Transcript Document

GeoSpatial Exposure Modeling in Ecological Risk Assessment: Whitewood Creek Site
William Thayer1, Dale Hoff2, Philip Goodrum1, Janet Burris3, Lynn Woodbury3
Environmental Science Center - Syracuse Research Corp., North Syracuse1, and Denver, CO3
and U.S. EPA Region 8, Denver, CO2
Introduction
A major source of uncertainty in risk assessment is often the exposure point concentration
(EPC): the chemical concentration to which a receptor may be exposed over a toxicologically relevant time
period within a geographic area called an exposure unit (EU). For terrestrial ecological risk assessments,
important factors in defining the EPC include: 1) the spatial distribution of concentrations within the EU; and 2)
the movement of the receptor. Biased sampling methods are often employed during site characterization to
identify potential hotspots. This, along with the assumption that concentrations are lognormally distributed, may
contribute to overly conservative estimates of the EPC. In addition, the assumption of equal and random access
to all areas of the site may not be appropriate, especially if a receptor’s home range is smaller than the site. By
using the spatial information present in the sample data, geostatistics can provide a more reliable measure of
uncertainty in the EPC. We developed the GeoSpatial Exposure Model (GeoSEM™), a software tool that
employs different geostatistical methods (Thiessen Polygons, Kriging, Sequential Gaussian Simulation), in a
GIS-based application, within a Microsoft Windows® environment. This poster illustrates an application of
GeoSEM™ to an ecological risk assessment for shrews exposed to arsenic in soil at the Whitewood Creek site
in Lead, SD. An individual’s movement is simulated as a random walk over a specified foraging area (Hope,
2000). The choice of geospatial method used to estimate concentrations at explicit locations is shown to be a
major source of model uncertainty. [Additional documentation is available at http://esc.syrres.com/geosem/.]
Risk Equation
Application of GeoSEM™ to the Whitewood
Creek Site
An ecological risk assessment was completed as part of a
required five year review of the record of decision for the
Whitewood Creek Superfund Site in Lead, South Dakota. Sitespecific data were collected to investigate arsenic concentrations
in soil, earthworms, plants (grasses and clover), grasshoppers,
aquatic invertebrates, fish, and small mammals. In addition, field
studies were implemented to collect site-specific data on
bioavailability and toxicity. Masked shrew (Sorex cinereus) was
chosen as the representative receptor to assess exposures and
risk to small insectivorous mammals, and to illustrate the effect of
assumptions regarding habitat area and foraging patterns. The
shrew has a relatively high metabolic rate and, thus, a larger
foraging area than other potential surrogate species. Home
ranges for shrews may range from 0.5 to 3.5 acres (Choate and
Fleharty, 1973).
For this analysis, a home range of
approximately 1 acre (0.39 ha = 3885 sq meters) was assumed.
Risks are expressed as a hazard quotient based on exposure via
ingestion of soil and invertebrates. The assessment endpoint is growth and survival of the
shrew population. An individual-based exposure modeling approach was used. The TRV
is based on a NOAEL of 0.12 (rather than the LOAEL of 0.36).
The concentration of lead in soil can be estimated at unsampled locations by employing
a variety of geostatistical methods. Figure 2 illustrates screen captures from GeoSEM™ for Ordinary Kriging.
Csoil  AUF  IR  AF
HQ 
BW  TRV
Table 1. Point estimate input values used to calculate HQ.
HQ
=
hazard quotient
Csoil
=
arsenic concentration in soil
mg/kg
AUF
=
1.0
IR
=
area use factor (portion of home range
that overlaps with impacted area)
ingestion rate (food plus soil)
AF
=
absorption fraction
1.0
BW
=
body weight
5.3 grams
TRV
=
toxicity reference value for arsenic
0.4 grams/day
0.12 mg/kg-day
Other exposure pathways (water, other dietary sources, inhalation)
were assumed to be relatively minor for this scenario.
Table 2. Arsenic concentrations in surface soil. Weighted
estimates are based on Thiessen Polygons.
Sample Statistic
Un-weighted
Weighted
Arsenic (ppm)
(Fig. 2)
Thiessen Polygons
Figure 1. Location of Whitewood Creek in
Lead, SD.
Geospatial Methods
n
mean
810
243
810
342
median
94
100
Standard deviation
455
614
minimum
1
1
maximum
6228
6228
Figure 2. Overview of Zone 1 showing sampling
locations (circles) and the spatial distribution of
predicted (kriged) arsenic concentrations in soil.
Higher concentrations are located nearer the flood
plain.
Receptor Movement
Figure 3. Examples of foraging areas simulated for
individual receptors. Results presented in this poster
reflect simulations for a population size of n=200
shrews, each with approximately 1 acre home ranges
of irregular shapes.
Exposures to an individual receptor (shrew) are modeled by choosing a random
starting point within the site and continuing a random walk until the foraging area is achieved. An example of
random walks for individual receptors is illustrated in Figure 3. At each location, an arsenic concentration is either
estimated based on the kriged surface or a measured concentration is used if a sample location coincides with the
receptor location. For each individual, the EPC can be estimated from the mean concentration encountered during
the random walk. Collectively, the simulations yield a measure of inter-individual variability in the EPC among a
potentially exposed population. Alternative scenarios could be simulated to reflect habitat suitability (Hope, 2000).
Kriging Results – Variability or Uncertainty in EPC?
A clear distinction needs to be made between modeling variability and
modeling uncertainty in the EPC in exposure assessments. When
estimating risks to a population, the statistic of interest is typically a
measure of the mean, or uncertainty in the mean concentration within
the EU. Figure 4 illustrates variability in the EPC, rather than
uncertainty in the EPC. Essentially, kriging produces a single map of
the “best” local estimates (using mean squared prediction error as the
criterion), and the estimates are “smoothed” such that low values will
tend to be overestimated and high values will tend to be
underestimated. The effect of smoothing is greatest in areas furthest
from sample locations. Deleterious consequences for risk assessment
include: 1) underestimate exposures by failing to reproduce areas of
extreme high concentrations; 2) provides unreliable estimates of the
probability that a given number of EUs exceed a risk-based action level
or soil concentration.
Prob
sample mean
= 243 ppm
0.30
0.20
0.10
0.00
0
200
400
600
800
1000
1200
1. Model uncertainty by generating a set of R
realizations (or maps) of the spatial
distribution of arsenic concentrations. For
this analysis, we generated 100.
2. Each realization is conditional to the original
data and approximately reproduces the
spatially
weighted
sample
frequency
distribution and the spatial autocorrelation
structure (i.e., variogram). Therefore, the set
of mean concentrations generated by SGS
will be similar to that of kriging (Figure 2).
3. The SGS algorithm used here assumes the
data are normally distributed, although other
probability models (including nonparametric)
could be selected.
4. For each individual receptor, generate a
unique pattern of movement within the site
(see Figure 3) and calculate a set of
exposure concentrations (yielding a mean
concentration).
Repeat the estimates of
concentration (not the random movement),
thereby yielding a distribution of mean
concentrations for each individual receptor.
n = 200
25th %ile = 85
mean = 267
90th %ile = 591
max = 1881
0.40
Sequential Gaussian Simulation (SGS)
1400
EPC (ppm) for Individual Receptors
Figure 4. Histogram of average arsenic concentrations
for 200 receptors each exposed to 39 cell locations.
Arsenic concentrations at each cell were estimated
from the kriged surface and measured concentrations
(see Figures 2 and 3).
Should kriging be used to estimate uncertainty in the EPC?
With GeoSEM™, kriging can be applied to exposure units with any shape.
Some software packages employ point or block kriging, where an estimate
for the mean concentration within an EU can be calculated by averaging
the block estimates. However, the kriging variances for the blocks cannot
be simply summed to assess uncertainty in the mean EU concentration
because the estimates for the block variances are not independent
(Journel and Huijbregts, 1978; Burger and Birkenshake, 1994). An
approach that avoids some of the restrictions of kriging is geostatistical
simulation.
5. For each individual receptor, calculate the
distribution of HQ’s that corresponds to the
distribution of EPC’s. Assuming a level of
concern for HQ is 1.0, calculate PA[HQ] (the
probability that HQ exceeds 1).
6. Two outcomes can be gained from this
analysis: a) a proportion of the population
expected to have HQ > 1; and b) a likelihood
that a proportion of the population will have
HQ > 1. How likely is it that more than
10% of the population
adversely affected?
will
be
Conclusions
Comparison of Kriging and SGS Results
For this
analysis, HQ for each simulation is expected to exceed 1 (HQ > 1),
suggesting adverse impacts to insectivorous mammals associated
with the incidental ingestion of soil at the Whitewood Creek Site.
Therefore, we have focused the comparison of the geostatistical
methods on the EPC rather than the risk characterization. Figure 5
gives the results of the SGS simulations for n=200 individual receptors.
The average EPC for each receptor is approximately the same for the
kriging and SGS approaches (compare Figures 4 and 5).
• GeoSEM™ provides a tool for applying a variety of
geostatistical approaches to data in which the exposure
unit can assume any shape and the receptors can move
randomly throughout the site.
Prob
n = 200
25th %ile = 0.10
mean = 0.21
90th %ile = 0.51
max = 0.57
0.40
0.30
0.20
Prob
n = 200
25th %ile = 89
mean = 268
90th %ile = 569
max = 1751
0.40
sample mean
= 243 ppm
0.30
0.10
0.00
0
0.1
0.2
0.3
0.4
0.5
0.6
p(EPC > 500ppm)
0.20
0.10
0.00
0
200
400
600
800
1000
1200
• In this example, all of the measured arsenic
concentrations predict an HQ > 1 for the shrew. Use of
un-weighted or area-weighted (Thiessen polygon)
approaches yield different results for the arithmetic
mean (243 and 342 ppm, respectively). Both
approaches suggest that if the EU is defined as the
entire site, risks to the shrew would be high. Yet, the
results also suggest that geostatistical methods may
improve upon un-weighted estimates given the distance
between sampling transect along the flood basin.
1400
EPC (ppm) for Individual Receptors
Figure 5. Histogram of SGS results showing the distribution
of average EPC for each of 200 individual receptors.
Results for the mean are similar to kriging (Figure 4).
Using SGS, for each individual, 100 realizations were simulated,
yielding 200x100 estimates of EPC. This approach allows for a
unique calculation of the probability that the EPC exceeds a level
of concern (LOC) such as a preliminary remediation goal. Thus,
for each individual, we can obtain an estimate of P(EPC> LOC).
Figure 6 illustrates results for a hypothetical scenario in which the
LOC was set at 500 ppm for arsenic. SGS provides a quantitative
measure of uncertainty in the estimates of exceedence
probabilities based on uncertainty in the soil concentrations within
each individual’s foraging area. For example, there is a 75%
likelihood that 10 percent of the shrew population will contact soils
with concentrations exceeding the LOC of 500 ppm. Similarly,
there is a 10% likelihood that greater than 50 percent of the
population will contact soils exceeding the this LOC.
Figure 6. Histogram of exceedence probabilities for
a hypothetical level of concern (LOC) of 500 ppm for
arsenic in soil. Using SGS, we can evaluate this
criterion for each individual receptor based on the
uncertainty in the soil concentrations within each
foraging area in order to make inferences about
potential population effects.
Variance in Arsenic Concentrations
Variance can be used to estimate uncertainty in
the EPC.
One of the limitations of kriging
variance is that it does not consider the sample
concentrations which are used in the prediction.
It relies solely on the geographic configuration of
the observations.
Non-constant variance
(heteroscedasticity),
together
with
sample
clustering in areas with high concentrations,
results in the overestimation of the variance at
short distances (Goovaerts,1997).
Kriging
variances can be an order of magnitude greater
than those of simulation.
• When the home range is taken into consideration and
an individual-based modeling approach is used, kriging
and SGS yield similar estimates for the mean
concentration among the exposed population. However,
when characterizing uncertainty, the two approaches will
differ significantly due to the difference estimates of
prediction variance.
• Using a hypothetical example of 500 ppm as a level of
concern for arsenic in soil, SGS allows for a quantitative
uncertainty analysis of the EPC. Risk managers could
be provided information about not only the probability of
exceeding an LOC on average, but also the likelihood
that specific fractions of the populations will be exposed
to concentrations exceeding the LOC.
Acknowledgements The authors would like to thank Edzer J.
Pebesma for the use of the gstat program. The development of
GeoSEM™ has been funded by Syracuse Research Corporation.
References
Burger, H. and Birkenshake, F. 1994. Geostatistics and the Polygonal-Method:
A Re-Examination. International Association for Mathematical Geology Annual
Conference, Mount Tremblant, Quebec, Canada, October 3-5, Papers and
Etended Abstracts, pp. 50-55.
Choate, J.R. and Fleharty, E.D. 1973. Habitat preference and spatial relations of
shrews in a mixed grassland in Kansas. Southwestern Naturalist. 18: 110-112.
Goovaerts, P. 1997. Geostatistics for Natural Resources Evaluation. New York:
Oxford University Press.
Hope, B.
2000.
Generating probabilistic spatially-explicit individual and
population exposure estimates for ecological risk assessments. Risk Anal. 20 (5):
573-589.
Journel, A.G. and Huijbregts, C.J. 1978. Mining Geostatistics. Academic Press,
New York.