DERR-2024-007746

2020
Environmental Toxicology and Chemistry, Vol. 22, No. 9, pp. 2020–2029, 2003
q 2003 SETAC
Printed in the USA
0730-7268/03 $12.00 1 .00
ANALYSIS OF FIELD AND LABORATORY DATA TO DERIVE SELENIUM TOXICITY
THRESHOLDS FOR BIRDS
WILLIAM J. ADAMS,*† KEVIN V. B RIX,‡ MELANIE EDWARDS,§ LUCINDA M. TEAR,§ DAVID K. DEFOREST,§ and
ANNE FAIRBROTHER§
†Kennecott Utah Copper Corporation, 8315 West 3595 South, P.O. Box 6001, Magna, Utah 84044, USA
‡EcoTox, 721 Navarre Avenue, Coral Gables, Florida 33134, USA
§Parametrix, 5808 Lake Washington Boulevard Northeast, Suite 200, Kirkland, Washington 98033, USA
(Received 11 March 2002;Accepted 19 December 2002)
Abstract—In this paper, we critically evaluate the statistical approaches and datasets previously used to derive chronic egg selenium
thresholds for mallard ducks (laboratory data) and black-necked stilts (field data). These effect concentration thresholds of 3%,
10% (EC10), or 20% have been used by regulatory agencies to set avian protection criteria and site remediation goals, thus the
need for careful assessment of the data. The present review indicates that the stilt field dataset used to establish a frequently cited
chronic avian egg selenium threshold of 6 mg/kg dry weight lacks statistical robustness (r2 5 0.19–0.28 based on generalized linear
models), suggesting that stilt embryo sensitivity to selenium is highly variable or that factors other than selenium are principally
responsible for the increase in effects observed at the lower range of this dataset. Hockey stick regressions used with the stilt field
dataset improve the statistical relationship (r2 5 0.90–0.97) but result in considerably higher egg selenium thresholds (EC10 5
21–31 mg/kg dry wt). Laboratory-derived (for mallards) and field-derived (for stilts) teratogenicity EC10 values are quite similar
(16–24 mg/kg dry wt). Laboratory data regarding mallard egg inviability and duckling mortality data provide the most sensitive
and statistically robust chronic threshold (EC10) with logit, probit, and hockey stick regressions fitted to laboratory data, resulting
in mean egg selenium EC10 values of 12 to 15 mg/kg dry weight (r2 5 0.75–0.90).
Keywords—Selenium toxicity thresholds Avian Egg
INTRODUCTION
The toxic effects of selenium were first described in South
Dakota (USA), where chickens (Gallus domesticus) were ob-
served to have poor reproductive success as a result of chick
teratogenesis and mortality when fed grains from specific
sources [1]. Subsequent studies identified elevated selenium
content in the grains as the cause of the observed effects [2].
During the 1980s, a severe reduction in reproductive success
of aquatic birds at Kesterson Reservoir (CA, USA) prompted
a series of studies that identified seleniferous water from sub-
surface agricultural drainage as the causative factor. Again,
teratogenesis and chick mortality were the primary toxicolog-
ical effects, although selenium concentrationsweresufficiently
elevated to also cause effects in adult birds [3].
Like other metals and metalloids, the primary avian ex-
posure pathway for selenium is the diet [4,5]. Aquatic inver-
tebrates and plants bioaccumulate selenium via the water and
their diet, and aquatic birds that feed on these invertebrates
and plants during the breeding season transfer a significant
fraction of their dietary intake to the eggs [6,7]. At sufficiently
high egg selenium concentrations, teratogenic effects on de-
veloping embryos can result [8,9].
Selenium interferes with embryo development and, at suf-
ficiently high concentrations, can cause gross abnormalities,
such as anophthalmia, incomplete beak development, brain
defects (hydrocephaly and exancephaly), foot defects, and oth-
er terata [8,9]. Moresubtleteratogeniceffects,suchasenlarged
hearts, edema, liver hypoplasia, and gastroschisis, also occur
* To whom correspondence may be addressed
([e-mail]).
but rarely are recorded in field studies. Any of these effects
can lead to reduced embryo and chick survival.
Several toxicological endpoints are relevant for selenium,
including teratogenesis, egg inviability, clutch inviability (i.e.,
one or more inviable eggs in a clutch), and chick mortality.
These endpoints have been characterized as a function of in-
dividual egg selenium concentrations, mean egg selenium con-
centrations for a clutch, and clutch or hen responses. Differ-
ences in the way the dose–response relationship is character-
ized are important in interpreting the thresholds proposed by
different researchers. The effects of these differences will be
discussed later.
Since the events at Kesterson, several laboratory and field
studies have been conducted to estimate egg selenium toxicity
thresholds for birds. The importance of the threshold is that
it often is used as a concentration of concern for regulatory
decisions. The U.S. Fish and Wildlife Service has reviewed
the extant data several times during the past few years and
consistently recommended an egg selenium threshold of 6 to
7 mg/kg dry weight based on field data for black-necked stilt
(Himantopus mexicanus) clutch inviability [10,11].
More recently, Fairbrother et al. [12,13] developed mean
egg selenium thresholds for teratogenesis and duckling mor-
tality by pooling toxicity data from several laboratory studies
with mallards (Anas platyrhynchos). Fairbrother et al. also
developed dose–response relationships for these pooled data.
They concluded that the mean egg selenium threshold for
duckling mortality was 16 mg/kg dry weight. They further
questioned whether the field-based threshold (6 mg/kg dry wt)
developed by the U.S. Fish and Wildlife Service was reliable,
noting that field- and laboratory-based thresholds (10% effect
Deriving egg selenium toxicity thresholds for birds Environ. Toxicol. Chem.22, 2003 2021
Table 1. Study areas for field data located in the United States
Study area
a
Teratogenesis
Avocet Stilt Duck
Invia-
bility
Stilt
Bowdoin NWR, MT
Grasslands Water District, CA
Gunnison River Area, CO
Honey Lake Valley, CA
Kendrick Irrigation Project, WY
Kesterson Reservoir, CA
Ouray NWR, UT
Red Rock Ranch, CA
Salton Sea, CA
San Juan River Area, NM
Sun River Area, MT
Tulare Basin, CA
Volta State Wildlife Area, CA
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
a NWR 5 National Wildlife Refuge.
concentration [EC10]) for the less-sensitive threshold of ter-
atogenesis were similar but that the chick/duckling mortality
endpoints were dissimilar. They hypothesized that the data
regarding teratogenesis were in agreement because it was a
selenium-specific endpoint that could be measured accurately
in the field. In contrast, the chick/duckling mortality endpoint
was not selenium-specific and was subject to confounding fac-
tors, such as weather, starvation, and other contaminants,when
measured in the field. They further hypothesized that the lower
threshold predicted from field data was confounded by one or
more of these factors and concluded that until these data could
be thoroughly scrutinized, the laboratory-derived threshold of
16 mg/kg dry weight should be used.
Since the publication of Fairbrother et al. [13], the raw data
for the field-based thresholds have been released to the public
for review [10]. Additionally, Ohlendorf [14] recently reex-
amined the available laboratory toxicity data using logistic
regression and concluded that the mean egg seleniumthreshold
(EC10) is 12.8 mg/kg dry weight. Ohlendorf pointed out that
Fairbrother et al. [13] failed to include three available data
points from laboratory mallard studies that could affect the
results of their threshold analysis. The purpose of our study
was to critically review the field data, update the analysis of
Fairbrother et al. to include the missing data, and compare and
contrast the field and laboratory data.
MATERIALS AND METHODS
Toxicological endpoints
Before comparing field and laboratory data, it is necessary
to define the different toxicological endpoints that have been
used to characterize the effects of selenium on aquatic birds.
Chick mortality is thought to be the most sensitive chronic
toxicological endpoint [10,13]. This endpoint evaluates net
reproductive success, taking into account chicks affected by
teratogenesis, egg inviability, and mortality up to 7 d post-
hatch. However, because of practical limitations in the field,
evaluation of chick mortality typically is not feasible because
it is difficult to accurately monitor chicks posthatch. Hence,
available field data do not consider this endpoint.
Egg inviability is slightly less sensitive than chick mortality
[15], whereas teratogenesis tends to beless sensitivethanchick
mortality by a factor of 1.5- to 3-fold [10,12]. A fourth end-
point, clutch inviability, is toxicologically equivalent to egg
inviability but is evaluated on a henwise (i.e., clutchwise)basis
rather than on an individual-egg basis. This endpoint has been
used only for field data. It is based on the premise that indi-
vidual sibling chicks within a clutch are not independent sam-
ples, so the hen (i.e., clutch) represents the lowest statistical
unit. In developing clutch inviability data, a clutch is consid-
ered to be inviable (i.e., affected) if at least one egg did not
hatch [10].
Teratogenicity field data are available for ducks, black-
necked stilts, and American avocets (Recurvirostra ameri-
cana). The duck data are a composite of information on Anas
spp. (gadwalls, mallards, pintails, and shovelers),Aythya spp.
(redheads and canvasbacks), and the ruddy duck (Oxyura ja-
maicensis). In contrast, field clutch inviability data are avail-
able only for black-necked stilts. A limited number of field
studies have also monitored clutch inviability and egg invia-
bility in concert. Skorupa [10] evaluated individual egg data
from eight nesting neighborhoods, each of which contained
from 18 to 42 nests. Using these data, Skorupa developed a
relationship between the probability of a given egg selenium
concentration resulting in an inviable clutch or an inviable
egg. Using this relationship, Skorupa developed the following
equation to relate the probability of an inviable egg to the
probability of an inviable clutch:
AE:AH ratio 5 0.047 1 0.423 3 log(AH) (1)
where
AE 5 % affected eggs
AH 5 % affected hens (or clutch)
We used the above equation in our analyses of field data
on stilt clutch inviability to calculate the AE:AH ratio. The
AE:AH ratio for a given egg selenium concentration was then
multiplied by the probability of an inviable clutch to estimate
the associated probability of an inviable egg. This effectively
normalized the field and laboratory data to the same endpoint.
This technique was used only in the hockey stick regression
analysis for reasons that will be discussed later.
The source of the field datasets was the raw data published
in the appendices of Skorupa. These binomial data were col-
lected by various researchers from 1983 through 1996 from
study areas listed in Table 1. Some of the locations were sys-
tematically sampled; others were not. All data were pooled
regardless of sampling approach, as in Skorupa.
Statistical analysis of laboratory and field data
The analysis of Fairbrother et al. [13] was updated by in-
cluding the three data points from Heinz and Hoffman [16,17]
that were previously excluded. Other than inclusion of these
additional data, the analysis was performed here as previously
described by Fairbrother et al. [12,13], except that the dataset
was also analyzed using hockey stick regression, as described
later in this paper. Table 2 provides the raw laboratory data
used in this reanalysis.
Initially, we evaluated the field data using thesamemethods
and datasets as Skorupa [10] for comparative purposes. Sko-
rupa used generalized linear models (GLiMs) and a three-point
moving average in separate analyses to evaluate goodness-of-
fit and to estimate toxicity thresholds (e.g., EC10). We used
a similar approach and then considered an alternative method
that involved binning the field data and using hockey stick
regressions to estimate the toxicity threshold. For comparison,
2022 Environ. Toxicol. Chem.22, 2003 W.J. Adams et al.
Table 2. Laboratory egg selenium toxicity data used in reanalysis of Fairbrother et al. [12,13]
Mean egg
selenium
(mg/kg
dry wt)
a
% Terato-
genesis in
eggs that did
not hatch % Fertility
b
% Fertile
eggs that
hatchedc
% All eggs
that hatched
% Duckling
survival
% Duckling
mortality for
all eggs laid
d
% Duckling
mortality for
all eggs laid
(Abbott’s-
corrected) Reference
0.6
0.93
1.16
1.35
1.4
2.8
5.3
0.6
NAe
6.1
1.3
1
0.9
0.5
98.6
88
96.0
96.7
100
96.3
97.0
59.6
62
44.2
41.3
91.4
70.7
60.0
58.8
55
42.4
39.9
91.4
68.1
58.2
99.4
74.2
96.2
98.4
82.5
97.7
98.9
41.6
59.5
59.2
60.7
24.6
33.5
42.4
0
0
0
0
0
0
1.5
[15]
[36]
[17]
[16]
[36]
[15]
[15]
11.3
12.1
24.5
25.1
29.4
30.4
1.4
NA
NA
36.2
28.2
24.6
97.9
90
89
99.1
99.1
95.6
53.4
61
41
24.0
6.4
7.6
52.3
55
36
23.8
6.3
7.3
99.7
78.1
60.9
66.0
20.0
47.2
47.9
57.1
77.8
84.3
98.7
96.6
10.8
0
45.0
62
97
91
[15]
[37]
[37]
[17]
[16]
[16]
36.7
42
60
6.8
57.5
67.9
96.6
100
99.6
36.9
8.5
2.2
35.6
8.5
2.2
80.6
10.0
0
71.3
99.2
100
50.8
98.9
100
[15]
[36]
[15]
a If mean egg selenium (MES) was reported as wet weight, then MES as dry weight was calculated as MESwet weight
4 fraction solids (using %
moisture from study).
b Percentage fertility not reported in Stanley et al. [36], who only reported that selenium had no effects on fertility. Assumed fertility was 100%.
c In Stanley et al. [36], it is unclear if hatching success is reported for all eggs or for fertile eggs. Currently, it is assumed it is for fertile eggs.
d Duckling mortality assessed after 6 d by Heinz et al. [15] and by Heinz and Hoffman [16], after 7 d by Heinz and Hoffman [17], and after 14
d by Stanley et al. [36,37]. Ducklings in the Heinz et al. [16] study were not fed selenium diets; ducklings in the Stanley et al. [36,37] studies
were fed the same selenium diets as their parents.
e Percentage of teratogenic embryos not reported; it was only stated that teratogenesis was not significantly different than in controls.
hockey stick regressions were also fitted to the laboratory data.
The methods for each of these models are described below.
Generalized linear models.For stilts, avocets, and ducks,
each measured egg selenium concentration (mg/kg dry wt)was
paired witha0ora1toindicate the absence (0) or thepresence
(1) of teratogenesis in the embryo. For stilt clutch inviability,
egg selenium was measured in one egg, and 0 or 1 indicated
whether one or more sibling eggs did (0) or did not (1) hatch.
Binary logistic regression models (logistic link, binomial er-
ror) were used to fit the relationship between egg selenium
concentrations and the binary responses, teratogenesis, and
clutch inviability. We did not apply the AE:AH ratio equation
to the stilt clutch inviability data, because as discussed below,
the r2 for this model was very low and further extrapolation
of the data was not considered to be appropriate. When data
indicated teratogenesis or clutch inviability at background se-
lenium levels, the methods of Bailer and Oris [18] were used
to estimate the percentage decrease in clutch viability relative
to the background level of effects. Models were fitted using
both SPlus [19] and SPSS [20] as a quality control, and results
were compared to those reported by Skorupa [10].
In binary logistic regression, maximum likelihood methods
are used to compare the fit of a model with an independent
variable of interest to the fit of a model without that variable
[21]. In maximum likelihood methods, the significance of the
proposed independent, explanatory variable is assessed by
comparing the likelihood of the modelwiththeterm(expressed
as deviance) to the likelihood of the model without the term.
A difference in deviances is assumed to be chi-squared dis-
tributed, with degrees of freedom being equal to the difference
between the number of parameters used in the two models.
In this case, the significance of selenium in explaining ter-
atogenicity and clutch inviability was assessed by comparing
the difference between the model deviance with an intercept
only and the model deviance with an intercept and a slope to
a chi-squared statistic with one degree of freedom. If the model
choice is correct, then maximum likelihood procedures si-
multaneously assess both the significance of the independent
variable and the goodness-of-fit of the model, in that the ad-
dition of an independent variable to the model will not be
significant if it does not represent a significant improvement
of fit compared to the less-complicated model. In the more
familiar cases of analysis of variance and general linear re-
gression, maximum likelihood and least-squares methods are
identical, and goodness of fit is expressed as the familiar r2
statistic that describes the percentage of variance that is ex-
plained by the model. For GLiMs, a variety of summary sta-
tistics have been developed that can be considered as analo-
gous to the r2 statistic (referred to as pseudo r2 by Hagle and
Mitchell [22]). Because none of these statistics is exactly
equivalent to the percentage of variance explained and each
reveals a slightly different perspective on the data, we report
two measures: First the Nagelkerke r2 ( in Menard [23]),
2rM
and then the Dhrymes measure (Hagle and Mitchell [22] from
Dhrymes [24]; also called in Menard [23]). These two mea-
2rL
sures can be expressed as
2/N12(L /L )012Nagelkerker5 (2)2/N12(L )0
ln L12Dhrymesr512 (3)ln L0
where
L 5 likelihood of model with intercept only0
L 5 likelihood of full model (intercept and slope)1
We did not utilize the frequently used Hosmer and Le-
Deriving egg selenium toxicity thresholds for birds Environ. Toxicol. Chem.22, 2003 2023
meshow [25] goodness-of-fit test, which is based on the chi-
squared test, because small numbers of teratogenic eggs for
all species meant that many of the cells of the contingency
table had counts of less than five, and this violated the as-
sumptions of the test. Discussion of these and other methods
of assessing model fit can be found in any number of texts,
articles, and software manuals about GLiMs (e.g., [19–23,25–
28]).
Three-point moving average.In addition to the GLiM anal-
ysis, we also analyzed the black-necked stilt field clutch in-
viability data using a three-point moving average, as in Sko-
rupa [29]. Egg selenium concentrations were rounded to whole
numbers, and a response rate was calculated for each. The
response rate was calculated as the number of inviableclutches
divided by the number of clutches assessed at a given (round-
ed) selenium concentration. A three-point moving average re-
sponse rate was calculated as the average of the calculated
response rates for egg selenium concentrations at 1 mg/kg (dry
weight) lower than the given level, at the given level, and at
1 mg/kg higher than the given level. Skorupa used this method
to further support a threshold of 6 mg/kg but only presented
results for selenium concentrations from 4 to 9 mg/kg. We
applied similar and additional analyses to the full range of
consecutive whole-mg/kg selenium egg concentrations (2–160
mg/kg dry wt) reported in the Appendix of Skorupa [10].
Hockey stick regressions.As another approach for evalu-
ating and interpreting data regarding the effects of selenium,
we fit hockey stick regression models to the field (stilt) and
laboratory (mallard) dose–response relationships. This type of
model has been used to define a threshold when an underlying
background level of response is unrelated to the dose, as is
the case for many toxicological datasets [30–32].
The response endpoints in the field and the laboratory stud-
ies were equated for ease of comparison. The field data for
clutch inviability were based on a clutchwise (i.e., henwise)
response, whereas the laboratory data were basedonindividual
responses. For the clutchwise response, a positive response
was recorded when one or more eggs were observed not to
hatch, which is consistent with Skorupa [10]. We used the AE:
AH ratio equation described earlier (Eqn. 1) to normalize the
data to individual egg responses [10].
The laboratory studies combined by Fairbrother et al.
[12,13] measured duckling survival up to 7 d posthatch,where-
as the field data reported by Skorupa [10,29] did not consider
chick survival posthatch. In the present study, we used the egg
inviability endpoint for the laboratory data withoutconsidering
posthatch survival so that we could make direct comparisons
between the field and laboratory data. However, our final anal-
ysis of the laboratory data (independent of the comparison to
the field data) was based on posthatch duckling survival to
provide an appropriately conservative threshold.
A background or control response rate was observed in
both field and laboratory datasets. This is particularly impor-
tant for the egg inviability endpoint, because the response rate
was significantly greater than zero. To account for control
responses, the datasets were normalized using Abbott’s for-
mula [33]. For the field data, the response observed in eggs
with selenium levels less than 3 mg/kg dry weight was con-
sidered as background for the normalization procedure. Use
of Abbott’s formula is necessary to separate the selenium-
related response from the background response, but it also
reduces response variance at background and low-exposure
concentrations (any response less than the mean background
response is set to zero). Consequently, to some extent,Abbott’s
formula artificially enhances the ability of the hockey stick
regression to define a threshold as compared to a noncorrected
dataset.
A final difference between the field and laboratory hatch-
ability data is that the field data were based on measured
responses in stilts, whereas the laboratory studies measured
mallard responses. Based on field data regarding teratogenesis
collected on both species, mallards are approximately 50%
more sensitive than stilts [10]. Whether the same relative sen-
sitivities apply to other endpoints (e.g., clutch inviability) is
unknown.
In the field, birds are exposed to an uncontrolled range of
dietary selenium concentrations, as opposed to the laboratory
studies, which exposed multiple hens to specific concentra-
tions. To model the field data similarly to the laboratory data,
the field selenium concentrations were binned into treatment
categories, and a percentage response was calculated for each
bin. Five different binning schemes (A through E) were used
to evaluate the sensitivity of the dose–response relationship to
a specific binning scheme. The rate of teratogenicity in ducks,
stilts, and avocets was calculated as the number of teratogenic
eggs divided by the number of eggs assessed within each
binned selenium range. In all binning schemes, the response
rate was paired with the median egg selenium concentration
within the respective bin. Data evaluations using binning
schemes were performed on the field and laboratory datasets
for both teratogenicity and clutch inviability.
Binning scheme A grouped egg selenium concentrations
according to a geometric sequence: 0 to ,2,2to,4,4to
,8, 8 to ,16 mg/kg dry weight, etc. Binning scheme B
grouped responses by every 5 mg/kg dry weight in measured
egg selenium, resulting in bins of 0 to ,5,5to,10, 10 to
,15, and 15 to ,20 mg/kg dry weight, etc. Binning scheme
C grouped a similar number of responses per bin, approxi-
mately 11 eggs (12 bins) for teratogenesis and 21 eggs (19
bins) for inviability. Scott’s normal approximation was used
to calculate the most informative number of bins from each
dataset [34]. The resulting binning scheme for stilt teratogen-
esis used median selenium concentrations of 1.6, 2.25, 3.9,
5.1, 7.3, 9.65, 11, 16, 20, 23, 27, and 35 mg/kg dry weight.
The binned median selenium concentrations for stilt egg in-
viability were 2.1, 2.85, 3.4, 4.2, 5.1, 6.5, 7.5, 9.65, 16, 18,
22, 28, 33, 37, 41, 48.5, 54, 59, and 72.5 mg/kg dry weight.
Binning scheme D was similar to binning scheme A but in-
cremented initially by 3 mg/kg dry weight (i.e., 0 to ,3, 3 to
,6,6to,12, 12 to ,24 mg/kg dry wt, etc.). This scheme
results in a bin break at the threshold of 6 mg/kg dry weight
proposed by Skorupa [10,29]. Binning scheme E used the
breakpoints identified by the U.S. Department of the Interior
[11], namely median selenium concentrations of 2.7, 9.2, 21.5,
31, and 76.5 mg/kg dry weight for teratogenesis and of 3.2,
7.65, 22, 42.5, 65, 74, and 97.5 mg/kg dry weight for egg
inviability. Each binning scheme has benefits and drawbacks,
so they should be interpreted collectively.
Using all binning schemes, considerable variance was seen
in the relationship between egg selenium concentration and
the measured response. A true dose–response relationship did
not appear to begin until higher concentrations were reached,
although the exact point at which the relationship began de-
pended on the binning scheme. Hockeystickregressionmodels
were fitto describe the varianceintheresponseatlowselenium
concentrations, to estimate the inflection concentration at
2024 Environ. Toxicol. Chem.22, 2003 W.J. Adams et al.
Fig. 1. Logit and probit models fit to mallard duck laboratory tera-
togenicity data obtained from several studies (h, Heinz et al. [15];
#, Heinz and Hoffman [16];n, Heinz and Hoffman [17];m, Stanley
et al. [36]).
Fig. 2. Logit and probit models fit to laboratory mallard duckling
survival at 7 d posthatch (h, Heinz et al. [15];#, Stanley et al. [36];
n, Stanley et al. [37];m, Heinz and Hoffman [16];l, Heinz and
Hoffman [17]).
which the dose–response relationship began, and to describe
the dose–response relationship above the inflection point.
Hockey stick is a form of nonlinear model in which the
independent variable is broken into two nonoverlappingranges
and separate linear regression models are fit to the response
variable in these two subsets [30,32]. The two models are
connected at an inflection point (quantification of the break-
point [t]) by constraining the second intercept to be a function
of the first model and t. This allows separate relationships to
be fit to different subsets of the independent variable, but all
data are used to estimate the slope and intercept of each line
as well as the inflection point at which they connect. The
general form for this model can be expressed as
a 1 bx x #t
y 5 (4)5c 1 dx x $t
When c 5 a 1 t(b 2 d),y will be continuous at all values
of x,and the final model will be of the form
a 1 bx x #t
y 5 (5)5a 1 b t1d (x 2t)x $t
Hockey stick models are frequently used to describe data
with a threshold response [31,32]. As part of model fitting,
maximum likelihood methods are used to determine whether
a significant linear relationship exists at values of the inde-
pendent variable that are less than the inflection point. If no
relationship exists between the independentand dependentvar-
iable when the independent variable is less than t (i.e.,b 5
0), then the model can be simplified to
ax#t
y 5 (6)5a 1 d (x 2t)x $t
In this application, the intercept estimates the mean re-
sponse associated with selenium concentrations less than t or,
in other words, the background level of response (teratogenic
or inviable eggs) that is not related to the presence of selenium.
The value of t is the selenium concentration at which a sig-
nificant relationship begins between selenium and response.
As a result, the minimum detectable effects concentration can
be estimated as the ECt.
In all cases, the selenium concentrations were analyzed on
a log scale to distribute the data more uniformly over the range
of concentrations assessed, allowing each data point to con-
tribute more equally to the fit of the curve [32]. The datasets
were truncated at the lowest concentration at which the max-
imum observed response occurred to represent the true slope
of the dose–response relationships. Models were fit in SPSS
[20]. An r2 value that assumed no intercept was used to assess
the fit of the models:
n
2eOi
i512 2r512(r for a no-intercept model) (7)n
2yOi
i51
RESULTS
Updated analysis of laboratory data for mallards
Inclusion of three data points from Heinz and Hoffman
[16,17] in the mallard laboratory dataset previously analyzed
by Fairbrother et al. [13] did make a slight difference in es-
timates of the EC10 for teratogenesis and duckling survival 7
d posthatch. Previously, Fairbrother et al. [13] reported an
estimated EC10 of 26 mg/kg dry weight for teratogenesis and
of 16 mg/kg dry weight using both probit and logit models
for duckling survival. Analysis of the revised dataset resulted
in an EC10 of 23 mg/kg dry weight for teratogenesis and of
14 to 15 mg/kg dry weight (logit, probit models) for mallard
duckling survival (Figs. 1 and 2). One data point reported by
Heinz et al. [35] was not included in the analysis (as discussed
by Fairbrother et al. [12]), because the mean egg selenium
concentration (15.9 mg/kg dry wt) associated with duckling
mortality (i.e., 76%, corrected for control mortality) was high-
er, by a factor of seven- or eightfold, than would be expected
as compared to all other data in the dataset. When the data
point is included in the dataset, the calculated mallardduckling
mortality EC10s for mean egg selenium are slightly lower and
range from 13 to 14 mg/kg dry weight using probit and logit
models, respectively. In addition to the probit and logitmodels,
we also fitted the data to a hockey stick regression (Fig. 3).
Analysis of the dataset via this method resulted in a slightly
lower mean egg selenium EC10 of 12 mg/kg dry weight (r2
5 0.90).
GLiMs applied to field data for black-necked stilts
Analysis of the field data showed teratogenesis to be a less-
sensitive endpoint than stilt egg inviability, as expected from
previous studies. The proportions of teratogenic eggs in the
field samples were much lower than the proportion of inviable
eggs (1–11% for teratogenesis and 29% for egg inviability).
It is interesting to note that the range of concentrations over
which effects were found forteratogenicitywasmuchnarrower
Deriving egg selenium toxicity thresholds for birds Environ. Toxicol. Chem.22, 2003 2025
Fig. 3. Hockey stick regression of laboratory mallard duckling sur-
vival at 7 d posthatch. EC10 5 10% effect concentration.
Fig. 4. Logistic model fit to duck teratogenicity field data. EC10 5
10% effect concentration. Dataset from Skorupa [29].
Table 3. Binary logistic model summary statistics for teratogenesis (duck, stilt, and avocet) and clutch inviability (stilt) using field data
Teratogenesis
Duck egg Stilt egg Avocet egg
Inviability
Stilt clutch
No. of cases 135 608 572 409
Not affected (0)
Affected (1)
123
12
541
67
564
8
290
119
Model statistics
Intercept (standard error)28.97 (2.34)26.13 (0.58)27.48 (1.18)22.32 (0.22)
Slope (standard error) 0.30 (0.09) 0.11 (0.01) 0.07 (0.01) 0.05 (0.01)
22 Ln(likelihood)
Intercept only in model 80.99 421.86 84.20 493.26
Intercept and slope in model 35.11 224.54 42.46 402.54
Reduction 45.88 197.32 41.74 90.72
Goodness of fit
Nagelkerke r2
Dhrymes r2
0.639
0.567
0.554
0.468
0.514
0.496
0.284
0.184
than the range for the egg inviability data. In fact, egg invi-
ability occurred over almost the entire range of egg selenium
concentrations, including very low concentrations.Thiscaused
the slope of the logistic model for inviability to be more shal-
low, and the goodness-of-fit measures to be lower, than the
respective statistics for any of the species teratogenicity mod-
els (Table 3).
In the present analysis, the data as well as the model curves
are shown, and the goodness-of-fit measures, Nagelkerke r2
and Dhrymes r2, are presented for all three models (Figs. 4–
6). This yielded interesting results, in that the goodness-of-fit
estimates (r2) were much lower for the inviability logisticmod-
el than for the teratogenesis logistic models. Nagelkerke r2
ranged from 0.514 (avocet) to 0.639 (duck) for the teratogen-
esis models but was only 0.284 for the stilt clutch inviability
data (Table 3). Dhrymes r2 ranged from 0.468 (stilt) to 0.567
(duck) for the teratogenesis models but was only 0.184 for the
stilt clutch inviability model (Table 3). The low r2 values for
the stilt chick inviability model goodness-of-fit measures in-
dicate either that the fitted logistic curve does not characterize
the data well or that considerable scatter around the model
exists.
Three-point moving average applied to field data
Three-point moving average responses were calculated for
each whole-number egg selenium concentration (2, 3, 4, . . .
mg/kg dry wt). Initially, a plot of the dose–response relation-
ship for concentrations up to 9 mg/kg dry weight forstiltclutch
inviability (Fig. 7) [29] showed an increase in the response
from 6 to 17% between 6 and 7 mg/kg dry weight, providing
support for the proposed egg selenium threshold of 6 mg/kg
dry weight. However, extending the analysis beyond 9 mg/kg
dry weight puts this increase in perspective and demonstrates
thatit is withinthebackgroundvariabilityofthedose–response
relationship for concentrations less than 30 mg/kg dry weight
(Fig. 7). For example, the increase in response between 6 and
7 mg/kg dry weight is identical to that between 27 and 28 mg/
kg dry weight (i.e., from 6–17%). In general, the response is
variable for low selenium concentrations, with no consistent
relationship beginning until concentrations near 30 mg/kg dry
weight (Fig. 7). Presentation of the entire dataset allows for
an interpretation of the data that differs from that published
by Skorupa [29].
Hockey stick regressions applied to field and laboratory
data (egg inviability data)
Constrained hockey stick models, with zero slope at con-
centrations less than t, fit all the datasets and produced r2
values generally greater than 0.90. The r2 values adjusted for
the number of model parameters supported use of the hockey
stick model with only an intercept for selenium concentrations
less than t, as opposed to a separate, additional slope estimate.
Tables 4 and 5 present the ECt, the egg selenium concentration
(mg/kg dry wt) at t, the model EC10, and the r2 value for the
teratogenic and inviability data, respectively. The ECt rep-
resents the minimum effect concentration the model could dis-
tinguish from background for that binning scheme.
For the teratogenicity endpoint, which is selenium-specific,
the ECt occurred at approximately the 1% effect level (EC1)
for all binning schemes except for the laboratory data, for
2026 Environ. Toxicol. Chem.22, 2003 W.J. Adams et al.
Fig. 5. Logistic model fit to stilt teratogenicity field data. Dataset
obtained from Skorupa [29]; EC10 5 10% effect concentration.
Fig. 7. Comparison of clutch inviability with egg selenium concen-
trations using the complete dataset from Skorupa [29]. EC10 5 10%
effect concentration.
Fig. 6. Logistic model fit to stilt clutch inviability field data. Dataset
obtained from Skorupa [29]. EC10 5 10% effect concentration.
Table 4. Constrained hockey stick models fit to laboratory and field
teratogenesis data
Binning scheme r2
Egg sele-
nium at t
(mg/kg
dry wt) ECt (%)a
EC10
(mg/kg
dry wt)
b
Ac
Bd
Ce
Df
Eg
Lab (mallard)
1.00
0.98
0.99
1.00
0.83
0.79
21.9
23.4
20.9
18.6
13.5
17.4
0.67
1.11
0.96
0.81
1.11
0.15
23.0
24.3
23.3
21.3
15.5
21.3
a ECt5percentage effects predicted to occur at quantification of the
breakpoint (t).
b EC10 5 10% effect concentration.
c Binning scheme A 5 0to,2,2to,4,4to,8,8to,16 mg/kg
dry weight, etc.
d Binning scheme B 5 0to,5,5to,10, 10 to ,15, 15 to ,20 mg/
kg dry weight, etc.
e Binning scheme C used 11 eggs/bin and the median egg selenium
concentration for 12 bins for teratogenicity (see Materials andMeth-
ods).
f Binning scheme D 5 0to,3,3to,6,6to,12, 12 to ,24 mg/
kg dry weight, etc.
g Binning scheme E used the break points identified by the U.S. De-
partment oftheInterior[11],namelymedianseleniumconcentrations
of 2.7, 9.2, 21.5, 31, and 76.5 for stilt teratogenesis.
which it was approximately 0.1%. Egg selenium concentra-
tions corresponding to ECt ranged from 13.5 to 23.4 mg/kg
dry weight for the different binning schemes applied to the
field data and was 17.4 mg/kg dry weight for the laboratory
data. The calculated EC10s for the field data ranged from 15.5
to 24.3 mg/kg dry weight, and the laboratory EC10 was 21
mg/kg dry weight (Table 4). In each case, the EC10s were
greater than ECt, indicating that the EC10s do not fall within
the range of background responses. A representative example
of the hockey stick model (binning scheme A) is shown in
Figure 8. Several inferences can be made from these data.
First, the laboratory teratogenicity data appear to be slightly
less sensitive than the field data. Second, once egg selenium
concentrations are sufficiently high to elicit a teratogenic re-
sponse, the response is very steep, because the ECt, which
approximates the EC1 for these data, is nearly the same as the
estimated EC10.
The stilt clutch inviability data were adjusted to account
for effects on individual eggs using the AE:AH ratio equation
(Eqn. 1). Figure 9 provides a representative example of the
dose–response relationship observed for the egg inviability
endpoint using hockey stick regression. Selenium egg con-
centrations at t ranged from 19.5 to 44.7 mg/kg dry weight
for the different binning schemes, and EC10 values for stilt
egg inviability field data ranged from 20.9 to 31.0 mg/kg dry
weight (Table 5). The ECt was in the range of 7.3 to 15.6%,
with binning schemes B and E having ECt values greater than
10%, which precluded estimation of an EC10. In comparison
to the field data, the egg selenium concentration at t was 10.2
mg/kg dry weight for the laboratory data, and the EC10 was
12.3 mg/kg dry weight. Inferences from these analyses are that
the laboratory egg inviability data are more sensitive than the
field data, the slope of the dose–response line forthelaboratory
data is much steeper than the field data, and the egg selenium
concentration associated with field egg inviability is fairlyhigh
because of the variability in the field dataset (EC10s of 21–
31 mg/kg dry wt, depending on the binning scheme).
DISCUSSION
The use of mean egg selenium thresholds for regulating
point and nonpoint sources of selenium in water has become
increasingly common in the western United States. The state-
of-the-science and regulatory approaches for protecting wild-
life and aquatic life resources have been evolving in the di-
rection of incorporating tissue residue–based thresholds or cri-
teria. Toxicologically, a tissue residue–based endpoint appears
to be the most appropriate, because it integrates the effects of
multiple factors (e.g., selenium speciation, biotransformation,
species feeding habits) that ultimately influence exposure to
sensitive receptors (e.g., birds and fish). It is critical, therefore,
that the mean egg selenium threshold used for regulatory pur-
poses be scientifically robust, as has been the case in devel-
oping regulatory approaches for water-quality and wildlife cri-
teria.
Deriving egg selenium toxicity thresholds for birds Environ. Toxicol. Chem.22, 2003 2027
Table 5. Constrained hockey stick models fit to black-necked stilt
field egg inviability data
Binning scheme r2
Egg sele-
nium at t
(mg/kg dry
wt) ECt (%)a
EC10
(mg/kg
dry wt)
b
A
B
C
D
E
Lab (mallard)
0.94
0.93
0.90
0.97
0.95
0.88
19.5
44.7
30.9
30.2
43.7
10.2
7.3
15.6
9.8
9.5
14.7
0.0
20.9
NEc
31.0
30.7
NE
12.3
a ECt5percentage effects predicted to occur at quantification of the
breakpoint (t).
b EC10 5 10% effect concentration.
c NE 5 not evaluated.
Fig. 9. Hockey stick model fit to laboratory mallard and field stilt egg
inviability data (binning scheme A).
Fig. 8. Hockey stick model fit to laboratory mallard and field stilt
teratogenicity data (binning scheme A). EC10 5 10% effect concen-
tration.
In this paper, we have critically reviewed all available data
that could be used to derive an avian mean egg seleniumchron-
ic toxicity threshold. A threshold of 6 mg/kg dry weight based
on a clutch inviability endpoint for black-necked stilts using
field data primarily collected from central California has been
proposed [10]. This threshold was defined using a GLiM ap-
proach to calculate the EC3 and also employed a three-point
moving average analysis as supportive evidence. We have
compared and contrasted the statistical reliability of this
threshold with results determined using a variety of models
(GLiM, three-point moving average, and hockey stick regres-
sions) for both field and laboratory data.
Examination of the clutch inviability endpoint using GLiM
analysis indicates that using a binary logistic model, a rela-
tively limited amount of the variability observed in field re-
sponses (19–28%) is explained by selenium concentrations in
eggs. The low r2 is not unexpected for an endpoint generated
from field data in which the effect (reduced hatching success)
is not selenium-specific and can be confounded by other fac-
tors, such as weather, disease, and other contaminants. In con-
trast, for the selenium-specific teratogenicity endpoint,r2 val-
ues in the range of 60% were observed for both species of
birds using field data and GLiMs. The r2 valuesweresomewhat
higher using hockey stick regression analysis. These values
are much more consistent with the level of precision expected
from field data when a relationship is known to exist. Given
the low r2 observed for the field clutch inviability data, we
believe that these data lack the necessary robustness on which
to derive a chronic threshold.
A second analysis of the clutch inviability data used by
Skorupa [29] to support a threshold of 6 mg/kg dry weight
involved plotting inviability as a function of a three-pointmov-
ing average of egg selenium concentrations. The present anal-
ysis shows a significant increase in clutch inviability between
6 and 7 mg/kg dry weight, as reported by Skorupa [29]. How-
ever, plotting the entire dataset shows an equally significant
decrease in clutch inviability at 12 mg/kg dry weight, with
continued oscillations in the response variable until approxi-
mately 30 mg/kg dry weight, when a consistent increase in
clutch inviability is observed. Analysis of the dataset across
all egg selenium concentrations does not support 6 mg/kg dry
weight as a threshold.
As a final analysis, we binned the egg selenium data in
multiple ways, converted the data from a clutch-based to an
egg-based response, and used hockey stick regression to de-
termine the egg selenium concentration at which egg invia-
bility increased above background levels. Several important
inferences can be made from this analysis. Most important is
quantification of the breakpoint (t) in the regression. This is
the point at which the dose–response relationship becomes
distinguishable from background and, as a result, defines the
effect concentration (i.e., ECx defined here as ECt) that can
be estimated reliably for the dataset. This is somewhat anal-
ogous to determining the minimum significant difference in
hypothesis testing.
Results from this analysis indicate t to be in the range of
the EC7 to EC16 for the field egg inviability data. This is
important, because the proposed threshold from Skorupa [10]
is 6 mg/kg dry weight based on an EC3. Our analysis indicates
that the EC3 is less than t and cannot be reliably distinguished
from background for any of the response scenarios. The EC10,
which is a more commonly used toxicity threshold, can be
estimated reliably for three of the five binning schemes. We
believe this analysis adds to the weight of evidence that an
EC3 of 6 mg/kg dry weight cannot be reliably distinguished
from naturally occurring effects based on the available data.
A comparison of the EC10 values for egg inviability to the
same effect level for the teratogenicity data also is of interest.
As has been well documented in the laboratory, egg inviability
is a significantly more sensitive endpoint than teratogenesis
[13,15]. However, evaluation of the hockey stick regressions
for the field data shows a different trend, with two of the three
binning schemes for which an egg inviability EC10 could be
estimated indicating teratogenesis to be more sensitive than
egg inviability. We do not interpret this to mean that terato-
genesis is actually the more sensitive endpoint, because lab-
2028 Environ. Toxicol. Chem.22, 2003 W.J. Adams et al.
Table 6. Summary of laboratory and field-derived selenium thresholds for avian species
Endpoints
Lab/field
data
EC10 (mg/kg
dry wt)
a Type of analysis r2 Reference
Teratogenicity
Stilt teratogenicity
Stilt teratogenicity
Duck teratogenicity
Mallard teratogenicity
Mallard teratogenicity
Mallard teratogenicity
Field
Field
Field
Lab
Lab
Lab
37
16–24
23
26
23
21
Logitb
Hockey stick
c
Logitb
Weibullc
Logitd & Probit
b
Hockey stick
c
0.55
0.83–1.0
0.64
0.84
0.76
0.79
[10]
Present study
[10]
[13]
Present study
Present study
Inviability/mortality
Stilt clutch inviability
Stilt egg inviability
Stilt egg inviability
Mallard egg inviability
Mallard duckling
mortality
Field
Field
Field
Lab
Lab
16
21–31
24
12
16
Logitb
Hockey stick
c
Logitb
Hockey stick
c
Probite
0.28
0.90–0.97
N/A
0.88
0.86
[10]
Present study
Present study
Present study
[13]
Mallard duckling
mortality
Mallard duckling
mortality
Mallard duckling
mortality
Lab
Lab
Lab
14
15
12
Probitb
Logitb
Hockey stick
c
0.75
0.79
0.90
Present study
Present study
Present study
a EC10 5 10% effect concentration.
b Generalized linear model (GLiM) with logit or probit link and binomial error.
c Nonlinear model.
d EC10.
e GLiM with probit transformation.
oratory data and sound underlying toxicological principles
support why egg inviability would be more sensitive. Rather,
this highlights the considerable variability in the field egg
inviability data and the subsequent inability to distinguish se-
lenium-related effects from background until excessively high
egg selenium concentrations are observed.
Finally, a comparison of EC10 values for the mallard lab-
oratory data evaluated using hockey stick regression, probit,
and logit analyses showed similar results for egg inviability
and duckling mortality (12–16 mg/kg dry wt) (Table 6). These
analyses indicate that the laboratory egg inviability and duck-
ling mortality endpoints are more sensitive than teratogenesis,
which is consistent with previously drawn conclusions.
In summary, a review of the key teratogenicity endpoints
for stilts and ducks using laboratory and field data indicates
fairly good agreement between the EC10 for all groups. The
EC10 ranges from 16 to 24 mg/kg dry weight (present study)
across species groups and laboratory and field data (Table 6).
The hockey stick regression analyses indicate that the thresh-
old egg concentration of selenium for teratogenic effects in
the field for stilts might be as low as 15.5 mg/kg dry weight.
In contrast, the hockey stick regression analyses for stilt egg
inviability using field data indicate that the mean egg selenium
threshold (EC10) lies in the range of 21 to 31 mg/kg dry
weight, which is considerably higher than has been previously
reported. These elevated EC10 values reflect the variability
that exists in the field dataset for egg inviability; they do not
represent the true potential for effects on egg inviability be-
cause of selenium exposure. Evaluation of the egg inviability
and duckling mortality endpoints for mallard laboratory data
indicates that the mean egg selenium EC10 ranges from 12
mg/kg dry weight based on hockey stick regression analysis
to 14 to 15 mg/kg dry weight based on logit and probit and
analyses.
Given that a mean egg selenium threshold in the range of
12 to 14 mg/kg dry weight appears to be the most appropriate,
the question of how such a threshold should be used requires
comment. First, it should be noted that this threshold is based
on data for mallards, the most sensitive species of the few that
have been tested to date. Application of this threshold to stilts
or avocets may not be appropriate. This is particularly true for
avocets, which appear to be substantially less sensitive than
mallards.
Second, the question is raised regarding what exposure data
should be used for comparison to the threshold. Given that the
threshold is a mean egg selenium value, it will be most ap-
propriate to compare the threshold against a mean egg sele-
nium value for a given study area. However, caution must be
used in defining a study area, because significant heterogeneity
in terms of dietary exposure and subsequent egg selenium
concentrations could lead to inaccurate assessments of poten-
tial selenium impacts. An evaluation of the Heinz et al. [15]
dataset, which provides a significant portion of the data used
to derivethethreshold,indicatesameancoefficientofvariation
in egg selenium concentrations within treatment levels of 25%
(range, 20–29% for all treatments). We believe itis appropriate
for field data to exhibit the same level of homogeneity (i.e.,
coefficient of variation #30%) for comparison to the mean
egg selenium threshold. For study locations with variability
in egg selenium concentrations greater than 30%, it would be
more appropriate to evaluate subgroups, as the high variability
indicates subgroups may be at greater risk than indicated by
the mean egg selenium concentration for the group as a whole.
In conclusion, considering the weight of evidencedescribed
above, we believe that the field data used to derive an egg
selenium threshold of 6 mg/kg dry weight are seriously con-
founded by factors other than selenium. This leads toexcessive
variability; consequently, the field clutch/egg inviability data
are not sufficiently reliable for establishing a chronic egg se-
lenium threshold for effects. In contrast, evaluation of the lab-
Deriving egg selenium toxicity thresholds for birds Environ. Toxicol. Chem.22, 2003 2029
oratory chronic data for mallards using probit, logit, and hock-
ey stick regressions provide an EC10 mean egg selenium
threshold that can be estimated reliably. Given the controlled
conditions under which the laboratory data were derived and
the similarity of EC10 values obtained using multiple statis-
tical analyses, a mean egg selenium threshold for effects of
12 to 14 mg/kg dry weight based on laboratory data appears
to be both appropriate and conservative.
Acknowledgement—The authors acknowledge Kennecott Utah Cop-
per for funding this study, an anonymous reviewer for providing help-
ful comments, and Debbie Fetherston for assistance in preparing this
manuscript.
REFERENCES
1. Franke KW, Tully WC. 1935. A new toxicant occurring naturally
in certain samples of plant foodstuffs. V. Low hatchability due
to deformities in chicks.Poult Sci 14:273–279.
2. Poley WE, Moxon AL, Franke KW. 1937. Further studies on the
effects of selenium poisoning on hatchability.Poult Sci 16:219–
225.
3. Ohlendorf HM, Kilness AW, Simmons JL, Stroud RK, Hoffman
DJ, Moore JF. 1988. Selenium toxicosis in wild aquatic birds.J
Toxicol Environ Health 24:67–92.
4. Adams WJ, Brix KV, Cothern KA, Tear LM, Cardwell RD, Fair-
brother A, Toll JE. 1998. Assessment of selenium food chain
transfer and critical exposure factors for avian wildlife species:
Need for site-specific data. In Little EE, Delonay AJ, Greenberg
BM, eds,Environmental Toxicology and Risk Assessment,Vol
7. American Society for Testing and Materials, Philadelphia, PA,
pp 312–342.
5. Lemly AD, Smith GJ. 1987. Aquatic cycling of selenium: Im-
plications for fish and wildlife. U.S. Fish and Wildlife Service,
Washington, DC.
6. Hothem RL, Ohlendorf HM. 1989. Contaminants in foods of
aquatic birds at Kesterson Reservoir, California, 1985.Arch En-
viron Contam Toxicol 18:773–786.
7. Ohlendorf HM, Kilness AW, Simmons JL, Stroud RK, Hoffman
DJ, Moore JF. 1993. Food-chain transfer of trace elements in
wildlife. In Allen RG, Neale CMU, eds,Management of Irri-
gation and Drainage Systems: Integrated Perspectives.Ameri-
can Society of Civil Engineers, Park City, UT, pp 596–803.
8. Hoffman DJ, Ohlendorf HM, Aldrich TW. 1988. Selenium tera-
togenesis in natural populations of aquatic birds in Central Cal-
ifornia.Arch Environ Contam Toxicol 17:519–525.
9. Hoffman DJ, Heinz GH. 1988. Embryotoxic and teratogenic ef-
fects of selenium in the diet of mallards.J ToxicolEnvironHealth
24:477–490.
10. Skorupa JP. 1998. Risk assessment for the biota database of the
National Irrigation Water Quality Program. U.S. FishandWildlife
Service, Sacramento, CA.
11. U.S. Department of the Interior. 1998. Guidelines for interpre-
tation of the biological effects of selected constituents in biota,
water, and sediment Selenium. National Irrigation Water Quality
Program Information Report 3. Denver, CO, pp 139–184.
12. Fairbrother A, Brix KV, Toll JE, McKay S, Adams WJ. 1999.
Egg selenium concentrations as predictors of avian toxicity.Hum
Ecol Risk Assess 5:1229–1253.
13. Fairbrother A, Brix KV, De Forest DK, Adams WJ. 2000. Egg
selenium thresholds for birds: A response to J. Skorupa’s critique
of Fairbrother et al. 1999.Hum Ecol Risk Assess 6:203–212.
14. Ohlendorf HM. 2002. Ecotoxicology of selenium. In Hoffman
DJ, ed,Handbook of Ecotoxicology.Lewis, Boca Raton, FL,
USA, pp 465–500.
15. Heinz GH, Hoffman DJ, Gold LG. 1989. Impaired reproduction
of mallards fed an organic form of selenium.J Wildl Manag 53:
418–428.
16. Heinz GH, Hoffman DJ. 1996. Comparison of the effects of se-
leno-L-methionine, seleno-DL-methionine, and selenized yeast on
reproduction of mallards.Environ Pollut 91:169–175.
17. Heinz GH, Hoffman DJ. 1998. Methylmercury chloride and sel-
enomethionine interactions on health and reproduction in mal-
lards.Environ Toxicol Chem 17:139–145.
18. Bailer AJ, Oris JT. 1997. Estimating inhibition concentrations for
different response scales usinggeneralizedlinearmodels.Environ
Toxicol Chem 16:1554–1559.
19. Mathsoft. 1997.SPlus Software.Seattle, WA, USA.
20. SPSS. 1999.SPSS Software.Chicago, IL, USA.
21. McCullagh P, Nelder JA. 1989.Generalized Linear Models,2nd
ed. Chapman & Hall, New York, NY, USA.
22. HagleTM, MitchellGE.1992.Goodness-of-fitmeasuresforprob-
it and logit.Am J Pol Sci 36:762–784.
23. Menard S. 2000. Coefficients of determination for multiple lo-
gistic regression analysis.American Statistics 54:17–24.
24. Dhrymes PJ. 1986. Limited dependent variables. In Griliches Z,
Intrilagator MD, eds,Handbook of Econometrics.North-Holland,
Amsterdam, The Netherlands, pp 1567–1631.
25. Hosmer DW, Lemeshow S. 1989.Applied Logistic Regression.
John Wiley, New York, NY, USA.
26. Aitkin M, Anderson D, Francis B, Hinde J. 1989.Statistical
Modeling in GLIM.Oxford University Press, New York, NY,
USA.
27. Christensen R. 1990.Log-Linear Models.Springer-Verlag, New
York, NY, USA.
28. Hosmer DW, Hosmer T, Le Cessie S, Lemeshow S. 1997. A
comparison of goodness-of-fit tests for the logistic regression
model.Stat Med 16:965–980.
29. Skorupa JP. 1998. Selenium poisoning in fish and wildlife in
nature: Lessons from 12 real-world examples. In Frankenberger
WT, Engberg RA, eds,Environmental Chemistry of Selenium.
Marcel Dekker, New York, NY, USA, pp 315–354.
30. Cox C. 1987. Threshold dose–response models in toxicology.
Biometrics 43:511–523.
31. Horness BH, Lomax DP, Johnson LL, Meyers MS, Pierce SM,
Collier TK. 1998. Sediment quality thresholds: Estimates from
hockey stick regression of liver lesion prevalence in English sole
(Pleuronectes vetulus).Environ Toxicol Chem 17:872–882.
32. Piegorsch WW, Bailer AJ. 1997.Statistics for Environmental
Biology and Toxicology.Chapman & Hall, London, UK.
33. Finney DJ. 1971.Probit Analysis,3rd ed. Cambridge University
Press, London, UK.
34. Palisade. 1997.Best Fit.Newfield, NY, USA.
35. Heinz GH, Hoffman DJ, Krynitsky AJ, Weller DMG. 1987. Re-
production in mallards fed selenium.Environ Toxicol Chem 6:
423–433.
36. Stanley TR, Spann JW, Smith GJ, Roscoe R. 1994. Main and
interactive effects of arsenic and selenium on mallard reproduc-
tion and duckling growth and survival.Arch Environ Contam
Toxicol 26:444–451.
37. Stanley TR, Smith GJ, Hoffman DJ, Heinz GH, Rosscoe R. 1996.
Effects of boron and selenium on mallard reproduction and duck-
ling growth and survival.Environ Toxicol Chem 15:1124–1132.