Development of Land Use Regression Models for PM
2.5
,PM
2.5
Absorbance, PM
10
and PM
coarse
in 20 European Study Areas; Results
of the ESCAPE Project
Marloes Eeftens,
†,
* Rob Beelen,
†
Kees de Hoogh,
‡
Tom Bellander,
§
Giulia Cesaroni,
∥
Marta Cirach,
⊥,#,▽
Christophe Declercq,
○
Audrius De
dele
,
◆
Evi Dons,
¶,∞
Audrey de Nazelle,
⊥,#,▽
Konstantina Dimakopoulou,
⊗
Kirsten Eriksen,
⋈
Gre
goire Falq,
○
Paul Fischer,
☼
Claudia Galassi,
◎
Regina Graz
ulevic
iene
,
◆
Joachim Heinrich,
◘
Barbara Hoffmann,
☆,√
Michael Jerrett,
$
Dirk Keidel,
%,∀
Michal Korek,
§
Timo Lanki,
&
Sarah Lindley,
@
Christian Madsen,
+
Anna Mo
lter,
©
Gizella Na
dor,
¥
Mark Nieuwenhuijsen,
⊥,#,▽
Michael Nonnemacher,
⧳
Xanthi Pedeli,
⊗
Ole Raaschou-Nielsen,
⋈
Evridiki Patelarou,
£
Ulrich Quass,
Я
Andrea Ranzi,
Å
Christian Schindler,
%,∀
Morgane Stempfelet,
○
Euripides Stephanou,
⬕
Dorothea Sugiri,
☆
Ming-Yi Tsai,
%,∀,⬕
Tarja Yli-Tuomi,
&
Miha
ly J Varro
,
¥
Danielle Vienneau,
‡
Stephanie von Klot,
∑
Kathrin Wolf,
∑
Bert Brunekreef,
†,∮
and Gerard Hoek
†
†
Institute for Risk Assessment Sciences, Utrecht University, P.O. Box 80178, 3508 TD Utrecht, The Netherlands
‡
MRC-HPA Centre for Environment and Health, Department of Epidemiology and Biostatistics, Imperial College London,
London, United Kingdom
§
Institute of Environmental Medicine, Karolinska Institutet, Stockholm, Sweden
∥
Epidemiology Department, Lazio Regional Health Service, Rome, Italy
⊥
Center for Research in Environmental Epidemiology (CREAL), Barcelona, Spain
#
IMIM (Hospital del Mar Research Institute), Barcelona, Spain
▽
CIBER Epidemiología y Salud Pu
blica (CIBERESP), Spain
○
French Institute for Public Health Surveillance, Saint-Maurice, France
◆
Vytautas Magnus University, Kaunas, Lithuania
¶
VITO-MRG (Flemish Institute for Technological Research), Environmental Risk and Health unit, Mol, Belgium
∞
Hasselt University, Diepenbeek, Belgium
⊗
Department of Hygiene, Epidemiology & Medical Statistics, National and Kapodistrian University of Athens, Medical School, Athens, Greece
⋈
Danish Cancer Society Research Center, Copenhagen, Denmark
☼
Centre for Environmental Health, National Institute for Public Health and the Environment, Bilthoven, The Netherlands
◎
AOU San Giovanni Battista − CPO Piedmont, Turin, Italy
◘
HMGU Institute of Epidemiology I, Neuherberg, Germany
☆
IUF Leibniz Research Institute for Environmental Medicine, and
√
Medical Faculty, Heinrich-Heine, University of Du
sseldorf,
Du
sseldorf, Germany
$
School of Public Health, University of California, Berkeley, California, United States
%
Department of Epidemiology and Public Health, Swiss Tropical & Public Health Institute, Basel, Switzerland
∀
University of Basel, Basel, Switzerland
&
Department of Environmental Health, National Institute for Health and Welfare, Kuopio, Finland
@
School of Environment and Development (Geography), The University of Manchester, Manchester, England
+
Division of Epidemiology, Norwegian Institute of Public Health, Oslo, Norway
©
Centre for Occupational and Environmental Health, The University of Manchester, Manchester, England
¥
Department of Environmental Epidemiology, National Institute of Environmental Health, Budapest, Hungary
⧳
Institute for Medical Informatics, Biometry and Epidemiology, University of Duisburg-Essen, Essen, Germany
Received: May 22, 2012
Revised: August 30, 2012
Accepted: September 10, 2012
Published: September 10, 2012
Article
pubs.acs.org/est
© 2012 American Chemical Society 11195 dx.doi.org/10.1021/es301948k | Environ. Sci. Technol. 2012, 46, 11195−11205
*
S
Supporting Information
ABSTRACT: Land Use Regression (LUR) models have been used increasingly for modeling
small-scale spatial variation in air pollution concentrations and estimating individual exposure
for participants of cohort studies. Within the ESCAPE project, concentrations of PM
2.5
,PM
2.5
absorbance, PM
10
,andPM
coarse
were measured in 20 European study areas at 20 sites per area.
GIS-derived predictor variables (e.g., traffic intensity, population, and land-use) were
evaluated to model spatial variation of annual average concentrations for each study area. The
median model explained variance (R
2
) was 71% for PM
2.5
(range across study areas 35−94%).
Model R
2
was higher for PM
2.5
absorbance (median 89%, range 56−97%) and lower for
PM
coarse
(median 68%, range 32− 81%). Models included between two and five predictor
variables, with various traffic indicators as the most common predictors. Lower R
2
was related
to small concentration variability or limited availability of predictor variables, especially traffic
intensity. Cross validation R
2
results were on average 8−11% lower than model R
2
.Careful
selection of monitoring sites, examination of influential observations and skewed variable
distributions were essential for developing stable LUR models. The final LUR models are used to estimate air pollution
concentrations at the home addresses of participants in the health studies involved in ESCAPE.
1. INTRODUCTION
Epidemiological studies have shown adverse health effects of
long-term exposure to air pollution.
1,2
Air pollution from
motorized road traffic is a main public health concern in Europe.
3
Many studies have demonstrated large within-city contrasts in
traffic related air pollutants in European and U.S. cities.
3−11
Land
Use Regression (LUR) modeling has been used frequently to
explain these spatial contrasts, using predictor variables derived
from geographic information systems (GIS).
6,7,11
LUR models
make use of a spatially dense network of measured air pollution
concentrations. Each monitoring site is characterized by a set
of potential predictors such as population density, land use and
various traffic-related variables. Statistical modeling is used to
determine which predictors best explain the pollution concen-
trations.
6,7,11
LUR modeling has generally been able to explain
a large amount of spatial variability. An increasing number of
epidemiological studies make use of LUR models for estimating
outdoor air pollution concentrations at the home addresses of
cohort subjects.
12,13
Many LUR studies have used data on nitrogen oxides, usually
because these can be easily obtained using low-cost passive
samplers.
7
While health effects are probably more related to
particles,
14,15
LUR models for particulate matter and absorbance
are less numerous because they require a more intensive
monitoring effort.
7
Routine monitoring networks often do not
offer the required spatial density, do not measure all components
of interest (e.g., soot) or do not measure at sites relevant for
population exposure. Within Europe there is still a lack of PM
2.5
monitoring and PM monitoring is performed with continuous
monitors that require correction factors and differ per country.
16
So far, there are few LUR studies on the coarse fraction of
particulate matter,
17
while there is increasing epidemiological
evidence showing that coarse particles are associated with acute
respiratory health effects.
18
Long-term effects of PM
coarse
have
not been studied extensively, partly because of a lack of spatially
resolved data on coarse particle concentrations.
18
The ESCAPE project (European Study of Cohorts for Air
Pollution Effects, www.escapeproject.eu) was designed to study
the effects of long-term air pollution exposure on health. ESCAPE
makes use of health data from existing cohort studies. Exposures
to air pollution were assessed for study participants' individual
home address with LUR models based upon standardized specific
PM monitoring campaigns in each of the study areas.
This paper describes the development and performance of
the LUR models of 20 European study areas for PM
2.5
,PM
2.5
absorbance, PM
10
, and PM
coarse
. The ESCAPE database is
currently the largest database of spatially resolved PM data in
Europe, allowing development of LUR models. We will discuss
issues in LUR model development, such as influential observa-
tions, which have not often been addressed in the LUR literature.
Results of the ESCAPE PM pollution measurements were
recently accepted for publication.
19
2. MATERIALS AND METHODS
For 20 study areas across Europe (Figure 1), LUR models were
developed for PM
2.5
,PM
2.5
absorbance, PM
10
and PM
coarse
based
upon measured annual average concentrations. LUR models
were developed using a range of GIS-derived predictor variables,
from consistent European data sets compiled through ESCAPE
and local data sets. Models were developed using a supervised
stepwise method that maximized model explained variance,
with a priori specified signs of slopes (e.g., positive for traffic
intensity). Models were optimized locally with no attempt to
force a common model to all study areas. This decision was based
on the diversity of study areas and differences in available GIS
predictor variables. LUR models were developed locally at each
center, following a common manual (http://www.escapeproject.
eu/manuals/). A workshop was attended by all local centers to
£
Department of Social Medicine, Medical School, University of Crete, Greece
Я
Air Quality & Sustainable Nanotechnology, IUTA Institu
tfu
r Energie- und Umwelttechnik e.V., Duisburg, Germany
Å
Regional Reference Centre on Environment and Health, ARPA Emilia Romagna, Modena, Italy
¢
Environmental Chemical Processes Laboratory, University of Crete, Heraklion, Greece
⬕
Department of Environmental & Occupational Health Sciences, University of Washington, Seattle, Washington, United States
∑
HMGU Institute of Epidemiology II, Neuherberg, Germany
∮
Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, The Netherlands
Environmental Science & Technology Article
dx.doi.org/10.1021/es301948k | Environ. Sci. Technol. 2012, 46, 11195−1120511196
standardize GIS analyses and LUR model development.
Finalized LUR models were sent to the coordinating center for
evaluation.
Air Pollution Measurement Data. The ESCAPE measure-
ments and sampling site s election have been described
previously.
19
Briefly, particulate matter (PM) was measured
between October 2008 and April 2011. Twenty PM sampling
sites were selected in each study area. In the larger study areas of
The Netherlands and Belgium and Catalunya, forty sites were
measured. Study areas were defined to represent the spatial
distribution of the cohort addresses. We selected regional
background, urban background and traffic sites. Traffic sites were
overrepresented, and we selected a range of traffic intensities to
limit outliers in modeling. Measurements in traffic sites (>10 000
vehicles.day
−1
) were made at building facades, rather than the
kerbside. A detailed description of each study area is given in
the Online Supplement of Eeftens et al.
19
Most study areas
comprised a major city and surrounding smaller towns. Each
selected site was measured three times for 14 days, in the cold,
warm and intermediate seasons. Two fractions of particulate
matter (smaller than 2.5 μm (PM
2.5
) and smaller than 10 μm
(PM
10
)) were sampled using Harvard Impactors. The coarse
fraction (PM
coarse
) was calculated as the difference between PM
10
and PM
2.5
.Reflectance was measured on PM
2.5
filters and trans-
formed into absorbance.
19
For each site, results from the three
measurements were averaged to estimate the annual average,
adjusting for temporal variation using a centrally located back-
ground reference site, which was operated for a whole year.
8,19
A temporal correction was calculated as the difference of each
individual reference site measurement from the annual mean at the
reference site. The calculated correction was then subtracted from
all measurements that took place in that particular round.
GIS Predictor Data. Positioning of Measurement Sites.
Multiple GPS measurements were taken at every site, but all
positions were corrected manually to ensure an accurate position
relative to roads on the detailed local road maps. This was done
by someone who had personally visited the site.
Predictor Variables. Predictor variables were calculated for
each site, using the site coordinates and digital data sets within a
GIS. We used a combination of European data obtained centrally
and local data. Local source data were collected because some
data were not available on a European level or were more precise
or more up-to-date. For traffic variables, we calculated circular
buffers with radii of 25, 50, 100, 300, 500, and 1000 m around
each monitoring site. For land use and population, we calculated
buffers of 100, 300, 500, 1000, and 5000 m. A detailed description
and an overview of all calculated variables, is shown in Supporting
Information (SI) SI1.
The following GIS source data were available centrally:
1 Digital road network Road data were available at 1:10 000
resolution from Eurostreets version 3.1 digital road
network, derived from the TeleAtlas MultiNet data set
for the year 2008. The network includes road class but not
traffic intensity.
2 Land use dataCORINE (COoRdination of INformation
on the Environment) land cover data were available from
the European Environment Agency (EEA) for the year
2000.
20,21
We used six land use categories: high density
Figure 1. ESCAPE study areas.
Environmental Science & Technology Article
dx.doi.org/10.1021/es301948k | Environ. Sci. Technol. 2012, 46, 11195−1120511197
residential land, low density residential land, industry,
ports, urban green and natural land.
20,21
3 Population density data Population data modeled at a 100
m grid were based upon land cover and the 2001
population density available from the EEA.
22,23
4 Altitude Digital elevati on data (SRTM 90 m) were
obtained through the Shuttle Radar Topographic Mission,
and available globally from CGIAR-CSI GeoPortal
(http://srtm.csi.cgiar.org/). The map has a resolution of
90 m at the equator.
A detailed overview of the local GIS variables can be found in
SI SI2. We required a spatial resolution of at least 100 m. Local
GIS data included land use, population and household density,
altitude and study-area specific variables such as distance to the
sea. Detailed local road networks with linked traffic intensity
were available for most areas. To account for variation in regional
background in The Netherlands/Belgium, 10 regional back-
ground sites were measured, which allowed us to use an inverse
distance weighted regional background concentration.
24
In the
other (smaller) study areas, few regional background sites were
measured as we anticipated little variation in regional back-
ground. We evaluated whether adding geographical coordinates
to the final GIS model improved prediction, and if these trends
were consistent with known pollution patterns.
LUR Model Development. Linear regression models were
developed using a supervised stepwise selection procedure,
first evaluating univariate regressions of the corrected annual
average concentrations with all available potential predictors
following procedures used before.
21
The predictor giving the
highest adjusted explained variance (adjusted R
2
) was selected
for inclusion in the model if the direction of effect was as defined
a priori. We then evaluated which of the remaining predictor
variables further improved the model adjusted R
2
, selected the
one giving the highest gain in adjusted R
2
, and the right direction
of effect. Subsequent variables were not selected if they changed
the direction of effect of one of the previously included variables.
This process continued until there were no more variables with
the right direction of effect, which added at least 0.01 (1%) to the
adjusted R
2
of the previous model.
As final steps, variables with a p-value above 0.10 were
removed from the LUR model. If the Variance Inflation Factor
(VIF) was higher than 3 −indicating collinearity-, the variable
with the highest VIF was removed and the model re-evaluated.
Cook’s D statistics were used to detect influential observations.
Cook’s D values above 1 were further examined by assessing the
changes in model coefficients on excluding the responsible site.
If removal of this site caused large changes in a specific variable’s
coefficient, the modeling procedure was repeated using all sites,
but now without offering this variable.
Overall model performance was evaluated by leave-one-out
cross validation (LOOCV): each site was sequentially left out
from the model while the included variables were left unchanged.
The Moran’s I statistic was calculated to indicate spatial auto-
correlation of the model residuals.
3. RESULTS
Within-Area Concentration Contrasts. Pollutant ranges
are shown in Tables 1−3 for each study area and in more detail in
Eeftens et al.
19
For most areas, substantial variation was present
within the area. Within-area contrasts were largest for PM
coarse
and PM
2.5
absorbance. Within-area contrasts differed between
areas, for example, for PM
2.5
lower contrasts were found in
Manchester, Ruhr Area, Gyor and Turin.
Available Predictor Variables. In 18 of the 20 study areas
local traffic inten sity data was collected. Excep tion s were
Heraklion and Catalunya. In many study areas, few sites were
within 100, 300, or 500 m of a port, forest or industrial area,
resulting in many 0-values. Similarly, for several areas a large
number of 0-values occurred for major roads in small buffer
(25 or 50 m). Generally, variables with less than 4−5 nonzero
values were not offered in the modeling, but we evaluated the
stability of parameter estimates for each model.
Land Use Regression Modeling. The LUR models for
PM
2.5
,PM
2.5
absorbance, and PM
coarse
are described in Tables
1−3 and those for PM
10
in SI SI3, Table 1. Descriptive statistics
of the predictor variables used in the models can be found in
SI SI4. In four areas, one site was excluded from modeling
because only one successful measurement was available (Lugano,
Oslo) or the site was too influential and was considered a non-
representative site (Stockholm County, Manchester), further
discussed in the modeling experiences section in the Discussion.
PM
2.5
Models. In most study areas, a substantial fraction of
the measured spatial variability was explained by the available
GIS predictor variables (Table 1). The median model explained
variance (R
2
) was 71% and ranged from 35% (Manchester) to
94% (Stockholm County).The variation in R
2
is partly related
to the limited availability of relevant predictors, especially local
traffic intensity data. The two areas without local or limited traffic
intensity data (Heraklion, Catalunya) both had R
2
below the
median. In Barcelona (part of the Catalunya study area), local
traffic data was available and a much better model could be
developed. Small variation of measured concentrations may have
contributed to lower R
2
in some areas, such as Manchester, but
overall the association is not strong (Table 1). There was no clear
geographical pattern of the magnitude of R
2
across Europe.
For most models, the differences between the model R
2
and the
leave-one-out cross validation R
2
was less than 15%, indicating
stable models. Models included two to five predictor variables.
Traffic indicators were included in 18 of the 20 models, with
traffic intensity in various buffer sizes included in most models.
Less often included predictors were residential land use,
population density, industrial/port and natural land use.
PM
2.5
Absorbance Models. Model R
2
was higher for PM
2.5
absorbance (median 89%) than for PM
2.5
, probably related to the
larger spatially variability (Table 2). In Manchester, R
2
was high,
whereas no reliable model could be developed for PM
2.5
.
Explained variance differed across areas from 56% (Heraklion)
to 97% (Ruhr Area). The low value in Heraklion is likely due to
the lack of traffic intensity data. Differences between model R
2
and LOOCV R
2
were generally lower than 10%, indicating stable
models. The models included two to five predictors. In all models
traffic variables were present. With the exception of Heraklion,
all models included small-scale tra ffic variables, such as traffic
intensity in the nearest street, the product of traffic intensity on
the nearest major street and inverse distance and small buffers
(≤100 m) of traffic intensity. Models also included trafficin
larger buffers and land use predictors.
PM
coarse
Models. The median model R
2
was 68%, with a range
from 32% (Kaunas) to 81% (Munich/Augsburg) (Table 3).
Model R
2
was the lowest from the modeled PM metrics.
Differences between model explained variance and cross valida-
tion were generally larger for PM
coarse
than for the other PM
metrics. PM
coarse
models generally included two to three predictor
variables, fewer than for the other PM metrics. In all areas except
Environmental Science & Technology Article
dx.doi.org/10.1021/es301948k | Environ. Sci. Technol. 2012, 46, 11195−1120511198
Table 1. Description of Developed LUR Models for PM
2.5
, Including Descriptive Statistics of the Measured Concentrations
study area LUR model
a
R
2
of
model
R
2
validation
RMSE
(validation)
(μg/m
3
)
number
of sites
b
Moran’sI
(p-value)
measured
concentration
(μg/m
3
)
c
Oslo, Norway 8.08 + 1.30 × 10
−3
× HHOLD_500 + 9.28 × 10
−5
× TRAFNEAR − 5.95 × 10
−8
× NATURAL_5000 74% 68% 1.2 19 −0.05 (0.56) 8.6 [5.0−12.9)
Stockholm County, Sweden 7.95 − 8.96 × 10
−6
× WATER_500 − 1.48 × 10
−7
× WATER_500_5000 + 1.37 × 10
−5
×
HEAVYTRAFLOAD_50 + 3.66 × 10
−4
× ROADLENGTH_500
87% 78% 0.8 19 −0.02 (0.28) 8.3 [4.4−11.3]
Helsinki/Turku, Finland 9.25 − 6.75 × 10
−6
× NATURAL_500
d
+ 6.34 × 10
−7
× TRAFMAJORLOAD_50 67% 53% 1.0 20 −0.30 (0.03) 8.6 [5.3−12.3]
Copenhagen, Denmark 9.12 + 1.96 × 10
−4
× ROADLENGTH_500 − 2.20 × 10
−3
× GREEN_100
d
62% 55% 1.1 20 −0.02 (0.68) 11.1 [8.4−14.0]
Kaunas, Lithuania 14.74 + 1.92 × 10
−2
× POP_100 + 1.67 × 10
−4
× TRAFMAJOR 60% 45% 2.6 20 −0.05 (0.45) 21.1 [16.6−30.3]
Manchester, UK 9.41 + 1.24 × 10
−6
× HDRES_1000 35% 21% 0.8 19 −0.08 (0.50) 9.8 [8.1−11.9]
London/Oxford, UK 7.19 + 1.38 × 10
−3
× INTMAJORINVDIST + 2.65 × 10
−4
× ROADLENGTH_500 82% 77% 1.4 20 −0.19 (0.20) 11.2 [7.0−21.2]
Netherlands/Belgium 9.46 + 0.42 × REGIONALESTIMATE + 0.01 × MAJORROADLENGTH_50 + 2.28 × 10
−9
×
TRAFMAJORLOAD_1000
67% 61% 1.2 40 0.02 (0.77) 17.7 [12.7−21.5]
Ruhr Area, Germany 81.73 + 5.61 × 10
−8
× HEAVYTRAFLOAD_1000 + 1.20 × 10
−7
× INDUSTRY_5000 + 1.04 × 10
−4
×
POP_1000 − 2.57 × 10
−5
× XCOORD
88% 79% 0.9 20 −0.02 (0.64) 18.5 [15.5−21.6]
Munich-Augsburg, Germany 11.90 + 1.94 × 10
−2
× MAJORROADLENGTH_50
d
+ 4.95 × 10
−4
× ROADLENGTH_300
d
− 14.30 ×
URBGREEN_5000
d
+ 7.41 × 10
−9
× TRAFMAJORLOAD_1000
d
78% 62% 1.0 20 −0.13 (0.49) 14.3 [9.7−17.6]
Vorarlberg, Austria 25.44 + 0.11 × BUILDINGS_100 − 0.65 × SQRALT 57% 42% 1.5 20 0.09 (0.06) 13.3 [8.8−17.3]
Paris, France 10.38 + 5.34 × 10
−4
× MAJORROADLENGTH_500 + 2.75 × 10
−7
× INDUSTRY_5000 + 1.46 × 10
−4
×
TRAFMAJOR
89% 73% 1.8 20 −0.11 (0.83) 16.0 [11.9−30.6]
Gyor, Hungary 23.98 − 1.71 × 10
−2
× URBGREEN_5000 + 7.52 × 10
−5
× ROADLENGTH_1000 + 5.90 × 10
−8
×
TRAFMAJORLOAD_500
64% 46% 1.2 20 −0.25 (0.05) 22.6 [20.6−26.2]
Lugano, Switzerland 46.30 + 2.25 × 10
−4
× HEAVYTRAFLOAD_50 − 0.57 × SQRALT − 6.90 × 10
−7
× NATURAL_5000 83% 77% 1.1 19 −0.12 (0.10) 17.2 [13.7−22.5]
Turin, Italy 24.90 − 7.03 × 10
−6
× NATURAL_1000 + 9.40 × 10
−7
× TRAFMAJORLOAD_50 + 1.63 × 10
−7
×
LDRES_5000
71% 59% 2.0 20 −0.09 (0.45) 29.3 [22.7−36.3]
Rome, Italy 16.08 + 4.56 × 10
−6
× TRAFLOAD_25 + 3.81 × 10
−3
× ROADLENGTH_100 71% 60% 1.9 20 0.02 (0.30) 19.8 [14.2−27.0]
Barcelona, Spain 16.21 − 4.08 × 10
−6
× GREEN_1000 + 2.04 × 10
−7
× TRAFLOAD_100 + 6.82 × 10
−3
× INTINVDIST2 83% 71% 2.1 20 0.01 (0.46) 16.3 [8.4−24.4]
Catalunya, Spain 14.88 + 9.91 × 10
−4
× INTMAJORINVDIST − 3.27 × 10
−6
× GREEN_1000 + 5.36 × 10
−7
× PORT_5000 62% 51% 2.4 40 −0.06 (0.38) 15.6 [8.4−24.4]
Athens, Greece 13.98 + 2.04 × 10
−8
× TRAFLOAD_500 − 1.77 × 10
−7
× NATURAL_5000 + 0.017 × ROADLENGTH_25
+ 1.52 × 10
−5
× INDUSTRY_300 + 1.80 × 10
−2
× MAJORROADLENGTH_50
86% 69% 1.7 20 −0.10 (0.30) 20.9 [13.7−25.7]
Heraklion, Greece 12.95 + 0.03 × ROADLENGTH_25 + 9.06 × 10
−6
× HDRES_300 49% 25% 2.1 20 −0.07 (0.98) 14.7 [11.3−21.0]
a
See SI SI1 and Table 1 for detailed explanation of the variable names. Some variables are buffers with _X indicating the radius of the buffer in meters. The following predictors were derived for all sites:
the surface area (m
2
) of high density residential land (HDRES_X), low density residential land (LDRES_X), all residential land (HLDRES_X), industry (INDUSTRY_X), port (PORT_X), urban green
space (URBGREEN_X), natural land (NATURAL_X), urban green and natural land combined (GREEN_X), water (WATER_X), the number (N) or surface area (m
2
) of buildings (BUILDINGS_X),
population (N) (POP_X) or number (N) of households (HHOLD_X), the square root of altitude (SQRALT), a regional concentration estimate (μg/m
3
or 10
−5
m
−1
), X-coordinate (XCOORD), Y-
coordinate (YCOORD), total length (m) of all road and all major road segments (ROADLENGTH_X, MAJORROADLENGTH_X), inverse distance (m
−1
) and inverse squared distance (m
−2
) to the
nearest road of the central road network (DISTINVNEARC1, DISTINVNEARC2) and the nearest major road in the central network (DISTINVMAJORC1, DISTINVMAJORC2), traffic intensity on the
nearest road (TRAFNEAR) and nearest major road (TRAFMAJOR), heavy traffic intensity on the nearest (HEAVYTRAFNEAR) and nearest major road (HEAVYTRAFMAJOR), inverse distance (m
−1
)
and inverse squared distance (m
−2
) to the nearest road of the local network (DISTINVNEAR1, DISTINVNEAR2) and the nearest major road in the local network (DISTINVMAJOR1,
DISTINVMAJOR2), the product of inverse/inverse squared distance to the nearest road and the traffic intensity on this road (vehicles·day
−1
m
−1
/vehicles·day
−1
m
−2
) (INTINVDIST, INTINVDIST2),
equivalent for major roads (INTMAJORINVDIST, INTMAJORINVDIST2), and for heavy traffic (HEAVYINTINVDIST, HEAVYINTINVDIST2), the sum of (traffic intensity × the length of all road
segments) within a buffer (vehicles· day
−1
·m) for all roads (TRAFLOAD_X), for major roads (TRAFMAJORLOAD_X), for heavy traffic (HEAVYTRAFLOAD_X) and heavy traffic on major roads
(HEAVYTRAFMAJORLOAD_X. See SI SI4 for description of distributions of included variables.
b
Number of sites that have been used for model development. Failed measurements explain fewer than
20 sites for Oslo and Lugano. Two sites in Stockholm County and Manchester were excluded from model building, see also SI SI6 for details.
c
Mean [min − max]
d
Local data, SI SI1 and SI2.
Environmental Science & Technology Article
dx.doi.org/10.1021/es301948k | Environ. Sci. Technol. 2012, 46, 11195−1120511199