Available via license: CC BY-NC-ND
Content may be subject to copyright.
(This is a sample cover image for this issue. The actual cover is not yet available at this time.)
This article appeared in a journal published by Elsevier. The attached
copy is furnished to the author for internal non-commercial research
and education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elsevier’s archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/copyright
Author's personal copy
Arsenic in North Carolina: Public Health Implications☆
Alison P. Sanders
a
, Kyle P. Messier
a
, Mina Shehee
b
, Kenneth Rudo
b
, Marc L. Serre
a
, Rebecca C. Fry
a,
⁎
a
Department of Environmental Sciences and Engineering, Gillings School of Global Public Health, University of North Carolina, 1213 Michael Hooker Research Building, Chapel Hill,
NC 27599, United States
b
Medical Evaluation and Response Assessment Unit, North Carolina Department of Health and Human Services, 1912 Mail Service Center, Raleigh, NC 27699, United States
abstractarticle info
Article history:
Received 20 June 2011
Accepted 7 August 2011
Available online xxxx
Keywords:
Arsenic
Drinking water
Well water
Spatial analysis
North Carolina
Carolina slate belt
Arsenic is a known human carcinogen and relevant environmental contaminant in drinking water systems.
We set out to comprehensively examine statewide arsenic trends and identify areas of public health concern.
Specifically, arsenic trends in North Carolina private wells were evaluated over an eleven-year period using
the North Carolina Department of Health and Human Services database for private domestic well waters. We
geocoded over 63,000 domestic well measurements by applying a novel geocoding algorithm and error
validation scheme. Arsenic measurements and geographical coordinates for database entries were mapped
using Geographic Information System techniques. Furthermore, we employed a Bayesian Maximum Entropy
(BME) geostatistical framework, which accounts for geocoding error to better estimate arsenic values across
the state and identify trends for unmonitored locations. Of the approximately 63,000 monitored wells, 7712
showed detectable arsenic concentrations that ranged between 1 and 806 μg/L. Additionally, 1436 well
samples exceeded the EPA drinking water standard. We reveal counties of concern and demonstrate a
historical pattern of elevated arsenic in some counties, particularly those located along the Carolina terrane
(Carolina slate belt). We analyzed these data in the context of populations using private well water and
identify counties for targeted monitoring, such as Stanly and Union Counties. By spatiotemporally mapping
these data, our BME estimate revealed arsenic trends at unmonitored locations within counties and better
predicted well concentrations when compared to the classical kriging method. This study reveals relevant
information on the location of arsenic-contaminated private domestic wells in North Carolina and indicates
potential areas at increased risk for adverse health outcomes.
© 2011 Elsevier Ltd. All rights reserved.
1. Introduction
Ingestion of arsenic through drinking water is implicated in heart
disease, neurological abnormalities, as well as cancers of the skin,
lung, kidney, and bladder (NRC, 2001). The United States Environ-
mental Protection Agency (EPA) regulates arsenic in public drinking
water supplies at a maximum contaminant level (MCL) of 10 μg/L
(EPA, 2010). Although this standard is enforceable in public water
systems via the Safe Drinking Water Act, there is no federal regulatory
standard for domestic well waters. Approximately 14% (about
42 million people) of the U.S. population obtains water from
unregulated private domestic wells (Kenny et al., 2009). It is
estimated that domestic well users in the U.S. carry an excess lifetime
risk of bladder and lung cancer of 66 people per million people, almost
five times higher than that estimated for public well users (Kumar
et al., 2010).
Although arsenic exposure through drinking water is documented
worldwide (Mukherjee et al., 2006), there is a paucity of data in the
U.S. For example, in a survey of U.S. drinking water supplies, many of
the Mid-Atlantic states had insufficient data with less than 10% of
counties represented (Welch et al., 1999). To help fill this void,
regional evaluations of groundwater arsenic in Appalachia (Shiber,
2005), Idaho (Hagan, 2004), Maine (Yang et al., 2009), Michigan (Kim
et al., 2002), Nevada (Shaw et al., 2005; Walker et al., 2006), New
England (Ayotte et al., 2003), and New Hampshire (Peters et al., 1999)
have provided data beyond those collected in nationwide studies and
have demonstrated contamination of drinking sources at spatial scales
finer than the county level. In addition, while the USGS provides
routine ambient well monitoring nationwide, data gathered currently
represent only a small fraction of groundwater sources. Domestic
wells are not often monitored in such nationwide programs and may
more adequately reflect exposure to unregulated contaminated
water.
It is known that areas of North Carolina contain naturally occurring
arsenic deposits including the geological region of Carolina terrane (or
Environment International 38 (2011) 10–16
☆We assess arsenic levels in over 63,000 domestic wells to identify areas where
levels exceed the EPA standard and residents may be at risk.
⁎Corresponding author at: Department of Environmental Sciences and Engineering,
Gillings School of Global Public Health, University of North Carolina, 1213 Michael
Hooker Research Building Chapel Hill, NC 27514, United States. Tel.: +1 919 843 6864;
fax: +1 919 966 7911.
E-mail addresses: apsander@unc.edu (A.P. Sanders), kmessier@email.unc.edu
(K.P. Messier), mina.shehee@dhhs.nc.gov (M. Shehee), ken.rudo@dhhs.nc.gov
(K. Rudo), marc_serre@unc.edu (M.L. Serre), rfry@unc.edu (R.C. Fry).
0160-4120/$ –see front matter © 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.envint.2011.08.005
Contents lists available at SciVerse ScienceDirect
Environment International
journal homepage: www.elsevier.com/locate/envint
Author's personal copy
Carolina slate belt) making it an ideal area for further investigation
and public health efforts (Foley et al., 2001). An initial study by Pippin
(2005) characterized arsenic occurrence in North Carolina ground-
water between 1996 and 2004 and employed classic kriging
techniques to map the probability of detecting elevated arsenic levels
across the state. More recently, Kim et al. (2011), characterized
geologic determinants of arsenic in Orange County, North Carolina.
Our work builds upon earlier characterization of arsenic in North
Carolina wells by developing methods to map historical contamina-
tion for the purposes of protecting public health. Importantly, the
number of individuals in North Carolina currently using domestic
wells for drinking water is estimated at 2.3 million (Kenny et al.,
2009). North Carolina has the fourth-largest state population in the
U.S. using private groundwater wells as a drinking source and is
exceeded only by Michigan, California, and Pennsylvania (Kenny et al.,
2009). Many states, such as North Carolina, still remain understudied
for the presence of arsenic in drinking water at a statewide level. Here,
we present results obtained from a statewide program to monitor
unregulated private domestic wells.
Given the known health risks and occurrence of arsenic in drinking
water, we set out to identify areas of concern and quantitatively assess
concentrations in domestic wells throughout North Carolina. To assess
arsenic trends in monitored private domestic wells, we applied a novel
geocoding scheme and mapped arsenic levels in wells using standard
Geographical Information System (GIS) techniques useful for public
health policy (Bellander et al., 2001; Holton, 2002; Miranda et al., 2002;
Nuckols et al., 2004). To estimate arsenic values at unmonitored
locations, we then applied a novel Bayesian Maximum Entropy (BME)
framework (Christakos, 1990, 2000; De Nazelle et al., 2010; Serre and
Christakos, 1999) to predict arsenic contamination across the state and
to examine areas of interest. These analytical techniques were applied to
more than60,000 domestic well water arsenic measures collected by the
North Carolina Department of Health and Human Services (NCDHHS)
dating back to 1998. In this work, we identify spatial and temporal
arsenic trends in North Carolina domestic wells and indicate specific
locations and populations of concern. Importantly, the geocoding and
geostatistical methods presented here can be applied to track contam-
inant trends in other states. Our results indicate a large number of
contaminated wells in North Carolina and suggest that ongoing
monitoring of well water contaminants is prudent. Moreover, these
data provide new information of specific areas in North Carolina where
targeted well monitoring programs can be used in a cost-effective
manner.
2. Materials and methods
2.1. Data collection
Domestic well water samples were collected by the NCDHHS
Division of Public Health (DPH) State Laboratory of Public Health and
Epidemiology Section which provides groundwater monitoring assistance
to North Carolina homeowners. Following DPH guidelines, local health
department officials collected water from homeowners' indoor, outdoor,
or well tap after allowing the water to run for 5–10 min. Arsenic analyses
were performed by the NCDHHS Laboratory for Environmental Inorganic
Chemistry. Samples were transported to the DPH State Laboratory of
Public Health and analyzed within 48 h as per EPA Method 200.8 Revision
5.4 via inductively coupled plasma mass spectrometry (ICP-MS) with
adherence to formal quality assurance/quality control (QA/QC) protocols
(EPA, 1994). Sample aliquots were acidified with nitric acid to below a pH
of 2.0 for at least 16 h prior to ICP-MS analysis. A 50-mL subsample was
then digested at 95 °C for at least 2 h. For samples with high amounts of
undissolved particulates, the digestate was filtered through a 0.45 μm
filter to prevent damage to the analytical instrumentation. This method
detects for total arsenic; species of As(III) and As(V) were not
differentiated. The NCDHHS detection limit for arsenic had shifted in the
past decade. In early 2000, the detection limit decreased from 10.0 to
1.0 μg/L. More recently, the detection limit was raised to 5.0 μg/L —the
level at which the laboratory presently operates. The laboratory maintains
QC requirements of Fortified Blank recoveries of 95–105% and reagent
blanks must show no contamination. In all lab analyses yttrium, rhodium,
lutetium were included as internal standards to account for instrument
drift and physical interferences. Every QC requirement must be met for the
sample analysis to be approved by the laboratory manager. Analyses were
then entered into an extensive electronic database maintained by the
State Laboratory of Public Health.
2.2. Electronic database formatting
Results from arsenic well water analyses were housed in an
electronic database containing informational fields for arsenic concen-
tration, well location ID, county name, global positioning system (GPS)
location (if available), well owner address, and date collected. We
analyzed 68,836 electronically available well water records of measured
arsenic concentrations collected between October 19, 1998 and
February 25, 2010. Data cleaning of the arsenic database excluded
entries with insufficient information on location or those with
improper/incomplete chemical analysis. We required that entries
included in this study provided the following minimum information:
a county name, a valid sampling date, and an approved laboratory
chemical analysis for arsenic. The resulting comprehensive database of
63,856 well measures was used for all subsequent analyses.
2.3. Descriptive statistics, heat map generation, and county ranking
For all analyses, arsenic measurements below the detection limit
(DL) were treated as 0.5 times the DL (0.5 ×DL). Descriptive statistics
were calculated using Spotfire DecisionSite 8.1 (TIBCO, Palo Alto, CA)
for each of the 100 North Carolina counties. Heat maps to visualize
temporal county trends were prepared using Partek Genomics Suite
6.4 (St. Louis, Missouri). Hierarchical cluster analysis using Euclidean
distance as a measure was used to identify relationships over time
among the top 35 counties exceeding the standard (Eisen et al., 1998).
Furthermore, counties were ranked by 1) the percentage of wells
exceeding the EPA standard over the full time period examined, and
2) the percentage of wells exceeding the EPA standard multiplied by
the percentage of county population using self-supplied domestic
wells (data reported by Kenny et al., 2009). We considered the current
EPA MCL of 10 μg/L as the threshold, although the regulatory standard
(originally 50 μg/L) was lowered over the course of data collection in
this study.
2.4. Four-class geocoding algorithm and error validation scheme
A geocoding algorithm was developed to extract spatial coordinates
and associated location error for the 63,856 private well measurements
contained in the database. A four-class strategy, detailed below, was
used to assign each data entry, l, with a spatial coordinate s
l
=(s
1
,s
2
)
l
,
where s
1
and s
2
were the longitude and latitude best-representing the
well location given the level of recorded information: GPS, street
address, zip code, or county. The first of the four classes (Class I) was
represented by data entries with available GPS coordinates, s
l(GPS)
.For
wells with reported GPS locations, geographical coordinates were
standardized to decimal degrees format. Standardized GPS coordinates
were visualized in ESRI ArcGIS™software Version 9.0 (Redlands, CA).
Well locations were classified as Class I when GPS coordinates were
available and within the longitude/latitude boundaries of North Carolina.
The second class (Class II) assigned geocoded owner address (GOA)
coordinates, s
l(GOA)
, to data entries based on street address. To geocode
the data,we applied a multi-stage geocoding process in which a series of
local and national reference files were used sequentially in order from
most comprehensive spatial detail to least as follows: a North Carolina
11A.P. Sanders et al. / Environment International 38 (2011) 10–16
Author's personal copy
point reference file (courtesy of NCDHHS Spatial Analysis Group),
followed by a North Carolina Department of Transportation line
reference file, then followed by a U.S. Street Address line reference file
(Tele Atlas DynamapTransportation). Resulting from this process,match
scores of 0–100 were associated with each successfully geocoded
address coordinate, s
l(GOA)
. To determine a match score threshold at
which each reference file provided acceptable coordinates, we devel-
oped a method using one-way analysis of variance (ANOVA) to select
acceptable match scores based on the distance between GPS and
geocoded coordinates, d
l
=||s
l(GPS)
−s
l(GOA)
||. We calculated the average
distance, d
l
, for wells with match score of 100 (perfectly matched
addresses) to serve as the referent group. The remaining wells with a
geocoded location were binned into deciles according to match score
(e.g. 99–90, 89–80, 79–70and so on) and the average distance d
l
was also
calculated for each match score decile. We used ANOVA to compare the
average distance d
l
to identify which match score deciles had an average
distance d
l
that was not statistically significantly different (α=0.05)
from the average distance d
l
obtained with a match score of 100. Using
this criterion, point file match scores of 70 and higher and line file match
scores of 60 and above were considered acceptable for describing
geocoded well locations. A data point was classified as Class II when it
was not a Class I, was successfully geocoded by a given reference file with
a match score above the match score threshold, and the county name
included in the owneraddress matched the recorded county of sampling
location. Class II data entries were assigned a single coordinate pair
representing the geocoded owner address coordinates s
l(GOA)
resulting
from the multi-stage geocode process. The third class (Class III) included
zip code centroid coordinates calculated and assigned using ArcGIS™.A
well location was categorized as Class III when it was neither a Class I or
Class II, was inside the North Carolina boundary and county of well
location and a zip code was available to geocode. Class III data entries
were assigned zip code centroid coordinates. The fourth class (Class IV)
included county centroid coordinates for each of the 100 counties in
North Carolina. A well location was considered Class IV when it did not
meet the requirements of any of the previous classes, but a county
centroid coordinate was available. Class IV data entries were assigned
county centroid coordinates. Each entry in the database wascategorized
as one of the four aforementioned classes and the resulting four-class
geocoded data was used in all subsequent analyses. In summary, the
geocoding process assigned one of each of the following four classes to
every arsenic measure in the database: Class I (GPS location), Class II
(street address), Class III (zip code centroid), and Class IV (county
centroid). Wells with assigned geographic coordinates and correspond-
ing arsenic concentration data were visualized using ArcGIS™ESRI
version 9.0 software.
The four-class geocoding scheme enabled maximum incorporation
of spatial information contained in the private well database. Next, we
assessed error associated with each of the four classes to account for
uncertainty introduced by the geocoding process. For Class I, GPS
device positional error resulted from inaccuracies in satellite
triangulation. Positional error associated with GPS instrumentation
found by others was between 5 and 25 m (Hulbert and French, 2001;
Wing and Eklund, 2007). Without rigorous quantification of GPS
device error by the NCDHHS, a conservative Class I error of 50 m was
estimated for the GPS coordinates s
l(GPS)
. For Class II, the location error
of geocoded street address coordinates s
l(GOA)
was considered to be a
combined effect of two error sources: character-matching error as
captured by the match score and inaccuracy in reference file
coordinates. For Class II data entries we estimated the location error
as the distance d
l
=||s
l(GPS)
−s
l(GOA)
|| between the well GPS location
and the street address geocoded location. Class II location error was
described as a function of match score grouped into deciles (e.g. 100,
90–99, 80–89, 70–79). The location error for a geocoded owner
address coordinate s
l(GOA)
corresponding to a given match score was
assigned the median of the distances d
l
in the corresponding match
score bin (Supplemental Fig. S3). For Classes III and IV, entries
provided limited locational information, thus data entries were
assigned an error estimate of the median radius of zip code or county,
respectively.
2.5. Spatiotemporal geostatistical estimation of arsenic concentrations
In addition to mapping the actual historical arsenic measures
contained in the database, a BME geostatistical framework was
applied to estimate concentrations for locations at which no data were
available. We let X(s′,t) be a space/time random field (S/TRF)
representing the yearly arsenic concentration at location s′and year
t.Wedefined the yearly county arsenic concentration at location sand
time tas the areal average of X(s′,t) over an area the size of a county
around s, i.e.
ZRs;tðÞ=1
‖AR‖∫
ARsðÞ
ds0Xs0
;t
ð1Þ
where A
R
(s) was an area of radius Raround s, and the subscript Rin Z
R
emphasized the county level observation scale over which arsenic was
estimated, which in this work corresponded to the median county
radius in North Carolina (approximately 11 km).
First, kriging principles were applied to estimate arsenic concen-
trations across space and time using only the county averages. The
average arsenic concentrations were calculated within the boundary
of any county i. This county average provided an exact (hard) value
z
hard(i,t)
for the S/TRF Z(s
i
,t) at the centroid s
i
of county i. These hard
data were processed with the well-documented kriging method
(Cressie, 1990) to model the mean trend (assumed constant) and
covariance c
Z
(p,p′) of the S/TRF Z(s,t) where p=(s,t) was the
space/time coordinate and obtain kriging estimates of county arsenic
concentrations at a grid of unmonitored locations.
To incorporate the geocoded information and refine the spatial
resolution of our arsenic estimate, we developed a BME framework to
account for errors associated with geocoded classes by generating soft
data. The soft data for yearly arsenic county concentrations were
constructed at soft data points located on a static fine resolution grid
across the state. The mean μ
j
(Eq. (2)) and variance σ
j
2
(Eq. (3)) for the
yearly county arsenic concentration incorporated geocoding distance
error and assigned weight w
l
(Eq. (4)). As such, the arsenic value at a
given soft data point s
j
was assigned a mean and variance calculated as
the weighted sample average and sample variance, respectively, of the
geocoded private well data that were within a distance Rof s
j
, where
the weights decreased as a function of increased geocoding location
error, i.e.
μj=∑wlAsl
∑wl
ð2Þ
σ2
j=
∑wlAsl−μj
2
∑wl
nð3Þ
where As
l
and nserved as the l-th and total number of measured
arsenic values within Rof s
j
, respectively, and w
l
described a weight
that was inversely related to the location error ε
l
of the l-th arsenic
data point. Weights were calculated as
wl=R−εl
Rð4Þ
which captured the chance that well lwas correctly placed within a
circle of radius Rdespite its location error ε
l
.To prevent the probability
of estimating a negative concentration, we defined the soft data using
a Gaussian probability distribution function truncated below zero,
with mean and variance calculated from Eqs. (2) and (3). Eqs. (2–4)
12 A.P. Sanders et al. / Environment International 38 (2011) 10–16
Author's personal copy
provided values for the soft data points s
j
, which together with the
hard data z
hard(i,t)
at the county centroids s
i
, constituted the site
specific knowledge Sthat was incorporated to produce refined
estimates of yearly county arsenic concentrations using the BME
method (Christakos, 1990, 2000; De Nazelle et al., 2010) and its
BMElib numerical implementation (Christakos et al., 2002; Serre and
Christakos, 1999). The estimated values were mapped using ArcGIS™
ESRI version 9.0 software.
Lastly,to quantitativelyevaluate the difference in performance of the
kriging and BMEmethods for accurate estimation of arsenic concentra-
tions, we applied a cross-validation approach. Each data point was
sequentially removed from the estimation scheme and then re-
estimated using the remaining space/time data points (Money et al.,
2009). Mean square error (MSE) was then derived from this cross-
validation process and calculated as the sum of the squared differences
between the re-estimated and original values.
2.6. USGS and EPA database retrieval
Archived arsenic monitoring data from the online Water Quality
section of the National Water Information System (NWIS) and STORET
database were obtained electronically (USGS, 2010). Basic statistics
were compared from the NCDHHS database with the field sample
dataset from the NWIS in North Carolina.
3. Results
3.1. Arsenic levels exceed EPA standard in North Carolina monitored wells
The historical database of monitored well data revealed increased well sampling in
recent years. Prior to 2008 the NCDHHS sampled over 4000 wells per year with a
notable increase to over 10,000 wells per year from 2008 to present (Supporting
information Table S1). Across the state, a total of 1436 well measurements (2.25%)
exceeded the current standard of 10 μg/L and 233 exceeded 50 μg/L (Table 1;
Supporting information Table S1 and Table S2). Over the 11-year period, 7713 samples
measured above the detection limit, representing approximately 12% of all private
wells tested (Table 1). The remaining domestic well water records were below the
detection limit (Supporting information Table S2).
To systematically determine counties with elevated arsenic levels, counties were
ranked by the percentage of wells that exceeded the EPA standard across the eleven-
year period (Fig. 1;Table 1; Supporting information Table S1). The top ten counties
with the highest percentage of wells containing elevated arsenic levels were in order:
Stanly, Union, Anson, Montgomery, Dare, Randolph, Davidson, Alexander, Cleveland,
and Currituck (Supporting information Table S1). Of the measured wells that exceeded
the EPA standard, nearly 70% were within these ten counties. The remaining 30% of
wells that exceeded the EPA standard were collected from the remaining ninety North
Carolina counties (Table 1). To identify counties that might exhibit similar arsenic
trends over time, cluster analysis was performed. Members of the top ten counties (e.g.
Union, Stanly, Alexander) had statistical temporal relationships in arsenic levels across
the 11-year period (Supporting information Fig. S1). This comprehensive temporal
assessment revealed a historical pattern of arsenic levels in counties along the Carolina
terrane, demonstrating that some counties appear to have been high for over a decade.
Table 1 also reports the demographics of county and state population (and percent
of total) using private domestic well water (Kenny et al., 2009). Notably, in some of the
highly ranked counties, such as Randolph County, nearly 48% of the county population
uses private domestic wells as a primary water source. Counties were assigned a second
ranking based on the percentage of population using self-supplied domestic wells
multiplied by the percentage of wells exceeding the EPA standard. The following top
ten counties were identified: Stanly, Union, Montgomery, Randolph, Lincoln, Pender,
Dare, Moore, Person, and Transylvania.
3.2. Four-class geocoding algorithm increased spatial information
The geocoding scheme developed in this work categorized the data into four
classes (Supporting information Fig. S2). Approximately 3.6% (2295 well measures) of
the database had original GPS coordinates available and represent Class I. Geocoded
well locations representing Class II comprised 68.9% (43,991 well measures) of the
database. The remaining well locations were categorized as Class III (13.3%) and Class IV
(14.2%) by assigning zip code and county centroid coordinates, respectively.
A geocoding location error was assessed for each geocoding class (Supporting
information Table S3). A conservative estimate of 50 m location error for Class I was
established based on previouslyreported quantification of GPS error (Hulbert andFrench,
2001; Wing and Eklund, 2007). Class II location errors were determined as a function of
match score. Acceptable geocoding match scores were established at 60–100 and the
corresponding median location error ranged between 78 m and 758 m (Supporting
information Fig. S3, Table S3).Location error forClass III and Class IV wasapproximated as
half the median radius of a zip code or county, 3500 and 11,000m, respectively.
Table 1
Top 25-ranked
a
North Carolina counties.
Rank
a
County Total no. of wells
sampled
No. of wells that exceed
EPA standard (%)
No. of wells above
detect (%)
Pop. using domestic wells,
in thousands (%)
b
Pop. at risk
rank
c
1 Stanly 849 176 (20.73) 485 (57.13) 25.947 (44.00) 1
2 Union 3250 634 (19.51) 1454 (44.74) 49.197 (30.20) 2
3 Anson 98 10 (10.20) 34 (34.69) 2.704 (10.60) 11
4 Montgomery 372 34 (9.14) 120 (32.26) 8.213 (30.06) 3
5 Dare 572 36 (6.29) 137 (23.95) 8.091 (23.87) 7
6 Randolph 1595 72 (4.51) 394 (24.70) 66.845 (48.31) 4
7 Davidson 552 23 (4.17) 106 (19.20) 36.893 (23.86) 13
8 Alexander 128 5 (3.91) 30 (23.44) 6.818 (19.21) 15
9 Cleveland 269 9 (3.35) 37 (13.75) 9.234 (9.39) 29
10 Currituck 153 5 (3.27) 24 (15.69) 3.492 (15.11) 23
11 Lincoln 990 31 (3.13) 139 (14.04) 39.858 (57.06) 5
12 Moore 1093 33 (3.02) 206 (18.85) 34.920 (42.75) 8
13 Gaston 1697 47 (2.77) 278 (16.38) 57.915 (29.53) 14
14 Cabarrus 626 15 (2.40) 98 (15.65) 37.367 (24.87) 21
15 Pender 800 19 (2.38) 115 (14.38) 33.179 (71.46) 6
16 Watauga 671 15 (2.24) 77 (11.48) 11.970 (28.18) 20
17 Nash 1137 25 (2.20) 145 (12.75) 4.870 (5.33) 43
18 Transylvania 424 9 (2.12) 24 (5.66) 15.156 (51.16) 10
19 Chatham 1404 26 (1.85) 455 (32.41) 32.080 (55.31) 12
20 Person 847 15 (1.77) 165 (19.48) 25.605 (68.80) 9
21 Catawba 454 8 (1.76) 42 (9.25) 62.709 (41.35) 17
22 Bladen 114 2 (1.75) 7 (6.14) 13.898 (42.19) 16
23 Duplin 240 4 (1.67) 13 (5.42) 15.305 (29.44) 24
24 Avery 261 4 (1.53) 28 (10.73) 8.291 (47.00) 18
25 New Hanover 1449 22 (1.52) 248 (17.12) 16.371 (9.12) 41
–Other counties 43,811 157 (0.36) 2852 (6.51) 1668.398 (24.95) –
Total –63,856 1436 (2.25) 7713 (12.08) 2295.33 (26.43) –
a
Rank based on tendency of wells to exceed the EPA standard throughout an 11-year period.
b
Data from (Kenny et al., 2009).
c
Rank based on composite of percentage of wells that exceed the EPA standard through the 11-year period and percentage of county residents using private domestic well water.
13A.P. Sanders et al. / Environment International 38 (2011) 10–16
Author's personal copy
3.3. Mapping of arsenic in monitored private domestic wells
We applied the results of our four-class geocoding process to map arsenic levels in
monitored wells and identify regions of arsenic contamination in North Carolina
(Fig. 2). As an example, we show locations of the geocoded wells and those that
exceeded the EPA standard in 2009 (Fig. 2A). Notably, wells exceeding the EPA
standard were located primarily in the south-central region of the state. The calculated
county averages for 2009 are also provided (Fig. 2B). The highest county average was
observed in Stanly County, where the average of 89 domestic well records approached
the 10 μg/L EPA standard.
3.4. Spatiotemporal modeling of estimated arsenic concentrations
Next, we set out to refine the spatial scale and apply the results of the geocoding
process using two geostatistical estimation methods, namely 1) the classical kriging
method and 2) our novel BME framework. Using the classical method that incorporates
no distance error information, the spatial distribution of kriging estimates of county
arsenic concentrations across North Carolina is represented (Fig. 2C). The kriging
estimates correspond to the spatial interpolation of county averages assigned to their
county centroid. In comparison, the BME framework with a county-level observation
scale (Fig. 2D) interpolated data in between county averages. The county observation
scale enabled estimates of aggregated arsenic concentrations across a ~11 km radius.
This map incorporated estimated distance errors associated with the geocoded data to
obtain less biased predicted arsenic values. From this BME map we identified regions
within counties where elevated arsenic is endemic. For example, southeastern Union
County and the border between Stanly and Montgomery Counties are areas of special
concern (Fig. 2-D1). The BME estimates reveal that in these areas the arsenic
concentration may reach 18 μg/L. Cross-validation analysis indicated that the BME
framework better estimated arsenic concentrations than the kriging method. The MSE
for the BME method was 41% lower than that of kriging (Supporting information Table
S4). In total, the geocoded data and the rigorous processing of location errors through
the BME method significantly improved our understanding of arsenic distributions
across unsampled areas of North Carolina.
4. Discussion
Arsenic is a known human carcinogen and relevant environmental
contaminant in drinking water systems. Publicly available data at the
NCDHHS represent a volume of historical arsenic analyses in North
Carolina domestic well waters performed under stringent EPA
guidelines that remain largely unanalyzed. A primary goal of this
research was to identify trends in areas of North Carolina with
elevated arsenic concentrations in monitored domestic well waters.
We revealed arsenic trends by county in monitored wells over time as
well as estimated concentrations in unmonitored locations. The
geocoding methods developed in this study data enabled a compre-
hensive report of over 4000 yearly arsenic measurements with
1. Stanly
2. Union
3. Anson
4. Montgomery
5. Dare
6. Randolph
7. Davidson
8. Alexander
9. Cleveland
10. Currituck
11. Lincoln
12. Moore
13. Gaston
14. Cabarrus
15. Pender
16. Watauga
17. Nash
18. Transylvania
19. Chatham
20. Person
21. Catawba
22. Bladen
23. Duplin
24. Avery
25. New Hanover
26. Rutherford
27. Rowan
28. Wilkes
29. Halifax
30. Harnett
31. Orange
32. Lee
33. Cumberland
34. Johnston
35. Wake
Total 1998 2010
Fig. 1. The top thirty-five counties that exceed the EPA standard (10 μg/L). Counties are
ranked by the percent of wells that exceed the EPA standard and are represented in a
heat map. Counties with percent of wells exceeding the statewide 2.25% appear in red-
scale, while those below the statewide percent appear in blue-scale. Counties with no
information available appear in white. (For interpretation of the references to color in
this figure legend, the reader is referred to the web version of this article.)
B
C
D
*
273
89 32
A
Arsenic concentration (µg/L)
High: 17
Low: 0
High: 8
Low: 0
Arsenic concentration (µg/L)
Arsenic concentration (µg/L)
0.5 - 1.0
1.1 - 2.5
2.6 – 5.0
5.1 – 8.0 0 40 80 160 240 320 Km 0 20 40 80 Km
EPA standard or below
Above EPA standard
County boundary
Average As conc. (µg/L)
C1
D1
C1
D1
Fig. 2. Geocoded arsenic concentrations in 2009. (A) Samples exceeding the EPA standard are shown in black. Well locations of samples below the standard appear in gray.
(B) County averages are displayed in grayscale and the number of arsenic analyses in 2009 appears within each county. *No wells were sampled in Chowan County in 2009. (C) A
classical kriging method estimated arsenic distribution across the state at unmonitored locations. (D) The Bayesian Maximum Entropy framework estimated arsenic distribution
across the state at unmonitored locations.
14 A.P. Sanders et al. / Environment International 38 (2011) 10–16
Author's personal copy
geographical coordinates from 1998 to 2007 and over 10,000 from
2008 to present, a substantial increase relative to the USGS and EPA
ambient monitoring systems. Specifically, the number of records
analyzed represents a 600-fold increase from samples collected by the
USGS (USGS, 2010) and more than three times the number of records
analyzed in other studies of North Carolina wells (Kim et al., 2011;
Pippin, 2005). The substantial increase in recent well sampling is
likely due to state legislation adopted in 2008 that requires every
newly constructed well be tested. These types of monitoring
programs, as evidenced here, are essential to ensuring increased
awareness of well water contaminants and protected health of
individual homeowners.
A notable result of this study is the surprisingly high levels of
arsenic (up to 806 μg/L) that were detected in some homeowners'
domestic wells. In addition, more than 1436 (2.25%) of wells exceeded
the EPA standard. Some of the top-ranked counties identified here as
most frequently exceeding the EPA MCL have not previously been
highlighted in nation- or statewide studies (Welch et al., 1999)
including Anson, Montgomery, Dare, Alexander, Cleveland, and
Currituck Counties. We identify historical trends in counties along
the Carolina terrane and through comprehensive temporal assess-
ment reveal that arsenic levels have been elevated for over a decade.
In some of these counties, greater than 50% of the population use
domestic wells (Kenny et al., 2009). Importantly, some of the
identified counties of concern have rapidly growing populations (US
Census Bureau, 2006), which compounded by an arsenic-prone
geology may have public health implications for residents. Simulta-
neously, rural areas are more likely to lack connection to a municipal
drinking water system. By ranking based on percentage of population
at risk we identify counties where county-level well monitoring
programs may be cost-effective. By integrating information on
population size in these counties, we show Union and Stanly continue
to rank among the most at-risk county populations. Our data confirm
increased concentrations of arsenic in counties located along the
Carolina terrane and highlight elevated levels over a decade-long
period. In addition, the counties of Stanly and Union have large
populations (roughly 26,000 and 49,000 individuals) relying on
private well water sources. Currently, no epidemiologic literature
has investigated the health impact of arsenic in these populations.
This study presents a new approach to assigning geographical
coordinates when sample locations are described by inconsistent
recorded information. It was evident from the spatial locations of GPS
data that GPS device use was not uniform across the state. It was
necessary, therefore, to increase the number of geocoded locations
using additional information (Classes II–IV) to avoid bias and provide
more accurate spatial representations. As such, we applied a four-step
geocoding process and error estimation scheme that increased the
available geographic coordinates of ~3000 to more than 63,000 well
analyses. We increased the knowledgebase using available locational
information to assign geographic coordinates of domestic wells based
on four spatial classes: GPS, street address, zip code, and county. As an
example, we tripled the number of successfully geocoded points used
in previous analyses over comparable geographic areas and time-
frames (Kim et al., 2011; Pippin, 2005). Additionally, while others
have shown that multi-stage geocoding methods improved the match
rate compared to single-step methods (Lovasi et al., 2007), to our
knowledge, the present study is one of the first to use GPS locations to
systematically quantify and account for geocoding location error. We
present a widely applicable method of systematically determining
acceptable match scores resulting from the multi-stage address
geocoding procedure that serves as an alternative to a minimum
match score determined a priori (Yang et al., 2004).
The general BME framework has been applied to arsenic levels in
Bangladesh groundwaters to estimate aqueous concentrations where
data are not available (Serre et al., 2003). In this study, we apply the
BME framework to U.S. groundwaters and incorporate location error
into a novel arsenic estimation process, which aggregates monitored
arsenic levels to a county level observation scale (~ 11 km). To the best
of our knowledge, this is the first study that accounts for locational
error. The cross-validation analysis shows that the BME approach
presented in this work more accurately predicts arsenic than
conventional modeling approaches. Within this framework, the
county observation scale refines well-to-well variation and we
would not expect to find the high concentrations seen in individual
monitored wells (e.g. up to 806 μg/L). By narrowing the scope to
counties of interest in the southwestern Piedmont we identified
southeastern Union County and the border between Stanly and
Montgomery Counties as areas of special concern. The levels
documented in this study indicate areas of arsenic contamination at
nearly twice the EPA standard—a level estimated to double the risk of
bladder and lung cancer (NRC, 2001). The local observation scale
enables our predictive maps to be useful for public health screening
purposes by identifying areas where the risk of arsenic exposure is
high and by providing a science-based criterion for cost-effective
monitoring of wells.
Through our analyses of over 60,000 geocoded well locations we
were able to more accurately assess spatial and temporal arsenic
trends in both monitored wells and estimates at unmonitored
locations across North Carolina. We found that areas near the eastern
coast and along the Carolina terrane have high prevalence of arsenic
contamination in private wells. The presence of arsenic in the Carolina
terrane has been confirmed by geological studies in this area (Foley et
al., 2001) and is supported by models that incorporate geologic
determinants (Kim et al., 2011). Abundant literature details the
sources of groundwater arsenic contamination through anthropogen-
ic factors including agricultural or industrial practices (Appleyard et
al., 2004; Embrick et al., 2005; Foley and Ayuso, 2008; Jackson et al.,
2006), yet, much of arsenic contamination highlighted in this study is
thought to be naturally occurring due to the underlying geology
(Duker et al., 2005; Foley et al., 2001; Welch et al., 2000). The Carolina
terrane, however, does not underlie each of the top ten counties
(Pender, Dare, and Currituck Counties for instance) and it is possible
that a combination of anthropogenic and natural sources may
contribute to arsenic contamination, however additional studies are
warranted. Currently, no EPA Superfund National Priorities List or
Toxic Release Inventory sites are listed in these three counties that
would indicate substantial anthropogenic contribution to environ-
mental arsenic concentrations. Moreover, while an EPA superfund site
is located in Haywood County with reported historical use of arsenical
pesticides, no wells in that county exceeded the EPA standard.
Studies like this one represent a major step towards arsenic
surveillance in contaminated areas. To lessen the risk of exposure to
arsenic in drinking water, recommended preventative measures
include point-of-use removal, modification of well depth, and/or use
of an alternate water source (Alaerts and Khouri, 2004; Pratson et al.,
2010). These solutions are rarely cost-effective, however, and may not
be feasible in rural areas. Simple cost-effective technologies for
mitigating arsenic are not widely available, and, in lieu of federal or
state water quality regulation of domestic wells, the best mitigation is
in the form of well water testing programs. Trivalent (As3+: arsenite)
and pentavalent (As5+: arsenate) arsenic most commonly occur in
natural waters (Duker et al., 2005; Feng et al., 2001). Due to the
variable toxicity of arsenic species (As3+ being more toxic),
additional studies are warranted to determine the distribution of
individual arsenic species in drinking water. Targeted monitoring is
crucial in reducing the financial cost of testing for speciated arsenic in
every monitored well and the methods developed here can be applied
to this end towards arsenic in other regions as well as to other
contaminants of concern to public health.
The compounding of environmental and social factors means
residents could be at increased risk for health effects from arsenic.
Additional studies are warranted to further ascertain the sources of
15A.P. Sanders et al. / Environment International 38 (2011) 10–16
Author's personal copy
and biogeochemical behavior of arsenic in unstudied regions of North
Carolina and to help reduce exposures among at risk populations. By
identifying regions of contamination through studies such as this one,
cost-effective monitoring programs can target at risk populations to
protect public health and help to shape state and local water
monitoring policies (Miranda and Edwards, 2011). Moving forward
we anticipate research that will integrate these findings with
biomonitoring and health outcome data to substantiate risks posed
to populations in arsenic endemic areas. The next steps to protecting
individuals include community education in at risk areas and
biomonitoring of those populations most at risk including children
and pregnant women.
Acknowledgments
We thank Leslie Wolf, Patrick Fleming, Dianne Enright, and John
Neal at the NCDHHS for their valuable contributions to this study. We
thank Kathleen Gray, Brennan Bouma, Tracey Slaughter, and Fred
Pfaender with UNC Research Translation Core for their ongoing
contributions. This research was supported in part by the NIEHS (P42-
ES005948, T32-ES007017, and P30-ES010126).
Appendix A. Supplementary data
Supplementary data to this article can be found online at doi:10.
1016/j.envint.2011.08.005.
References
Alaerts GJ, Khouri N. Arsenic contamination of groundwater: mitigation strategies and
policies. Hydrogeol J 2004;12(1):103–14.
Appleyard S, Wong S, Willis-Jones B, Angeloni J, Watkins R. Groundwater acidification
caused by urban development in Perth, Western Australia: source, distribution, and
implications for management. Aust J Soil Res 2004;42(5–6):579–85.
Ayotte JD, Montgomery DL, Flanagan SM, Robinson KW. Arsenic in groundwater in
eastern New England: occurrence, controls, and human health implications.
Environ Sci Technol 2003;37(10):2075–83.
Bellander T, Berglind N, Gustavsson P, Jonson T, Nyberg F, Pershagen G, et al. Using
geographic information systems to assess individual historical exposure to air
pollution from traffic and house heating in Stockholm. Environ Health Perspect
2001;109(6):633–9.
Christakos G. A Bayesian Maximum-Entropy view to the spatial estimation problem.
Math Geology 1990;22(7):763–77.
Christakos G. Modern spatiotemporal geostatistics. Oxford University Press; 2000.
Christakos G, Bogaert P, Serre ML. Temporal GIS: advanced functions for field-based
applications. New York, NY: Springer-Verlag; 2002.
Cressie N. The origins of kriging. Math Geol 1990;22(3):239–52.
De Nazelle A, Arunachalam S, Serre ML. Bayesian maximum entropy integration of
ozone observations and model predictions: an application for attainment
demonstration in North Carolina. Environ Sci Technol 2010;44(15):5707–13.
Duker AA, Carranza EJM, Hale M. Arsenic geochemistry and health. Environ Int 2005;31
(5):631–41.
Eisen MB, Spellman PT, Brown PO, Botstein D. Cluster analysis and display of genome-
wide expression patterns. Proc Natl Acad Sci USA 1998;95(25):14863–8.
Embrick LL, Porter KM, Pendergrass A, Butcher DJ. Characterization of lead and arsenic
contamination at Barber Orchard, Haywood County, NC. Microchem J 2005;81(1):
117–21.
EPA. Method 200.8 Revision 5.4 —determination of trace elements in water and wastes
by inductively coupled plasma-mass spectrometry. Cincinnati, OH; 1994.
EPA. Arsenic in drinking water; 2010. Available: http://epa.gov/safewater/arsenic/
index.html [accessed May 24 2010].
Feng ZM, Xia YJ, Tian DF, Wu KK, Schmitt M, Kwok RK, et al. DNA damage in buccal
epithelial cells from individuals chronically exposed to arsenic via drinking water
in Inner Mongolia, China. Anticancer Res 2001;21(1A):51–7.
Foley NK, Ayuso RA. Mineral sources and transport pathways for arsenic release in a
coastal watershed, USA. Geochem Explor Environ Anal 2008;8:59–75.
Foley N, Ayuso RA, Seal R. Remnant colloform pyrite at the Haile gold deposit, South
Carolina: a textural key to genesis. Econ Geol Bull Soc Econ Geol 2001;96(4):
891–902.
Hagan EF. Ground Water Quality Technical Brief: statewide ambient groundwater
quality monitoring program arsenic speciation results (2002 & 2003); 2004.
Holton WC. Locating lead —mapping leads to intervention. Environ Health Perspect
2002;110(9):A533.
Hulbert IAR, French J. The accuracy of GPS for wildlife telemetry and habitat mapping.
J Appl Ecol 2001;38(4):869–78.
Jackson BP, Seaman JC, Bertsch PM. Fate of arsenic compounds in poultry litter upon
land application. Chemosphere 2006;65(11):2028–34.
Kenny JF, Barber NL, Hutson SS, Linsey KS, Lovelace JK, Maupin MA. Estimated use of
water use in the United States in 2005; 2009.
Kim MJ, Nriagu J, Haack S. Arsenic species and chemistry in groundwater of southeast
Michigan. Environ Pollut 2002;120(2):379–90.
Kim D, Miranda ML, Tootoo J, Bradley P, Gelfand AE. Spatial modeling for groundwater
arsenic levels in North Carolina. Environ Sci Technol 2011;45(11):4824–31.
Kumar A, Adak P, Gurian PL, Lockwood JR. Arsenic exposure in US public and domestic
drinking water supplies: a comparative risk assessment. J Expo Sci Environ
Epidemiol 2010;20(3):245–54.
Lovasi GS, Weiss JC, Hoskins R, Whitsel EA, Rice K, Erickson CF, et al. Comparing a
single-sta ge geocoding method to a m ulti-stag e geocodin g method: how much
and where do they disagree? Int J Health Geogr 2007:6.
Miranda ML, Edwards SE. Use of spatial analysis to support environmental health
research and practice. NC Med J 2011;72(2):132–5.
Miranda ML, Dolinoy DC, Overstreet MA. Mapping for prevention: GIS models for
directing childhood lead poisoning prevention programs. Environ Health Perspect
2002;110(9):947–53.
Money E, Carter GP, Serre ML. Using river distances in the space/time estimation of
dissolved oxygen along two impaired river networks in New Jersey. Water Res
2009;43(7):1948–58.
Mukherjee A, Sengupta MK, Hossain MA, Ahamed S, Das B, Nayak B, et al. Arsenic
contamination in groundwater: a global perspective with emphasis on the Asian
scenario. J Health Popul Nutr 2006;24(2):142–63.
National Research Council (NRC). Arsenic in drinking water: 2001 Update. Washington,
DC: National Academy Press; 2001.
Nuckols JR, Ward MH, Jarup L. Using geographic information systems for exposure
assessment in environmental epidemiology studies. Environ Health Perspect
2004;112(9):1007–15.
Peters SC, Blum JD, Klaue B, Karagas MR. Arsenic occurrence in New Hampshire
drinking water. Environ Sci Technol 1999;33(9):1328–33.
Pippin CG. Distribution of total arsenic in groundwater in the North Carolina Piedmont.
NGWA Naturally Occurring Contaminants Conference: Arsenic, Radium, Radon, and
Uranium; 2005. p. 89–102.
Pratson E, Vengosh A, Dwyer G, Pratson L, Klein E. The effectiveness of arsenic
remediation from groundwater in a private home. Ground Water Monit Rem
2010;30(1):87–93.
Serre ML, Christakos G. Modern geostatistics: computational BME analysis in the light
of uncertain physical knowledge —the Equus Beds study. Stoch Environ Res Risk
Assess 1999;13(1–2):1–26.
Serre ML, Kolovos A, Christakos G, Modis K. An application of the holistochastic human
exposure methodology to naturally occurring arsenic in Bangladesh drinking
water. Risk Anal 2003;23(3):515–28.
Shaw WD, Walker M, Benson M. Treating and drinking well water in the presence of
health risks from arsenic contamination: results from a US hot spot. Risk Anal
2005;25(6):1531–43.
Shiber JG. Arsenic in domestic well water and health in Central Appalachia, USA. Water
Air Soil Pollut 2005;160(1–4):327–41.
U.S. Census Bureau. Population estimates for the 100 fastest-growing U.S. counties by
percentage growth from July 1, 2004 to July 1, 2005. Washington, DC; 2006.
USGS. National water information system. Available: http://waterdata/usgs.gov/nc/
nwis/qwdata. 2010.
Walker M, Shaw WD, Benson M. Arsenic consumption and health risk perceptions in a
rural western US area. J Am Water Resour Assoc 2006;42(5):1363–70.
Welch AH, Helsel DR, Focazio MJ, Watkins SA. Arsenic in ground water supplies of the
United States. New York: Elsevier Science; 1999.
Welch AH, Westjohn DB, Helsel DR, Wanty RB. Arsenic in ground water of the United
States: occurrence and geochemistry. Ground Water 2000;38(4):589–604.
Wing MG, Eklund A. Performance comparison of a low-cost mapping grade global
positioning systems (GPS) receiver and consumer grade GPS receiver under dense
forest canopy. J Forest 2007;105(1):9–14.
Yang D, Bilaver L, Hayes O, Goerge R. Improving geocoding practices: evaluation of
geocoding tools. J Med Syst 2004;28(4):361–70.
Yang Q, Jung HB, Culbertson CW, Marvinney RG, Loiselle MC, Locke DB, et al. Spatial
pattern of groundwater arsenic occurrence and association with bedrock geology in
Greater Augusta, Maine. Environ Sci Technol 2009;43(8):2714–9.
16 A.P. Sanders et al. / Environment International 38 (2011) 10–16
