Land Use Regression Model for Ultrafine Particles in Amsterdam

Dec 15, 2010 - A land use regression model has been developed for the city of Amsterdam that well predicts measured spatial variability in particle nu...
3 downloads 0 Views 248KB Size
Environ. Sci. Technol. 2011, 45, 622–628

Land Use Regression Model for Ultrafine Particles in Amsterdam G E R A R D H O E K , * ,† R O B B E E L E N , † GERARD KOS,‡ MARIEKE DIJKEMA,§ SASKIA C VAN DER ZEE,§ PAUL H FISCHER,| AND B E R T B R U N E K R E E F †,⊥ Institute for Risk Assessment Sciences (IRAS), Division of Environmental Epidemiology, Utrecht University, Utrecht, The Netherlands, Energy research Centre of the Netherlands (ECN), Municipal Health Service Amsterdam, Department of Environmental Health, Amsterdam, The Netherlands, National Institute for Public Health and Environment, and Julius Center for Health Sciences and Primary Care, Utrecht University, Utrecht, The Netherlands

Received July 9, 2010. Revised manuscript received November 9, 2010. Accepted November 14, 2010.

There are currently no epidemiological studies on health effects of long-term exposure to ultrafine particles (UFP), largely because data on spatial exposure contrasts for UFP is lacking. The objective of this study was to develop a land use regression (LUR) model for UFP in the city of Amsterdam. Total particle number concentrations (PNC), PM10, PM2.5, and its soot content were measured directly outside 50 homes spread over the city of Amsterdam. Each home was measured during one week. Continuous measurements at a central urban background site were used to adjust the average concentration for temporal variation. Predictor variables (traffic, address density, land use) were obtained using geographic information systems. A model including the product of traffic intensity and the inverse distance to the nearest road squared, address density, and location near the port explained 67% of the variability in measured PNC. LUR models for PM2.5, soot, and coarse particles (PM10, PM2.5) explained 57%, 76%, and 37% of the variability in measured concentrations. Predictions from the PNC model correlated highly with predictions from LUR models for PM2.5, soot, and coarse particles. A LUR model for PNC has been developed, with similar validity as previous models for more commonly measured pollutants.

Introduction Epidemiological and toxicological studies have suggested acute health effects related to short-term exposure to ultrafine particles (UFP) (1). Epidemiological studies have found associations between daily average UFP concentrations measured at central monitoring locations and daily (cardiorespiratory) mortality, hospital admissions, and respiratory symptoms (2). A recent expert panel elicitation on the health effects of ultrafine particles concluded that health effects of short- and long-term exposure to UFP at realistic outdoor * Corresponding author phone: 31 30 2539498; fax: 31 30 2539499; e-mail: [email protected]. † Institute for Risk Assessment Sciences (IRAS). ‡ Energy research Centre of the Netherlands (ECN). § Municipal Health Service Amsterdam. | National Institute for Public Health and Environment. ⊥ Julius Center for Health Sciences and Primary Care. 622

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 45, NO. 2, 2011

concentrations were likely (3). The major source of uncertainty in the estimated health effects of ultrafine particles identified by the experts was the lack of epidemiological studies on health effects of long-term exposure to UFP (3). An important reason for this gap is the lack of spatially resolved exposure data, related to the high costs of UFP monitoring equipment and the lack of reliable UFP dispersion models. Land-use regression (LUR) is increasingly used to develop empirical models for the long-term average concentration of outdoor air pollution (4). LUR models have been developed for NO2, NOx, VOC, and to a lesser extent PM2.5 and soot or elemental carbon (4). LUR models require monitoring data for typically between 20 and 80 locations in a reasonably confined study area to establish the spatial variability of ambient concentrations. These data however are currently not readily available for UFP. For the city of Amsterdam, total particle number concentration (PNC) data were collected directly outside 50 homes in the framework of a study of the variability of particle number concentrations and particle mass (5). The study reported high temporal correlations of all components including PNC, PM2.5, and soot content of PM2.5 across Amsterdam (5). PNC has been used as a more easily measured indicator for ultrafine particles. At an urban background site in Amsterdam in another study, ultrafine particles accounted for more than 80% of the total particle number concentration and the temporal correlation between ultrafine and total particle number concentration at that site was 0.94 (6). The aim of this paper is to explore the possibility to model the spatial variation of PNC concentrations in the city of Amsterdam, The Netherlands with land use regression. The second aim was to evaluate the correlation of predictions of a PNC model with predictions from a PM2.5, soot, and coarse particle model developed from measurements taken at the same time at the same locations.

Experimental Section Study Design. Monitoring data were taken from the RUPIOH (relationship between ultrafine and fine particulate matter in indoor and outdoor air and respiratory health) study. Total particle number, PM10, PM2.5, and the soot content of PM2.5 were measured from October 2002 until March 2004 directly outside 50 homes spread over the city of Amsterdam. In each week seven 24 h average measurements were made at one home and at a central urban background site in the city center, where measurements were during the entire study period. The average difference of the concentrations measured at the home outdoors and the continuous measurement site was used to develop the land use regression model. Predictor variables were obtained using geographic information systems (GIS) from the city of Amsterdam and a European land-use database. Linear regression was used to develop models for PNC, PM2.5, coarse particles and soot. Monitoring Data. Details about the sampling campaign, monitoring methods and quality assurance have been published before (5). Briefly, continuous total particle number concentrations (PNC) were measured with condensation particle counters (CPC3022a). The CPC3022a counts particles from 7 nm and above. Twenty-four hour average concentrations of PM2.5 and PM10 were measured with Harvard impactors, operated at 10 L/minute. The soot content of the PM2.5 filters was assessed from measurement of reflectance of the filter and expressed as the absorbance according to ISO 9835. NO2 was not measured in the study. The continuous urban background site was located in the city center of 10.1021/es1023042

 2011 American Chemical Society

Published on Web 12/15/2010

TABLE 1. Distribution of Measured Home-Specific Average Concentrations Stratified by Traffic (n = 20) and Urban Background (n = 26) -3

PNC (cm ) 3

PM2.5 (µg/m ) soot (10-5 m-1) coarse (µg/m3)

site

min

P10

median

P90

max

traffic background traffic background traffic background traffic background

22 064 12 248 22 19 2.2 1.6 9 8

26 436 13 289 22 20 2.3 1.8 10 9

40 353 22 359 24 22 3.1 2.1 12 10

70 543 32 179 27 24 4.2 2.6 15 12

86 902 46 633 28 25 4.7 3.1 17 13

Amsterdam on top of a museum (sampling height 24 m). Measurements directly outside the home were made at the fac¸ade of the street side of the home for the 22 traffic locations, typically at balconies at the first floor. At the 28 background locations, measurements were made near the fac¸ade as well, but not necessarily at the street side. Standard geographical coordinates of the addresses of the 50 monitoring locations were obtained from the Dutch Kadaster. 99.5% of the coordinates is located in the correct parcel (7), which in the compact city of Amsterdam probably produces an uncertainty of at most 5-10 m. Predictor Data. Information on potential predictor variables was obtained from four data sources using GIS. Data on land-use for the year 2001 from the European database CORINE were used. The large number of land use data classes were condensed into low density residential, industry, port, urban green, and seminatural land use (8). Buffer sizes of 100, 200, 300, 500, 1000, and 3000 m were evaluated. From a national database, we obtained data on address density in buffers of 300, 1000, 3000, and 5000 m (8). The third database included more detailed land use data from the municipality of Amsterdam, including proximity to water surfaces, a potential source of shipping emissions (Dijkema et al. submitted). None of the sampling sites was located along the canals, hence our study could not assess contributions of small-scale local shipping emissions related to tour and pleasure boats. The fourth database included detailed traffic intensity on the nearest road, distance to roads and traffic intensity in 100 and 250 m buffers (Dijkema et al., submitted). The road network (NWB) is geographically precise within 5-10 m (7). Traffic intensity was also available for heavy and light duty vehicles separately. However, total traffic intensity and heavy duty traffic intensity were highly correlated (R ) 0.99 for the intensity on the nearest road), so we focused on total traffic intensity in this paper. Finally we added data taken from direct observation during the field campaigns: distance of the fac¸ade of the home to the side of the road, width of the street, and height of the buildings. Width of the street and height of the buildings was used to characterize street configuration. Five of the 22 traffic locations were defined as canyon streets (9). We further added distance to a traffic light, distance to an intersection (surrogates for congestion) and distance to small industrial sources and large parking lots. Data Analysis. Measurements near the 50 homes were not made simultaneously. To take temporal variation into account we calculated the difference of the measured concentration and the simultaneously measured concentration at the continuous urban background site. The average difference per home was used in model development. To obtain meaningful concentrations we added the overall mean concentration at the urban background (19 272 cm-3 for PNC, 21.6 µg · m-3 for PM2.5, 9.9 µg · m-3 for coarse particles and 2.1 × 10-5m for soot) to all differences. Only homes with three or more valid measurement days for both particle number and PM2.5 were used (n ) 46). We used the absolute

difference with the continuous site for adjustment instead of the ratio based on better performance in a previous Dutch study (10). In the current study, the correlation between averages adjusted with the two methods was 0.95, with slightly smaller within-site adjusted concentration variability for the difference method. A supervised stepwise regression was used to develop LUR models (8). In the first step we entered all individual predictors separately and assessed which predictor explained the largest percentage of variability in measured concentrations. We then evaluated whether adding the other variables increased the adjusted R2 with more than 1%. This procedure was repeated until no more variables entered the model. Variables were only entered in the model if the sign of the slope was in the a priori expected direction, for example, positive for traffic intensity and negative for distance to a major road. All variables were entered as linear variables, with the exception of distance for which we entered the natural logarithm of distance to account for the welldocumented nonlinear decrease of concentrations with distance to the road (11). In addition to traffic intensity and log distance as two separate variables, we also included the product of traffic intensity and the inverse of distance, distance squared and log distance as potential predictors. These variables resemble more closely the specification of dispersion models where emissions (linked to traffic intensity) are dispersed inversely with distance (12). We used leave-one out cross-validation to test the validity of the developed models. This involved successively leaving out one data point and estimating the model on the remaining N-1 sites. In this procedure, the variables in the model were the same as identified using the full data set, only the coefficients of the model changed. This is the common procedure in LUR model evaluation (4).

Results Concentration Variability. Table 1 illustrates that substantial variability was present in the home-specific average concentrations adjusted for temporal variability. The variability was larger for particle number concentration and soot than for PM2.5. The spatial variability for coarse particles was intermediate between soot and PM2.5. Higher concentrations for all four components were found for traffic locations. Measured PNC had a moderate to high correlation with the measured average concentrations of three PM metrics (Figure 1). Land Use Regression Models. Distance to the axis of the road as determined from the GIS was not a significant predictor of measured PNC concentrations. Increased distance to the side of the road as determined during field observations was associated with lower concentrations. The product of traffic intensity and the inverse of distance from field observation squared was the most predictive single predictor variable (R2 0.59). There were only small differences between the three intensity-distance product variables investigated. Two more variables (address density within a VOL. 45, NO. 2, 2011 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

623

FIGURE 1. Relationship between measured and modeled mean PNC (cm-3) and soot (10-5 m-1), PM2.5 and coarse particles (µg · m-3) concentrations. Pearson correlations between measured PNC with measured PM2.5, soot, and coarse was 0.66, 0.85, and 0.47, respectively. Pearson correlations between modeled PNC with modeled PM2.5, soot, and coarse was 0.90, 0.90, and 0.83, respectively. 300 m buffer and fraction of port area within a 3000 m buffer) entered the final model, raising the R2 to 0.67 (Table 2). The traffic variable remained associated with the largest impact on PNC. The identified model predicted PNC much better than other traffic representations, including traffic intensity and log distance as two separate variables (multiple R2 ) 0.49) and traffic in a 100 m buffer (multiple R2 ) 0.28). Leaveone out cross-validation showed modestly lower R2 values than for the model development: R2 ) 0.57, comparing measured and predicted concentrations (Figure 2). Figure 2 shows that the model performs better for traffic sites than for urban background sites. When the variables obtained from field observations were removed, substantially less variability was explained (R2 ) 44%). The model included the same traffic intensity distance variable and urban green in a 3000 m buffer. Weather data 624

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 45, NO. 2, 2011

were available from Schiphol airport, located approximately 10 km northeast from the city center for the entire study period. The regression slopes for the traffic variable did not differ significantly between conditions with low and high wind speed or low and high temperatures. This may be, however, related to the lack of urban data. Weather conditions during the 46 sampling weeks were representative for the 1.5 year full study period, for example, mean temperature was 9 °C (range of daily values -6 to +25) for both the full period and the sampling weeks. Mean wind speed was 4.8 m/s (1.2 to 14.9) for the full period and 4.9 m/s (1.2 to 14.4) for the sampling weeks. The product of traffic intensity and the inverse of distance predicted the largest contrast in concentrations for PM2.5 and soot (Table 3). In both models urban green space in a 3000 m buffer made a small additional contribution. In the

TABLE 2. Land Use Regression Model for Particle Number Concentration (cm-3) regression coefficienta intercept product T.I. and inverse distance squared address density, 300 m port, 3000 m

TABLE 3. Land Use Regression Model for PM2.5 (µg · m-3), Soot (10-5 m-1)

standard error

14491

(3165)

29523

(3795)

10266 6059

(3839) (3421)

a regression slopes multiplied by the difference between the 10th and 90th percentile for each of the three predictors (1102, 2653, and 4 149 780), intercept directly from model. The R2 of the model was 0.67 (adjusted R2 ) 0.65). T.I. is traffic intensity.

soot model, traffic in a 100 m buffer added further prediction. As for PNC, other traffic representations predicted PM2.5 and soot less well. Cross-validation R2 were 0.50 and 0.71 for PM2.5 and soot, compared to 0.57 and 0.76 in model development. The final model for coarse particles included four variables (Table 4), with the product of traffic intensity and the inverse

PM2.5

soot

regression standard regression standard coefficienta error coefficienta error intercept product T.I. and inverse distance urban green, 3000 m traffic in 100 m buffer

23.3

(0.8)

2.40

(0.24)

3.4

(0.6)

1.46

(0.15)

-1.4

(0.7)

-0.42

(0.20)

0.24

(0.14)

a

multiplied by the difference between the 10th and 90th percentile for each of the predictors (3640, 2 885 380, and 519 242). The R2 of the PM2.5 model was 0.57 (adjusted R2 ) 0.54). The R2 of the soot model was 0.76 (adjusted R2 ) 0.75).

of distance squared explaining the largest fraction of variability. Port area in a 3000 m buffer, density of residential land use in a 100 m buffer and playground in a 100 m buffer added additional prediction. The performance of the model

FIGURE 2. Predicted versus measured PNC and coarse particle concentrations, from leave-one out cross validation. R2 was 0.57 for PNC and 0.22 for coarse particles. Open circles are traffic sites; closed circles background sites. VOL. 45, NO. 2, 2011 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

625

TABLE 4. Land Use Regression Model for Coarse Particles (µg · m-3) regression coefficienta intercept product T.I. and inverse distance squared port, 3000 m buffer low density residential, 100 m buffer playground, 100 m buffer

standard error

8.89

(0.96)

2.02

(0.67)

1.40

(0.62)

1.01

(0.55)

-0.71

(0.39)

The R of the model was 0.37 (adjusted R ) 0.31). Playground is defined as an area with public access equipped with playing devices (often located in green areas such as parks). Coarse particle concentration calculated as PM10-PM2.5 multiplied by the difference between the 10th and 90th percentile for each of the predictors (1102, 4 149 780, 1 242 510,and 1.6, respectively). a

2

2

was poorer than for the other three pollutants, with an R2 of 0.37 in model development and 0.22 in cross-validation. Compared to the other pollutants the predicted impact of the traffic variable on concentrations was less dominant for coarse particles (Tables 2-4). The predicted values of the four LUR models were highly correlated (Figure 1). The correlations between modeled concentrations were higher than we observed for measured concentrations (Figure 1), as one would expect from the fact that the major explanatory variables were similar for all pollutants. The difference in the PNC-soot correlation between modeled and measured concentrations was not large (0.90 versus 0.85). The relationship between PNC and coarse particles was nonlinear, with virtually no increase in the predicted coarse particle concentration at the highest PNC.

Discussion A land use regression model has been developed for the city of Amsterdam that predicted measured spatial variability in particle number concentrations well. The product of traffic intensity of the nearest road and inverse of distance squared was the most important single predictor. Predictions from the LUR model for PNC correlated highly with predictions from LUR models for PM2.5, soot, and coarse particles. Particle Number LUR Model. For the first time, a land use regression model was developed for the spatial variation of ultrafine particle concentration within a large urban area. The performance of the model (R2) was similar to the performance of previously reported and widely used land use regression models for the pollutants NO2, PM2.5, and soot (4), illustrating that models can be developed for particle number concentrations using land use regression. The lack of spatially resolved monitoring data for particle numbers has limited the development of LUR models for PNC so far. We made use of particle number monitoring data at 50 locations spread over Amsterdam from a previous study. A limitation was that we did not have simultaneous measurements at all 50 locations, as was also the case with other targeted particulate matter sampling campaigns (10, 13). To adjust for temporal variation, concentration data from a continuous urban background monitoring location in the city center of Amsterdam was used following the methods of these previous studies. The correction for temporal variation is supported by the high temporal correlation between concentrations measured near the homes and the continuous urban background site for all pollutants (5). Median Pearson correlation coefficients between concentrations measured at the central and residential outdoor site were 0.72, 0.98, 0.94, and 0.89 for PNC, PM2.5, soot, and 626

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 45, NO. 2, 2011

coarse particles, respectively. Temporal variation may have played a larger role in our study compared to previous studies, as we measured simultaneously at only one outdoor site and the continuous urban background site. In the TRAPCA study, measurements at the 40 locations were performed in four groups of 10 sites that were measured in the same two week period. Though we cannot exclude some residual effect of temporal variation, this impact is likely to be small. This is supported by the observation that at the urban background location, there was a very low (temporal) correlation between PM2.5 and PNC (R ) 0.19) and between PNC and soot (R ) 0.38), consistent with a previous study in Amsterdam (6). The adjusted average PNC was highly correlated with PM2.5 (R ) 0.66, Figure 1) and soot (R ) 0.85, Figure 1), largely reflecting spatial covariation. The current database was available from a previous study, requiring a major monitoring effort, which is not easily repeated in other settings. A paper from Vancouver used mobile monitoring data to provide the spatial particulate matter concentration data for development of LUR models (14). This approach could be used to develop databases for PNC as well, using portable particle number counters in combination with a measuring van with CPC on board, thus enhancing a fast and complete scan of the whole city in short time, avoiding a data set that might be influenced by temporal variations due to changed meteorological conditions (15). The traffic variables were the most important predictors, consistent with studies documenting a large contrast in particle number concentrations with distance to major freeways (11) or large urban roads (16). We observed that distance to the side of the road as observed in the field was more predictive of concentrations of all pollutants than distance to the road calculated from a GIS. Two factors might explain the poorer performance of the models using GISbased distances. First, in the GIS a road is represented as a one-dimensional line. Thus, an address may be 20 m removed from the axis of the road and at the same time a few meters from the side of the road. Second, in the densely built city of Amsterdam the small inaccuracy of the road network and the address coordinates may combine into relatively large errors compared to the true distance contrast (mostly between 3 and 30 m). Both, address coordinates and the road network have an accuracy of about 5-10 m (7). The Spearman correlation between GIS-derived distance and field observations was 0.33. An alternative explanation for the better performance of distance to the side of the road is that the impact of mopeds and motorcycles may be better represented because of their position on the road. Investigators have used a large number of approaches to represent the influence of traffic on pollutant concentrations, including traffic intensity on the nearest (major) road, distance to the nearest (major) road or traffic intensity in buffers of, for example, 50, 100, or 250 m. We observed that the product of traffic intensity and inverse distance outperformed previously used indicators substantially. The product variable represents better the actual processes of emission and dispersion than the addition of traffic intensity and distance in a linear regression model. The buffer variables may have performed worse, because they ignore the presence of buildings in between the receptor (sampling points) and source, an important issue in the densely built city of Amsterdam and many other European cities. The model further included address density in a 300 m buffer and port in a 3000 m buffer, reflecting the impact of nontraffic sources probably including home heating and emissions from shipping and cargo handling. Though road traffic is an important source of urban ultrafine particles, there is currently insufficient information about other sources (17, 18). Two recent source-apportoinment studies identified

multiple sources, including traffic, nucleation, industrial and home heating emissions (17, 18). Small point sources may also be important for local variations, as illustrated by high PNC near smokers on the pavement and construction works (19). The variability of PNC at urban background sites further supports the impact of urban sources. We offered detailed land-use information available from the city of Amsterdam as predictors, which included industry, port, urban green in addition to traffic. Information about small point sources such as (smokers outside) restaurants is however not well represented in GIS databases in Amsterdam and probably elsewhere. Excluding predictors derived from field observations reduced the percentage explained variability of the model substantially. The inclusion of field observations in the model clearly limits the ability of the model to estimate PNC at other locations for which no field observations have been collected. The importance of the model is more in the illustration of the potential to model PNC if detailed input data are available. As this is the first study using land use regression, we cannot assess the generalizability of the model to other locations. Based upon the diversity of potentially important sources, we should expect other sources to be important in other locations. More work is necessary identifying these sources beyond what is available in public GIS databases. Coarse Particle Model. LUR models have been published for PM10 and PM2.5, but not for the coarse fraction of PM separately (4). Coarse particles have been associated with short-term health effects, with little evidence for effects of long-term exposure (20). This could however partly be due to the lack of data on spatial variation of coarse particles, hampering the evaluation of coarse particles in epidemiological studies. The model we developed for coarse particles included traffic intensity and land use variables. The traffic variable probably reflects the impact of nontail pipe emissions. Nontail pipe emissions include brake wear, tire wear, and resuspended road dust (21). With stringent regulations for exhaust emissions, interest in nontailpipe emissions is growing (21). The predicted quantitative impact of traffic was of the same magnitude as that of the other nontraffic land use variables, expressed for a difference in the 10th-90th percentile of each predictor. This distinguishes the coarse particles model from especially the PNC and soot model, in which the traffic variable was more dominant. The performance of the coarse particle LUR model was worse than for the other components. This could be due to lower concentration data quality because coarse particle concentration was calculated as the difference between measured PM10 and PM2.5; poorer adjustment for temporal variation using a central monitoring location and/or a larger diversity of sources contributing to urban coarse particles. In a study in Philadelphia, the spatial variability of coarse particles was larger than that of fine PM and the temporal correlation was lower (22). The first explanation is unlikely to be important as the precision expressed as the coefficient of variation of the individual daily measurements determined from field duplicates was 9% (5). The temporal correlation in our study between measurements made at the central site and directly outside the homes was 0.89, only slightly lower than observed for the PM2.5 and soot (R ) 0.98 and 0.94), but higher than for PNC (R ) 0.72) (5), suggesting that this is an unlikely explanation. Probably, incomplete information available in GIS on urban sources of coarse particles explains the poorer performance. Comparison of PNC Model with PM2.5, Soot, and Coarse Particle Models. The performance of the LUR models for PM2.5 and soot agrees well with the limited number of previous studies (4). In the TRAPCA study (10) and the study in the Ruhr area (13), it was also observed that soot was

better explained than PM2.5. We developed these models to investigate how the predictions of the PNC model correlated with the predictions of other traffic-related pollutants. High correlations were found, as previous LUR studies also found high correlations between estimated concentrations between PM2.5, soot, and NO2 (10). The high correlation between PNC and PM2.5 predictions is partly a consequence of the modeling approach, as the correlation between measured concentrations was lower. The correlation between modeled concentrations might have been lower if we had been able to include more specific predictors than the publicly available GIS data, which despite their detail still are general variables. The relatively large number of traffic locations may also have contributed to the similarity of traffic-variable dominated models for PNC and PM2.5. However, we also measured at 26 urban background sites spread across the city, differing substantially in, for example, land use and residential density. In several models, nonroad traffic variables were identified as significant predictors. Industrial land use probably did not enter prediction models, because of the lack of major industries around Amsterdam, in contrast to two recent source apportionment studies in which industry was identified as a factor at urban background locations in two cites with important industrial sources (17, 18). As none of the sampling sites was located among the Amsterdam canals, we were unable to assess the local impact of emissions from largely unregulated pleasure and tour boats. The high correlation between PNC and soot predictions is probably real, as the correlation between measured concentrations was similar. Ambient soot particles are generally smaller than 1 µm and predominantly below 0.18 µm (17). Freshly emitted diesel particles are mostly within the ultrafine range. Therefore soot and particle number concentration partly represent the same particles. At short distances, the decrease in concentration for both pollutants is governed largely by dispersion and less by coagulation or other physicochemical removal processes, which could distinguish between ultrafine and fine particles (23). A further limitation is the lack of data on nitrogen dioxide (NO2), the pollutant for which the most LUR models have been developed. We were therefore unable to evaluate whether NO2 could be a surrogate for PNC. Recent trends however probably result in NO2 being a poorer surrogate of particulate emissions from motor vehicles in the future. Because of availability of data, we only evaluated the correlation of predicted values for the different pollutants at the monitoring sites used to develop the models. This correlation could be lower at other locations, depending on the correlation between the included predictor variables with the traffic intensity variable, which is included in all models. The high correlation of the LUR model predictions developed with publicly available GIS data does not allow separation of the health effects of these pollutants when applied in an epidemiological study. However, when models are applied in different locations with varying ratios in absolute levels of the three pollutants, an indication of the relevant component(s) may be obtained. For PM2.5 comparison of health effects related to intracity and intercity exposure contrasts may give an indication of the importance of different PM fractions.

Acknowledgments The study was performed within the framework of the “Relationship between Ultrafine and fine Particulate matter in Indoor and Outdoor air and respiratory Health”(RUPIOH)project. The project was funded by the EU ENVIRONMENT and CLIMATE Research Programme contract QLRT-2001VOL. 45, NO. 2, 2011 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

627

00452. The study was further funded by the Municipal Health Service Amsterdam (GG&GD), department of Environmental Medicine.

Literature Cited (1) Donaldson, K.; Stone, V. Current hypotheses on the mechanisms of toxicity of ultrafine particles. Ann. Ist. Super. Sanita 2003, 39 (3), 405–410. (2) Ibald-Mulli, A.; Wichmann, H. E.; Kreyling, W.; Peters, A. Epidemiological evidence on health effects of ultrafine particles. J. Aerosol Med. 2002, 15 (2), 189–201. (3) Hoek, G.; Boogaard, H.; Knol, A.; De Hartog, J.; Slottje, P.; Ayres, J. G.; Borm, P.; Brunekreef, B.; Donaldson, K.; Forastiere, F.; Holgate, S.; Kreyling, W. G.; Nemery, B.; Pekkanen, J.; Stone, V.; Wichmann, H. E.; Van der Sluijs, J. Concentration response functions for ultrafine particles and all-cause mortality and hospital admissions: Results of a European expert panel elicitation. Environ. Sci. Technol. 2010, 44 (1), 476–482. (4) Hoek, G.; Beelen, R.; Hoogh, C. de; Vienneau, D.; Gulliver, J.; Fischer, P.; Briggs, D. A review of land-use regression models to assess spatial variation of outdoor air pollution. Atmos. Environ. 2008, 42, 7561–7578. (5) Puustinen, A.; Hameri, K.; Pekkanen, J.; Kulmala, M.; De Hartog, J.; Meliefste, K.; Ten Brink, H.; Kos, G.; Katsouyanni, K.; Karakatsani, A.; Kotronarou, A.; Kavouras, I.; Meddings, C.; Thomas, S.; Harrison, R.; Ayres, J. G.; Van der Zee, S.; Hoek, G. Spatial variation of particle number and mass over four European cities. Atmos. Environ. 2007, 41 (31), 6622–6636. (6) de Hartog, J. J.; Hoek, G.; Mirme, A.; Tuch, T.; Kos, G. P.; ten Brink, H. M.; Brunekreef, B.; Cyrys, J.; Heinrich, J.; Pitz, M.; Lanki, T.; Vallius, M.; Pekkanen, J.; Kreyling, W. G. Relationship between different size classes of particulate matter and meteorology in three european cities. J. Environ. Monit. 2005, 7 (4), 302–310. (7) Beelen, R.; Hoek, G.; Fischer, P.; van den Brandt, P. A.; Brunekreef, B. Estimated long-term outdoor air pollution concentrations in a cohort study. Atmos. Environ. 2007, 41, 1343–58. (8) Vienneau, D.; de Hoogh, K.; Beelen, R.; Fischer, P.; Hoek, G.; Briggs, D. Comparison of land-use regression models between great britain and the netherlands. Atmos. Environ. 2010, 44 (5), 688–696. (9) Lianou, M.; Chalbot, M. C.; Kotronarou, A.; Kavouras, I. G.; Karakatsani, A.; Katsouyanni, K.; Puustinnen, A.; Hameri, K.; Vallius, M.; Pekkanen, J.; Meddings, C.; Harrison, R. M.; Thomas, S.; Ayres, J. G.; Brink, H.; Kos, G.; Meliefste, K.; de Hartog, J. J.; Hoek, G. Dependence of home outdoor particulate mass and number concentrations on residential and traffic features in urban areas. J. Air Waste Manage. Assoc. 2007, 57 (12), 1507–1517. (10) Brauer, M.; Hoek, G.; van Vliet, P.; Meliefste, K.; Fischer, P.; Gehring, U.; Heinrich, J.; Cyrys, J.; Bellander, T.; Lewne, M.; Brunekreef, B. Estimating long-term average particulate air pollution concentrations: Application of traffic indicators and geographic information systems. Epidemiology 2003, 14 (2), 228–239.

628

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 45, NO. 2, 2011

(11) Zhu, Y.; Hinds, W.; Kinm, S.; Shen, S.; Sioutas, C. Study of ultrafine particles near a major highway with heavy-duty diesel traffic. Atmos. Environ. 2002, 36, 4233–4235. (12) Vardoulakis, S.; Fisher, B. E. A.; Pericleous, K.; Gonzalez-Flesca, N. Modelling air quality in street canyons: A review. Atmos. Environ. 2003, 37 (2), 155–82. (13) Hochadel, M.; Heinrich, J.; Gehring, U.; Morgenstern, V.; Kuhlbusch, T.; Link, E.; Wichmann, H. E.; Kramer, U. Predicting long-term average concentrations of traffic-related air pollutants using GIS-based information. Atmos. Environ. 2006, 40, 542– 553. (14) Larson, T.; Henderson, S. B.; Brauer, M. Mobile monitoring of particle light absorption coefficient in an urban area as a basis for land use regression. Environ. Sci. Technol. 2009, J43 (13), 4672–4678. (15) Weijers, E. P.; Khlystov, A. Y.; Kos, G. P. A.; Erisman, J. W. Variability of particulate matter concentrations along roads and motorways determined by a moving measurement unit. Atmos. Environ. 2004, 38, 2993–3002. (16) Aalto, P.; Hameri, K.; Paatero, P.; Kulmala, M.; Bellander, T.; Berglind, N.; Bouso, L.; Castano-Vinyals, G.; Sunyer, J.; Cattani, G.; Marconi, A.; Cyrys, J.; von Klot, S.; Peters, A.; Zetzsche, K.; Lanki, T.; Pekkanen, J.; Nyberg, F.; Sjovall, B.; Forastiere, F. Aerosol particle number concentration measurements in five European cities using TSI-3022 condensation particle counter over a three-year period during health effects of air pollution on susceptible subpopulations. J. Air Waste Manage. Assoc. 2005, 55 (8), 1064–1076. (17) Kasumba, J.; Hopke, P.; Chalupa, D. C.; Utell, M. J. Comparison of sources of submicron particle number concentrations measured at two sites in Rochester, NY. Sci. Total Environ. 2009, 407, 5071–5084. (18) Fernandez-Camacho, R.; Rodrigueze, S.; Rosa de la, R.; Sanchez de la Campa, A. M.; Viana, M.; Alastuey, A.; Querol, X. Ultrafine particle formation in the inland seebreeze airflow in southwestern Europe. Atmos. Chem. Phys. Discuss. 2010, 10, 17753– 17788. (19) Kaur, S.; Clark, R.; Walsh, P.; Arnold, S. J.; Colville, R. N.; Nieuwenhuijsen, M. Exposure visualisation of ultrafine particle counts in a transport micro-environment. Atmos. Environ. 2006, 40 (2), 386. (20) Brunekreef, B.; Forsberg, B. Epidemiological evidence of effects of coarse airborne particles on health. Eur. Respir. J. 2005, 26 (2), 309–318. (21) Thorpe, A.; Harrison, R. M. Sources and properties of nonexhaust particulate matter from road traffic: A review. Sci. Total Environ. 2008, 400 (1-3), 270–282. (22) Burton, R. M.; Suh, H. H.; Koutrakis, P. Spatial variation in particulate concentrations within metropolitan Philadelphia. Environ. Sci. Technol. 1996, 30, 400–407. (23) Harrison, R. M.; Jones, A. M. Multisite study of particle number concentrations in urban air. Environ. Sci. Technol. 2005, 39 (16), 6063–6070.

ES1023042