Published March, 2012
Journal of Environmental Quality
TECHNICAL REPORTS
GROUNDWATER QUALITY
Regression Models for Estimating Concentrations of Atrazine plus Deethylatrazine
in Shallow Groundwater in Agricultural Areas of the United States
Paul E. Stackelberg,* Jack E. Barbash, Robert J. Gilliom, Wesley W. Stone, and David M. Wolock
T
he U.S. Geological Survey National Water-Quality
Assessment (NAWQA) Program has monitored the
occurrence of pesticides and other contaminants at
thousands of sites throughout the United States and in a variety of media since 1992 (Gilliom et al., 2006; Dubrovsky et al.,
2010). The resulting data provide a resource for the development of national- and regional-scale statistical models to estimate concentrations or frequencies of occurrence of pesticides
and other contaminants in groundwater (Kolpin et al., 2002;
Stackelberg et al., 2006a; Nolan and Hitt, 2006), stream water
(Larson et al., 2004; Stone et al., 2008; and Stone and Gilliom,
2009), and fish tissue (Nowell et al., 2009). Development of
these predictive capabilities is a vital component of a comprehensive national water quality assessment because, given the
infeasibility of monitoring everywhere, the models provide a
cost-effective means of understanding the chemical quality of
the United States’ water resources and for estimating water
quality conditions for unmonitored areas. This information is
vital to improving water quality management and for setting
efficient monitoring priorities.
Atrazine is one of the most extensively used agricultural
herbicides in the United States, with an average annual use
(1992–1995) of 29 million kg active ingredient applied to
about 231,000 km2 of cropland (primarily corn and sorghum)
(Thelin and Gianessi, 2000). National models of atrazine
occurrence in shallow groundwater were previously developed
by Kolpin et al. (2002) and Stackelberg et al. (2006a) to evaluate atrazine detection frequencies in well networks (typically
20–30 wells each) associated with NAWQA land-use studies
(Gilliom et al., 2006). Explanatory variables, values of which
were averaged across each study area, were used to represent
the natural setting, agricultural management practices, and the
types and amount of development in each area. The model of
Kolpin et al. (2002) was developed from data on the occurrence of atrazine in shallow groundwater underlying agricultural and urban settings, whereas the model of Stackelberg et
Tobit regression models were developed to predict the summed
concentration of atrazine [6-chloro-N-ethyl-N’-(1-methylethyl)1,3,5-triazine-2,4-diamine] and its degradate deethylatrazine
[6-chloro-N-(1-methylethyl)-1,3,5,-triazine-2,4-diamine]
(DEA) in shallow groundwater underlying agricultural settings
across the conterminous United States. The models were
developed from atrazine and DEA concentrations in samples
from 1298 wells and explanatory variables that represent the
source of atrazine and various aspects of the transport and fate of
atrazine and DEA in the subsurface. One advantage of these newly
developed models over previous national regression models is that
they predict concentrations (rather than detection frequency),
which can be compared with water quality benchmarks. Model
results indicate that variability in the concentration of atrazine
residues (atrazine plus DEA) in groundwater underlying
agricultural areas is more strongly controlled by the history of
atrazine use in relation to the timing of recharge (groundwater
age) than by processes that control the dispersion, adsorption, or
degradation of these compounds in the saturated zone. Current
(1990s) atrazine use was found to be a weak explanatory variable,
perhaps because it does not represent the use of atrazine at the
time of recharge of the sampled groundwater and because the
likelihood that these compounds will reach the water table is
affected by other factors operating within the unsaturated
zone, such as soil characteristics, artificial drainage, and water
movement. Results show that only about 5% of agricultural areas
have greater than a 10% probability of exceeding the USEPA
maximum contaminant level of 3.0 μg L−1. These models are
not developed for regulatory purposes but rather can be used to
(i) identify areas of potential concern, (ii) provide conservative
estimates of the concentrations of atrazine residues in deeper
potential drinking water supplies, and (iii) set priorities among
areas for future groundwater monitoring.
Copyright © 2012 by the American Society of Agronomy, Crop Science Society
of America, and Soil Science Society of America. All rights reserved. No part of
this periodical may be reproduced or transmitted in any form or by any means,
electronic or mechanical, including photocopying, recording, or any information
storage and retrieval system, without permission in writing from the publisher.
P.E. Stackelberg, U.S. Geological Survey, 425 Jordan Rd., Troy, NY 12180; J.E.
Barbash, U.S. Geological Survey, 934 Broadway, Suite 300, Tacoma, WA 98402; R.J.
Gilliom, U.S. Geological Survey, Placer Hall, 6000 J St., Sacramento, CA 95819-6129;
W.W. Stone, U.S. Geological Survey, 5957 Lakeside Blvd., Indianapolis, IN 462781996; D. M. Wolock, U.S. Geological Survey, 4821 Quail Crest Pl., Lawrence, KS
66049. Any use of trade, product, or firm names is for descriptive purposes only
and does not imply endorsement by the U.S. Government. Assigned to Associate
Editor Joe Kozak.
J. Environ. Qual. 41:479–494 (2012)
doi:10.2134/jeq2011.0200
Posted online 13 Jan. 2012.
Freely available online through the author-supported open-access option.
Received 7 June 2011.
*Corresponding author ([email protected]).
© ASA, CSSA, SSSA
5585 Guilford Rd., Madison, WI 53711 USA
Abbreviations: DEA, deethylatrazine; DO, dissolved oxygen; GC–MS, gas
chromatography/mass spectrometry; LTMDL, long-term method detection level;
NAWQA, National Water Quality Assessment; OM, organic matter; PC, principal
component; PI, prediction interval; pR2, pseudo coefficient of determination.
479
al. (2006a) focused solely on atrazine occurrence in agricultural settings. Both models used ordinary least-squares regression methods to identify and quantify the effects of factors that
were significantly correlated with atrazine detection frequencies. Stackelberg et al. (2006a) also applied their model to predict the frequency with which atrazine would be detected in
groundwater beneath unmonitored agricultural settings across
the conterminous United States.
This study modified and expanded the atrazine model of
Stackelberg et al. (2006a) by (i) using atrazine residue concentrations (i.e., the sum of the concentrations of atrazine and
one of its degradates, deethylatrazine [DEA]) as the response
variable rather than atrazine-detection frequencies and (ii)
incorporating data for explanatory variables derived from sitespecific information (such as water chemistry) that was available only for the locations at which the values were measured.
The reason for using atrazine residue concentrations as the
response variable, rather than the concentration of atrazine
alone, was to capture more information on factors that control
the transport and fate of atrazine-derived compounds in the
subsurface. Three other primary atrazine degradates—deisopropylatrazine, didealkylatrazine, and hydroxyatrazine—were
considered but were not examined because of considerations
related to analytical performance (Mark Sandstrom, USGS,
written communication, 2011). The specific objectives of this
study were (i) to develop regression models that can predict
atrazine residue concentrations as well as the probability that
selected concentration thresholds will be exceeded in agricultural settings across the conterminous United States and (ii)
to ascertain whether the incorporation of explanatory variables derived from site-specific information can improve our
understanding of the relative importance of different factors
that control the concentrations of atrazine and DEA in shallow
groundwater beneath agricultural areas.
Materials and Methods
Data for Model Development and Evaluation
Estimates of annual atrazine use, in kilograms of active
ingredient per square kilometer (kg km−2) were available
for agricultural settings with a wide range of use intensity,
although most use is on corn and sorghum. Usage in nonagricultural settings is less than in agricultural settings (Gilliom
et al., 2006), but reliable estimates of nonagricultural use
are not available in most parts of the United States. Given
the direct relation between atrazine use and occurrence in
groundwater, all model-development and model-evaluation
data were collected in areas dominated by agricultural land
use. Because the models were developed from information
collected solely through the NAWQA Program, the water
quality data and explanatory variables represent only those
settings monitored as part of the Program (Gilliom et al.,
2006). The water quality data used in model development
were collected between 1993 and 2001 from 1298 wells in 55
study areas representing regionally extensive combinations of
agricultural land and hydrogeologic conditions (Gilliom et
al., 2006). Model evaluation entailed the use of corresponding data from an additional 209 sites in 19 study areas that
were sampled as part of the NAWQA Program during 2001–
480
2008. Locations of the model-development and model-evaluation sites are shown in Fig. 1.
Sampling and Analytical Methods
All groundwater sampling sites were selected through
methods described by Scott (1990), and the wells were
installed in accordance with protocols described by Lapham
et al. (1995) for the collection of “recently recharged” (young,
shallow) groundwater. The median depth of the sampled wells
was 9.5 m, and the median distance from the water table to
the bottom of the wells was 4.6 m.
Water samples were collected in accordance with procedures and protocols described in Koterba et al. (1995).
All samples were analyzed at the USGS National WaterQuality Laboratory in Denver, CO, by gas chromatography/
mass spectrometry (GC–MS) with selected ion monitoring
(Zaugg et al., 1995). The GC–MS analytical method does
not have specified “detection limits” for each pesticide analyte; rather, compounds detected and conclusively identified
by retention time and spectral characteristics are quantified
and reported (Zaugg et al., 1995; Oblinger Childress et al.,
1999; Stackelberg et al., 2006b; Martin, 2009). Analyses
that do not meet identification criteria based on retention
time and spectral characteristics are deemed “nondetections”
and are reported as less than their reporting level. The types
and numerical values of reporting levels used to report nondetections analyzed by GC–MS changed during the period
represented by the samples used for this study (1993–2008).
Removal of the resulting temporal structure in the reporting levels entailed reassigning the concentration value for
each nondetection to the long-term method-detection level
(LTMDL) for water years 1994 to 2006 (Martin, 2009),
which for atrazine and DEA are 0.004 and 0.007 μg L−1,
respectively. (A water year is the period 1 October through
30 September and is named for the year in which the water
year ends.)
Analytical results for atrazine and DEA, including censored concentrations (those deemed to be nondetections) were
adjusted to account for differences in analytical recoveries. This
was done through methods described in Martin et al. (2009)
and was necessary because of a pronounced disparity in the
recoveries of the two analytes; the mean recovery of atrazine
from groundwater matrix-spike samples during 1992–2006
was 99.6%, whereas the corresponding recovery for DEA was
47.2% (Martin et al., 2009).
Explanatory Variables
A total of 88 factors (explanatory variables) deemed likely
to affect the concentration of atrazine or DEA in shallow
groundwater were considered for model development. Data on
each of these variables were compiled for a 500-m radius surrounding each well site; exceptions were (i) variables derived
from site-specific information (such as water chemistry data)
that can be measured only at a discrete location and (ii) the
variable “artdrn” (prevalence of artificial drainage), for which
the data were compiled for grid cells that were 1 km2 in area.
Explanatory variables were transformed as needed to accommodate the assumption that the regressions were linear and to
ensure that the model residuals were normally distributed with
Journal of Environmental Quality • Volume 41 • March–April 2012
constant variance. Transformations
that were considered included (i)
raising individual explanatory variables to powers of 0.33 (cube root),
0.5 (square root), and 2 (square) and
(ii) converting to their logarithm. (All
logarithms used during this study
were base 10 values.) The explanatory variables considered for use in
development of the final models were
selected using methods described
in the next section. Those that were
selected are grouped into seven categories in Table 1. Several of these
variables are discussed below.
Intensity of Atrazine Use
The estimated use of atrazine
for agricultural purposes in the
500-m-radius area around each well
site (use9200 and use97 in Table 1) was
obtained by combining county-level
estimates of applications of atrazine Fig. 1. Atrazine use distribution throughout the conterminous United States and locations of the
to agricultural crops with information 1298 model-development and 209 model-evaluation sites.
on the spatial distribution of agriRecharge
cultural land within each radius using methods described by
Mean annual groundwater recharge (recharge in Table 1)
Larson et al. (2004) and Thelin and Gianessi (2000). Estimates
was computed from estimates of base flow and mean annual
of atrazine use between 1992 and 2000 (use9200) were derived
runoff as described by Wolock (2003) and modified to include
from proprietary-use data obtained from the Doane Marketing
groundwater withdrawn for use in irrigation (Karen Burow,
Research-Kynetec (DMRKYNETEC) Agro Trak database
USGS, written communication, May 2009).
(Thelin and Stone, 2010), and estimates for 1997 use (use97)
were derived from data from the National Center for Food
Water Quality
and Agricultural Policy (Thelin and Gianessi, 2000) (Table 1).
Spatial patterns in groundwater quality were identified
The source for the spatial distribution of agricultural land was
through principal component (PC) analysis of water qualthe 2001 National Land Cover Data set, which was obtained
ity data (Table 2). Principal components analysis was run on
at 30-m resolution (NLCD01; Vogelmann et al., 2001).
ranked data to accommodate censored values as described in
“Agricultural land” in the calculation of atrazine use consists of
Ator (2008). Two PCs that explain more than 50% of the varithe NLCD01 classifications of pasture/hay and the combined
ability in water quality data were considered in this study. The
grouping of row crops/small grains/fallow land with orchards/
first PC explained 32% of the variability and was interpreted to
vineyards. The intensity of atrazine use (kg km−2) was calcurepresent overall ionic strength (IS in Table 1). This PC showed
lated as the total amount of atrazine applied to the crops of
strong positive loadings (which indicate the correlation coefinterest divided by the radius area.
ficient between a variable and a PC) for specific conductance
Land Use
The land-use variable agrcveg represents the combined categories of row crops and vegetables, and agcc represents cover
crops (Table 1). These two variables represent ancillary data
that are required for wells from which water quality data are
collected for the NAWQA Program. Details on methods used
to collect these ancillary data were provided by Koterba (1998).
Agricultural Management Practices
Data on subsurface tile drains and surface ditches were
obtained from the 1992 National Resources Inventory of
the Natural Resources Conservation Service. The variable
artdrn (Table 1) represents the percentage of 1-km2 grid cells
with subsurface drains and graded ditches as represented by
National Resources Inventory conservation practice codes
CP606, CP607, and CP608 (NRCS, 1995).
www.agronomy.org • www.crops.org • www.soils.org
and the concentrations of major cations (Ca2+, Mg2+, and Na+)
and anions (Cl− and SO42–) (Table 2). The second PC explained
20% of the variability and was interpreted to be an indicator of groundwater residence time in the saturated zone (RTI
in Table 1). A detailed discussion of RTI is presented in the
Results section.
Model-Development Approach
Measured atrazine and DEA concentrations, adjusted to
approximate 100% analytical recovery, ranged from <0.004 to
4.7 μg L−1 for atrazine and from <0.007 to 8.2 μg L−1 for DEA.
The sum of the concentrations of atrazine and DEA, referred to
herein as “atrazine residue,” was transformed to its logarithm to
make its relation with explanatory variables more linear, reduce
heteroscedasticity, and shift the distribution of model residuals
closer to normality.
481
Regression Analysis
In this study, 52% of the atrazine and DEA concentrations
were censored (i.e., deemed to be nondetections); therefore, a
Tobit model, which provides parameter estimates for a censored linear model (Judge et al., 1985; Tobin, 1958), was used
to develop the regression models as described in Larson et al.
(2004), Stone et al. (2008), and Nowell et al. (2009). Model
parameters were estimated using maximum likelihood methods
implemented by the SURVREG procedure (Therneau, 1999)
in the statistical analysis program S-PLUS (TIBCO Software
Inc., 2008). Maximum likelihood computations require lower
and upper bounds be defined for left and interval censored
samples (Helsel, 2005). The response variable in this study represents the summed concentration of atrazine and DEA (after
both were adjusted to 100% analytical recovery, as described
earlier). For samples in which neither atrazine nor DEA were
detected, the response variable was left censored with a lower
bound at 0 and the upper bound established as the sum of
the LTMDLs (0.011 μg L−1). For samples in which atrazine or
DEA (but not both) was detected, the response variable was
interval censored with the lower bound established by summing concentrations with the assumption that the censored
value was equal to zero and the upper bound established by
summing concentrations with the assumption that the censored value was equal to its LTMDL. For a sample in which
atrazine concentration was reported as 0.009 μg L−1 and DEA
was not detected (and reported as less than the LTMDL of
0.007 μg L−1), the response variable would be considered interval censored, and the lower and upper bounds of the interval
would be 0.009 and 0.016 μg L−1, respectively.
Explanatory variables for each model were screened through
a stepwise, interval-censored Tobit regression procedure that
used the Akaike Information Criterion statistic to select variables. The Akaike Information Criterion statistic balances the
desire to explain as much variability as possible in the response
variable (here, atrazine residue concentration) with the
Table 1. Explanatory variables used for final model development.
Variable
use9200†
use97‡
pasthay§
agrcveg‡
agcc‡
pden90†
artdrn¶
awcup§
awcave†
omup§
wtdep¶
permin¶
bdup†
bd§
siltup†
runoff†
perdun†
recharge¶
welldepth‡
airtemp¶
wtotl§
wswto†
IS‡
RTI‡
pH‡
Description
Intensity of atrazine use
average use within 500-m radial area for 1992–2000 on row crops, small grains, fallow land, pastures,
hay, orchards, and vineyards (kg)/area of 500-m radial buffer (km2)
sum of use (1997) within 500-m radial area on row crops, small grains, fallow land, pastures, hay, orchards,
and vineyards (kg)/area of 500-m radial buffer (km2)
Land use and population
pasture and hay, % (Price et al., 2003)
agricultural row crops and vegetables (fraction)
agricultural cover crops (fraction)
population density from the 1990 census (people km−2) (U.S. Bureau of the Census, 1991)
Agricultural management practices
subsurface tile drains and surface trenches in 1-km2 grid cells (%) × 100
Soil and aquifer characteristics and well depth
average water-holding capacity of uppermost soil layer (fraction) (NRCS, 1994)
average water-holding capacity of soil column (fraction) (NRCS, 1994)
average organic matter content of uppermost soil layer, % w/w (NRCS, 1994)
average depth to seasonally high water table, m (NRCS, 1994)
permeability of least permeable layer, cm h−1 (NRCS, 1994)
average bulk density of uppermost soil layer, g cm−3 (NRCS, 1994)
average bulk density of soil column, g cm−3 (NRCS, 1994)
average silt content of uppermost soil layer, % (NRCS, 1994)
mean annual runoff (1951–1980), cm (Gebert et al., 1987)
Dunne overland flow from TOPMODEL, %
mean annual recharge, cm yr−1
depth of well, m
Climate
mean annual air temperature (1980–97), °C (Thorton, 2003)
Water use
withdrawals for irrigation (1995), total, Mgal d−1 (Solley et al., 1998)
withdrawals for irrigation (1995), surface water, Mgal d−1 (Solley et al., 1998)
Water quality
ionic strength
residence time indicator
pH, standard pH units
† Nationally available variables considered in the development of the national-variable model.
‡ Site-specific variables considered in the development of the site-variable model (Table 3).
§ Nationally available variables considered in the development of the site-variable model.
¶ Nationally available variables considered in the development of both the national-variable model and the site-variable model.
482
Journal of Environmental Quality • Volume 41 • March–April 2012
practicality of including as few explanatory variables as possible
(Helsel and Hirsch, 1992). Groups of explanatory variables were
sequentially added before each step of the stepwise procedure,
and variables were retained if their p values remained ≤0.25. A
p value of 0.25 was used at this stage of model development
to ensure that all explanatory variables that might contribute
significantly to a final model formulation would be considered.
Remaining variables were dropped from further consideration
if they were redundant (derived from the same ancillary data
set), highly correlated with one another (correlation coefficient
>0.75), or dependent (as indicated by a condition index near
30 and variance decomposition proportions >0.5) (Belsley et
al., 1980). In these cases, the variable with the lowest p value
was retained, and the redundant, correlated, or dependent variables were eliminated from further consideration.
Optimal model formulations were identified by maximizing
pseudo coefficient of determination (pR2) values (Laitila, 1993)
and minimizing the scale parameter. The pR2 and scale parameter are alternatives to the R2 and the standard deviation of
the residual error, respectively, which are used in conventional
least-squares analysis. As with conventional R2, the pR2 ranges
from 0 to 1 and is an estimate of the percentage of the variation in the response variable that is explained by the regression
model. Thus, a pR2 value of zero indicates that no variation
is explained, and a value of unity indicates that 100% of the
variation is explained. The pR2 was calculated using methods
described by Laitila (1993). Estimates of the standard deviation of the residual error from maximum-likelihood methods
provide only asymptotically unbiased estimates of the standard
deviation of the residual error when estimated from sample
data (Aitkin, 1981). These estimates, on average, underestimate the true standard deviation. The bias is a function of the
sample size and degree of censoring and is expected to be minimal for models with low percentages of censored observations.
In this report, biased estimates of the standard deviation of the
residual error are referred to as “scale.”
Before final model formulations were established, all explanatory variables that had previously been dropped were reevaluated
to confirm that they did not contribute significantly to the final
model (i.e., that they were associated with a p value >0.05 and/or
an increase in model pR2 of <1%). Finally, each pairwise combination of explanatory variables in the final models was considered as
a cross-product term to assess whether interaction of explanatory
variables contributed significantly to the final models.
Calculation of Prediction Intervals and Probabilities of Exceedance
Prediction intervals (PIs), which are a function of the fit of
a model and the amount of variability explained by the model,
represent the likelihood that a particular value for an individual
site falls within a specified interval of the predicted value. The
confidence level used for PIs in this study was 95%. Prediction
intervals can be used to estimate the probability that the actual
concentration of atrazine residues at a site exceeds a specified
concentration, as described in Nowell et al. (2009).
A PI is computed as follows:
PI = [log(Cest) ± SEest × tα/2]
[1]
where Cest is the estimated atrazine residue concentration, SEest
is the standard error of log(Cest), and tα/2 is the point on the
Student’s t distribution with a probability of exceedance of one
half of α. (For the purpose of this study, α = 0.05.)
As described previously, standard errors were estimated
from the scale parameter through an adjustment suggested by
Aitkin (1981). The effect of censoring of atrazine residue concentrations cannot be fully accounted for; therefore, PIs calculated for this study represent approximate values.
The probability (Pexc) that the actual concentration at a site
will exceed a specified concentration (Cb) is approximated as
follows:
⎡ log (Cb ) − log (Cest )⎤
⎦
Pexc = ⎣
SEest
[2]
Table 2. Loadings on the first two principal components from principal component analysis.
Constituent or physical property
Principal component interpretation (percentage of variance explained)
Component 1: Ionic strength (32%)
Specific conductance
−0.79
Dissolved oxygen
pH
Component 2: Residence time indicator (20%)
0.92†
0.38
-0.27
NH3
0.68
NO3−
−0.74
P
Ca2+
2+
0.73
Mg
0.76
Na+
0.76
K
0.57
Cl−
0.73
SO4
2−
0.83
Silica
0.44
Fe2+
0.79
Mn2+
0.81
† Numbers in italics denote loadings with absolute values greater than 0.60. Loadings with absolute values less than 0.25 are omitted.
www.agronomy.org • www.crops.org • www.soils.org
483
When the regression residual error is normally distributed, the
error distribution about an estimated value of the dependent
variable follows the Student’s t distribution. After antilogarithms have been computed, however, the error distribution
around the regression line is no longer symmetric, and the variance is no longer constant across the full range of the independent variable. As a result, the magnitude of the uncertainty in
Cest is greater for values above the estimated concentration than
for those below it and is greater for high concentrations than
for lower concentrations.
Results
groundwater beneath agricultural areas in this study (Table 3).
Thus, the inclusion of site-specific variables increased the percentage of variance explained by about 22% (increased from
41 to 50%). The following sections examine the physical significance of each of the explanatory variables listed in Table 3.
Explanatory Variables
The 10 explanatory variables used in the two models (Table
3) are grouped below into three categories depending on
whether they are associated with the sources, transport, or fate
of atrazine and DEA in groundwater.
Source Variables
Many models could be created from subsets of the final set
of explanatory variables (Table 1). The two optimal formulations that were obtained by maximizing pR2 and minimizing
scale, as described above, are presented in Table 3. One of the
models was constructed using explanatory variables whose
values are available nationwide (termed the “national-variable”
model), and the second model was derived from a combination
of nationwide and site-specific explanatory variables (termed
the “site-variable” model). Inclusion of any of the previously
discarded explanatory variables resulted in negligible improvement in the model fits, as indicated by an increase in pR2 of
<1%. Interaction between the variables artdrn (artificial drainage) and wtdep (depth to the seasonally high water table) was
found to be significant in the national-variable model and thus
required that both variables remain in the final model formulation (Kleinbaum and Klein, 2002), despite a p value for wtdep
that exceeded 0.05 (Table 3).
The optimal model formulations explain 41% (nationalvariable model) and 50% (site-variable model) of the variability
in the concentrations of atrazine residues measured in shallow
Three variables are related to amounts of atrazine applied
within the 500-m radial areas around the sampled wells. Two
of these variables are atrazine-use intensities (use9200 and
use97 in Table 1), and the third represents the organic-matter
(OM) content of the uppermost soil layer (omup in Table 1),
which, as explained below, has an influence on use.
The presence of atrazine in groundwater is assumed to result
solely from its application for agricultural purposes because it is
an anthropogenic compound used almost exclusively for agriculture in most parts of the United States and is thus unlikely
to have come from any other source. Furthermore, although
DEA may also be produced from the in situ dealkylation of
the herbicides propazine and cyprazine, the use of these two
compounds in the United States was negligible in the 1990s
(Gianessi and Marcelli, 2000). Thus, the source of atrazine residues at each site of interest in this study was represented by the
estimated intensity of atrazine use for agricultural purposes in
that location. The intensity of atrazine use was calculated as
the total amount of the herbicide applied to the crops of inter-
Table 3. Statistics for Tobit regression models based on nationally available explanatory variables (National-variable model) and nationally available
and site-specific explanatory variables (Site-variable model).
Variable
Coefficient
p value
Contribution to overall pR2†
by addition of variable
N
Overall
pR2
Scale
Censored data
Interval
Left censored
censored
——— % ———
wtdep
use92001/3
airtemp
artdrn1/3
artdrn1/3 * wtdep
log10(1 + recharge)
permin1/3
intercept
0.15
0.46
−0.05
−0.1
0.04
0.62
0.23
−3.0
0.11
<0.0001
<0.0001
<0.0001
0.0002
<0.0001
<0.0001
RTI
log10(1 + use97)
artdrn1/3
wtdep
log10(1 + omup)
permin1/3
log10(1 + recharge)
log10(1 + welldepth)
intercept
−0.42
0.69
−0.06
0.36
1.18
0.21
0.38
−0.24
−3.64
<0.0001
<0.0001
<0.0001
<0.0001
<0.0001
<0.0001
0.0002
0.008
National-variable model
0.13
1298
0.09
0.08
0.06
0.03
0.01
0.01
Site-variable model
0.25
0.095
0.045
0.04
0.03
0.02
0.01
0.01
1272
0.41
0.82
51.6
10.2
0.50
0.76
50.9
10.4
† Pseudo coefficient of determination.
484
Journal of Environmental Quality • Volume 41 • March–April 2012
est divided by the area of the 500-m radius around each well
(Table 1) (Thelin and Gianessi, 2000; USGS, 2010).
The variable for intensity of atrazine use (use9200 or use97)
was the second most important with respect to the incremental
amount of variability explained by the addition of these variables to each model (Table 3). However, the amount of variability explained by use9200 or use97 for atrazine residues in
groundwater was markedly less than the amount explained by
similar watershed-based, use-intensity estimates in models of
atrazine in stream water. Indeed, use intensity has been found
to be the most important explanatory variable in regression
models developed to estimate atrazine concentrations in rivers
and streams throughout the United States (Larson et al., 2004;
Stone et al., 2008; Stone and Gilliom, 2009). Pesticide-use
intensity accounted for as much as 64% of the spatial variations in pesticide concentrations explained by these models
for stream water, in contrast to the 9% of observed variability in groundwater concentrations accounted for by atrazineuse intensity in this study (Table 3) and the 7% of observed
variability in atrazine-detection frequencies in groundwater accounted for by atrazine-use intensity in the study by
Stackelberg et al. (2006a).
The greater explanatory power of pesticide-use intensity for
stream concentrations compared with groundwater concentrations is consistent with the difference between contaminant
transport and transformation processes in the two media.
Surface waters tend to integrate water chemistry over space and
time to a much greater extent than does groundwater because
they exhibit relatively short flow paths, brief soil contact time,
and rapid mixing. As a result, the concentrations of dissolved
contaminants in surface waters respond rapidly to changing
conditions within the watershed. In contrast, groundwater,
which generally represents water entering the subsurface at a
particular location and time (especially water from monitoring wells and other low-capacity wells), typically shows a more
delayed response to changing conditions due to greater soil
contact time as it percolates through the unsaturated zone to
the water table and through the saturated zone to discharge
areas (Kauffman et al., 2001). The longer residence times associated with groundwater systems than with surface runoff and
the greater opportunity for physical, biological, and chemical
effects make groundwater chemistry more difficult, compared
with surface-water chemistry, to correlate with human activities (such as pesticide use) at land surface.
An unexpected observation from this study was that atrazine residues in groundwater were positively correlated with
the OM content of the uppermost soil layer omup (Table
3). This is counter to the finding by other studies that lower
amounts of soil OM tend to be associated with lower atrazine
degradation rates (e.g., Fenner et al., 2007) and less sorption of
organic compounds (Hamaker and Thompson, 1972), both of
which should allow higher concentrations of atrazine residues
to reach shallow groundwater. The intensity of atrazine use,
however, depends in part on the soil characteristics of a given
area. Loamy soils, which contain higher amounts of OM than
sandy soils and are generally preferred for the cultivation of
most crops, typically receive 25 to 50% greater applications of
atrazine per unit area than sandy soils (Wisconsin Department
of Agriculture, Trade and Consumer Protection, 2010). The
www.agronomy.org • www.crops.org • www.soils.org
significant correlation between omup and use97 may therefore
reflect an elevated use of atrazine in loamy areas (Spearman ρ
= 0.39; p < 0.0001; n = 1298), which may explain why omup
is a significant explanatory variable in the site-variable model
(Table 3). Inclusion of omup in the model appears to refine
source estimates by identifying preferred areas for crop cultivation and indicating where atrazine might have been applied
at higher rates. It is unlikely that OM content reflects preferential flow processes because it was found to be only weakly
correlated with clay content (Spearman ρ = 0.17; p < 0.0001;
n = 1298), which has been found to be associated with structured soils and the rapid, downward transport of contaminants
through preferential flow paths within the subsurface (e.g.,
Jarvis et al., 2009).
Transport Variables
Five variables are related to factors that may affect the transport of atrazine residues to the water table and to the sampled
well. These are (i) prevalence of artificial drainage (artdrn in
Table 1), (ii) depth of the seasonally high water table (wtdep in
Table 1), (iii) rate of recharge (recharge in Table 1), (iv) permeability of the least permeable soil layer (permin in Table 1), and
(v) depth of the well (welldepth in Table 1).
Artificial drainage (artdrn) and depth of the seasonally high
water table (wtdep) interact significantly in the national-variable model to affect the concentrations of atrazine residues in
shallow groundwater (Table 3). The water table beneath poorly
drained agricultural soils tends to be close to the land surface
during the growing season, often requiring the use of artificial
drainage systems for reducing soil water content to facilitate
agricultural activities. In this study, atrazine residue concentration was positively correlated with depth to the seasonally
high water table and was inversely correlated with artificial
drainage. This pattern reflects the efficacy of surface ditches
and subsurface drains for capturing soil water and diverting
it to streams rather than allowing it to enter the groundwater system. This diversion reduces the transport of atrazine to
groundwater underlying the drained soils and thus may result
in concentrations of atrazine residues in shallow groundwater beneath poorly drained areas that are lower than in well
drained areas that do not require drainage systems; this pattern has been observed in previous studies (e.g., Keim et al.,
1989; Kladivko et al., 2001). The significant interaction that
was observed between artdrn and wtdep reflects the fact that
artificial drainage systems are typically installed only in areas
where soils are poorly drained.
The permeability of the least permeable soil layer, permin,
and the rate of recharge, recharge, were positively correlated
with atrazine residue concentrations (Table 3). This indicates
that, as expected, high soil permeability and large water flux
through the unsaturated zone favor the transport of atrazine
residues to the water table.
The depth of wells, welldepth, was inversely correlated with
atrazine residue concentration (Table 3), indicating that shallower wells tended to have higher concentrations than deeper
wells. Because the wells used in this study were selected or
installed with the goal of providing samples of “recently
recharged” (shallow, young) groundwater, the screened or open
interval in most of the wells is located a short distance below
485
the water table. As a result, the median distance from the water
table to the bottom of the wells was relatively shallow (4.6 m).
Given the fact that the wells were generally installed a short distance below the water table, well depth served as a surrogate for,
and was significantly correlated with, unsaturated-zone thickness (Spearman ρ = 0.80; p < 0.0001; n = 1267). The thickness
of the unsaturated zone is directly related to the amount of
time available for atrazine residue concentrations to be diminished through sorption or degradation (Barbash and Resek,
1996; Tesoriero et al., 2007). The inverse correlation observed
between atrazine residue concentrations and welldepth during
this study thus appears to support the hypothesis that the concentrations of atrazine and DEA are likely to diminish with
increasing residence time within the unsaturated zone.
Fate Variables
Two variables reflect influences of fate processes in the subsurface: (i) residence-time indicator (RTI in Table 1), and (ii)
mean annual air temperature (airtemp in Table 1).
Of the explanatory variables considered in the present study,
RTI is by far the most powerful single variable for predicting
the concentration of atrazine residues in groundwater beneath
agricultural areas. The incremental inclusion of this variable in
the site-variable model accounted for 25% of the variability in
atrazine residue concentrations, representing half of the total
variability in atrazine residue concentrations accounted for by
the full site-variable model (50%) (Table 3). The variable RTI
was developed through a PCA of water quality data (Table 2)
and provides an indication of the effect of time on geochemical
conditions along groundwater flowpaths in the saturated zone.
The PC identified as RTI was significantly correlated with the
concentrations of several redox-active solutes—negatively with
NO3− and dissolved oxygen (DO) and positively with Fe2+,
Mn2+, and NH3 (Table 2). Negative RTI values thus represent
oxic conditions, and positive values represent reducing conditions. The relation to time derives from the gradual depletion
of DO through microbial respiration, the decrease of NO3−
through denitrification, and the increase in concentration of
reduced species (i.e., Fe2+, Mn2+, and NH3) as groundwater
becomes more reducing with increasing residence time in the
saturated zone.
Although RTI was derived from redox-active solutes, it does
not appear to reflect the effect of redox conditions on the persistence of atrazine residues. Results from several laboratory
investigations indicate that atrazine is less persistent under oxic
conditions than under reducing conditions (e.g., Kaufman and
Kearney, 1970; Nair and Schnoor, 1992; Rügge et al., 1999;
Larsen et al., 2000; Accinelli et al., 2001), although field evidence has provided mixed results (Barbash and Resek, 1996).
For example, Druliner (1989) observed an inverse correlation between atrazine and DO concentrations in groundwater
from the High Plains aquifer in Nebraska, which is consistent
with the above-cited laboratory study results. Similarly, atrazine has been reported to be relatively unreactive under reducing conditions in situ (e.g., Spillmann, 1989; Papiernik and
Spalding, 1998; Rügge et al., 1999). In contrast, four national
or regional monitoring programs have reported positive correlations between atrazine and DO concentrations (or other
measures of oxidizing conditions) in groundwater (Kolpin et
486
al., 1997, 2002; Ator, 2008; Saad, 2008), as was observed in
this study. An explanation for this disparity was provided by
Kolpin et al. (1997, 2002), who suggested that, within the
saturated zone, DO concentration is a surrogate for residence
time because oxygen tends to be consumed through biotic and
abiotic processes as water moves downgradient through the
saturated zone. Thus, a positive correlation between atrazine
concentrations and DO concentrations reflects a simultaneous loss of both solutes as water moves downgradient, rather
than a causal relation between redox condition and atrazine
concentration. Kolpin et al. (1997, 2002) were unable to test
this hypothesis, however, because estimates of groundwater age
were not available at the time of their study. More recent studies that have found statistically significant correlations between
DO concentrations in groundwater and residence time in the
saturated zone have also observed inverse relations between the
two variables (Tesoriero et al., 2005, 2007; Saad, 2008; Denver
et al., 2010). This is consistent with the hypothesis that, within
the saturated zone, DO concentration may often be viewed as
a surrogate for residence time. The present study obtained estimates of groundwater age for 259 samples through an analysis of the concentrations of chlorofluorocarbons, SF6, and/or
3
H/3He (Hinkle et al., 2010) (246 of these samples were in the
model-development data set; the remaining 13 were not used
for model development because they lacked values for one or
more of the explanatory variables). These age estimates were
found to be positively correlated with RTI values (Spearman ρ
= 0.46; p < 0.0001; n = 259). The correlation between RTI and
atrazine residue concentrations therefore appears to be primarily a reflection of the influence of residence time in the saturated zone on atrazine and DEA concentrations rather than the
effect of redox conditions on the persistence of the two solutes.
The explanatory power of the RTI appears to derive from
the fact that the residence time of sampled groundwater represents the combined influence of factors that can affect the concentrations of atrazine residues in groundwater through time.
Another time-related factor that has been shown to be correlated with the concentrations of atrazine residues in groundwater is historical changes in atrazine use intensity (Tesoriero
et al., 2007). The relation between atrazine residue concentration and the estimated year in which the sampled groundwater
recharged the groundwater system is depicted in Fig. 2, along
with boxplots showing the range and distribution of atrazine
concentrations for each decade from 1950 through 2000. A
bivariate Tobit model indicated a statistically significant relation between atrazine residue concentration and the estimated
year of recharge (pR2 = 0.15; p < 0.0001; scale = 0.93; n = 259),
suggesting that historical changes in atrazine use intensity may
explain some of the lower concentrations of atrazine residues in
older samples, relative to those in younger samples (Fig. 2). To
evaluate the possibility that factors other than historical changes
in atrazine use intensity (such as degradation, adsorption, and
dispersion) diminish the concentration of atrazine residues
through time, a second bivariate Tobit model was developed
from only those samples of groundwater that entered the system
between 1975 and 2000 (n = 208), when national atrazine use
was relatively stable, varying between 30 and 40 million kg
yr−1 (Fig. 2). This model was significant at the 95% confidence
level (p = 0.05) but exhibited a weak correlation (pR2 = 0.02),
Journal of Environmental Quality • Volume 41 • March–April 2012
revealing little evidence that atrazine
residue concentrations are diminished
in shallow groundwater by processes
occurring in the saturated zone such as
degradation, adsorption, or dispersion.
This conclusion is consistent with the
findings from several laboratory studies that have observed negligible rates of
atrazine mineralization in saturated, oxic
aquifer sediments (e.g., Konopka and
Turco, 1991; McMahon et al., 1992).
Similarly, the mole fraction of atrazine
plus DEA that was present as DEA did
not vary as a function of groundwater
age, in agreement with Tesoriero et al.
(2007). These findings indicate that, at
the national scale, residence time in the
saturated zone is predictive of atrazine
concentrations in groundwater primarily because it reflects temporal changes
in the concentration of atrazine residues Fig. 2. Concentrations of atrazine residues in shallow groundwater beneath agricultural areas
in recharge rather than the influence of throughout the conterminous United States, 1945–2005; national atrazine-use totals, 1964–2004;
and boxplots of concentration distributions for each decade, 1950–2000. n, number of samples;
redox controls on the persistence of these DF, detection frequency (%).
compounds or the effects of processes
such as degradation, adsorption, or distively. For a prediction with a censored observed concentration,
persion within the saturated zone.
the residual was computed in the following manner (Gregory
Mean annual air temperature (airtemp) during 1980–1997
Schwarz, USGS, written communication, 2010):
was inversely correlated with the concentration of atrazine resi⎡ ⎛ log(LTMDL i ) − log(C ) ⎞⎤
dues (Table 3). Air temperature has been shown to correlate
est ⎟
i
⎟⎟⎥⎥
[3]
ec = σs Φ−1 ⎢⎢μΦ⎜⎜⎜
with soil temperature at the national scale (Toy et al., 1978;
⎟⎥
⎜
σ
⎝
⎠
s
⎣⎢
⎦
Zheng et al., 1993), indicating that airtemp may be viewed as
a surrogate for soil temperature. As with all nonphotochemical
where σs is the scale parameter estimated from the Tobit
reactions, the rate of atrazine transformation in soil decreases
model, Φ−1 is the inverse of the cumulative normal integral,
with decreasing temperature (e.g., Rocha and Walker, 1995), as
μ is a random number between zero and unity drawn from
does volatilization from soil (Thomas, 1990). Consequently, if
a uniform (equiprobable) distribution, and LTMDL and Cest
all other factors are held constant, atrazine and DEA are more
are as defined earlier. The values of ec are randomly generated,
likely to persist long enough to reach groundwater if the atranormally distributed residuals that adhere to the lower and
zine is applied in cool climates or at cool times of the year than
upper bounds of the censored observation and conform with
if the herbicide is applied in warm climates or during warm
the distribution of the residuals used to obtain the Tobit model
seasons. Several authors have suggested that the effect of temestimates. These residuals were added to the predicted values
perature may be reinforced by the effects of recharge because
for censored samples to obtain estimates of the observed concold seasons often coincide with periods of greater precipitacentrations for assessing model performance.
tion than warm seasons (e.g., Wyman et al., 1985; Jones et al.,
Model Performance
1986; Priddle et al., 1989, 1992; Pickett et al., 1990; Porter
et al., 1990; Traub-Eberhard et al., 1994). However, the presWhen all sites were examined, both models were found
ent study did not observe a statistically significant interacto be unbiased, with residuals that were normally distributed
tion between airtemp and recharge (p = 0.94), indicating that
about zero across all predicted concentration values (Fig. 3A,
the effects of temperature on atrazine residue concentrations
3B). Boxplots of the residuals from both models (Fig. 3C, 3D)
in groundwater are independent of the amount of recharge
suggest a similar conclusion when all sites were considered,
received.
but both models tended to overpredict concentrations at sites
where neither atrazine nor DEA were detected (left-censored
Model Performance and Evaluation
samples). Residuals from these sites were based on synthetic
Residual errors were calculated by subtracting the logaestimates rather than measured concentrations. At sites where
rithm of the predicted atrazine residue concentration for each
both atrazine and DEA were detected (uncensored samples),
site from the logarithm of the observed value. A residual error
both models tended to underpredict concentrations (Fig. 3C,
of less than zero thus indicated overprediction, and a residual
3D). Observed atrazine residue concentrations are plotted in
error greater than zero indicated underprediction. Residual
relation to predicted concentrations in Fig. 3E and 3F, which
values of −1 and +1 indicate predicted concentrations that are
indicate that 77% of the concentrations predicted by the
10 times and one tenth of the observed concentration, respecnational-variable model (Fig. 3E) and 81% of those predicted
www.agronomy.org • www.crops.org • www.soils.org
487
atrazine residue concentrations at the
model evaluation sites. However, the
degree of underprediction (Fig. 4C,
4D) was less than that observed for
the model-development sites (Fig.
3C, 3D). Overall, more than 70%
of the predicted concentrations for
either model were within a factor
of 10 of the observed values at the
model-evaluation sites (Fig. 4E, 4F).
Uncertainty in Model Predictions
Uncertainty in a predicted concentration is expressed in terms of the PI
(Eq. [1]) in Fig. 5, which shows the
atrazine residue concentrations for
the model-development sites ordered
by increasing predicted concentration. Comparison of Fig. 5A and 5B
indicates that the PIs are larger for the
national-variable model than they are
for the site-variable model. This was
expected because PIs are a function
of the amount of variability explained
by each model, and the site-variable
model explained a higher proportion
of variability in the observed atrazine
residue concentrations than did the
national-variable model (Table 3).
The levels of uncertainty in model
predictions are also compared in Fig.
6, in which the size of the PI at each
site is represented as the ratio of the
upper PI to the predicted atrazine residue concentration (which is the same
as the ratio of the predicted atrazine
residue concentration to the lower
PI). Outliers in Fig. 6 represent sites
where one or more explanatory variables were relatively extreme in value
Fig. 3. Distribution of residuals and observed and predicted atrazine residue concentrations for
when compared with values from the
model-development sites. All, all samples; IC, interval censored; LC, left censored; UC, uncensored.
rest of the sites. These outlier sites,
therefore, exhibited greater uncerby the site-variable model (Fig. 3F) were within a factor of 10
tainty
in
their
predicted
atrazine residue concentrations than
of the observed values.
did the others. Figure 6 also indicates that, as anticipated from
Model Evaluation
the results presented in Table 3 and Fig. 5, predictions from
The models were applied to the independent set of 209
the national-variable model exhibited greater uncertainty than
agricultural sites described earlier (Fig. 1), and model predicthose from the site-variable model.
tions and residuals were computed. Residuals for these sites
were normally distributed but, in general, are biased toward
Model Limitations
overprediction (Fig. 4A–4D). The bias was primarily caused
The amount of variability in atrazine residue concentraby overprediction of left-censored samples for which synthetic
tions explained by the optimal Tobit model that used the siteresiduals were calculated (as noted earlier). The source of this
specific variables was substantial. However, the fact that 50%
overprediction is unknown but is not likely to have been overof the observed variability in these concentrations remains
estimates in atrazine use because atrazine use appears to have
unexplained (Table 3) indicates that important factors related
been relatively stable from 1992 to 2006 (Fig. 2).
to the sources, transport, or fate of atrazine may not be well
As was seen among the model development sites (Fig. 3C,
represented by the explanatory variables for which data are
3D), for locations where both atrazine and DEA were detected
available on a national basis. In addition, the national scale
(uncensored samples), the model tended to underpredict the
at which these models were developed could result in erro488
Journal of Environmental Quality • Volume 41 • March–April 2012
neous predictions if the models are
applied at more local scales because
they may not adequately reflect local
variations in some of the explanatory
variables that strongly affect atrazine
residue concentrations in shallow
groundwater. Local factors that may
affect the transport of atrazine and
DEA to wells require consideration
before results from these models
can be interpreted for any particular location. Also, the application
of the models to sites with fundamental differences in hydrogeologic
conditions or to sites whose values
for explanatory variables differ from
those used for model development
will result in greater uncertainty in
model predictions.
In this study, atrazine use estimates
were for agricultural purposes only
because estimates for nonagricultural
use of pesticides are not available
in most states. Although atrazine is
known to have nonagricultural uses,
they are considered to be minor relative to agricultural uses, especially
in agricultural settings. Uncertainty
arises, however, when examining correlations between atrazine residue
concentrations in groundwater and
estimates of atrazine use that were
averaged for periods during which use
was known to have changed, particularly for sites where groundwater age
estimates are not available.
The observed uncertainty in the
predictions from the models constructed for this study, as characterized in the previous section, indicate
that these models should not be used Fig. 4. Distribution of residuals and observed and predicted atrazine residue concentrations for
to characterize absolute concentra- model-evaluation sites. All, all samples; IC, interval censored; LC, left censored; UC, uncensored.
tions at specific locations. Instead,
represented at least 50% of the total area. Only 4% of these
the models should be used for identigridded areas had values for one or more explanatory variables
fying areas where atrazine residues are likely to be highest and
that were outside the range of values used in the model develmay therefore merit further investigation.
opment dataset. This observation illustrates the high degree to
which the groundwater quality assessment conducted by the
Model Applications
NAWQA Program encompassed conditions that were represenThe national- and site-variable models (Table 3) were used
tative of groundwater settings across the country. Virtually all
to make two types of predictions for agricultural areas throughof the explanatory variable values that were outside the model
out the conterminous United States. The national-variable
development dataset range were mean annual air temperature
model was applied to predict concentrations in groundwater
(airtemp) values—primarily in small areas of northern North
with a range of characteristics and residence times similar to
Dakota and northwestern Minnesota where airtemp values are
those in the model-development data set. By contrast, the sitelower
than the minimum value used for model development.
variable model was used to predict concentrations in recently
Because
only a small portion of the agricultural areas in the
recharged groundwater by setting the RTI variable to the low
United
States
was affected, the overall effect of the airtemp outend of its range. The models were applied by quantifying
liers
on
the
model
estimates described in the following section
2
explanatory variables for every 1-km grid cell of a gridded map
was
likely
to
have
been
minimal.
of the conterminous United States in which agricultural land
www.agronomy.org • www.crops.org • www.soils.org
489
Fig. 5. Prediction intervals for atrazine residue concentration statistics predicted by (A) the Nationalvariable model and (B) the Site-variable model.
National-Variable Model
The national-variable model predicted atrazine residue
concentrations in groundwater to be less than the sum of
their LTMDL values (0.011 μg L−1) in 53% of the gridded agricultural areas across the United States (Fig. 7A). The
highest predicted concentrations were in parts of the High
Plains aquifer system in Nebraska, Kansas, and Texas; the
Driftless Area in Illinois, Iowa, Minnesota, and Wisconsin;
and in southeastern Pennsylvania (Fig. 7A). Although these
are areas where atrazine is known to be used, they do not
generally represent the locations with the most intensive
use of atrazine (Fig. 1). This suggests that, as noted previously by Barbash et al. (1999), Kolpin et al. (2002), and
Stackelberg et al. (2006a), factors other than application
intensity appear to exert a substantial effect on the distributions and concentrations of atrazine residues in shallow groundwaters beneath agricultural areas. Such factors
include soil conditions, climate, recharge rates, and agricultural management practices. For example, some of the
most intensive atrazine use during the study period was in
parts of Iowa, Illinois, and Indiana, where soils tends to be
poorly drained and where agricultural fields commonly have
subsurface tile drains that capture soil water and divert it to
Fig. 6. Potential prediction errors among concentration statistics
predicted using the National- and Site-variable models.
490
nearby streams. Consequently, the
concentrations of atrazine residues
in groundwater would be expected
to be lower in these locations than
in areas with conditions that did
not require tile drains, as suggested
by previous researchers (e.g., Keim
et al., 1989; Kladivko et al., 2001).
In contrast, the national-variable
model (Table 3) predicts that areas
with extensive atrazine use on soils
that do not require artificial drainage (e.g., parts of the High Plains
aquifer and Driftless Area and
southeastern Pennsylvania) would
have the highest concentrations of
atrazine residues in shallow groundwater (Fig. 7A).
Site-Variable Model
Atrazine-residue concentrations predicted by the nationalvariable model were expected to be lower than the actual
concentrations in recently recharged, young groundwater
because some of the samples used for model development had
a decades-long residence time that predated the period of most
intensive atrazine use, which began in the mid-1970s (Fig. 2).
Predictions from the site-variable model cannot be obtained at
unmonitored locations without assuming values for RTI and
welldepth because national datasets for these variables are not
available. To address this limitation, estimates of the concentration of atrazine residues in recently recharged groundwater
were obtained using specific values for RTI and welldepth in
the site-variable model. As previously described, RTI provides
an indication of residence time in the saturated zone, where
low values represent oxic, young groundwater and higher
values represent older, more reduced water. Assigning the fifth
percentile of RTI and the median depth of wells (welldepth)
to all areas made it possible to use the site-variable model to
predict atrazine concentrations for young, oxic groundwater
drawn from shallow wells. The resulting atrazine residue concentrations in groundwater were lower than the sum of their
LTMDLs (0.011 μg L−1) in 10% of the agricultural areas across
the United States (Fig. 7B), a percentage that was considerably lower than the estimate of 53% for the national-variable
model. Predictions from the site-variable model are appropriate for comparison with those from process-based models of
atrazine transport through the unsaturated zone that are not
calibrated to field observations (e.g., Goss et al., 1998; Barbash
and Voss, 2007). Such predictions are also useful for providing initial atrazine residue concentrations at the water table for
models of atrazine transport and fate in the saturated zone. In
the absence of a predictive model developed from supply-well
data, predictions from the site-variable model may also provide
conservative estimates of the potential concentrations of atrazine residues in deeper groundwater supplies that may be used
for drinking. This is supported by the finding, noted here and
by previous studies, that processes such as degradation, adsorption, or dispersion in the saturated zone do not appear to cause
substantial decreases in atrazine residue concentrations over
Journal of Environmental Quality • Volume 41 • March–April 2012
time in groundwater underlying agricultural settings.
Predicting the Probability of Exceeding
a Specific Concentration
The national- and site-variable models
were also used to predict the probability
that atrazine residues might exceed a
specified concentration in groundwater beneath agricultural areas across the
conterminous United States. The concentration thresholds considered were
3.0 and 0.3 μg L−1. The 3.0 μg L−1 value
was selected as an exceedance threshold
because it represents the maximum contaminant level established by the USEPA
for atrazine in drinking water (USEPA,
2003). The 0.3 μg L−1 value was selected
to identify areas that may warrant further
evaluation if groundwater is to be used
as a drinking water supply. The USEPA
has not established a regulatory concentration for DEA in drinking water supplies but has ascertained that atrazine
and DEA share a common mechanism
of toxicity and that their binary mixture
Fig. 7. Predicted atrazine residue concentration, and probability of exceeding 0.3 μg L−1 in shallow
therefore would have additive toxicity groundwater
underlying agricultural settings across the conterminous United States.
(USEPA, 2006).
Results from this analysis using
the national-variable model indicate
that atrazine residue concentrations in
groundwater with characteristics and
residence times similar to those examined for the model-development data set
have greater than a 10% probability of
exceeding the 0.3 and 3.0 μg L−1 thresholds in 25% (Fig. 7C) and 2% (Fig. 8A)
of the agricultural areas, respectively. In
contrast, the results from this analysis
using the site-variable model indicate
that atrazine residue concentrations in
recently recharged, young groundwater Fig. 8. Predicted probability of exceeding 3.0 μg L−1 in shallow groundwater underlying agriculhave a considerably greater probability tural settings across the conterminous United States.
of exceeding these thresholds: 60% of
50% or greater probability of exceeding a concentration of 0.3
agricultural areas are predicted to have
−1
μg L−1 in 16 and 9% of agricultural areas, respectively. In cona greater than 10% probability of exceeding 0.3 μg L (Fig.
trast, 12% and 0.4% of agricultural areas had a 50% or greater
7D), while 5% are predicted to have a greater than 10% prob−1
probability of exceeding a concentration of 0.3 μg L−1 when
ability of exceeding 3.0 μg L (Fig. 8B). In general, only in
using the 10th and 90th percentile values, respectively, for RTI.
the High Plains aquifer, the Driftless Area, and southeastern
Viewed with respect to the overall ranges spanned by these two
Pennsylvania do atrazine residue concentrations have a greater
−1
parameters, predictions from the site-variable model are thus
than 10% probability of exceeding 0.3 μg L . Overall, howmore sensitive to variations in RTI than they are to variations
ever, these results indicate that, in more than 90% of the agriin welldepth.
cultural areas across the conterminous United States, atrazine
residue concentrations have less than a 50% probability of
Conclusions
exceeding 0.3 μg L−1 (Fig. 7C, 7D).
Tobit regression models developed from explanatory variThe sensitivity of predictions from the site-variable model
ables
whose values are available nationwide (the nationalto variations in welldepth and RTI was evaluated by adjusting
variable
model) or from a combination of nationwide and
the values of these parameters. Application of the model using
site-specifi
c explanatory variables whose values are available
the 10th and 90th percentile values for welldepth revealed a
www.agronomy.org • www.crops.org • www.soils.org
491
only for the locations at which they were measured (the sitevariable model) were found to explain up to half of the variability observed in the concentrations of atrazine and DEA
in groundwater beneath agricultural areas across the conterminous United States. Explanatory variables that were found
to be statistically significant were related to three factors that
control atrazine residue (atrazine plus DEA) concentrations in
groundwater: atrazine sources, transport processes, and degradation. The inclusion of explanatory variables derived from sitespecific information (such as water chemistry data) improved
predictive capability and suggested that, at the national scale,
historical changes in use intensity have a greater effect on the
distribution of atrazine residue concentrations in groundwater
underlying agricultural settings than do the processes that control their dispersion, sorption, or degradation in the saturated
zone. Results from this study indicate that the highest atrazine
residue concentrations in groundwater underlying agricultural
areas are likely to be in parts of the High Plains aquifer system,
the Driftless Area of Illinois, Iowa, Wisconsin, and Minnesota
and southeastern Pennsylvania. The degree of explanatory
power exhibited by these newly developed models is notable,
given the inherent difficulties associated with correlating the
concentrations of pesticide compounds in a single groundwater sample with the land use, hydrologic, physical, biological,
and chemical factors that can affect the concentrations of the
compounds between their point of entry at the land surface (or
formation in the subsurface) and the screened or open interval
of the well from which they may be extracted.
These limitations notwithstanding, the Tobit models
described here represent an approach that provides the most
reliable predictions of atrazine residue concentrations in
shallow agricultural groundwater across the United States.
Predicted concentrations in individual wells are generally of
greater value than network detection frequencies (Kolpin et al.,
2002; Stackelberg et al., 2006a) because they can be directly
compared with water quality benchmarks or with health-based
screening levels (Toccalino and Norman, 2006).
The site-variable model allowed predictions to focus on
recently recharged (young) groundwater that is representative
of water quality conditions near the water table. Such predictions may be useful for comparison with similar predictions
from process-based models of transport through the unsaturated zone that are not calibrated to field observations and for
providing initial concentration estimates at the water table for
models of atrazine transport through the saturated zone. By
focusing on recently recharged groundwater, the predictions
from the site-variable model provide conservative estimates
of the concentrations of atrazine residues in deeper portions
of the aquifer that are being used, or considered, for drinking
water supplies. Such comparisons can provide an early indication to water resource managers of whether or not pesticide
concentrations in source waters in certain areas may warrant
further study. The Tobit models can also be used to identify
areas where the concentrations of atrazine residues, or their
probability of exceeding specified concentrations, are likely to
be relatively high or low. Such predictions can be used for the
design of future monitoring programs or for setting priorities
among different areas for further research.
492
References
Accinelli, C., G. Dinelli, A. Vicari, and P. Catizone. 2001. Atrazine and
metolachlor degradation in subsoils. Biol. Fertil. Soils 33:495–500.
doi:10.1007/s003740100358
Aitkin, M. 1981. A note on the regression of censored data. Technometrics
23:161–163. doi:10.2307/1268032
Ator, S.W. 2008. Natural and human influences on water quality in a shallow
regional unconsolidated aquifer, Northern Atlantic Coastal Plain. USGS
Investigations Rep. 2008-5190. USGS, Reston, VA.
Barbash, J.E., and E.A. Resek. 1996. Pesticides in ground water: Distribution,
trends, and governing factors. Lewis, Boca Raton, FL.
Barbash, J.E., G.P. Thelin, D.W. Kolpin, and R.J. Gilliom. 1999, Distribution
of major herbicides in ground water of the United States. U.S. Geological Survey Water-Resources Investigations Report 98-4245. USGS,
Reston, VA.
Barbash, J.E., and F.D. Voss. 2007. Use of a geographic information system
with process-based simulation modeling to predict atrazine concentrations in shallow groundwater across the United States: Simulation approach and testing against nationwide observations. 233rd American
Chemical Society National Meeting, Chicago, IL, 27 Mar. 2007. Abstract available at http://oasys2.confex.com/acs/233nm/techprogram/
P1057615.HTM (verified 8 Dec. 2011).
Belsley, D.A., E. Kuh, and R.E. Welsch. 1980. Regression diagnostics. John
Wiley & Sons, New York.
Denver, J.M., A.J. Tesoriero, and J.R. Barbaro. 2010. Trends and transformation of nutrients and pesticides in a Coastal Plain aquifer system, United
States. J. Environ. Qual. 39:154–167. doi:10.2134/jeq2009.0107
Druliner, A.D. 1989. Overview of the relations of nonpoint-source agricultural chemical contamination to local hydrogeologic, soil, land-use, and
hydrochemical characteristics of the High Plains Aquifer of Nebraska. p.
411–435. In G.E. Mallard ed. U.S. Geological Survey Toxic Substances
Hydrology Program: Proceedings of the technical meeting, Phoenix, AZ,
Sept. 26–30, 1988: U.S. Geological Survey Water-Resources Investigations Report 88–4220.USGS, Reston, VA.
Dubrovsky, N.M., K.R. Burow, G.M. Clark, J.M. Gronberg, P.A. Hamilton,
K.J. Hitt, D.K. Mueller, M.D. Munn, B.T. Nolan, L.J. Puckett, M.G.
Rupert, T.M. Short, N.E. Spahr, L.A. Sprague, and W.G. Wilber. 2010.
The quality of our Nation’s waters: Nutrients in the Nation’s streams
and groundwater, 1992–2004: U.S. Geological Survey Circular 1350.
USGS, Reston, VA.
Fenner, K., V.A. Lanz, M. Scheringer, and M.E. Borsuk. 2007. Relating atrazine degradation rate in soil to environmental conditions: Implications for global fate modeling. Environ. Sci. Technol. 41:2840–2846.
doi:10.1021/es061923i
Gebert, W.A., D.J. Graczyk, and W.R. Krug. 1987. Average annual runoff in
the United States: U.S. Geological Survey Hydrologic Atlas 710. USGS,
Reston, VA.
Gianessi, L.P., and M.B. Marcelli. 2000. Pesticide use in U.S. crop production:
1997 National summary report. National Center for Food and Agricultural Policy, Washington, DC.
Gilliom, R.J., J.E. Barbash, C.G. Crawford, P.A. Hamilton, J.D. Martin, N.
Nakagaki, L.H. Nowell, J.C. Scott, P.E. Stackelberg, G.P. Thelin, and
D.M. Wolock. 2006. Pesticides in the Nation’s streams and ground water, 1992–2001. U.S. Geological Survey Circular 1291. USGS, Reston,
VA.
Goss, D.W., R.L. Kellog, J. Sanabria, S. Wallace, and W. Kniesel. 1998. The
National Pesticide Loss Database: A tool for management of large watersheds. Poster presented at 53rd annual meeting of the Soil and Water
Conservation Society, San Diego, CA, 5–9 July 1998.
Hamaker, J.W., and J.M. Thompson. 1972. Adsorption. p. 49–143. In C.A.I.
Goring and J.W. Hamaker (ed.) Organic chemicals in the soil environment. Vol. 1. Marcel Dekker, New York.
Helsel, D.R., and R.M. Hirsch. 1992. Statistical methods in water resources.
Elsevier Science, Amsterdam, the Netherlands.
Helsel, D.R. 2005. Nondetects and data analysis: Statistics for censored environmental data. John Wiley & Sons, Hoboken, NJ.
Hinkle, S.R., S.D. Shapiro, L.N. Plummer, E. Busenberg, P.K. Widman, G.C.
Casile, and J.E. Wayland. 2010. Estimates of tracer-based piston-flow
ages of groundwater from selected sites: National Water-Quality Assessment Program, 1992–2005. U.S. Geological Survey Scientific Investigations Rep. 2010-5229. USGS, Reston, VA.
Jarvis, N.J., J. Moeys, J.M. Hollis, S. Reichenberger, A.M.L. Lindahl, and I.G.
Dubus. 2009. A conceptual model of soil susceptibility to macropore
flow. Vadose Zone J. 8:902–910. doi:10.2136/vzj2008.0137
Jones, R.L., R.V. Rourke, and J.L. Hansen. 1986. Effect of application
Journal of Environmental Quality • Volume 41 • March–April 2012
methods on movement and degradation of aldicarb residues in Maine
potato fields. Environ. Toxicol. Chem. 5:167–173. doi:10.1002/
etc.5620050207
Judge, G.G., W.E. Griffiths, R.C. Hill, H. Lütkepohl, and T.C. Lee. 1985.
The theory and practice of econometrics, 2nd ed. John Wiley & Sons,
New York.
Kauffman, L.J., A.L. Baehr, M.A. Ayers, and P.E. Stackelberg. 2001. Effects of
land use and travel time on the distribution of nitrate in the KirkwoodCohansey aquifer system in southern New Jersey. U.S. Geological Survey
Water-Resources Investigations Rep. 01-4117. USGS, Reston, VA.
Kaufman, D.D., and P.C. Kearney. 1970. Microbial degradation of s-triazine
herbicides. Residue Rev. 32:235–265.
Keim, A.M., L.C. Ruedisili, D.B. Baker, and R.E. Gallagher. 1989. Herbicide
monitoring of tile drainage and shallow groundwater in northwestern
Ohio farm fields: A case study. p. 62–78. In Pesticides in terrestrial and
aquatic environments: Proceedings of a national research conference.
Virginia Polytechnic Institute and State Univ., Blacksburg.
Kladivko, E.J., L.C. Brown, and J.L. Baker. 2001. Pesticide transport to subsurface tile drains in humid regions of North America. Crit. Rev. Environ. Sci. Technol. 31:1–62. doi:10.1080/20016491089163
Kleinbaum, D.G., and M. Klein. 2002. Logistic regression: A self-learning
text. In K. Dietz, M. Gail, A. Krickelberg, and T.J. Samet (ed.) Statistics
for biology and health. 2nd ed. Springer, New York.
Kolpin, D.W., S.J. Kalkhoff, D.A. Goolsby, D.A. Sneck-Fahrer, and E.M.
Thurman. 1997. Occurrence of selected herbicides and herbicide degradation products in Iowa’s ground water, 1995. Ground Water 35:679–
688. doi:10.1111/j.1745-6584.1997.tb00134.x
Kolpin, D.W., J.E. Barbash, and R.J. Gilliom. 2002. Atrazine and metolochlor occurrence in shallow ground water of the United States, 1993 to
1995: Relations to explanatory factors. J. Am. Water Resources Assoc.
38:301–311.
Konopka, A., and R. Turco. 1991. Biodegradation of organic compounds
in vadose zone and aquifer sediments. Appl. Environ. Microbiol.
57:2260–2268.
Koterba, M.T., F.D. Wilde, and W.W. Lapham. 1995. Ground-water datacollection protocols and procedures for the National Water-Quality Assessment Program: Collection and documentation of water-quality samples and related data: U.S. Geological Survey Open-File Rep. 95–399.
USGS, Reston, VA.
Koterba, M.T. 1998. Ground-water data-collection protocols and procedures
for the National Water-Quality Assessment Program: Collection, documentation, and compilation of required site, well, subsurface, and landscape data for wells. U.S. Geological Survey Water-Resources Investigations Rep. 98-4107. USGS, Reston, VA.
Laitila, T. 1993. A pseudo-R-square measure for limited and qualitative dependent variable models. J. Econom. 56:341–356.
doi:10.1016/0304-4076(93)90125-O
Lapham, W.W., F.D. Wilde, and M.T. Koterba. 1995. Ground-water datacollection protocols and procedures for the National Water-Quality Assessment Program: Selection, installation, and documentation of wells,
and collection of related data. U.S. Geological Survey Open-File Report
95–398. USGS, Reston, VA.
Larsen, L., S.R. Sørensen, and J. Aamand. 2000. Mecoprop, isoproturon, and
atrazine in and above a sandy aquifer: Vertical distribution of mineralization potential. Environ. Sci. Technol. 34:2426–2430. doi:10.1021/
es9911723
Larson, S.J., C.G. Crawford, and R.J. Gilliom. 2004. Development and application of watershed regressions for pesticides (WARP) for estimating
atrazine concentration distributions in streams. U.S. Geological Survey
Water-Resources Investigations Rep. 03-4047. USGS, Reston, VA.
Martin, J.D. 2009. Sources and preparation of data for assessing trends in
concentrations of pesticides in streams of the United States, 1992–2006.
U.S. Geological Survey Scientific Investigations Rep. 2009-5062. USGS,
Reston, VA.
Martin, J.D., W.W. Stone, D.S. Wydoski, and M.W. Sandstrom. 2009. Adjustment of pesticide concentrations for temporal changes in analytical
recovery, 1992–2006: U.S. Geological Survey Scientific Investigations
Report 2009–5189. USGS, Reston, VA.
McMahon, P.G., F.H. Chapelle, and M.L. Jagucki. 1992. Atrazine mineralization potential of alluvial-aquifer sediments under aerobic conditions.
Environ. Sci. Technol. 26:1556–1559. doi:10.1021/es00032a010
Nair, D.R., and J.L. Schnoor. 1992. Effect of two electron acceptors on atrazine mineralization rates in soil. Environ. Sci. Technol. 26:2298–2300.
doi:10.1021/es00035a032
Nolan, B.T., and K.J. Hitt. 2006. Vulnerability of shallow groundwater and
drinking-water wells to nitrate in the United States. Environ. Sci. Techwww.agronomy.org • www.crops.org • www.soils.org
nol. 40:7834–7840. doi:10.1021/es060911u
Nowell, L.H., C.G. Crawford, R.J. Gilliom, N. Nakagaki, W.W. Stone, G.P.
Thelin, and D.M. Wolock. 2009. Regression model for explaining and
predicting concentrations of organochlorine pesticides in fish from
streams in the United States. Environ. Toxicol. Chem. 28:1346–1358.
doi:10.1897/08-508.1
NRCS. 1994. State Soil Geographic (STATSGO) data base for the United
States and Puerto Rico [CD-ROM]: Ft. Worth, TX. Available at http://
soildatamart.nrcs.usda.gov/USDGSM.aspx (verified 13 Dec. 2011).
NRCS. 1995. 1992 National Resources Inventory [CD-ROM reissued May
1995]. Iowa State Univ., Statistical Laboratory, Ames.
Oblinger Childress, C.J., W.T. Foreman, B.F. Connor, and T.J. Maloney. 1999.
New reporting procedures based on long-term method detection levels
and some considerations for interpretations of water-quality data provided by the U.S. Geological Survey National Water-Quality Laboratory:
U.S. Geological Survey Open-File Report 99–193. USGS, Reston, VA.
Papiernik, S.K., and R.F. Spalding. 1998. Atrazine, deethylatrazine, and deisopropylatrazine persistence measured in groundwater in situ under lowoxygen conditions. J. Agric. Food Chem. 46:749–754. doi:10.1021/
jf9706235
Pickett, C.H., L.S. Hawkins, J.E. Pehrsom, and N.V. O’Connell. 1990. Herbicide use in citrus production and ground water contamination in Tulare
County. Rep. PM 90-1. California Department of Food and Agriculture,
Sacramento, CA.
Porter, K.S., R.J. Wagenet, R.L. Jones, and T.E. Marquardt. 1990. Field research on aldicarb management practices for upstate New York. Environ.
Toxicol. Chem. 9:279–287. doi:10.1002/etc.5620090304
Price, C.V., N. Nakagaki, K.S. Hitt, and R.M. Clawges. 2003. Mining GIRAS: Improving on a national treasure of land use data: Proceedings of
the ESRI 2003 International User Conference. Available at http://www.
esri.com/uc (verified 6 Dec. 2011).
Priddle, M.W., R.E. Jackson, and J.P. Mutch. 1989. Contamination of
the sandstone aquifer of Prince Edward Island, Canada, by aldicarb and nitrogen residues. Ground Water Monit. Rev. 9:134–140.
doi:10.1111/j.1745-6592.1989.tb01022.x
Priddle, M.W., J.P. Mutch, and R.E. Jackson. 1992. Long-term monitoring of
aldicarb residues in groundwater beneath a Canadian potato field. Arch.
Environ. Contam. Toxicol. 22:183–189. doi:10.1007/BF00213283
Rocha, F., and A. Walker. 1995. Simulation of the persistence of atrazine in soil at different sites in Portugal. Weed Res. 35:179–186.
doi:10.1111/j.1365-3180.1995.tb02031.x
Rügge, K., P.L. Bjerg, H. Mosbaek, and T.H. Christensen. 1999. Fate
of MCPP and atrazine in an anaerobic landfill leachate plume
(Grindsted, Denmark). Water Res. 33:2455–2458. doi:10.1016/
S0043-1354(98)00444-8
Saad, D.A. 2008. Agriculture-related trends in groundwater quality of the Glacial Deposits Aquifer, Central Wisconsin. J. Environ. Qual. 37:S209–
S225. doi:10.2134/jeq2007.0053
Scott, J.C. 1990. Computerized stratified random site-selection approaches
for design of a ground-water-quality sampling network. U.S. Geological Survey Water-Resources Investigations Rep. 90-4101. USGS, Reston,
VA.
Solley, W.B., R.R. Pierce, and H.A. Perlman. 1998. Estimated use of water in
the United States in 1995: U.S. Geological Survey Circular 1200. USGS,
Reston, VA.
Spillmann, P. 1989. Decomposition and elimination of typical pollutants from
sanitary landfills in porous aquifers. p. 227–233. In Groundwater management: Quantity and quality. Vol. 188. International Association of
Hydrological Sciences.
Stackelberg, P.E., R.J. Gilliom, D.M. Wolock, and K.J. Hitt. 2006a. Development and application of a regression equation for estimating the occurrence of atrazine in shallow ground water beneath agricultural areas of
the United States. U.S. Geological Survey Scientific Investigations Rep.
2005–5287. USGS, Reston, VA.
Stackelberg, P., E. Furlong, M. Meyer, S. Zaugg, A. Henderson, and D. Reissman. 2006b. Response to comment on “Persistence of pharmaceutical
compounds and other organic wastewater contaminants in a conventional drinking-water-treatment plant.” Sci. Total Environ. 354:93–97.
doi:10.1016/j.scitotenv.2005.04.012
Stone, W.W., R.J. Gilliom, and C.G. Crawford. 2008. Watershed regressions
for pesticides (WARP) for predicting annual maximum and maximum
moving-average concentrations of atrazine in streams. U.S. Geological
Survey Open-File Rep. 08-1186. USGS, Reston, VA.
Stone, W.W., and R.J. Gilliom. 2009. Update of watershed regressions for pesticides (WARP) for predicting atrazine concentrations in streams. U.S.
Geological Survey Open-File Rep. 2009-1122. USGS, Reston, VA.
493
Tesoriero, A.J., T.B. Spruill, H.E. Mew, Jr., K.M. Farrell, and S.L. Harden.
2005. Nitrogen transport and transformations in a coastal plain watershed: Influence of geomorphology on flow paths and residence
times. Water Resour. Res. 41:W02008 10.1029/2003WR002953.
doi:10.1029/2003WR002953
Tesoriero, A.J., D.A. Saad, K.R. Burow, E.A. Frick, L.J. Puckett, and J.E.
Barbash. 2007. Linking ground-water age and chemistry data along
flow paths: Implications for trends and transformations of nitrate
and pesticides. J. Contam. Hydrol. 94:139–155. doi:10.1016/j.
jconhyd.2007.05.007
Thelin, G.P., and L.P. Gianessi. 2000. Method for estimating pesticide use for
county areas of the conterminous United States. U.S. Geological Survey
Open-File Rep. 00-250. USGS, Reston, VA.
Thelin, G.P., and W.W. Stone. 2010. Method for estimating annual atrazine
use for counties in the conterminous United States, 1992–2007: U.S.
Geological Survey Scientific Investigations Rep. 2010-5034. USGS, Reston, VA.
Therneau, T.M. 1999. A package of survival routines for S. Mayo Clinic Department of Health Sciences Research, Rochester, MN.
Thomas, R.G. 1990. Volatilization from soil. p. 16-1–16-50. In W.J. Lyman et
al. (ed.) Handbook of chemical property estimation methods. American
Chemical Society, Washington, DC.
Thorton, P.E. 2003. DAYMET climatological summaries for average air temperatures and total precipitation (18-year mean for 1980–1997). Available at http://daymet.org (verified 8 Dec. 2011).
TIBCO Software Inc. 2008. TIBCO Spotfire S+ 8.1 for Windows user’s guide.
TIBCO, Palo Alto, CA.
Tobin, J. 1958. Estimation of relationships for limited dependent variables.
Econometrics 26:24–36. doi:10.2307/1907382
Toccalino, P.L., and J.E. Norman. 2006. Health-based screening levels to
evaluate U.S. Geological Survey ground-water quality data. Risk Anal.
26:1339–1348. doi:10.1111/j.1539-6924.2006.00805.x
Toy, T.J., A.J. Kuhaida, Jr., and B.E. Munson. 1978. The prediction of mean
monthly soil temperature from mean monthly air temperature. Soil Sci.
126:181–189. doi:10.1097/00010694-197809000-00008
Traub-Eberhard, U., W. Kördel, and W. Klein. 1994. Pesticide movement
into subsurface drains on a loamy silt soil. Chemosphere 28:273–284.
doi:10.1016/0045-6535(94)90124-4
494
U.S. Bureau of the Census. 1991. Census of population and housing, 1990:
Public Law 94-171 data (United States). [CD-ROM]. U.S Census Bureau, Washington, DC.
USEPA. 2003. Interim reregistration eligibility decision for atrazine. Case
no. 0062: Office of Prevention, Pesticides and Toxic Substances. Available at http://www.epa.gov/oppsrrd1/REDs/atrazine_ired.pdf (verified
8 Dec. 2011).
USEPA. 2006. Cumulative risk from triazine pesticides. Office of Pesticide
Programs, Health Effects Division. Available at http://www.epa.gov/
oppsrrd1/REDs/triazine_cumulative_risk.pdf (verified 8 Dec. 2011).
USGS. 2010. Pesticide National Synthesis Project. 1992 pesticide use maps:
Atrazine Available at http://water.usgs.gov/nawqa/pnsp/usage/maps/
show_map.php?year=92&map=m1980 (verified 8 Dec. 2011).
Vogelmann, J.E., S.M. Howard, L. Yang, C.R. Larson, B.K. Wylie, and N. Van
Driel. 2001. Completion of the 1990s national land cover data set for
the conterminous United States from landsat thematic mapper data and
ancillary data sources. Photogramm. Eng. Remote Sens. 67:650–652.
Wisconsin Department of Agriculture, Trade and Consumer Protection. 2010.
Available at http://datcp.wi.gov/Environment/Water_Quality/Atrazine/
Use_Restrictions/index.aspx (verified 13 Dec. 2011).
Wolock, D.W. 2003. Estimated mean annual natural ground-water recharge in
the conterminous United States: U.S. Geological Survey Open-File Rep.
03-311. USGS, Reston, VA.
Wyman, J.A., J.O. Jensen, D. Curwen, R.L. Jones, and T.E. Marquardt. 1985.
Effects of application procedures and irrigation on degradation and
movement of aldicarb residues in soil. Environ. Toxicol. Chem. 4:641–
651. doi:10.1002/etc.5620040508
Zaugg, S.D., M.W. Sandstrom, S.G. Smith, and K.M. Fehlberg. 1995. Methods of analysis by the U.S. Geological Survey National Water-Quality
Laboratory: Determination of pesticides in water by C-18 solid-phase
extraction and capillary-column gas chromatography/mass spectrometry
with selected-ion monitoring: U.S. Geological Survey Open-File Rep.
95-181.USGS, Reston, VA.
Zheng, D., E.R. Hunt, Jr., and S.W. Running. 1993. A daily soil temperature
model based on air temperature and precipitation for continental applications. Clim. Res. 2:183–191. doi:10.3354/cr002183
Journal of Environmental Quality • Volume 41 • March–April 2012