paul, Minnesota; and Winston-Salem, North Carolina, in 2009 for the Multi-Ethnic Study of Atherosclerosis and Coarse Airborne Particulate Matter (MESA Coarse). In each city, samples were collected simultaneously outside the homes of approximately 40 participants over 2 weeks in the winter and/or summer. City-specific and combined prediction models were developed using land use regression (LUR) and universal kriging (UK). Model performance was evaluated by cross-validation (CV).

    Results: PM10–2.5 mass and species varied within and between cities in a manner that was predictable by geographic covariates. City-specific LUR models generally performed well for total mass (CV R2, 0.41–0.68), copper (CV R2, 0.51–0.86), phosphorus (CV R2, 0.50–0.76), silicon (CV R2, 0.48–0.93), and zinc (CV R2, 0.36–0.73). Models pooled across all cities inconsistently captured within-city variability. Little difference was observed between the performance of LUR and UK models in predicting concentrations.

    Conclusions: Characterization of fine-scale spatial variability of these often heterogeneous pollutants using geographic covariates should reduce exposure misclassification and increase the power of epidemiological studies investigating the long-term health impacts of PM10–2.5.

  • Citation: Zhang K, Larson TV, Gassett A, Szpiro AA, Daviglus M, Burke GL, Kaufman JD, Adar SD. 2014. Characterizing spatial patterns of airborne coarse particulate (PM10–2.5) mass and chemical components in three cities: the Multi-Ethnic Study of Atherosclerosis. Environ Health Perspect 122:823–830;

    Address correspondence to K. Zhang, Division of Epidemiology, Human Genetics and Environmental Sciences, University of Texas School of Public Health, Houston, TX 77030 USA. Telephone: (713) 500-9249. E-mail address:

    We acknowledge the other investigators, staff, and participants of MESA (Multi-Ethnic Study of Atherosclerosis), MESA Coarse, and MESA Air for their valuable contributions to this work. A full list of MESA investigators and institutions is located at

    This work was supported by STAR research grants RD833741010 and RD83169701 from the U.S. Environmental Protection Agency (EPA). The MESA study is supported by contracts N01-HC-95159 through N01-HC-95169 from the National Heart, Lung, and Blood Institute.

    This research has not been formally reviewed by the U.S. EPA. The views expressed in this document are solely those of the authors, and the U.S. EPA does not endorse any products or commercial services mentioned in this publication.

    The authors declare they have no actual or potential competing financial interests.

    Received: 28 June 2013
    Accepted: 14 March 2014
    Advance Publication: 18 March 2014
    Final Publication: 1 August 2014

  • PDF icon Supplemental Material (330 KB) PDF
  • In the original online version of this article, Figures 1 and 2 were presented incorrectly as Figures 2 and 1, respectively. In addition, Tables 2 and 3 presented values for silicon in nanograms per cubic meter. The figures have been corrected, and the values for silicon in the two tables have been rescaled to micrograms per cubic meter.

    The authors regret the errors.


Although considerable evidence has linked adverse health with long-term exposures to fine particulate matter (PM2.5; ≤ 2.5 μm in aerodynamic diameter) (Brook et al. 2010), there has been little epidemiological research examining long-term exposures to coarse particulate matter (PM10–2.5; 2.5–10 μm in aerodynamic diameter). Toxicological studies have shown that PM10–2.5 can induce reactive oxygen species and initiate inflammatory responses in vivo and in vitro (Becker et al. 2005; Monn and Becker 1999; Pozzi et al. 2003; Schins et al. 2004; Shi et al. 2003), suggesting a plausible biological mechanism for long-term health effects. The studies that have investigated such relationships with PM10–2.5, however, have generally found weak and nonstatistically significant or null associations [U.S. Environmental Protection Agency (EPA) 2006]. One possible explanation for differences between the toxicological and epidemiological evidence is that previous epidemiological studies have had a limited ability to characterize spatial variations in PM10–2.5. This can be important because PM10–2.5 has relatively short residence times in atmosphere due to high gravitational settling (U.S. EPA 2009), and spatial heterogeneity has been shown to be large (Burton et al. 1996; Chen et al. 2007; Eeftens et al. 2012a; Houthuijs et al. 2001; Wilson and Suh 1997). In addition, there has been limited characterization of spatial differences in PM10–2.5 chemical composition, which may help to differentiate key sources of PM10–2.5 mass (e.g., mineral and roadway dust, sea spray, pollen, and mechanical grinding including vehicular brake and tire wear) (U.S. EPA 2009). An improved understanding of the spatial variation of PM10–2.5 mass and chemical components is therefore expected to be critical in quantifying the long-term effects of PM10–2.5 exposures.

Because some pollutants, such as those from traffic, vary over small spatial scales (i.e., 10–100 m), there is an increasing emphasis on estimating individual-level exposures (Health Effects Institute 2010). Regression models with geographic information system (GIS)–derived covariates such as land use, nearby emission sources, and distance to roadways, termed “land use” regression (LUR) models, are a common approach. Universal kriging (UK) is an extension of this methodology that further incorporates spatial correlations. Although spatial prediction models are commonly employed for PM2.5 and oxides of nitrogen (Health Effects Institute 2010), very few investigations have generated covariate-based spatial prediction models for PM10–2.5, and none, to our knowledge, have created covariate-based, spatial interpolation models for PM10–2.5 species.

As part of the Multi-Ethnic Study of Atherosclerosis and Coarse Airborne Particulate Matter (MESA Coarse), we characterized fine-scale spatial differences in PM10–2.5 mass and chemical components within three American cities using data from intensive monitoring campaigns. MESA Coarse builds on the MESA cohort of 6,814 adults from six metropolitan areas (Bild et al. 2002) and the MESA Air Pollution project (MESA Air), which investigates the impacts of PM2.5 on the progression of atherosclerosis (Cohen et al. 2009; Kaufman et al., 2012). Here we present the MESA Coarse field study design and development of spatial prediction models for PM10–2.5 mass, copper, zinc, phosphorus, and silicon. These four chemical components were selected because they were shown to be good indicators of brake wear, tire wear, agriculture, and mineral dust, respectively, across all three cities, using positive matrix factorization (Sturtz et al. 2014).


Sampling design. PM10–2.5 concentrations were measured simultaneously over 2-week periods in two seasons outside the homes of approximately 40 MESA participants residing in Chicago, Illinois (8–22 April 2009; 20 August–3 September 2009), St. Paul and Minneapolis, Minnesota (17–31 January 2009; 27 May–10 June 2009), and Winston-Salem, North Carolina (25 February–11 March 2009; 6–20 July 2009). Homes were selected in a targeted approach that aimed to maximize geographic coverage as well as variability of features believed to be predictive of coarse particles and selected source-specific components. Specifically, we targeted vegetation, distance to major roads, as well as rural, commercial, and industrial land use. Although most homes were sampled during one season only, we collected samples during both seasons at approximately one-third of homes to assess the stability of concentrations over time. Homes with more unique geographic features were oversampled during the second round to ensure sufficient variability for modeling. Other repeats were selected at random. Institutional review boards at each site approved the study, and all participants provided written informed consent.

PM10–2.5 sampling. We collected 2-week integrated samples using Harvard personal environmental monitors (HPEMs; Thermo Environmental Instruments, Franklin, MA) with Medo VP0125 pumps (Medo, Hanover Park, IL) calibrated to a flow rate of 1.8 L/min, which has been evaluated in ambient field tests against the Harvard impactor operating at 10 L/min (Lee et al. 2006). To prevent overloading and minimize the number of pumps required, air flow was cycled between paired HPEMs with cut points for PM10 and PM2.5 every 5 min over the 2-week sampling periods. Programmable timers allowed for the simultaneous collection of samples across all locations in a city.

All Teflon filters were preconditioned for ≥ 24 hr at 22.3 ± 1.9°C and 34.7 ± 2.5% relative humidity, before weighing by microbalance (model UMT2; Mettler-Toledo Inc., Highstown, NJ) (Allen et al. 2001). Samples were analyzed for elements by X-ray fluorescence spectroscopy by Cooper Environmental Services (Portland, OR). Concentrations were estimated by subtracting PM2.5 from PM10 based on research by Chen et al. (2011).

Many quality control procedures were performed including voiding samples with insufficient durations (< 9 days), out-of-range air flows (± 20%), damaged filters, extreme concentrations over the 2-week sampling period (> 5 SDs from the mean), and high sulfur levels in the PM10–2.5 fraction (> 0.2 μg/m3) because sulfur should be limited primarily to PM2.5. Overall, the precision of duplicate PM10, PM2.5, and PM10–2.5 samples was 2%, 10%, and 18%, respectively. Concentrations were also compared with measurements reported by the U.S. EPA’s Air Quality System (AQS) for corresponding time periods during the same year (U.S. EPA 2010).

Geographic covariates. Table 1 illustrates the covariates derived in ArcGIS 9.3 (ESRI, Redlands, CA) that were considered for our spatial prediction models. They include five major categories: a) land use such as commercial, industrial, and residential; b) local transportation including roadways, railways, truck routes, airports, and a traffic dispersion model output; c) population density; d) ground cover including impervious surface and vegetation; e) PM10–2.5 emission sources; and f) positional information (MESA Air 2011). Briefly, land use data included U.S. Geological Survey satellite-derived raster images from 2000 (Price et al. 2006) and aerial photography from the 1970s and 1980s.Transportation variables were derived from data from TeleAtlas (Lebanon, NH), National Transportation Atlas Database 2009 (Bureau of Transportation Statistics 2009), and the CALINE3 line source dispersion model (Wilton et al. 2010). PM10–2.5 emissions were derived from the National Emission Inventory Database (U.S. EPA 2008), population density was obtained from the 2000 U.S. Census (U.S. Census Bureau 2000), imperviousness was downloaded from the National Land Cover Database 2006 (U.S. Geological Survey 2011), and vegetation was estimated by the Normalized Difference Vegetation Index (NDVI) (Carroll et al. 2008).

Table 1 - See HTML for full tableTable 1 – Variables considered for spatial prediction models for PM10–2.5 mass and chemical component concentrations.

View Table (HTML Version)
View larger image (TIF File)

Modeling approach. We used LUR and UK to estimate spatial patterns of PM10–2.5 mass, copper, zinc, phosphorus, and silicon using approaches previously described by Mercer et al. (2011). Our primary models were constructed separately for each city, with season included as a predictor and an effect modifier of other predictors as needed to account for seasonal variation. Pooled models across all cities were also explored in secondary analyses. All stages of our model selection procedure were fit using the glmnet package (Friedman et al. 2010) and R 2.7.2 software (R Development Core Team; The resulting prediction models are intended to reflect long-term exposures because preliminary data analyses using AQS monitors in our study regions previously suggested that the average of 2-week samples from two seasons was highly correlated with annual average concentrations.

Given the large number of potential predictors, we followed a process of variable screening that began with removal of predictors with insufficient variability (i.e., excluded if 85th percentile equaled the 15th percentile). Then, for variables with varying buffer radii, we selected the best short-range (50–500 m) and long-range (500 m to 5 km) buffer based on the highest univariate Pearson correlation coefficient with the exposure being modeled. Moreover, when multiple predictors were highly correlated with one another (ρ > 0.85), we selected the predictor that was most strongly correlated with the exposure, although we preferentially excluded latitude and longitude and selected raster-based land use data (collected in 2000) over older aerial photography-based data (collected in the 1970s and 1980s). Next, we applied the least absolute shrinkage and selection operator (LASSO) by changing a tuning parameter to reduce the number of variables down to ≤ 15 (Friedman et al. 2010). Finally, we conducted an exhaustive search to examine all possible combinations of these covariates, restricting to models with ≤ 6 parameters. When interaction terms were selected by either of these modeling steps without the main effects, we also forced the main effects into the model. No other interactions between predictors other than season or city were considered.

The final combinations of variables were selected that resulted in the lowest root-mean-square error (RMSE) and the highest R2 under 10-fold cross-validation (CV). In this method, the data set was randomly divided into 10 equal sub–data sets, where model fitting occurred for each selection of nine-tenths of the data while validating on the final tenth. Differences between the true and estimated values of PM10–2.5 concentrations for each validation set were then used to calculated RMSE and R2. This method is intended to avoid overfitting of the models to the observed data.

Sensitivity analyses were conducted to assess the impacts of a) excluding outliers (> 3 SDs from city-specific means); b) different data sources of land use variables; c) selecting buffers in a repeated step-wise manner, as recommended by Su et al. (2009); and d) natural log-transforming concentrations.

Visualization. Maps were generated by kriging provided by the spatial analysis package in ArcGIS 9.3 at a lattice grid over our three cities, with spacing of 0.25 km in urban areas and 1–2 km in rural areas.


Measured PM10–2.5 mass concentrations. Between 17 January and 3 September 2009, we collected 235 collocated PM2.5 and PM10 samples. After our quality control procedures were applied, we had 207 (88%) and 195 (83%) valid PM2.5 and PM10 mass measurements, respectively, resulting in 191 (81%) valid sample pairs from 156 unique locations (61, 43, and 52 in Chicago, St. Paul, and Winston-Salem, respectively; Figure 1). A subset of 35 locations had samples collected during two seasons: 4, 18, and 13 in Chicago, St. Paul, and Winston-Salem, respectively.

Figure 1 - 5 sets of maps for Chicago, St. Paul, and Winston-Salem showing spatial distributions for PM10-2.5, copper, phosphorus, silicon, and zinc, respectively.Figure 1 – Predicted concentrations for PM10–2.5 mass (A), copper (B), phosphorus (C), silicon (D), and zinc (E) in three U.S. cities. Bins were chosen to highlight within-city contrasts and are not even. Participants’ locations have been jittered.

View larger image (TIF File)

Table 2 summarizes PM10–2.5 mass and species concentrations by city and season (see Supplemental Material, Table S1, for detailed descriptive statistics). Mean (± SD) PM10–2.5 concentrations across seasons were 5.7 ± 2.0, 5.3 ± 3.3, and 3.6 ± 1.4 μg/m3 in Chicago, St. Paul, and Winston-Salem, respectively. A strong seasonal difference was seen in St. Paul (3.3 ± 2.2 and 6.7 ± 3.3 μg/m3 in winter and summer, respectively) but not in the other two cities (Chicago: 5.5 ± 2.0 and 5.9 ± 2.1 μg/m3 in winter/early spring and summer, respectively; Winston-Salem: 3.5 ± 1.2 and 3.8 ± 1.6 μg/m3 in winter and summer, respectively).

Table 2 - See HTML for full tableTable 2 – Summary of statistics (mean ± SD) for PM10–2.5 mass (μg/m3) and chemical component (ng/m3)a concentrations in each sampling city by season.

View Table (HTML Version)
View larger image (TIF File)

Silicon had the largest concentrations of the species investigated, with levels 12 to 260 times larger than the three other species (Table 2). Especially high concentrations of silicon were observed in St. Paul during the summer (720 ± 188 ng/m3), whereas the other two cities had lower levels that peaked in the winter. Phosphorus concentrations in all three cities were highest in the summer, with an approximate doubling of concentrations compared with the winter in St. Paul and Winston-Salem. In contrast to the other pollutants that had similar concentration ranges across cities, copper and zinc concentrations differed by location, with the highest levels of both observed in Chicago, and the lowest levels in Winston-Salem. Zinc was the most variable species in each of the three cities during both the winter (coefficient of variation defined as SD over mean: 0.75–0.81) and summer (coefficient of variation: 0.69–1.27).

Spatial modeling results. Data reduction procedures reduced the overall number of potential predictors from 802 to 64–94, and the LASSO procedure further reduced the number of candidate predictors to approximately 15 for each city and species. Final models included seven or eight main effect predictors and up to four interactions (with season in the city-specific models, and with season or city in the all-city models) (see Supplemental Material, Tables S2–S6, for lists of the predictors included in each final model according to city and exposure.).

Overall, our models fit the data well (Figure 2), explaining between 36% and 93% of the variability in PM10–2.5 mass and species concentrations under cross-validation (Table 3). Our models performed generally better in Chicago and St. Paul than in Winston-Salem, and had the most consistent predictive ability for silicon when including data for all cities. The other models pooled across all cities demonstrated inconsistent performance at capturing within-city variability, with the best within-city CV R2 ranging from 0.54 to 0.66 compared with 0.0 to 0.34 for cities with the worst predictions. UK models for individual cities as well as pooled across all cities generally had similar or lower model performance than their corresponding LUR models.

Figure 2 - 15 scatterplots with regression lines showing predictors (y-axes) according to measurements (x-axes) for PM10-2.5 mass, copper, phosphorus, silicon, and zinc in Chicago, St. Paul, and Winston-Salem, respectively.Figure 2 – Scatter plots showing observations and the predictions from the “best” LUR ­models by city and species: PM10–2.5 (A), copper (B), phosphorus (C), silicon (D), and zinc (E). R2 = square of correlations between measurements and predictions.

View larger image (TIF File)

Table 3 - See HTML for full tableTable 3 – Model performance (cross-validated R2 and RMSE) for PM10–2.5 mass (μg/m3) and species concentrations (ng/m3)a using land use regression (LUR) and universal kriging (UK) models.

View Table (HTML Version)
View larger image (TIF File)

Figure 1 shows the spatial patterns and distributions of predicted concentrations by city. PM10–2.5 mass, silicon, and phosphorus had somewhat similar prediction ranges across all cities, whereas copper and zinc showed much higher predictions in Chicago and St. Paul than in Winston-Salem. The highest PM10–2.5 mass predictions were for the urban centers of Chicago and St. Paul, but in Winston-Salem higher levels were predicted outside of the urban core. Across all three cities, copper and zinc exhibited higher concentrations in urban areas, with elevations of copper focused in areas of high-intensity land use and along major roadways. Silicon was highest in St. Paul and most variable in Chicago, where high concentrations were focused along an industrial corridor to the west of the city and lowest concentrations were found in the outlying areas. Phosphorus was similarly patterned in Chicago but less concentrated downtown. Predictions were highly variable for phosphorus in St. Paul but consistently high in Winston-Salem, with higher predictions in the outlying areas.

The estimated coefficients of the LUR models by city and pooled across cities are available in the Supplemental Material, Tables S2–S6. In general, PM10–2.5 mass, copper, and zinc concentrations were associated with traffic-related features across the three cities (e.g., land use, distance to roads, sum of road lengths, and sum of truck route lengths). Models for phosphorus and silicon concentrations consistently included vegetation features as predictors in St. Paul and Winston-Salem, but also included water features in Chicago. Vegetation appeared predictive across all mass and species models in Winston-Salem. Medium or high intensity of urban land use was predictive among all models for total mass, and sum of road lengths was consistently included in all pooled models, as were season and city.

Sensitivity analyses indicated that two extreme measurements from St. Paul were influential, with weaker predictive performance without the outliers (CV R2, 0.51 compared with 0.65). Slightly improved model performance was found using land use from both the raster data and aerial photography, compared with either alone (see Supplemental Material, Table S7). Models based on the natural log of PM10–2.5 levels and the iterative covariate selection approach proposed by Su et al. (2009) showed generally poorer prediction performance as denoted by lower CV R2 and higher RMSE than those presented here (data not shown).


In this study we conducted spatially intensive PM10–2.5 sampling and developed two types of spatial prediction models (LUR and UK) for PM10–2.5 mass and chemical components (copper, phosphorus, silicon, and zinc) for use in the MESA Coarse epidemiology study. To our knowledge, this is one of the first studies to develop fine-scale spatial prediction models for PM10–2.5 mass and chemical components, which can serve as tracers for different sources of pollution. We demonstrated that geographic covariates can explain within-city variations in concentrations (mean CV R2, 0.61), although the predictive power varies across cities and species (range of CV R2, 0.36–0.93). By capturing fine-scale spatial variability of these often heterogeneous pollutants, this work is expected to substantially reduce measurement error and improve our ability to investigate the long-term health impacts of PM10–2.5 over traditional approaches that rely on a limited number of central monitoring stations.

Across all models, there was some similarity in the spatial features predictive of concentrations, including high- and medium-intensity land use and indicators of traffic and vegetation, yet there was limited commonality across cities in the key predictors. For example, variables related to traffic, high/medium development, and residential areas were consistently selected in final models in Chicago, but indicators of vegetation were more consistently included in final models in Winston-Salem. This may reflect the fact that Chicago is a more industrialized city, with higher concentrations of copper and zinc, than Winston-Salem, which is a smaller city. Similarly, water features appeared to be a stronger predictor of PM10–2.5 particles in St. Paul than the other cities likely because this is a prominent feature of this region.

Season modified the influence of certain predictors, as would be expected given the presence of snow cover during winter periods in St. Paul and Chicago and changes in vegetation across all areas. Although other characteristics of seasonality may influence PM10–2.5 concentrations, we did not have fine-scale spatial meteorology data to explicitly explore these associations. An analysis of wind speed and direction, however, failed to show a strong prevailing wind direction during our sampling periods, suggesting that wind is likely not a strong predictor or modifier of concentrations. In addition, we did not identify important differences between our sampling periods and other typical weeks, suggesting that our results should be representative of other time periods.

Given the observed area-specific differences, we were unable to identify robust predictions of within-city variability based on the same models for all cities (see Supplemental Material, Tables S2–S6). Although some of our pooled models showed good performance in more than one city, no model worked well in all locations. Thus, though it is possible that these models may be generalized to other cities with similar characteristics, further validation is warranted before broad application.

Predictive performance of our models was consistent with the one study from Europe (Eeftens et al. 2012a) and generally better than the few PM10–2.5 models in the United States. Using 20–40 monitoring stations in each of 20 European areas, similar CV R2 values were reported as in our study (Eeftens et al. CV R2, 0.03–0.73; our MESA Coarse study CV R2, 0.41–0.68.) (Eeftens et al. 2012b). In a recent study from Ohio, Mukerjee et al. (2012) reported an R2 of 0.78 for their LUR model. Because this was not derived from cross-validation in which some data are withheld from the model-building step for validation, however, the predictive power would be inflated due to overfitting. Also in the United States, Yanosky et al. (2009) reported models with CV R2 values for PM10–2.5 of 0.39 and 0.33 after and before 1999 for cities across the northeastern and midwestern United States, respectively. Their lower predictive performance can be likely be explained by their use of regulatory monitors only, which are more sparse and reflect less variation in geographical predictors than the intensive campaigns of this, the European, and Ohio studies. Predicting PM10–2.5 indirectly by modeling PM2.5 and PM10 separately may also have influenced their models. Independent of the source of data, we generally found similar predictors to past work including proximity to roadways, land use, and vegetation. As in our study, Eeftens et al. (2012b) also indicated that key covariates differed by city.

UK performed similarly to LUR for all cities, suggesting limited spatial correlation of PM10–2.5 after control for geographic covariates. An alternative explanation is that our relatively small sample size (n ~ 30–40 locations in each city) may have limited the ability of our models to characterize fine-scale spatial structure of PM10–2.5. Nevertheless, our models should have been able to capture at least some of any small-scale correlation structure, given that distances between some of the sampled residences were small (minimum distances of 8, 64, and 174 m in Chicago, St. Paul, and Winston-Salem, respectively).

Measured PM10–2.5 levels as well as spatial variability were generally lower than expected, possibly because our study samples were collected exclusively at residual locations where there may be fewer sources of PM10–2.5. The average levels of PM10–2.5 measured at study locations in both Chicago and St. Paul (5–6 μg/m3) were comparable with those reported in residential neighborhoods in Detroit, Michigan (6–7 μg/m3) (Thornburg et al. 2009), but lower than reported for a variety of sites in Los Angeles, California (5–14 μg/m3) (Pakbin et al. 2010), Philadelphia, Pennsylvania (5–9 μg/m3) (Burton et al. 1996), Denver, Colorado (9–16 μg/m3) (Clements et al. 2012), Research Triangle Park, North Carolina (1–13 μg/m3) (Chen et al. 2007), and Central and Eastern European countries (12–40 and 6–24 μg/m3) (Eeftens et al. 2012a; Houthuijs et al. 2001). Although our means were lower, study locations with the highest measured concentrations (12 μg/m3 in Chicago; 17 μg/m3 in St. Paul) were in line with concentrations reported by these other studies.

Spatial variability also was lower than expected, which may also be attributable to sampling from residential locations only. Average PM10–2.5 concentrations were, however, generally comparable with those estimated by concurrent PM10 and PM2.5 AQS samples during the same 2-week sampling periods. For example, the 2-week averages of four available Chicago AQS sites were 4.5 ± 5.3 and 6.1 ± 2.1 μg/m3 when we reported 5.5 ± 2.0 μg/m3 and 5.9 ± 2.1 μg/m3 in matched time periods. In Winston-Salem, one available AQS site had one 2-week average of 3.7 ± 1.5 μg/m3 when we observed 3.8 ± 1.6 μg/m3, and 2.4 ± 0.2 μg/m3 when we observed 3.5 ± 1.2 μg/m3. In St. Paul, two available AQS sites had 2-week mean levels of 6.6 ± 4.7 and 8.6 ± 4.6 μg/m3 when our reported concentrations were 3.3 ± 2.2 and 6.7 ± 3.3 μg/m3, respectively.

This study has a few important strengths. First, we collected spatially intensive samples of PM10–2.5 using the same sampling protocol in three U.S. cities. Using a snapshot style campaign, we were able to predict PM10–2.5 mass at unmeasured locations based on geographic characteristics of that precise location. This represents a substantial improvement for predicting long-term concentrations at unmeasured locations compared with assigning the concentration of the nearest monitor or simple interpolation methods that do not consider the characteristics of a location other than latitude and longitude (e.g., inverse distance weighted method and ordinary kriging). By including chemical speciation of these particles, we are also the first, to our knowledge, to predict the spatial distribution of PM10–2.5 components. This is important because components can be used as indicators of different source types in the related health study.

This study has a few limitations. First, we used repeated 2-week samples over 1 year to assess long-term exposures to PM10–2.5. Although not an annual average, our sampling duration should be sufficiently long to mitigate transient perturbations such as transient meteorological fluctuations and reflect average conditions. Comparisons with available data from AQS monitors suggest that average PM10–2.5 mass during our 2-week sampling periods were highly correlated with, and within 15% of, the annual averages at the AQS stations. However, we cannot confirm our estimates for PM10–2.5 components because measured values are not available for comparison. Additionally, our models were derived based on data collected during a single year, and thus may not be accurate for other time periods if spatial patterns vary over time. Spatial stability has been recently demonstrated for traffic-related gases in Vancouver, Ontario, Canada, over 7 years (Wang et al. 2013) but is not guaranteed in other locations or for other pollutants. Other limitations pertain to our sampling locations. Although we attempted to capture different land uses, samples were collected at the homes of MESA participants, so our models may be more appropriate for predicting concentrations in residential areas than in industrial or commercial areas. In addition, our models may not have captured very small-scale spatial correlations, because only a few of the sampling locations were located within 100 m of one another. Finally, although each component targeted for this analysis was intended to be a common predictor of brake wear, tire wear, agriculture, or mineral dust across all cities, some caution is warranted in the strict interpretation of these indicator species because our tracers were not always unique to a single source (Sturtz et al. 2014). Research is ongoing to explore the spatial patterning of individual source contributions in further detail.


In summary, we demonstrated that a spatially intensive monitoring campaign was useful in predicting fine-scale spatial variability of PM10–2.5 mass and chemical component concentrations within and across three U.S. cities. This research and the resulting prediction models represent a substantial improvement for epidemiology over studies of PM10–2.5 that have previously assigned pollutant concentrations from a central site to an entire city. That there are some differences in predictive performance by city and species, however, implies that caution should be taken in epidemiology studies when inferring the comparative health impacts of pollutants estimated by these types of models. Therefore, investigators should be mindful that any observed differences in apparent toxicity may also be at least partly attributable to differential accuracy in estimating concentrations.


Allen R, Box M, Liu LJS, Larson TV. 2001. A cost-effective weighing chamber for particulate matter filters. J Air Waste Manag Assoc 51(12):1650–1653.

Becker S, Dailey LA, Soukup JM, Grambow SC, Devlin RB, Huang YCT. 2005. Seasonal variations in air pollution particle-induced inflammatory mediator release and oxidative stress. Environ Health Perspect 113:1032–1038; doi: 10.1289/ehp.7996.

Bild DE, Bluemke DA, Burke GL, Detrano R, Diez Roux AV, Folsom AR, et al. 2002. Multi-Ethnic Study of Atherosclerosis: objectives and design. Am J Epidemiol 156(9):871–881.

Brook RD, Rajagopalan S, Pope CA III, Brook JR, Bhatnagar A, Diez-Roux AV, et al. 2010. Particulate matter air pollution and cardiovascular disease: an update to the scientific statement from the American Heart Association. Circulation 121:2331–2378.

Bureau of Transportation Statistics. 2009. National Transportation Atlas Database (NTAD) 2009 DVD. Available:​cts/src/products.xml?p=2994 [accessed 2 February 2009].

Burton RM, Suh HH, Koutrakis P. 1996. Spatial variation in particulate concentrations within metropolitan Philadelphia. Environ Sci Technol 30(2):400–407.

Carroll ML, DiMiceli CM, Sohlberg RA, Townshend JRG. 2008. 250m MODIS Normalized Difference Vegetation Index. Available: [accessed 23 April 2010].

Chen FL, Vanderpool R, Williams R, Dimmick F, Grover BD, Long R, et al. 2011. Field evaluation of portable and central site PM samplers emphasizing additive and differential mass concentration estimates. Atmos Environ 45(26):4522–4527.

Chen FL, Williams R, Svendsen E, Yeatts K, Creason J, Scott J, Terrell D, et al. 2007. Coarse particulate matter concentrations from residential outdoor sites associated with the North Carolina Asthma and Children’s Environment Studies (NC-ACES). Atmos Environ 41(6):1200–1208.

Clements N, Piedrahita R, Ortega J, Peel JL, Hannigan M, Miller SL, et al. 2012. Characterization and nonparametric regression of rural and urban coarse particulate matter mass concentrations in northeastern Colorado. Aerosol Sci Tech 46(1):108–123.

Cohen MA, Adar SD, Allen RW, Avol E, Curl CL, Gould T, et al. 2009. Approach to estimating participant pollutant exposures in the Multi-Ethnic Study of Atherosclerosis and Air Pollution (MESA Air). Environ Sci Technol 43(13):4687–4693.

Eeftens M, Beelen R, de Hoogh K, Bellander T, Cesaroni G, Cirach M, et al. 2012a. Development of land use regression models for PM2.5, PM2.5 absorbance, PM10 and PMcoarse in 20 European study areas: results of the ESCAPE project. Environ Sci Technol 46(20):11195–11205.

Eeftens M, Tsai M-Y, Ampe C, Anwander B, Beelen R, Bellander T, et al. 2012b. Spatial variation of PM2.5, PM10, PM2.5 absorbance and PMcoarse concentrations between and within 20 European study areas and the relationship with NO2—results of the ESCAPE project. Atmos Environ 62(0):303–317.

Friedman J, Hastie T, Tibshirani R. 2010. Regularization paths for generalized linear models via coordinate descent. J Stat Softw 33(1):1–22.

Health Effects Institute. 2010. Traffic-related Air Pollution: A Critical Review of the Literature on Emissions, Exposure, and Health Effects. Special Report 17. Boston, MA:Health Effects Institute.

Houthuijs D, Breugelmans O, Hoek G, Vaskövi E, Miháliková E, Pastuszka JS, et al. 2001. PM10 and PM2.5 concentrations in Central and Eastern Europe: results from the CESAR study. Atmos Environ 35(15):2757–2771.

Kaufman JD, Adar SD, Allen RW, Barr RG, Budoff MJ, Burke GL, et al. 2012. Prospective study of particulate air pollution exposures, subclinical atherosclerosis, and clinical cardiovascular disease: the Multi-Ethnic Study of Atherosclerosis and Air Pollution (MESA Air). Am J Epidemiol 176(9):825–837.

Lee SJ, Demokritou P, Koutrakis P, Delgado-Saborit JM. 2006. Development and evaluation of personal respirable particulate sampler (PRPS). Atmos Environ 40(2):212–224.

Mercer LD, Szpiro AA, Sheppard L, Lindstrom J, Adar SD, Allen RW, et al. 2011. Comparing universal kriging and land-use regression for predicting concentrations of gaseous oxides of nitrogen (NOx) for the Multi-Ethnic Study of Atherosclerosis and Air Pollution (MESA Air). Atmos Environ 45(26):4412–4420.

MESA (Multi-Ethnic Study of Atherosclerosis) Air. 2011. Data Organization and Operating Procedures. Available:​MESAAirDOOP.pdf [accessed 18 December 2012].

Monn C, Becker S. 1999. Cytotoxicity and induction of proinflammatory cytokines from human monocytes exposed to fine (PM2.5) and coarse particles (PM10–2.5) in outdoor and indoor air. Toxicol Appl Pharm 155(3):245–252.

Mukerjee S, Willis RD, Walker JT, Hammond D, Norris GA, Smith LA, et al. 2012. Seasonal effects in land use regression models for nitrogen dioxide, coarse particulate matter, and gaseous ammonia in Cleveland, Ohio. Atmos Poll Res 3(3):352–361.

Pakbin P, Hudda N, Cheung KL, Moore KF, Sioutas C. 2010. Spatial and temporal variability of coarse (PM10–2.5) particulate matter concentrations in the Los Angeles area. Aerosol Sci Tech 44(7):514–525.

Pozzi R, De Berardis B, Paoletti L, Guastadisegni C. 2003. Inflammatory mediators induced by coarse (PM2.5–10) and fine (PM2.5) urban air particles in RAW 264.7 cells. Toxicology 183(1–3):243–254.

Price CV, Nakagaki, N, Hitt KJ, Clawges RC. 2006. Enhanced historical land-use and land-cover data sets of the U.S. Geological Survey. U.S. Geological Survey digital data series 240. Available: [accessed 24 January 2012].

Schins RPF, Lightbody JH, Borm PJA, Shi TM, Donaldson K, Stone V. 2004. Inflammatory effects of coarse and fine particulate matter in relation to chemical and biological constituents. Toxicol Appl Pharm 195(1):1–11.

Shi T, Knaapen AM, Begerow J, Birmili W, Borm PJA, Schins RPF. 2003. Temporal variation of hydroxyl radical generation and 8-hydroxy-2′-deoxyguanosine formation by coarse and fine particulate matter. Occup Environ Med 60(5):315–321.

Sturtz TM, Adar SD, Gould T, Larson TV. 2014. Constrained source apportionment of coarse particulate matter and selected trace elements in three cities from the Multi-Ethnic Study of Atherosclerosis. Atmos Environ 84:65–77.

Su JG, Jerrett M, Beckerman B. 2009. A distance-decay variable selection strategy for land use regression modeling of ambient air pollution exposures. Sci Total Environ 407(12):3890–3898.

Thornburg J, Rodes CE, Lawless PA, Williams R. 2009. Spatial and temporal variability of outdoor coarse particulate matter mass concentrations measured with a new coarse particle sampler during the Detroit exposure and aerosol research study. Atmos Environ 43(28):4251–4258.

U.S. Census Bureau. 2000. Redistricting Census 2000 TIGER/Line Files: Technical Documentation. Available:​a/pdfs/tiger/rd_2ktiger/tgrrd2k.pdf [accessed 11 December 2011].

U.S. EPA (U.S. Environmental Protection Agency). 2006. Provisional Assessment of Recent Studies on Health Effects of Particulate Matter Exposure. Available:​060720.pdf [accessed 17 February 2014].

U.S. EPA (U.S. Environmental Protection Agency). 2008. 2002 National Emissions Inventory Data & Documentation. Available:​entory.html [accessed 1 August 2008].

U.S. EPA (U.S. Environmental Protection Agency). 2009. Integrated Science Assessment for Particulate Matter (Final Report). Available:​c2009/PM_ISA_full.pdf [accessed 17 February 2014].

U.S. EPA (U.S. Environmental Protection Agency). 2010. Technology Transfer Network (TTN) Air Quality System (AQS). Available: [accessed 11 June 2010].

U.S. Geological Survey. 2011. National Land Cover Database (NLCD). Available: [accessed 10 May 2012].

Wang R, Henderson SB, Sbihi H, Allen RW, Brauer M. 2013. Temporal stability of land use regression models for traffic-related air pollution. Atmos Environ 64:312–319.

Wilson WE, Suh HH. 1997. Fine particles and coarse particles: concentration relationships relevant to epidemiologic studies. J Air Waste Manag Assoc 47(12):1238–1249.

Wilton D, Szpiro A, Gould T, Larson T. 2010. Improving spatial concentration estimates for nitrogen oxides using a hybrid meteorological dispersion/land use regression model in Los Angeles, CA and Seattle, WA. Sci Total Environ 408:1120–1130.

Yanosky JD, Paciorek CJ, Suh HH. 2009. Predicting chronic fine and coarse particulate exposures using spatiotemporal models for the northeastern and midwestern United States. Environ Health Perspect 117:522–529; doi: 10.1289/ehp.11692.

WP-Backgrounds Lite by InoPlugs Web Design and Juwelier Schönmann 1010 Wien