Journal of Exposure Analysis and Environmental Epidemiology (1999) 9, 535 ± 545

# 1999 Stockton Press All rights reserved 1053- 4245/99/$12.00

http://www.stockton-press.co.uk

Estimating environmental exposures to sulfur dioxide from multiple

industrial sources for a case±control study
J. FELIX ROGERS,a GEORGE G. KILLOUGH,b SHIRLEY J. THOMPSON,c CHERYL L. ADDY,c
ROBERT E. MCKEOWNc AND DAVID J. COWENd

a
Centers for Disease Control and Prevention, National Center for Environmental Health, Radiation Studies Branch, 4770 Buford Highway, NE, Atlanta,
Georgia 30341 - 3742
b
Hendecagon Corporation, 105 Netherlands Road, Oak Ridge, Tennessee 37830
c
Department of Epidemiology and Biostatistics, School of Public Health, University of South Carolina, Columbia, South Carolina 29298
d
Department of Geography, University of South Carolina, Columbia, South Carolina 29208

This paper first discusses how population exposures to environmental pollutants are estimated from environmental monitoring data and the problems that are
encountered in estimating risk from pollutants on the basis of ecologic studies. We then present a technique of estimating individualized exposures to an
atmospheric pollutant, sulfur dioxide ( SO2 ) , through atmospheric transport modeling for a case ± control study. The transport model uses the quantities of SO2
released from 30 geographically identified industrial facilities and meteorological data ( wind speed and direction ) to predict the downwind ground - level
concentrations of SO2 at geographically identified residences, receptors, of 797 study subjects. A distribution of facility SO2 emissions, uncertainties in
effective stack height, and model uncertainty are incorporated to examine the uncertainty in the predicted versus ambient monitoring SO2 levels, and to
generate an exposure uncertainty distribution for both the cases and controls. The transport model's accuracy is evaluated by comparing recorded ambient
measurements of SO2 with the model's predicted SO2 estimates at geographically identified ambient monitoring stations.

Keywords: atmospheric pollution, environmental, exposure, SO2, transport modeling.

Introduction product of fossil fuel combustion. SO2 is a common air

Air pollution has been linked to increased mortality and
respiratory symptoms in a number of studies both in the
United States and Europe ( Abbey et al., 1993, 1995; Ito et
al., 1993; PoÈnkaÈ and Virtanen, 1994; Xiping et al., 1995;
Ballester et al., 1996; Wojtyniak and Piekarski, 1996 ).
Sulfur dioxide (SO2 ) is one of the components of air
pollution resulting from industrial processes and a source of
sulfate ( SO4 ) in the atmosphere. Atmospheric sulfates
promote the production of acid rain and reduce atmospheric
visibility ( Shprentz et al., 1996 ). SO2 itself is a non-
flammable, nonexplosive, irritant colorless gas that is a
1. Abbreviations: AOD, airway obstructive disease; CO, carbon mon-
oxide; EPA, Environmental Protection Agency; GAEPD, Georgia
Environmental Protection Division; GHD9, Georgia Health District 9;
km2, square kilometer; g / m3, micrograms / cubic meter; NAAQS,
National Ambient Air Quality Standards; NCDC, National Climatic Data
Center; NO2, nitrogen dioxide; SO2, sulfur dioxide; SO4, sulfate; TSP,
total suspended particulate matter; UTM, Universal Transverse Mercator;
VOCs, volatile organic compounds.
2. Address all correspondence to: Felix Rogers, PhD, Centers for Disease
Control and Prevention, National Center for Environmental Health,
Radiation Studies Branch ( MS F - 35 ) , 4770 Buford Highway, NE,
Atlanta, GA 30341 - 3742. Tel.: ( 770 ) 488 - 7099. E-mail: fxr3@cdc.gov
Received 12 November 1997; accepted 24 January 1999.
pollutant in areas surrounding coal -fired electric generating
plants, smelters, sulfuric acid factories and other industries.
Electric generating plants are the primary source of SO2 in
industrialized countries (Petruzzi et al., 1994 ).
Numerous federal agencies in the U.S. have established
standards for air pollutants based upon the pollutants'
potential for adverse health effects or damage to the
environment (Sanders, 1986 ) . The U.S. Environmental
Protection Agency (EPA ) has set the primary National
Ambient Air Quality Standards ( NAAQS ) for SO2
exposure at 78 g/ m3. The annual average ambient
concentration of SO2 in the U.S. is 15 g / m3. Ambient
concentrations of SO2 tend to be highest in the summer
months (40 ±50 g/m3 ) when the demand for electricity is
greatest (National Air Quality and Emissions Trends
Report, 1995 ) .
This investigation reviews recent studies that estimate
environmental exposures to atmospheric pollutants and
demonstrates how misclassification errors in exposure
assessment occur and the impact these errors have on
inferences made from these studies. Finally, this study
presents a method that improves assessment of individual
exposures to atmospheric pollutants for case± control
epidemiological studies.
Rogers et al.
Background
Estimating environmental exposures to atmospheric pollu-
tants is difficult in epidemiologic studies, and, at present,
such studies rely on crude methodologies. Obtaining
information about the sources of the pollutants, including
the kinds of emissions and emission rates, determining the
effect that weather plays upon the dispersion of pollutants
from the sources, identifying the locations where an
individual exposure may occur, and quantifying individual
exposures are difficult and time consuming. As a result,
environmental studies generally have been designed as
inexpensive ecologic studies (Morgenstern and Thomas,
1993 ) that make no attempt to estimate individual
exposures. As surrogates of exposure, environmental
epidemiologic studies have used air quality monitoring data
obtained from sampling networks that were established for
purposes other than health -related research. In areas
without monitoring systems, researchers have attempted to
interpolate measurements collected from monitored areas to
unmonitored areas. The ecological design ignores the
locations of pollutant sources and the atmospheric influence
upon the dispersion and ultimate concentrations of these
pollutants. In ecologic studies researchers assign a crude
estimate of exposure to large population areas such census
tracts, counties, or postal zip code areas.
Several attempts have been made to associate sulfur-
containing pollutants with adverse health effects ( Abbey et
al., 1993, 1995; Ito et al., 1993; Xu et al., 1995; Ballester et
al., 1996; PoÈnkaÈ and Virtanen, 1996; Wojtyniak and
Piekarski, 1996 ). In studies of respiratory symptoms and
exposure to a variety of pollutants, Abbey et al. (1993,
1995 ) used SO2 cut -points from 10- year cumulative
ambient measurements that ranged from 2 to 14 parts -per-
hundred -million (pphm ) . Xu et al. ( 1995 ) used mean
ambient concentrations of SO2 measured in two residential
areas in Beijing, China, to study the effect that exposure to
SO2 had on gestational age in four residential areas. The
exposures, annual mean concentrations of SO2, were from
93 to 108 g / m3. In their study of mortality and air
pollution in the city of London, England, Ito et al. (1993 )
used daily averages of SO2 measurements from seven
sampling sites with ambient exposures of 100 to 1000
g/ m3 while in their study of low- level air pollution and
asthma in Helsinki, Finland, PoÈnkaÈ and Virtanen (1994 )
used eight ambient sampling sites with ambient exposures
ranging from 13 to 25 g / m3.
Another series of more recent ecologic studies that
evolved from the multinational European air pollution
health effects study also relied on ambient measurements
of pollutants for exposure estimates ( Ballester et al.,
1996; PoÈnkaÈ and Virtanen, 1996; Wojtyniak and
Piekarski, 1996 ). The air pollution± mortality study of
Ballester et al. ( 1996 ) and Wojtyniak and Piekarski
536
Estimating environmental exposures to SO2
(1996) included meteorological data such as daily mean
temperatures and humidity in addition to ambient 24 -
measurements of SO2 for generating an exposure
estimate.
Other environmental exposure investigations have ge-
nerated exposure estimates expressed as categories of
exposure based on census tracts (Shaw et al., 1992 ) or
exposure risk indexes incorporating distance from a waste
site and a hazard ranking score ( Geschwind et al., 1992 ).
Marshall et al. ( 1993 ) estimated potential exposure scores
for hazardous waste sites within a 1 -mile radius by using a
template originating at the waste site. The template was
divided into 25 sectors extending from the waste site. The
exposure score was based on the probability of exposure
(i.e., the degree to which a residence in a sector was in the
path of offsite migration of contamination ) . The exposure
metric used in these studies was not related to any specific
pollutant.
A weakness common to all ecologic studies is that the
researchers cannot identify study subjects as exposed or
unexposed. All study subjects may be regarded as exposed,
and those with health effects from the exposure are
generally reflected in increases in the number of hospital
admissions, reported deaths, and reported asthma attacks. In
essence, an ecologic study's use of ambient exposure
measures results in a nondifferential exposure misclassifica-
tion bias. This bias may lead to an overestimation of effect
(Brenner et al., 1992 ) , or an underestimation of effect
(Morgenstern and Thomas, 1993) depending on the study
design and response rates. Furthermore, any causal
inferences suggested from ecologic studies fail to demon-
strate the true effect at the individual level ( Morgenstern
and Thomas, 1993 ) . Attempts to associate exposure with
outcome at the individual level in ecologic studies is known
as the ecologic fallacy (Morgenstern, 1982 ) .
We found no study in which individualized environ-
mental exposures were modeled from a point source to a
geographically identified receptor, a residence, through the
use of an atmospheric transport model. A few studies,
however, have used an atmospheric transport model to
estimate exposures in regions that ranged in sizes from 1
km2 to zip code areas, from which exposure estimates were
interpolated across larger areas (Brown et al., 1984; Hatch
et al., 1990 ) . Although exposure estimates based on
residential locations have not been used in transport
modeling techniques, residential locations have been used
to estimate individual external exposures from nuclear
bomb test fallout (Lloyd et al., 1990 ). However, in the
Lloyd et al. ( 1990 ) study, residence was defined in terms of
a geographical area that contained the true residence. In the
studies by Brown et al. ( 1984 ) , Hatch et al. (1990) and
Lloyd et al. ( 1990 ) the study subjects that lived within
specific areas received the same exposure estimate and
therefore were subject to exposure misclassification.
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(6)
Estimating environmental exposures to SO2 Rogers et al.

The Gaussian environmental transport model (Meade

and Pasquill, 1958 ) presented below is used to estimate
ground -level SO2 concentrations at residences of case and
control subjects when SO2 originates from multiple
geographically identified SO2 release points. The ground -
level concentration estimate is a summation of the model
SO2 concentration estimates for the residence from each
release point. Exposure predictions obtained in this manner
will provide an individualized external exposure estimate
that will reduce the probability of exposure misclassifica-
tion. In this study, we used our estimates of residential SO2
concentrations to test the association between exposure to
environmental SO2 and the birth of very low birthweight
babies in Georgia Health District 9 (GHD9 ) between April
1, 1986 and March 31, 1988.
The model described below was chosen on the basis of
several general assumptions. First, we assumed that the
model is suitable for this type of study. This assumption is
based on the demonstrated reliability of the Gaussian
atmospheric transport model in predicting downwind
concentrations of material released into the atmosphere,
given certain parameter limitations ( Vesper, 1992 ). Second,
we assumed that the use of annual industrial emissions and
meteorological data in the model will provide good
estimates of annual SO2 concentrations at the locations of
interest. Third, we assumed that annual ambient monitoring
data that are used to validate the model estimates are
accurate and represent the true annual ambient SO2
concentrations for the monitored regions. The limitations
of using annual meteorological and monitoring data in
estimating exposures are discussed below.

Methods Figure 1. ( a ) Industrial sulfur dioxide emission locations ( o ) in

Study Area
The study area is Georgia Health District 9 ( GHD9 ), which
is composed of 24 counties in the southeastern coastal
region of Georgia (USA ) . This largely rural region of
Georgia consists primarily of small towns, although it
contains two moderately sized cities, Savannah and
Brunswick. The entire area encompasses approximately
38,000 km2 (Figure 1a ).
Emission Site Data
Emissions data for 1986 ±1988 (Table 1 ) on 30 industrial
facilities within GHD9 were obtained from the Georgia
Environmental Protection Division (GAEPD ) 1 of the
Georgia Department of Natural Resources. The emissions
from these industries, whose products include chemicals,
plastics, fertilizers, asphalt, wood, paper, and gypsum, are
1
Georgia Environmental Protection Division, Georgia Department of
Natural Resources, Atlanta, GA, 30354.
Georgia Health District 9 counties. ( b ) Sulfur dioxide emission sites
( o ) and ambient monitoring sites ( M ) in Chatham County,
Georgia.
monitored by the GAEPD. According to the GAEPD these
facilities account for between 90% and 95% of the 50,500
tons per year of SO2 emissions in the area. The GAEPD
emissions data included the quantity (in tons ) of the
estimated annual emissions of SO2 (the source term ),
facility geocodes, and stack data (stack height, stack
diameter, gas temperature, and gas flow rate ) . No informa-
tion was available to describe the variation in stack releases
over the reported time period. In order to establish some
degree in the uncertainty of the annual SO2 releases, we
incorporated a uniform distribution with a 40% variation,
above and below, the reported release estimates greater than
1000 tons per year (Table 1 ) for stochastic sampling. Of the
30 point sources located within the study area, 14 are
located in 10 of the 24 counties ( Figure 1a ), and 16 are
within a single county, Chatham (Figure 1b ).

Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(6) 537

Rogers et al. Estimating environmental exposures to SO2

Table 1. Point source emission data: county, annual release ( tons ) , Ambient Monitoring Data
geocode, Universal Transverse Mercator ( UTM ) , and physical stack
Data for five ambient monitoring stations within GHD9
heights ( meters ) .
were obtained from the GAEPD for 1986 ±1988. The data
included mean annual SO2 concentrations and geocodes of
County Release Northing Easting Stack
( UTM ) ( UTM ) height the monitoring stations. Four monitoring sites are located
within the city of Savannah (Chatham County, Figure 1b)
Ware 2 3457188.34 365306.95 8
and the fifth site is located 114 km southeast of Savannah in
Chatham 5 3548806.06 495101.77 10
Brunswick (Glynn County ). The monitoring data will be
Chatham 11 3551191.09 482987.58 15
used to test the accuracy of the atmospheric transport model
Evans 13 3558595.75 413411.23 17
in predicting ground -level concentrations of SO2 at
Chatham 20 3554880.15 484201.07 47
Chatham 34 3549989.55 485901.53 11
unmonitored residences.
Bryan 53 3572400.04 455820.06 20
Glynn 77 3443993.31 453598.88 32
Residence Data
Chatham 79 3552603.44 487395.77 18
The residences outside Savannah ( those of 140 case and
Chatham 90 3551191.09 482987.58 26
360 control subjects ) were geocoded ( latitude/longitude )
Chatham 102 3552292.35 487999.2 33
with USGS topographic maps and on- site surveillance.
Chatham 148 3552602.71 487999.57 42
Residences within Savannah ( those of 98 case and 199
Pierce 153 3464903.78 381900.58 10
control subjects ) were geocoded by street address matching
Glynn 218 3453185.44 447904.73 62
of residences to degrees latitude and longitude with the
Chatham 220 3557902.8 486506.55 17
geographic software Maptitude.3 All geographic latitude
Charlton 221 3407999.31 403510.98 20
and longitude coordinates were converted to universal
Chatham 427 3555985.66 486192.77 47
transverse Mercator ( UTM ) coordinates for modeling
Chatham 517 3496505.42 490808.09 59
purposes using the software Geocalc.4
Liberty 518 3510794.92 460999.74 49
Chatham 631 3548285.47 494393.65 20
Environmental Transport Model
Glynn 633 3447802.23 454416.2 60
The Gaussian plume atmospheric transport model used in
Glynn 645 3439591.38 440702.97 40 this study was first introduced by Holland (1953 ) and
Chatham 1946 3549490.62 486004.65 20 Meade and Pasquill ( 1958 ). The sector-averaged form of
Chatham 2336 3556917.28 485797.97 80 the model used in this study is given by:
p  
Wayne 3535 3502404.34 420416.46 80 Q 2=f h2e
Chatham 5860 3551704.32 488498.57 100
ˆ exp …1†
…2x=16†z u 22z
Effingham 6491 3579797.95 484108.21 40
Glynn 7343 3448718.11 450389.02 90 where u= the wind speed ( m /s ), f =the fraction of the time
Camden 7453 3401284.17 448204.93 90 the wind blows into the receptor sector, x =the distance (m )
Chatham 10762 3556794.95 486090.14 90 from the source to the receptor,  z = the vertical dispersion
coefficients (m ) as a function of downwind distance from
the source ( Hanna et al., 1982 ), 2 /16 = the angular width
Meteorological Data of a standard windrose sector of 22.58, h e = effective stack
Meteorological data for Savannah was obtained from the height of the source ( m ). Q = the release rate of SO2 from
National Climatic Data Center ( NCDC ) , Asheville, NC.2 the source ( g /s ), and  = the average ground - level
Savannah was the only location within GHD9 where concentration ( g / m 3).
meteorological data is collected by the NCDC. These data, This model is designed to estimate the average ground -
for 1986 ± 1988, were annualized and included a frequency level concentration over long periods of time such as a
distribution of six wind speed categories over a 16- sector season or a year. The model concentrations are estimated by
windrose for six atmospheric stability classes defined as weighting the predicted concentration at a receptor by the
follows: A Ð extremely unstable, B Ð moderately frequency f that the wind blows from a point source into the
unstable, C Ð slightly unstable, D Ð neutral, E Ð 22.58 sector containing the receptor according to climato-
slightly stable, and F Ð moderately stable conditions. This logical records (the windrose) . The Gaussian model
stability - typing classification is known as the Pasquill ± predicts the averaged concentration along an arc within
Gifford scheme (Hanna et al., 1982 ) . the sector at the calculated distance from a source. Two

2 3
National Climatic Data Center, Federal Building, Ashville, NC 28801 - Caliper Corporation, 1172 Beacon St., Newton, MA 02161.
4
2696. Blue Marble Geographics, 46 Water Street, Gardiner, ME 04345.

538 Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(6)

Estimating environmental exposures to SO2
residences at the same radial distance from a source within
the same sector would receive the same estimated exposure
from that source, but two residences at different distances
within the same sector or two residences in different sectors
would, in general, receive different exposures from the
source. This model allowed us to estimate individualized
exposures, since no two birth homes will fall along the same
arc.
Transport Model Uncertainty
We implemented uncertainty in the transport model, after
Miller and Hively (1987) , by multiplying the predicted
exposure concentrations for each source ± receptor pair by a
lognormally distributed factor with geometric mean 1.0 and
geometric standard deviation 1.52. The lognormal distribu-
tion with these parameters has its 5th percentile at 1 /2 and
its 95th percentile at 2; thus, the factor is between these two
values with probability 0.90) . This factor is assumed to be
stochastically independent of other sources of uncertainty,
such as the release rate and the parameters affecting plume
rise. The choice of a factor of 2 for model uncertainty is
discussed below.
The effective stack height h e ( Hanna et al., 1982 ) is
estimated from:
 v 1:4 
T

he ˆ h ‡ d 1‡ ; …2†
u T
where h =the physical stack height (m ), u= the wind speed
( m / s ), v =the stack gas velocity ( m /s ), T= the stack gas
temperature ( 8F) ,
T= the difference in stack gas tempera-
ture and ambient temperature (8F ), and d= the stack
diameter ( m ) .
The effective stack height thus takes into account plume
rise height as a result of the exit velocity and temperature of
the effluent gas. Uncertainties in the effective stack height
were propagated by sampling from a normal ambient
temperature distribution ( mean = 658, standard devia-
tion = 108) , and the meteorological wind speed distribution.
Exposure Calculations
For this study, estimates of ground -level concentrations of
SO2 at a receptor were obtained by summing the ground -
level concentrations estimated from Equation 1 over each of
the 30 SO2 point sources. The source term for the
deterministic model with plume rise used the reported
annual SO2 facility releases ( Table 1 ), an ambient
temperature of 658, and a wind speed of 1 km /s for
calculating the effective stack height. The stochastic model
sampled from the source term distributions described above
for the release points, incorporating the uncertainty in the
effective stack height discussed above. The program for the
exposure model was written in a Visual Basic module and
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(6)
Rogers et al.
executed on an Excel5 spread sheet. Stochastic sampling
was achieved with the software Crystal Ball6 to estimate the
exposure distributions.
Model Bias
To test the accuracy of the environmental transport model,
we compared the model -predicted concentrations with the
environmentally monitored (observed ) ambient concentra-
tions. When the annual source term ( SO2 release data ) from
the industrial facilities is known, model performance can be
validated by comparing how well the model predicts a
downwind concentration at a receptor to the observed
ambient monitoring data at that receptor. In order to validate
the model used in this study we obtained ambient
monitoring data (see Ambient Monitoring Data above )
for the five locations within the study area that recorded
ambient SO2 levels.
There are several ways to compare predicted with
observed SO2 concentrations. One simple way is to examine
the trend of concentrations over distances. However, it is
useful to quantify the difference between the model -
predicted and measured concentrations. Bias can be
expressed in terms of a predicted -to - observed ratio (P / O
ratio ) where:
P /O = 1 exact agreement,
P /O > 1 overprediction,
P /O < 1 underprediction.
When a number of locations are being examined, the
geometric bias is used. The geometric bias is the geometric
mean of the individual P /O ratios, given by
0P n 1
ln…Pi =Oi †
Biˆ1 C
Geometric bias ˆ expB @
C;
A …3†
n
where P i = predicted concentration at location i, O i =ob-
served concentration at location i, and n = the number of
locations being compared.
Results
Model Performance
Table 2a ( physical stack heights, h ) and b (effective stack
height ( plume rise ) , h e ) show both the measured
(observed ) and model - predicted SO2 concentrations at
the GAEPD ambient monitoring stations obtained from the
deterministic model. The predicted values are the sums of
the model -predicted downwind SO2 concentrations over
the 30 point sources. The effect of using plume rise to
establish the effective stack height, h e, is to move the model
5
Excel97, Microsoft Corporation.
6
Crystal Ball, Decisioneering, 1515 Araphahoe Street, Denver, Colorado.
539
Rogers et al. Estimating environmental exposures to SO2

Table 2. Observed ( measured ) and predicted ( estimated ) SO2 concen- per year ) located in Chatham County, estimated from the
trations at ambient monitoring sites with predicted - to - observed ratios,
deterministic model. The closest monitoring station at 3218
geometric bias, and correlation coefficient, r, for models with and without
plume rise ( deterministic model ) . m, Savannah Water Works, was predicted to receive
approximately two- thirds of GAEPD measured atmos-
Observed Predicted P/O pheric SO2 concentration from this single release site. The
g / m3 g / m3 other ambient monitoring stations within the city of
a. Model without plume rise a Savannah and Brunswick receive increasingly smaller
Savannah Water Works 17.40 50.15 2.88 predictions of SO2 concentrations from this single SO2
Savannah Farmers Market 15.70 116.33 7.71 source as distance from the release point increased.
Savannah Electric 16.31 51.68 3.17
Savannah E. President St. 13.07 49.73 3.81 Exposure Estimates
Brunswick, GA 5.23 14.91 2.85 The exposure estimates for the cases gave a lognormal
distribution with a mean of 7.64 and a standard deviation of
b. Model with plume rise ( effective stack height ) b 10.87 while the exposure estimates for the controls resulted
Savannah Water Works 17.40 18.75 1.08 in a lognormal distribution with a mean of 5.71 and a
Savannah Farmers Market 15.70 23.04 1.47 standard deviation of 7.92. Figure 2 shows the percentile
Savannah Electric 16.31 12.95 0.79 distribution for the case and control residential SO2. The 5th
Savannah E. President St. 13.07 11.44 0.88 percentile of the two groups were essentially the same, 0.88
Brunswick, GA 5.23 6.15 1.18 versus 0.89 g/ m3. The percentile estimates for the case and
a
Geometric bias = 3.74 ( r = 0.29 ) . control residences begin to separate starting at the 25th
b
Geometric bias = 1.05 ( r = 0.78 ) . percentile. Figure 3 shows the uncertainty in SO2 exposure
for the cases and controls from their exposure distributions.

predictions toward the GAEPD observed values (Table

2b ). The estimated bias of the model is 3.74 when physical
stack height is used in the transport model and 1.05 when
the effective stack height is used in the transport model.
Therefore, effective stack height was used in calculating the
exposure estimates described below.
Table 3 shows the uncertainty in the ambient monitor
estimates generated by the stochastic model and the
distribution in the model bias when estimating ground -
level concentrations at the monitoring stations for the 30
release sites. After 1000 simulations the model bias was
shown to fluctuate between a 5th percentile of 0.47 and a
95th percentile of 1.99, with the median estimate being
0.99.
Table 4 shows the ground -level SO2 concentrations at
the GAEPD ambient monitoring stations within the city of
Savannah and the monitoring station at Brunswick
attributable to the largest point source (10,762 tons SO2
Discussion
Model Accuracy
Hanna et al. ( 1977 ) described the Gaussian transport model
as ``the basic workhorse'' for dispersion calculations
because of the model's good agreement with experimental
data, ease of operation, conceptional appeal, and reflection
of the random nature of atmospheric turbulence. The
ground -level concentration estimates from the Gaussian-
plume model are responses to a number of clearly visible
parameters. The predicted concentration at a receptor from a
source is directly proportional to the release rate at the
source ( Equation 1) . Distance from the source to the
receptor also plays a major role in the exposure estimate. As
seen in Figure 4, the exposure estimate declines approxi-
mately logarithmically as distance from the source in-
creases. Meteorological parameters such as wind speed,

Table 3. Uncertainty distribution ( percentiles ) of ground - level concentrations of SO2 ( g / m3 ) at the ambient monitoring stations with predicted - to -
observed ratios ( P / O ) , and percentile geometric bias ( stochastic model ) .

Monitor location 5th 25th 50th 75th 95th

Water Works 8.05 ( 0.46 ) 12.77 ( 0.73 ) 17.49 ( 1.01 ) 23.40 ( 1.34 ) 35.84 ( 2.06 )
Farmers Market 10.10 ( 0.64 ) 15.91 ( 1.01 ) 21.58 ( 1.37 ) 28.61 ( 1.82 ) 43.26 ( 2.76 )
Savannah Electric 5.78 ( 0.35 ) 9.23 ( 0.57 ) 12.43 ( 0.76 ) 16.39 ( 1.00 ) 24.90 ( 1.53 )
Savannah East President 5.07 ( 0.39 ) 8.14 ( 0.62 ) 10.94 ( 0.84 ) 14.47 ( 1.11 ) 21.72 ( 1.66 )
Brunswick, GA 2.82 ( 0.54 ) 4.32 ( 0.83 ) 6.05 ( 1.10 ) 7.80 ( 1.49 ) 11.87 ( 2.27 )

Geometric bias 0.47 0.74 0.99 1.32 1.99

540 Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(6)

Estimating environmental exposures to SO2 Rogers et al.

Table 4. Contribution of the largest point source to the ambient monitoring stations. This table reflects the effect of distance ( meters ) on SO2 concentration
at ambient monitoring sites from the largest SO2 emission site ( deterministic model ) .

Source North East Height Monitor location North East Distance Direction g / m3
( tons ) ( UTM ) ( UTM ) (m) ( UTM ) ( UTM ) (m)
10762 3556794 486090 90 Water Works 3556814 482872 3218 SE 10.53
Farmers Market 3550696 485748 6109 SE 5.74
Savannah Electric 3546000 488600 11083 SE 2.54
Savannah East President 3547850 495375 12893 SE 2.08
Brunswick, GA 3446292 454522 114944 SSW 0.18

direction, and atmospheric stability class ( A ±F ) also affect
model estimates. For example, two locations equidistant
from a source but lying in different directional sectors may
experience different concentrations because of a combina-
tion of different average wind speeds, directional frequen-
cies, and stability classes for the two sectors.
Miller and Hively (1987 ) investigated the uncertainty of
the Gaussian -plume model by summarizing ratios of model
predictions with corresponding experimental observations
( P /O ratios ) from numerous studies. Their data indicated
that for annual average predictions for landscapes with flat
terrain and receptors located within 10 km of the source, the
uncertainty attributed to the Gaussian plume model was
approximately a factor of 2 ( i.e., P / O is between 1/2 and 2
with probability 0.90 ). This model uncertainty increased to
a factor of 4 when the distance exceeded 10 km (up to 150
km ) .
Based on a small number of studies for complex terrain
or coastal meteorology (the latter is the case for the region
under study ), the Gaussian -plume estimate of uncertainty
for long -term releases ranged from 0.1 to 10, or two orders
of magnitude. Because of the small variation in the P / O
ratios (Table 2b ), approximately 30%, obtained from local
monitoring data, and the fact that our study area was
basically flat terrain, we considered the factor of 10 unduly
pessimistic for this work. Our deterministic model appeared
to behave more like that seen in flat terrain validation
studies. We therefore assumed that the uncertainty that
could be attributable to the Gaussian -plume model in this
study would approximate that seen in flat terrain, or a factor
of 2.
The height of the release of a source also has a strong
impact on exposure estimates as shown in Table 2a and b.
When the physical stack height was modeled as the release
point of emissions (Table 2a) , predicted -to -observed
ratios at the ambient monitoring stations ranged from
approximately 3 to 7 with a model bias of 3.74 ( r= 0.29 ).
However, when plume rise was incorporated into the model
(Table 2b ), the predicted -to -observed ratios at the ambient
monitoring stations ranged from 0.79 to 1.47, with a model
bias of only 1.05 (r =0.78 ). Plume rise effectively raises the
release point of stack emissions, and the higher the release
point, the smaller the ground -level concentration at the
ambient monitoring sites. Data such as stack gas tempera-
ture, gas velocities, and diameters also affect the release
point as calculated by Equation 2. This additional informa-
tion improved the accuracy of the model in predicting SO2
concentrations at the known ambient monitoring stations.
Thus our stochastic model with a median geometric bias of
0.99 (Table 3 ) would be considered a good predictor of
SO2 concentrations at geographically known but unmoni-
tored locations.

Figure 2. Percentile distribution of estimated SO2 exposures for cases Figure 3. Uncertainty distribution of SO2 exposures for cases and
and controls. controls.

Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(6) 541

Rogers et al.
Figure 4. Logarithmic decline of exposure estimates with distance.
Exposure Estimate
The uncertainty distribution of exposures presented in
Figure 3 represents a broad range of estimated ground - level
SO2 concentrations that may be seen over a 24- county area
in south Georgia as a result of industrial emissions. Whereas
the annual averages of ambient SO2 from GAEPD
( observed ) data ranged from 5.23 to 17.4 g/ m3 (Table
2 ), with an average of approximately 14 g / m3, the
variation in the stochastic model predicted exposures for
both the case and control subjects ( Figure 3) , demonstrate
how misclassification of exposure occurs in an ecologic
study.
The exposure distribution shown in Figure 3 represent an
environmental transport model - predicted estimate of an
external exposure at a specific location, a study subject's
residence. However, the use of such modeling techniques is
usually only a first step in exposure assessment for
epidemiologic studies. Environmental pollutants to which
the individual is exposed may take many pathways, and
estimating an individual's exposure requires assessing other
potential pathways through which that person might be
exposed. For example, in the study of exposures to
radionuclides, geographic and lifestyle parameters such as
time spent indoors versus time spent outdoors, time spent at
specific locations such as work or home or school,
migration into and out of a study area, farming practices,
milk consumption, consumption of garden - fresh or canned
foods, and decay of the radionuclide all influence the
exposure estimate (Lloyd et al., 1990; Stevens et al., 1990 ).
Multipollutants/ Surrogates
In studying exposures from industrial releases, researchers
must discriminate between the effects of multiple pollutants.
The industrial -emissions data obtained that we obtained
from GAEPD included information on volatile organic
compounds (VOCs), nitrogen dioxide ( NO2 ) , carbon
monoxide (CO ), and total particulate matter (TSP ), as
well as information on SO2. To determine a pollutant's
effect upon an outcome, one needs to investigate the effects
542
Estimating environmental exposures to SO2
of other pollutants that are a part of the same continuum of
exposures. For example, because particulate matter has been
associated with adverse health outcomes (Abbey et al.,
1993; Ito et al., 1993; PoÈnkaÈ and Virtanen, 1996) , the
researcher must consider the potential of one pollutant
acting as a surrogate for another pollutant which has a direct
impact upon the health of individuals.
Lippmann and Thurston (1996 ) demonstrated that the
quantity of total particulate matter (TSP ) in the atmosphere
is closely correlated with the sulfate ion (SO4 2 ) and
suggested that since TSP has a major impact on public
health and quality of life, SO4 2 should be monitored as a
metric for ambient TSP. Moreover, in studies in which
associations between an exposure and outcome are
significant, investigators may question the association as
one representing that of a surrogate exposure. This problem
was seen by Abbey et al. ( 1993 ), when, using ambient data,
they found a significant association between SO4 2 and
airway obstructive disease ( AOD ). However, when the
association between TSP and AOD, and SO4 and AOD were
tested together, only TSP was found to be associated with
AOD. Abbey et al. ( 1993 ) concluded that SO4 2 was a
surrogate pollutant that whose apparent effect was actually
attributable to TSP. These kinds of problems involving
multiple exposures seen by Abbey et al. ( 1993 ) and
Lippmann and Thurston ( 1996 ) could be resolved with
better descriptions of an individual's exposure to multiple
pollutants.
Model Limitations
There are five major limitations to the model estimates
developed for this study. First, the exposure estimates are
based on annual source term data ( i.e., they assume that
releases are constant over a year ) , and annual meteo-
rological data. We were not able to consider peak daily
emission releases and daily changes in meteorology that
may have increased a subject's chance for higher or lower
exposure. We did, however, attempt to incorporate some
uncertainty in our model that would express some
extremes in exposures. In addition, we are not able to
define a critical time of exposure. Such information about
time of exposure would strengthen the association
between exposure and outcome in an epidemiologic study
(Rothman and Poole, 1988 ).
Second, factors such as chemical conversions of SO2 to
SO4 2 and wet and dry deposition for chemicals such as
SO2, which generally take a few days ( Hanna et al., 1982 ),
were not modeled. Atmospheric transport models that
include chemical conversions and deposition of SO2 have
been written and tested (Andresen and Tha Paw, 1985 ) and
could be used with enhanced data. However, as shown
above, the model we used proved to be a good estimator of
SO2 concentrations at the ambient monitoring stations
without these additional parameters.
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(6)
Estimating environmental exposures to SO2
Third, parameters that were not modeled in this study
may also account for some uncertainty in the model. For
example, the model bias at the 50th percentile ranged about
30% ( Table 3 ). Part of the discrepancy between the
predicted and observed concentrations may be due to the
effect of city buildings on wind patterns Savannah. Our
model does not account for changes in dispersion patterns as
the plume moves alternately through urban and rural
regions. Neither does the model consider building wake
effects and downwash (Hanna et al., 1982 ). However, these
phenomena are of greater concern for predictions near a
source than for locations a few kilometers away. It would
have been advantageous to have more ambient monitoring
data from more than five ambient monitoring sites in the
study area. Additional monitoring data could have provided
a more complete evaluation of our model performance, but
we were limited to the number of available SO2 monitoring
sites located within the study area.
Fourth, the summary emissions data obtained from the
GAEPD suggest that emissions from the monitored point
sources in south Georgia have not fluctuated much in the
last decade. These source - term data, however, may contain
some error that would directly affect the modeling exposure
estimates. We have attempted to account for this potential
source of error by incorporating an uncertainty distribution
in the reported release estimate. Our source terms also are
based on data recorded for environmental monitoring
purposes. Thus, we have no information on SO2 entering
the environment from sources that are exempt from
maintaining data on emissions. Additional SO2 from
background sources such as automobiles and small
industries would presumably add to ambient SO2 concen-
trations and would, if incorporated in our model, increase
our estimates of SO2 concentrations to which residents were
exposed. However, SO2 concentrations that the model
predicts at the monitoring stations closely approximates
what is being recorded at these stations, and the monitored
concentrations should include any background contribu-
tions of SO2.
Fifth, the meteorological data used in this study come
from the only meteorological recording station in the area.
This limited set of meteorological data has been used to
portray the weather patterns of a large area of south Georgia.
There are no available monitoring data to test the accuracy
of the model for the inland areas of south Georgia. Also, the
Gaussian plume model is known to have less predictive
capability in coastal regions ( Miller and Hively, 1987 ) than
in inland areas with flat terrain. However, the stochastic
model predictions of all available monitoring locations, with
a geometric bias of 0.99, suggests that factor-of -ten
uncertainties generally applicable to a Gaussian plume's
usage in coastal regions might be overly pessimistic in this
case, and within this resolution we believe the model should
be considered reasonably accurate for the present domain of
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(6)
Rogers et al.
application, which includes Savannah and Brunswick,
Georgia.
Statistical Questions
The problem addressed here concerns the correlation matrix
that should be assigned to the Gaussian plume uncertainty
component when applied to multiple exposure locations.
The question is not of primary importance for the
predictions we have examined in this paper, but it could
play a major role in studies where uncertainty distributions
involving multiple locations are crucial to estimates
expected numbers of excess cases, given an exposure, or
test statistics.
Correlations among model uncertainties at different
locations are induced by the distributions of release rates
and plume rise parameters. These correlations will be
affected by what is assumed about the joint distribution of
the model uncertainty. Intuition suggests that model
uncertainties for nearby residences (e.g., next door ) should
be highly correlated. Therefore the model will tend to be in
error by about the same amount at the two locations when
estimating annual average exposures. Those correlations for
residences at great distances from each other may ( at least
in some circumstances ) be uncorrelated and the model error
differ for the two locations. Unfortunately we know of no
convincing quantification of such a correlation pattern.
However, it is possible to perform exposure calculations
twice to answer the above question by assuming first that all
pairs of model uncertainties have correlation coefficient
= 1 ( highly correlated hypothesis ) and second that all
pairs of model uncertainties are uncorrelated ( uncorrelated
hypothesis ). Sums of exposures have a larger variance
under the highly correlated hypothesis than under the
uncorrelated hypothesis ( Wilks, 1962 ). Presumably, it may
occur that under one hypothesis a test statistic gives
significant results, but fails to do so under the other
hypothesis.
The handling of correlations is an unresolved question
that requires attention for most environmental assessments
involving airborne transport of pollutants. If the structure of
the correlation matrix were known, and if the number of
receptors is moderately large, the size of the correlation
matrix can be a barrier to the Monte Carlo calculations. For
example, to implement the desired simulations for 1000
locations, a matrix of 500,000 pairs of correlated samples
must be generated. However, given the structure of the
matrix and the nature of the statistical estimates and tests to
be carried out, there may be approximations that can give
some information, if not everything the investigator might
desire. For example, if the purpose is to estimate numbers of
health effects as being proportional to collective exposure,
then the variable of interest is the sum of exposure over all
locations. Even if the correlation matrix is unknown, it is
possible to show that the coefficient of variation of the sum
543
Rogers et al.
cannot exceed the largest coefficient of variation of a term
( i.e., location ). It might be possible to determine an upper
bound for the coefficient of variation for the locations
without developing distributions for all of them. Thus one
would be able to give a conservative estimate of the variance
of the estimated total of health effects, together with a
deterministically estimated mean.
Summary
In this paper we have presented a method for estimating an
individual environmental exposure, with uncertainties, that
would be appropriate for a case ± control epidemiologic
study. By using the atmospheric transport model above,
researchers can approximate the level of environmental
pollutants at an individual's residence and thus assign
individual exposures to study subjects. As we have
demonstrated here, environmental exposure estimates
obtained in this manner, although not the true measure of
individual exposure, will be considerably more accurate
than exposures bases upon ambient measures.
It would have been advantageous to have more than five
ambient monitoring stations to evaluate our model perfor-
mance. However, the uncertainty in the exposure estimates
at the five ambient monitoring stations (Table 3 ) provides
the reader with some feeling about the information that can
be gained in estimating exposures using an environmental
transport model.
Using an atmospheric transport model of pollutants to
estimate individual exposures can substantially reduce the
problems of nondifferential exposure misclassification bias
seen in ecological studies. With these methods, researchers
can also overcome a weakness common to all ecologic
studies discussed above, and identify a study subject as
exposed or unexposed. If the environmental transport model
is reasonably accurate in predicting an individual exposure,
then associations between environmental exposures and
outcomes should be easier to estimate and more precise
( Hatch and Thomas, 1993 ) . Furthermore, by combining
exposure estimates obtained from a validated transport
model with geographic and lifestyle parameters, researchers
can improve the accuracy of individual exposure assessment
used in studies of the relationships between environmental
exposures and adverse health outcomes.
Acknowledgments
This study was supported in part by funds from the
Comprehensive Environmental Response, Compensation
and Liability Act Trust Fund, through an Interagency
Agreement between the Agency for Toxic Substances and
Disease Registry and the Centers for Disease Control and
Prevention, National Center for Environmental Health.
544
Estimating environmental exposures to SO2
References
Abbey D.E., Peterson F.F., Mills P.K., and Beeson W.L. Long - term
ambient concentrations of total suspended particulates, ozone, sulfur
dioxide and respiratory symptoms in a non smoking population. Arch.
Environ. Health 1993: 48 ( 1 ) : 33 ± 46.
Abbey D.E., Peterson F.F., Mills P.K., Beeson W.L., and Kittle L. Chronic
respiratory disease associated with long - term ambient concentrations
of fine particulates less than 2.5 microns in aerodynamic diameter and
other pollutants. J. Expo. Anal. Environ. Epidemiol. 1995: 5 ( 2 ) :
137 ± 159.
Andresen J., and Tha Paw U.K. Modeling sulfur dioxide emissions and
acidic precipitation at mesoscale distances. J. Air Pollut. Control
Assoc. 1985: 35 ( 11 ) : 1159 ± 1163.
Ballester G., Corella D., Perez - Hoyos S., and Hervas A. Air pollution and
mortality in Valencia, Spain: a study using the APHEA methodology.
J. Epidemiol. Community Health 1996: 50: 527 ± 533.
Brenner H., Savitz D.A., JoÈckel K., and Greenland S. Effects of
nondifferential exposure misclassification in ecologic studies. Am. J.
Epidemiol. 1992: 135 ( 1 ) : 85 ± 95.
Brown L.M., Pottern L.M., and Blot W.J. Lung cancer in relation to
environmental pollutants emitted from industrial sources. Environ. Res.
1984: 34: 250 ± 261.
Geschwind S.A., Stolwijk J.A., Bracken M., Fitzgerald E.F., Stark A.,
Oslen C., and Melius J. Risk of congenital malformations associated
with proximity to hazardous waste sites. Am. J. Epidemiol. 1992: 135:
1197 ± 1207.
Hanna S.R., Briggs G.A., Deardorf J., Eagan B.A., Gifford F.A., and
Pasquill F. American Meteorological Society workshop of stability
classification schemes and sigma curves: summary of recommenda-
tions. Bull. Am. Meteorol. Soc. 1977: 58 ( 12 ) : 1305 ± 1309.
Hanna S.R., Briggs G.A., and Hosker R.P. ( Eds. ) National Technical
Information Service handbook on atmospheric diffusion, U.S. Depart-
ment of Commerce, Springfield, VA, 1982.
Hatch M.C., and Thomas D. Measurement issues in environmental
epidemiology. Environ. Health Perspect. 1993: 101 ( suppl 1 ) : 49 ±
56.
Hatch M.C., Beyea J., Nieves J.W., and Susser M. Cancer near the Three
Mile Island nuclear plant: radiation emissions. Am. J. Epidemiol. 1990:
132 ( 3 ) : 397 ± 412.
Holland J.Z. A meteorological survey of the Oak Ridge area: final report
covering the period 1948 ± 1952. USAEC / report ORO - 99, Weather
Bureau, Oak Ridge, TN, 1953.
Ito K., Thurston G.D., Hayes C., and Lippman M. Associations of London,
England, daily mortality with particulate matter, sulfur dioxide, and
acidic aerosol pollution. Arch. Environ. Health 1993: 48 ( 4 ) : 213 ±
220.
Lippmann M., and Thurston G.D. Sulfate concentrations as an indicator of
ambient particulate matter for health risk evaluations. J. Expo. Anal.
Environ. Epidemiol. 1996: 6 ( 2 ) : 123 ± 146.
Marshall E.G., Geary N.S., Cayo M.R., and Lauridsen P.A. Residential
exposure summary methodology for a reproductive health study of
multiple hazardous waste sites. J. Expo. Anal. Environ. Epidemiol.
1993: 3 ( suppl 1 ) : 87 ± 97.
Meade P.J., and Pasquill F. A study of the average distribution of pollution
around Staythorpe. Int. J. Air Pollut. 1958: 1: 60 ± 70.
Miller C.W., and Hively L.M. A review of validation studies for the
Gaussian plume atmospheric dispersion model. Nucl. Saf. 1987: 28
( 4 ) : 522 ± 531.
Morgenstern H. Uses of ecologic analysis in epidemiologic research. Am.
J. Public Health 1982: 72: 1336 ± 1344.
Morgenstern H., and Thomas D. Principles of study design in
environmental epidemiology. Environ. Health Perspect. 1993: ( suppl
4 ) : 23 ± 38.
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(6)
Estimating environmental exposures to SO2
National Air Quality and Emissions Trends Report ( NAQETR ) , EPA -
454 / R - 96 - 005. October 1996, 1995.
Lloyd R.D., Gren D.C., Simon S.L., Wrenn M.E., Hawthorne H.A., Lotz
T.M., Stevens W., and Till J.E. Individual external exposures from
Nevada Test Site fallout for Utah leukemia cases and controls. Health
Phys. 1990: 59 ( 5 ) : 723 ± 737.
Petruzzi S., Musim B., and Bignami G. Acute and chronic sulfur dioxide
( SO2 ) exposure: an overview of its effects on humans and laboratory
animals. [ Review ] . Ann. Ist. Super. Sanita 1994: 30: 151 ± 156.
PoÈnkaÈ A., and Virtanen M. Chronic bronchitis, emphysema, and low level
air pollution in Helsinki, 1987 ± 1989. Environ. Res. 1994: 6: 207 ±
217.
PoÈnkaÈ A., and Virtanen M. Asthma and ambient air pollution in Helsinki.
J. Epidemiol. Community Health 1996: 50 ( suppl 1 ) : S59 ± S62.
Rothman K., and Poole C. A strengthening program for a weak
association. Int. J. Epidemiol. 1988: 17 ( suppl ) : 955 ± 959.
Sanders G.L. Toxicological aspects of energy production. Battelle Press,
Columbus, Ohio, 1986, p. 116.
Shaw G.M., Schulman J., Frisch J.D., Cummins S.K., and Harris J.A.
Congenital malformations and birthweight in areas with potential
Journal of Exposure Analysis and Environmental Epidemiology (1999) 9(6)
Rogers et al.
environmental contamination. Arch. Environ. Health 1992: 47 ( 2 ) :
147 ± 154.
Shprentz D.S., Bryner G.C., Shprentz J.S., and Hawkins D.G. National
Resources Defense Council, Breath taking: premature mortality due to
particulate air pollution in 239 American cities. National Resources
Defense Council, Washington, DC, 1996, p. 99.
Stevens W., Thomas D.C., Lyon J.L., Till J., Kerber R.A., Simon S.L.,
Lloyd F.D., Elghany N.A., and Preston - Martin S. Leukemia in Utah
and radioactive fallout from the Nevada Test Site: a case ± control
study. JAMA 1990: 264: 585 ± 591.
Vesper B.E. Estimating the uncertainty in dose calculation methods for
accidental atmospheric releases. Radiat. Prot. Manage. 1992: 9 ( 4 ) :
69 ± 79.
Wilks S.S. Mathematical Statistics. Wiley, New York, 1962.
Wojtyniak B., and Piekarski T. Short term effect of air pollution on
mortality in Polish urban populations: what is the difference? J.
Epidemiol. Community Health 1996: 50 ( suppl ) : S36 ± S41.
Xu X., Ding H., and Xiaobin W. Acute effects of total suspended particles
and sulfur dioxide on preterm delivery: a community - based cohort
study. Arch. Environ. Health 1995: 50 ( 6 ) : 407 ± 415.
545