Full text of DOI:10.1002/etc.5620210608

Environmental Toxicology and Chemistry, Vol. 21, No. 6, pp. 1168–1175, 2002
Printed in the USA
0
730-7268/02 $9.00 ϩ .00
PREDICTING LEVELS OF STRESS FROM BIOLOGICAL ASSESSMENT DATA:
EMPIRICAL MODELS FROM THE EASTERN CORN BELT PLAINS, OHIO, USA
S
USAN B. NORTON,*† SUSAN M. CORMIER,‡ MARC
ARY
CHUBAUER-BERIGAN
U.S. Environmental Protection Agency, National Center for Environmental Assessment, Washington, DC 20460
U.S. Environmental Protection Agency, National Exposure Research Laboratory, Cincinnati, Ohio 45268
Ohio Environmental Protection Agency, Ecological Assessment Section, Groveport, Ohio 43125, USA
George Mason University, Department of Environmental Science and Policy, Fairfax, Virginia 22030, USA
National Institute for Occupational Safety and Health, Department of Health and Human Services, Cincinnati, Ohio 45213, USA
SMITH,§ R. CHRISTIAN JONES,࿣ and
M
S
#
†
‡
§
࿣
#
(
Received 31 January 2001; Accepted 25 November 2001)
Abstract—Interest is increasing in using biological community data to provide information on the specific types of anthropogenic
influences impacting streams. We built empirical models that predict the level of six different types of stress with fish and benthic
macroinvertebrate data as explanatory variables. Significant models were found for six stressor factors: stream corridor structure;
siltation; total suspended solids (TSS), biochemical oxygen demand (BOD), and iron (Fe); chemical oxygen demand (COD) and
2
BOD; zinc (Zn) and lead (Pb); and nitrate and nitrite (NO ) and phosphorus (P). Model
factor and highest for TSS, BOD, and Fe. Model
R
values were lowest for the siltation
x
2
R values increased when spatial relationships were incorporated into the model.
The models generally performed well when applied to a random subset of the data. Performance was more mixed when models
were applied to data collected from a previous time period, perhaps because of a change in the spatial structure of these systems.
These models may provide a useful indication of the levels of different stresses impacting stream reaches in the Eastern Corn Belt
Plains ecoregion of Ohio, USA. More generally, the models provide additional evidence that biological communities can serve as
useful indicators of the types of anthropogenic stress impacting aquatic systems.
Keywords—Multivariate regression
Spatial autocorrelation
Streams
Macroinvertebrates
Fish
INTRODUCTION
The current effort investigated the feasibility of building
multiple linear regression models that predict the degree of
different types of stress based on characteristics of the bio-
logical community. These predictions may provide useful in-
formation for identifying which stressors should be the subject
of management action. In addition, the results can be used to
evaluate whether stream communities respond in distinctive
ways to different types of stress. If communities respond in
consistently different ways, then distinctive models should se-
lected for different types of stress.
Biological assessment is becoming an increasingly popular
tool in the evaluation of stream ecosystem integrity. However,
little progress has been made to date in developing tools to
relate assessment results to specific stressors. This paper con-
tinues the investigation of the feasibility of using fish and
benthic macroinvertebrate community structure to distinguish
among major types and degrees of anthropogenic stressors in
the Eastern Corn Belt Plains ecoregion of Ohio, USA.
This paper builds on a previous effort that constructed a data
set of spatially and temporally matched stressor and response
data, reduced the stressor data to six orthogonal factors, and
explored the ability of the biological community to discriminate
among the different types and degrees of stress [1]. That study
found that biological variables could significantly distinguish
higher and lower quality sites classified on the basis of six
different types of stress: quality of stream corridor structure;
degree of siltation; total suspended solids (TSS), iron (Fe), and
biochemical oxygen demand (BOD); chemical oxygen demand
MATERIALS AND METHODS
Data set and preliminary analyses
The Eastern Corn Belt Plains ecoregion of Ohio was se-
lected as a study location to take advantage of the large amount
of biological monitoring data that was collected in a consistent
manner and made available by the Ohio Environmental Pro-
tection Agency [1]. The study was confined to one ecoregion
to minimize some of the natural physical, biological, and geo-
logical factors that may confound responses to stressors.
The data set contained spatially and temporally matched
descriptors of fish and macroinvertebrate community structure,
and variables associated with potential stressors, including in-
stream chemistry and habitat. The complete data set encom-
passed the years 1988 to 1994 and included 179 sites, 42
biological variables, 18 variables associated with stressors, and
the stream gradient and drainage area size associated with each
sampling location. Descriptive statistics for the data set are
provided in Norton et al.[1] and Norton [2].
(COD) and BOD; lead (Pb) and zinc (Zn); and nitrate and nitrite
(
NO ) and phosphorus (P). Functions based on biological var-
x
iables could also discriminate between sites having different
predominant stressors (12 of 15 pairwise combinations).
*
To whom correspondence may be addressed
(
norton.susan@epa.gov).
Presented at the 20th Annual Meeting, Society of Environmental
Toxicology and Chemistry, Philadelphia, Pennsylvania, November
1
4–18, 1999.
The views expressed in this paper are those of the author and do
not necessarily reflect the views and policies of the U.S. Environ-
mental Protection Agency. Mention of trade names or commercial
products does not constitute endorsement or recommendation for use.
Before analysis, the variables were transformed to near nor-
mality on the basis of visual inspection of the transformation
1
168

Predicting stress from biological data
Environ. Toxicol. Chem. 21, 2002
1169
series q-q plots and results of the Shapiro–Wilkes test [3–5].
In addition, we identified variables that were significantly cor-
models because of the spatial structure would be not be in-
cluded in the Cochrane–Orcutt models.
related (
p
ϭ
0.05) with drainage area. To minimize the influ-
This procedure transforms the dependent and explanatory
variables by using an estimate of the correlation between ad-
jacent residuals
ence of this covariate, we fit linear and quadratic regression
models to the identified variables and used the model residuals
as replacement variables [1,2]. Variables that proved important
in the models are shown in Appendix 1.
YtЈ ϭ ␤ Ј ϩ ␤ ЈXtЈ ϩ ␮t
(2)
0
1
The 18 stressor variables and stream gradient were reduced
to a set of six factors by using principal components analysis
where
Y
tЈ ϭ Yt Ϫ ␳ tϪ1
Ј0 ϭ ␤0(1 Ϫ ␳
and ␮t is the uncorrelated error term
et ϭ ␳ tϪ1 ut
Y
X
tЈ ϭ Xt Ϫ ␳XtϪ1
with varimax rotation in SAS
௡ [1,4]. Stream corridor structure
(
factor 1) was highly correlated with ordinal scores for channel,
␤
)
␤Ј ϭ ␤1
1
cover, the riparian zone, and pool depth; higher scores along
this factor correspond to higher quality sites (e.g., more sin-
uous reaches). Siltation (factor 2) was correlated with ordinal
scores for riffle quality, substrate quality, and embeddedness;
high scores along this factor correspond to higher quality sites
e
ϩ
(3)
where et is the model residual associated with observation t,
etϪ is the model residual associated with the closest upstream
(
i.e., less siltation). The last four factors were highly correlated
with stream chemistry variables: TSS, Fe, and BOD (factor
), COD and BOD (factor 4); Zn and Pb (factor 5); and N and
1
observation, and ut is the uncorrelated error term. The spatial
correlation coefficient
␳
is estimated by r
3
P (factor 6). High scores along the stream chemistry factors
correspond to lower quality sites (i.e., higher chemical con-
centrations).
n
͸
e
t
Ϫ
1
e
t
t
ϭ
2
n
r
ϭ
(4)
2
͸
e
tϪ
1
Model fitting
tϭ2
The purpose of the models was to predict the site score
along each of the six factors representing different types of
stress. Multiple linear regression was used as the baseline mod-
eling approach (base models)
The second approach (Stream ID) included individual
stream identifiers as additional explanatory variables. In this
expanded version of the base model (Eqn. 1), the suite of
explanatory variables included indicator variables for the 40
individual streams in the data set in addition to the fish and
macroinvertebrate variables. Stream identifiers were based on
U.S. Geological Survey topographic maps (scale 1:250,000).
The last approach (Lag) included the closest upstream value
of the dependent variable as one of the explanatory variables
[7]. More complex models can incorporate the relationship
between distance and the correlation between adjacent points;
as the distance between adjacent observations increases, var-
iable values are expected to become less similar [6,7,10]. How-
ever, exploratory plots (i.e., variograms [5]) indicated no clear
structure, so we used the simplest approach and included a
first-order lagged value, unweighted for the distance between
observations
Yi ϭ ␤ ϩ ␤1X1
ϩ
· · · ϩ ␤pϪ1XpϪ1 ϩ ␧i
(1)
0
where Yi are the values of the stressor factor being modeled;
␤0 ␤1, . . . , ␤pϪ1 are parameters (estimated by B0 B1, . . . ,
,
,
BpϪ1); Xi1, . . . , Xi, pϪ1 are the values of the explanatory (i.e.,
biological) variables; and ␧i are errors that are independent
2
and normally distributed
N
(0,
F
), estimated by the model
residuals, ei
i
ϭ 1, . . . , n
An exploratory evaluation of the models (analysis not
shown) indicated that model residuals located adjacent to one
another on the same stream were significantly correlated. This
result was not surprising; the issue of spatial and temporal
correlation has been studied extensively in time series mod-
eling, economics, and geography [6–9], and distance decay
functions are commonly used in mechanistic models of water
quality [9,10]. Spatial correlation of residuals can result in
underestimates of the both variance of the error term of the
regression model and the standard deviation of the estimated
Yt ϭ ␤ ϩ ␣YtϪ1 ϩ ␤1X1
ϩ
· · · ϩ ␤pϪ1XpϪ1 ϩ ␧t
(5)
0
where Yt is the value of the dependent variable at location t,
is the first-order autoregressive coefficient, YtϪ1 is the closest
upstream value of Y, and the other variables are as defined in
Equation 1. By incorporating spatial information more ex-
plicitly into the models, these Stream ID and Lag approaches
would be expected to yield higher
Within each stressor factor and modeling approach, several
models were fit by using a modified stepwise variable selection
procedure that was designed to select models while controlling
for correlation between the biological explanatory variables.
Up to four seed biological variables were selected (based on
their high correlation with the stressor factor) to initiate the
modeling process. Standard stepwise variable selection pro-
cedures were then used; however, variables that were highly
␣
2
regression coefficients. As a result, the coefficient of multiple
R
values.
2
determination (
R
) can be overestimated, and variables can be
concluded to be significant when in reality they are not [11].
For this reason, we investigated approaches to mitigate the
correlation.
Three approaches for reducing the spatial correlation in
model residuals were pursued. Each modeling approach was
pursued for all six stressor factors and the Durbin–Watson test
was used to evaluate whether the spatial correlation of the
residuals was significant [6]. The first approach, the Cochrane–
Orcutt procedure, uses the correlation between adjacent values
correlated (i.e.,
p Ͻ 0.01) with variables that were already in
the model were not considered for addition [4]. This procedure
resulted in up to four models for each modeling approach–
stressor factor combination. Final models were selected based
on testing performance (described below) and consistency of
explanatory variables.
to remove the influence of the spatial correlation [11]. Because
2
the spatial correlation is removed before the model is fit,
R
values are expected to decrease with this procedure. In addi-
tion, variables that may have been included in the baseline

1
170
Environ. Toxicol. Chem. 21, 2002
S.B. Norton et al.
Table 3. Final models for TSS, Fe, and BODa
Table 1. Final models for stream corridor structurea
Parameter
estimates
Parameter
estimates
Modeling approach R2
Db
Variablesc
Modeling approach
Base
R2
Db
Variablesc
Base
0.21
0.18
0.53
0.29
0.068
0.05
CYPRINID
OLIGO
CYPRINID
OLIGO
CYPRINID
OLIGO
SUNFISH
CYPRINID
OLIGO
0.14
0.34
0.47
TOXTOL
HEADWATR
RELWT
TOXTOL
HEADWATR
TOXTOL
HEADWATR
HEADWATR
OTHDIP
RELWT
TOPCAR
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
1.89
1.05
0.069
1.18
0.26
1.29
0.41
0.30
0.31
0.023
0.06
Ϫ
Ϫ
Ϫ
1.49
0.12
1.35
0.149
1.42
0.13
0.11
1.34
Cochrane–Orcutt
Stream IDd
Ϫ
Ϫ
Cochrane–Orcutt
Stream ID
Lag
0.12
Ϫ
Ϫ
0.04
0.13
0.27
0
.38e
0.76
0.69
0.66
d
Lag
0.31
Ϫ
0.055
Ϫ
a
Complete specifications for the models are provided in Appendices
2
D
and 3.
b
a
denotes the correlation between model residuals from adjacent
Complete specifications for the models are provided in Appendices
2 and 3; TSS total suspended solids; Fe iron; BOD bio-
chemical oxygen demand; Stream ID Stream identifier.
denotes the correlation between model residuals from adjacent
stream locations. Correlations significant at
Variables are defined in Appendix 1.
p
ϭ
0.05 are shown in italic.
ϭ
ϭ
ϭ
c
d
e
ϭ
b
Stream ID
ϭ
Stream identifier.
D
2
The
R
value adjusted for the number of variables [5].
stream locations. Significant correlations are shown in italic.
c
Variables are defined in Appendix 1.
d
2
The
R
value adjusted for the number of variables [5].
Model testing
The models were tested against data from a previous time
period and against a random subset of the data. The first ap-
proach tested models that had been fit with all of the 179
evaluated by calculating the mean square prediction residuals
(MSPR) with the mean square error (MSE) of the training
model [11]. To aid in this comparison, the ratio of the MSPR
and the MSE was calculated. If the MSPR is less than or equal
to the MSE (i.e., ratios less than or equal to 1), the residuals
in the test data set are comparable to those of the training data
set. If the MSPR is greater than the MSE (i.e., ratios greater
than 1), errors associated with application of the model would
be larger than expected. The MSPR values substantially greater
than the MSE indicate that the models would have poor pre-
dictive performance.
observations available for years 1988 to 1994 (n ϭ179) against
data collected from 1980 to 1987. A limitation of the test data
set was that only 28 samples representing seven streams had
complete data for the variables of interest. When the additional
constraints of the lagged variable approach were included (i.e.,
the observations in the test data set must be downstream of
those in the training data set), the sample size decreased to
1
1
5. All of the instream chemistry variables in samples from
980 to 1987 had lower or similar concentrations as the 1988
RESULTS
to 1994 data set except for Fe. Iron concentrations were sub-
stantially higher in the earlier time period. They ranged from
Significant models were found for all factors and all model
approaches. Biological variables selected for each model are
shown in Tables 1 through 6 and complete specification of all
of the models is provided in Appendices 2 and 3. Cyprinids
and oligochaetes were important variables predicting the quality
of stream corridor structure (Table 1). Percent Glyptotendipes
and insectivorous fish were important for predicting siltation
3
7
00 to 3,380
␮g/L during 1980 to 1987 compared with 22 to
43 g/L from 1988 to 1994. Before testing, the variables in
␮
the 1980 to 1987 data set were transformed to near-normality
and regressed against drainage area by using the same rela-
tionships as used in the training data sets.
For the second testing approach, the 1988 to 1994 data set
was randomly split 80:20, resulting in a data set of
n ϭ 143
for model building. Models were fit by using the procedures
described above and then were tested against the 36 withheld
samples.
Table 4. Final models for COD and BODa
Parameter
The expected predictive performance of the models was
Modeling approach
R2
Db
Variablesc
estimates
Base
0.29
0.44
DEP
Ϫ
1.22
0.18
0.09
0.138
1.53
0.12
0.096
1.58
0.26
1.75
0.089
0.30
Table 2. Final models for siltationa
Parameter
PIONEERP
MAYFLY
TANY
DEP
PIONEERP
TANY
DEP
OTHDIP
DEP
PIONEERP
OTHDIP
Ϫ
Ϫ
Ϫ
Modeling approach
R2
Db
Variablesc
estimates
Cochrane–Orcutt
0.19
Ϫ
0.04
Base
0.10
0.30
GLYP
INSECT
GLYP
GLYP
INSECT
GLYP
Ϫ
1.05
0.011
1.28
1.52
0.009
1.00
Ϫ
Ϫ
Stream ID
Lag
0.68
0.58
0.52
Ϫ
Ϫ
0.030
Cochrane–Orcuttb
Stream IDd
0.07
0.46
0.11
0.038
Ϫ
Ϫ
d
Ϫ
Ϫ
0.20
Ϫ
e
0
.30
Lag
0.18
0.12
Ϫ
a
a
Complete specifications for the models are provided in Appendices
Complete specifications for the models are provided in Appendices
2 and 3; COD chemical oxygen demand; BOD biochemical
oxygen demand; Stream ID Stream identifier.
denotes the correlation between model residuals from adjacent
2
and 3.
ϭ
ϭ
b
D
denotes the correlation between model residuals from adjacent
ϭ
b
stream locations. Significant correlations are shown in italic.
Variables are defined in Appendix 1.
D
c
d
e
stream locations. Significant correlations are shown in italic.
c
Stream ID
ϭ
Stream identifier.
Variables are defined in Appendix 1.
2
d
2
The
R
value adjusted for the number of variables [5].
The
R
value adjusted for the number of variables [5].

Predicting stress from biological data
Table 5. Final models for Zn and Pba
Environ. Toxicol. Chem. 21, 2002
1171
Parameter
estimates
Modeling approach
Base
R2
Db
Variablesc
0.13
0.09
0.57
0.46
ALLINT
NUMTAXA
ALLINT
NUMTAXA
ALLINT
NUMTAXA
ALLINT
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
0.06
0.03
0.057
0.02
0.06
0.02
0.047
Cochrane–Orcutt
Stream ID
0.054
0.008
0.039
Ϫ
Ϫ
0
.44d
Lag
0.29
a
Complete specifications for the models are provided in Appendices
2
D
and 3; Stream ID
denotes the correlation between model residuals from adjacent
ϭ Stream identifier.
b
stream locations. Significant correlations are shown in italics.
c
Variables are defined in Appendix 1.
d
2
The
R
value adjusted for the number of variables [5].
(Table 2). Percent headwater fish species, toxic tolerant inver-
tebrate taxa, and noninsect and dipteran invertebrate taxa were
important variables in the TSS, Fe, and BOD models (Table 3).
The proportions of mayflies, midges in the tribe Tanytarsini,
and depositional insect species were important variables for the
COD and BOD models (Table 4). Percent intolerant fish and
invertebrate taxa richness were important for predicting Zn and
Pb concentrations (Table 5). Finally, percent round-bodied suck-
ers, carnivorous fish, numbers of fish, and shredding inverte-
brates were selected for the N and P models (Table 6).
Within each stressor factor, the explanatory variables se-
lected varied across the different modeling options. However,
when substitution occurred, the original and substituted vari-
ables were always strongly correlated with each other. For
example, the number of intolerant species was replaced by the
percent of round-bodied suckers in the Stream ID model for
N and P factor. For the Cochrane–Orcutt procedure, variables
that had high correlations between adjacent points tended to
be dropped from the models. For example, the percent of shred-
ders did not meet the criteria for inclusion in the N and P
model after it was corrected for spatial dependence. When the
same explanatory variable appeared in several model ap-
proaches for the same stressor factor, the directionality of the
parameter estimate remained consistent.
Fig. 1. Plots of observed versus predicted values for the six stressor
factors with the stream identifiers modeling approach. TSS total
suspended solids; BOD biochemical oxygen demand; COD
chemical oxygen demand; NOx nitrate and nitrite.
ϭ
ϭ
ϭ
ϭ
The R2 values were the highest for the Stream ID models,
ranging from 0.30 for the siltation factor to 0.69 for the TSS,
Fe, and BOD factor. Plots of predicted versus observed values
for these models are shown in Figure 1. Standard regression
diagnostic plots were produced for each of the final models
(residuals vs fitted values, quantiles of residuals vs standard
normal distribution, Cook’s distance) [5,11]. They indicated
that no issues (e.g., overly influential observations) required
mitigation.
Table 6. Final models for NO and Pa
x
Compared with the base approach, the Cochrane–Orcutt
2
Parameter
estimates
procedure generally decreased the
R
values, whereas the
Modeling approach
Base
R2
Db
Variablesc
Stream ID code and the Lag modeling approaches increased
2
the
Lag modeling approaches markedly decreased the correlation
between adjacent residuals ( ); the correlations associated with
most models were not significant at 0.05.
The application of the 179 training set to the 28 samples
R of the models. The Cochrane–Orcutt, Stream ID, and
0.13
0.45
TOPCARN
RDSUCKPC
RELNO
Ϫ
Ϫ
Ϫ
0.125
0.098
0.39
␳
SHRED
0.247
0.128
0.080
0.48
0.16
0.11
0.09
0.057
0.35
p ϭ
Cochrane–Orcutt
0.11
0.56
0
0.36
0.003
TOPCARN
RDSUCKPC
RELNO
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
n
ϭ
collected from earlier years showed mixed success. Almost
one half (11 of 24) of the model tests had MSPR:MSE ratios
greater than 1.5. The MSPR:MSE ratios (Table 7) were near
or less than 1 for the stream corridor structure factor but were
uniformly high for the siltation factor. At least one model under
each approach (except for Stream ID) had an MSPR:MSE ratio
near 1 for the TSS, Fe, and BOD factor. Only the base model
approach and the Cochrane–Orcutt procedure had MSPR:MSE
ratios that were near 1 for the COD and BOD factor. Only the
base model had MSPR:MSE ratios that were near 1 for the
Stream ID
Lag
0.0039 TOPCARN
INTOLS
0.0005 TOPCARN
RDSUCKPC
.42d
RELNO
a
Complete specifications for the models are provided in Appendices
and 3; NOx nitrate and nitrite; P phosphorus; Stream ID
Stream identifier.
2
ϭ
ϭ
ϭ
b
D
denotes the correlation between adjacent model residuals. Sig-
nificant correlations are shown in italic.
c
Zn and Pb factor and the NO and P factor.
Variables are defined in Appendix 1.
x
d
2
The
R
value adjusted for the number of variables [5].
Models constructed with the n ϭ 143 data set (not shown)

1
172
Environ. Toxicol. Chem. 21, 2002
S.B. Norton et al.
Table 7. The MSPR: MSE ratios of test data (
n
28 from 1980–1987)
Stream ID Lag
variables and parameter estimates are consistent with biolog-
ical expectation.
a
and final models (n ϭ 179)
Sites with high scores for quality of stream corridor struc-
ture (Table 1) were associated with more minnows and fewer
oligochaetes. The increase in minnows may be related to the
existence of refugia at sites scoring high for stream corridor
structure. The increased number of oligochaetes at sites scoring
low for stream corridor structure generally is consistent with
the classification of oligochaetes as tolerant organisms.
Sites with high siltation were associated with a higher percent
of Glyptotendipes and fewer insectivorous fish (Table 2). High
proportions of Glyptotendipes have been associated with agri-
cultural nonpoint sources and with conventional municipal
waste treatment plants, perhaps because of the nutrients con-
tributed by both of these sources [12]. The increase in Glyp-
totendipes may provide evidence that the primary source of
sediments in the Eastern Corn Belt Plains ecoregion is from
agricultural soils that contain high concentrations of nutrients.
The relationships between siltation and the biological variables
were quite weak. A potential explanation for this is that the use
of the Hester–Dendy samplers mitigated the impacts of siltation
on the macroinvertebrates by providing hard surface habitat.
Still, the finding is inconsistent with another study in the same
region that found that the substrate metric was an important
predictor variable of index of biotic integrity (fish community)
scores [13]. However, the final explanatory variables used in
that study did not include drainage area or a surrogate (e.g.,
stream order), which may explain the inconsistency.
Cochrane–
Orcutt
Factor
Base
Stream corridor structure
Siltation
TSS, Fe, and BOD
COD and BOD
Zn and Pb
0.67
1.48
0.80
1.01
0.95
0.93
0.64
2.02
1.05
0.92
2.08
1.71
0.88
2.10
2.40
1.74
2.76
2.82
0.63
1.96
1.03
1.33
2.70
1.90
NO and P
x
a
MSPR
error; Stream ID
BOD biochemical oxygen demand; COD
nitrate and nitrite.
ϭ
means square prediction residuals; MSE
Stream identifier; TSS total suspended solids;
Chemical oxygen
ϭ mean square
ϭ
ϭ
ϭ
ϭ
demand; NOx
ϭ
were similar to those constructed with the full data set, with some
substitution of strongly correlated variables. The models generally
performed well with the randomly selected subset of 36 obser-
vations; 20 of the 24 model tests had MSPR:MSE ratios less than
1.5 (Table 8). The Stream ID models performed very well for
the TSS, Fe, and BOD factor and COD and BOD factors. The
Lag models performed well across all of the stressor factors.
DISCUSSION
This paper presents a series of models that use biological
assessment data to predict the degree of six types of anthro-
pogenic stress. Significant models were found for all six stress-
or factors explored in this study. The overall consistency in
explanatory variables and paramater estimates across the dif-
ferent modeling approaches increases the confidence that the
The models for the TSS, Fe, and BOD factor (Table 3) had
2
the highest
R
values, but presented difficulties in interpreta-
tion. The proportion of headwater species decreased with high-
er concentrations of TSS, Fe, and BOD. Headwater species
are considered to be sensitive but are also strongly influenced
by stream gradient, and their inclusion in the model may in-
dicate a residual effect of this covariate. The number and
weight of fish decreased at higher concentrations. The percent
of toxic-tolerant invertebrate taxa and noninsect and dipteran
invertebrate taxa also decreased with increasing TSS, Fe, and
BOD. This finding is inconsistent with initial expectations,
because these organisms are considered to be tolerant of chem-
ical stress. High concentrations of suspended solids may re-
duce the bioavailability of toxic substances in the water col-
umn. The increase in suspended solids actually may reduce
the effective toxicity of chemicals in the water column, yield-
ing these counterintuitive results.
Increasing values for COD and BOD (Table 4) were asso-
ciated with decreasing proportions of mayflies, midges in the
tribe Tanytarsini, and depositional insect species. Mayflies and
midges in the tribe Tanytarsini appear frequently in the literature
as sensitive indicators of stressors such as nutrients, sources
such as sewer overflows and industrial outfalls, and land uses
such as agriculture and urbanization [14–18]. The percent of
pioneering species, considered to be tolerant of stress, increased
with increasing concentrations of COD and BOD.
models are biologically meaningful.
2
The models with the highest
R
values included either the
Stream ID or the lagged dependent variable. The great im-
2
provement in
R values seen in the last two modeling options
has several interpretations. One is that not all of the important
variables were included in the base models. Stream ID may
be acting as a surrogate for variables that might represent the
initial condition of the stream community, other stressor var-
iables, or attributes of the stream that might mitigate response
to stress. The lagged variable has a more physical, source-
related interpretation. The level of stress at any site logically
is affected by the level of stress at upstream sites, which reflects
underlying fate and transport processes within the stream.
The parameter estimates are difficult to interpret quanti-
tatively because of the data transformations and the regression
against drainage area. However, in terms of directionality, most
Table 8. The MSPR: MSE ratios of test data (
n ϭ 36, randomly
a
selected) and final models ( 143)
n
ϭ
Cochrane–
Orcutt
Factor
Base
Stream ID Lag
Stream corridor structure 1.06
1.27
0.77
1.10
0.62
1.23
1.08
2.23
1.62
0.49
1.10
1.63
1.87
1.09
0.77
1.00
1.38
1.10
1.19
Percent intolerant fish decreased with increasing Zn and Pb
concentrations (Table 5). The percent of round-bodied suckers
and intolerant fish species also decreased with increasing con-
Siltation
0.82
0.78
1.10
1.05
1.16
TSS, Fe, and BOD
COD and BOD
Zn and Pb
centrations of NO and P (Table 6). Although both of these
x
NO and P
results are consistent with expectations in general terms, seeing
such similar explanatory variables indicates that the effects of
these two stressor factors may be difficult to differentiate in
the Eastern Corn Belt Plains ecoregion. The percent of shred-
x
a
MSPR
error; Stream ID
BOD biochemical oxygen demand; COD
demand; NOx nitrate and nitrite.
ϭ
mean square prediction residuals; MSE
Stream identifier; TSS total suspended solids;
chemical oxygen
ϭ mean square
ϭ
ϭ
ϭ
ϭ
ϭ
ders increased with higher concentrations of NO and P. If
x

Predicting stress from biological data
Environ. Toxicol. Chem. 21, 2002
1173
these nutrients increased plant growth, the detritus that shred-
ders feed on may also increase, or become more nutritive
through increased microbial colonization.
models would be unlikely to accurately predict the degrees of
different types of stress when applied to other regions. How-
ever, the relationships seen should provide relevant insights
into the general patterns of biological responses that we can
expect in response to these stressors. Finally, the models pro-
vide additional evidence that biological communities can serve
as useful indicators of the types anthropogenic stress that are
impacting aquatic systems.
The results of this analysis were consistent with the dis-
criminant function results described previously [1]. In that
study, sites were grouped into low-, medium-, and high-stress
categories based on quartiles of each stressor factor distribu-
tion. In most cases, the variables selected in this study were
among the most strongly correlated with the discriminant func-
tion for the same stressor factor. In a few cases, additional
variables proved important in the models. These included per-
cent top carnivores in the TSS, Fe, and BOD factor models;
percent mayflies in the COD and BOD factor models; and the
Acknowledgement—We gratefully acknowledge the assistance of
Dennis Mishne, Ed Rankin, Jeff DeShon, Steve Gordon, Sarada Ma-
jumder, Florence Fulk, and Karen Blocksom, and the comments of
Michael Troyer, Rick Racine, and two anonymous reviewers.
REFERENCES
number of fish and percent shredders in the NO and P factor
x
1
2
. Norton SB, Cormier SM, Smith M, Jones RC. 2000. Can bio-
logical assessments discriminate among types of stress? A case
study from the Eastern Corn Belt Plains ecoregion. Environ Tox-
icol Chem 19:1113–1119.
. Norton SB. 1999. Using biological monitoring data to distinguish
among types of stress in the Eastern Corn Belt Plains ecoregion.
PhD thesis. George Mason University, Environmental Science
and Public Policy, Fairfax, VA, USA.
models. Some of these differences can be attributed to the
clustering of sites into groups that was necessary for the dis-
criminant analysis. The regression approach has the advantage
of treating the stressor factor scores as continuous.
The strong testing results found by randomly partitioning
the 1988 to 1994 data set indicates that these models will have
predictive value for estimating the degree of stress at locations
where biological samples are available, but stream chemistry
or habitat variables are not, within this time period. In cases
where predictions are desired in streams that have at least one
existing sample, the use of the Stream ID or Lag models would
provide more accurate predictions.
3
4
5
6
. Cleveland WS. 1993. Visualizing Data. Hobart Press, Summit,
NJ, USA.
. SAS Institute. 1990. SAS௡ User’s Guide: Statistics, Ver 6. Cary,
NC, USA.
. Statistical Sciences. 1993. S-Plus for Windows User’s Manual:
Ver 3. 1. Statistical Sciences, Seattle, WA, USA.
. Diggle PJ. 1990. Time Series: A Biostatistical Introduction. Ox-
ford University Press, New York, NY, USA.
In contrast, the mixed testing results and the generally high
MSPRs for the Lag and Stream ID models when tested against
the observations collected from 1980 to 1987 indicate that
these models will have little predictive value for additional
observations from that time period. The particularly high
MSPRs for the Lag and Stream ID models indicate that the
spatial structure of the data changed from the earlier time
period to the later. This conclusion is supported by an analysis
of the historical trends in one of the streams—the Big Darby
Creek [19]. In that study, a higher degree of spatial correlation
was seen in the time periods of 1986 to 1993, which would
roughly correspond to the period of time of the training data
set, as compared to an earlier time period of 1979 to 1981.
Differences in spatial structure may be attributable to contin-
ued refinement of the biological sampling methods during the
early 1980s (M. Smith, personal communication). However,
management actions, including the removal of low-head dams
and the institution of additional wastewater treatment, also
occurred in the mid- to late 1980s. These changes likely also
changed the spatial structure of the data by increasing con-
nectivity in the streams.
7. Anselin L. 1988. Spatial Econometrics: Methods and Models.
Kluwer, Dordrecht, The Netherlands.
. Edbon D. 1985. Statistics in Geography. Blackwell, Cambridge,
8
MA, USA.
9. Thomann RV, Mueller JA. 1987. Principles of Surface Water
Modeling and Control. Harper Collins, New York, NY, USA.
0. Cressie N. 1991. Statistics for Spatial Data. John Wiley, New
York, NY, USA.
1
1
1
1. Neter J, Kutner MH, Nachtsheim CJ, Wasserman W. 1996. Ap-
plied Linear Statistical Models. Irwin, Chicago, IL, USA.
2. Yoder CO, Rankin ET. 1995. Biological response signatures and
the area of degradation value: New tools for interpreting multi-
metric data. In Davis WS, Simon TP, eds, Biological Assessment
and Criteria: Tools for Water Resource Planning and Decision
Making. Lewis, Boca Raton, FL, USA, pp 263–286.
3. Majumder S. 1998. A spatial empirical analysis of stressor–re-
sponse relationships for prospective ecological risk assessment
in the Eastern Cornbelt Plains of Ohio. PhD thesis. Ohio State
University, Columbus, OH, USA.
1
1
4. Lenat DR, Crawford JK. 1994. Effects of land use on water quality
and aquatic biota of three North Carolina piedmont streams. Hy-
drobiologia 294:185–199.
15. Quinn JM, Hickey CW. 1990. Magnitude of effects of substrate
particle size, recent flooding, and catchment development on ben-
thic invertebrates in 88 New Zealand rivers. N Z J Mar Fresh-
water Res 24:411–427.
1
6. Townsend CR, Arbuckle CJ, Crowl TA, Scarsbrook MR. 1997.
The relationship between land use and physicochemistry, food
resources and macroinvertebrate communities in tributaries of the
Taieri River, New Zealand: A hierarchically scaled approach.
Freshwater Biol 37:177–191.
Finally, the modeling results provide insight into whether
biological communities respond in distinctive ways to different
types of stress. In this study, very different biological variables
and parameter estimates best explained the variability for four
of the stressor factors: stream corridor structure; siltation; TSS,
Fe, and BOD; and COD and BOD. This indicates that the
biological communities may be responding differently to these
types of stress. The model variables and parameter estimates
17. Jones RC, Clark CC. 1987. Impact of watershed urbanization on
stream insect communities. Water Resour Bull 23:1047–1055.
8. Yoder CO, Rankin ET. 1998. The role of biological indicators in
1
a state water quality management process. Environ Monit Assess
5
1:61–88.
19. Schubauer-Berigan MK, Smith M, Hopkins J, Cormier SM. 1999.
Using historical biological data to evaluate status and trends in
the Big Darby Creek (Ohio) watershed. Environ Toxicol Chem
that fit for the Zn and Pb factor and the NO and P factor were
x
very similar, with both relying on intolerant fish species. This
indicates that these two types of stress may be difficult to
distinguish with these models.
The models produced in this effort are product of the spe-
cific stressors, processes, and biological communities present
in the Eastern Corn Belt Plains ecoregion of Ohio. The same
2
0:1097–1081.
20. Ohio Environmental Protection Agency. 1989. Biological Cri-
teria for the Protection of Aquatic Life: Vol III: Standardized
Biological and Field Sampling and Laboratory Methods for As-
sessing Fish and Macroinvertebrate Communities. Procedure
WQPA-SWS-3. Division of Water Quality Planning and Assess-
ment, Ecological Assessment Section, Columbus, OH, USA.

1
174
Environ. Toxicol. Chem. 21, 2002
S.B. Norton et al.
APPENDIX 1. Biological variables used in modelsa
Regressed against
Variable
ALLINT
Description
Transformation used
drainage area?
Number of sensitive fish sp. (i.e., ‘‘I’’ and ‘‘M’’ in Ohio En-
vironmental Protection Agency [Ohio EPA] species file)
Number of minnow species
Number of headwater fish species in a sample
Percent insectivorous fish
None
Yes: quadratic
CYPRINID
HEADWATR
INSECT
None
Log
None
None
Yes: Quadratic
Yes: linear
Yes: quadratic
Yes: quadratic
INTOLS
Number of intolerant fish sp. (i.e., ‘‘I’’ in Ohio EPA species
file)
PIONEERP
RDSUCKPC
RELNO
Percent individuals of pioneering fish species in a sample
Percent round-bodied suckers in a sample
Number of fish per unit distance
Weight of fish per unit distance
Number of sunfish species
Percent carnivorous fish in a sample
Percent of total number of invertebrates that are depositional
taxa
Square root
Square root
Log
Square root
None
Yes: linear
Yes: linear
Yes: linear
Yes: linear
Yes: quadratic
Yes: linear
Yes: linear
RELWT
SUNFISH
TOPCARN
DEP
Square root
0
.25
x
GLYP
Percent of total number of invertebrates that are Glyptoten-
x0.25
Yes: linear
dipes
MAYFLY
NUMTAXA
OLIGO
Percent of total number of invertebrates that are mayfly taxa
Total number of invertebrate taxa
Percent of the total number of invertebrates that are oligo-
chaetes
Square root
None
x
Yes: quadratic
No
Yes: quadratic
0
.25
OTHDIP
SHRED
TANY
Percent of total number of invertebrates that are dipterans
and noninsects
Percent of total number of invertebrates that are shredding
insect taxa
Percent of total number of invertebrates that are midges in
the tribe Tanytarsini
Percent of the total number of invertebrates that are toxic
tolerant
Log
Yes: linear
No
x0.25
Square root
x0.25
Yes: linear
Yes: linear
TOXTOL
a
Fish were sampled by electrofishing a standard length of stream, and invertebrates were sampled with Hester–Dendy artificial substrate samplers
20].
[
APPENDIX 2. Parameter estimates for final models: linear regression, Cochrane–Orcutt, and lag variable modelsa
Linear Cochrane–Orcutt Lag
Variable Estimate
Factor
Variable
Estimate
Variable
Estimate
Stream corridor structure
Intercept
CYPRINID
OLIGO
Ϫ
Ϫ
0.02645255 Intercept
0.13964855 CYPRINID
1.49003922 OLIGO
0.01999269 Intercept
0.12275532 Lag
1.35352723 CYPRINID
OLIGO
0.00337758 Intercept
1.27996383 Lag
GLYP
0.01394814 Intercept
1.18575477 Lag
0.26274274 HEADWATR
OTHDIP
0.02237040
0.26907438
0.11159972
1.33683716
0.00127199
0.31097454
1.00714717
0.01816708
0.59953495
0.30390340
0.03121970
0.02352797
0.05957139
0.03241357
0.50199922
1.75314809
0.08922815
0.29757691
0.01735244
0.46533381
0.04703134
0.06733056
0.44340646
0.09052676
0.05662102
0.34991299
Ϫ
Ϫ
Ϫ
Siltation
Intercept
GLYP
0.03597585 Intercept
1.04670787 GLYP
0.01135981
0.09890619 Intercept
1.89898569 TOXTOL
1.04728530 HEADWATR
0.06888591
Ϫ
Ϫ
Ϫ
INSECT
Intercept
TOXTOL
HEADWATR
RELWT
Ϫ
TSS, Fe, and BOD
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
RELWT
TOPCAR
COD and BOD
Zn and Pb
Intercept
DEP
PIONEERP
MAYFLY
TANY
Intercept
ALLINT
NUMTAXA
Intercept
TOPCARN
RDSUCKPC
RELNO
0.07216969 Intercept
1.22068150 DEP
0.18158845 PIONEERP
0.09324662 TANY
0.13841174
0.96872162 Intercept
0.05394964 ALLINT
0.02599705 NUMTAXA
0.14333429 Intercept
0.12507637 TOPCARN
0.09813103 RDSUCKPC
0.38632408 RELNO
0.24740824
Ϫ
Ϫ
0.01540381 Intercept
1.53665446 Lag
0.12083238 DEP
0.09621573 PIONEERP
OTHDIP
0.42454920 Intercept
0.05711383 Lag ALLINT
0.01761057
0.05159821 Intercept
0.12849238 Lag
0.07958748 TOPCARN
0.47811403 RDSUCKPC
RELNO
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
NO and P
x
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
SHRED
a
Variables are defined in Appendix 1; TSS
NOx nitrate and nitrite.
ϭ
total suspended solids; BOD
ϭ
biochemical oxygen demand, COD ϭ chemical oxygen demand;
ϭ

Predicting stress from biological data
Environ. Toxicol. Chem. 21, 2002
1175
APPENDIX 3. Parameter estimates for final models: Stream identifier (Stream ID) modelsa
Parameter estimates
Stream corridor
structure
Variable
Intercept
Siltation
0.73621
TSS, Fe, and BOD COD and BOD
Zn and Pb
NO and P
x
Ϫ
0.13832
0.23258
0.01478
0.45842
0.40057
0.21284
0.18139
0.30138
0.00533
0.72438
0.96912
0.82854
0.22640
0.71712
0.61610
0.53285
0.69266
0.18251
0.93667
0.87734
0.60381
0.31761
0.58400
0.96371
0.50360
0.06205
0.49696
1.19484
0.90821
0.25000
0.37439
0.14308
0.65938
0.67062
0.72319
0.76114
1.56217
1.23322
0.48859
0.27234
1.49138
0.14653
1.44217
0.12931
—
Ϫ
0.47070
0.07938
0.26025
0.51276
0.89464
0.43742
1.27789
0.64115
0.01615
0.36015
0.12643
0.17033
0.26492
0.55552
0.87093
1.50590
0.51196
0.03550
0.36231
0.11114
1.49075
1.23285
0.23014
0.49322
0.65607
0.45680
0.50141
0.55843
1.35892
0.32586
1.22770
1.43522
1.30825
1.41493
2.51133
1.24876
2.38426
0.62501
0.29473
1.01598
1.27986
—
Ϫ
Ϫ
0.17244
0.58324
0.26896
0.33583
0.46736
1.78086
0.34144
0.10655
0.07644
2.04947
0.25846
0.11336
0.16343
0.38431
0.78360
1.00706
0.30670
0.75101
1.34094
1.40664
1.97845
0.30789
0.11212
0.88966
1.40150
0.19528
0.65504
0.26939
0.01798
0.39862
1.70213
0.05287
0.37230
0.14842
0.87223
0.38875
0.55406
1.19479
0.04476
1.18842
2.57288
—
Ϫ
1.11789
0.49212
2.54876
1.41227
2.42228
3.51575
0.52495
1.69185
1.18871
2.23704
2.99707
0.99936
0.11115
1.54499
1.76447
1.00263
2.34784
2.19421
1.56977
1.59064
1.53047
2.42373
1.82671
2.74914
1.02993
1.41889
1.53569
0.85883
2.31634
1.80362
1.29954
1.65923
2.00498
1.80288
2.26296
1.01038
2.88834
2.92226
2.84301
1.12120
1.62348
—
Ϫ
2.38698
1.98291
2.95151
2.21760
1.41381
3.15929
2.91421
2.95129
2.98486
1.91627
2.54316
3.17029
2.87123
2.21479
3.56693
1.22028
3.24677
0.66129
2.20911
2.90545
1.83830
1.88610
2.83103
3.50999
1.14505
2.93159
2.45251
1.92109
1.88917
2.71530
2.60600
2.77636
3.00604
3.17934
2.29024
3.13625
3.24118
1.68723
2.78308
2.89603
1,57901
—
R01
R02
R02
R02
R02
R02
R02
R02
R02
R02
R02
R02
R02
R02
R04
R04
R04
R04
R04
R04
R05
R11
R11
R14
R14
R14
R14
R14
R14
R14
R14
R14
R14
R14
R14
R14
R14
R14
R14
R14
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
࿞
001
001
069
100
109
138
200
204
207
210
211
245
400
500
100
160
168
200
221
500
001
001
040
001
043
100
110
120
130
139
200
208
220
226
227
235
400
410
600
804
Ϫ
2.87121
1.40565
0.24102
0.63096
0.57209
0.80117
0.18803
1.39285
0.94588
0.68045
0.56214
0.59981
1.09356
0.51990
0.32446
0.39174
0.13204
0.21687
0.30182
0.19250
1.35300
0.10206
0.23852
1.44231
1.05509
0.89857
0.82392
1.77609
0.85427
0.50203
0.19281
0.76332
0.59810
0.08962
1.11363
0.33275
0.81730
0.41887
2.20998
0.36390
—
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
CYPRINID
OLIGO
SUNFISH
GLYP
INSECT
TOXTOL
HEADWATR
DEP
OTHDIP
ALLINT
NUMTAXA
TOPCARN
INTOLS
Ϫ
—
—
1.53833
0.00892
—
—
—
—
—
—
—
—
—
1.26984
0.37400
—
—
—
—
—
—
—
—
—
1.46668
0.26632
—
—
—
—
—
—
—
—
—
—
—
—
—
—
—
—
—
—
—
0.16093
0.11649
Ϫ
—
—
—
—
—
—
—
—
Ϫ
Ϫ
Ϫ
Ϫ
Ϫ
0.05523
0.01726
—
—
—
—
—
—
—
—
—
—
Ϫ
Ϫ
—
—
a
Variables beginning with R refer to Stream ID. Other variables are defined in Appendix 1. TSS
oxygen demand; COD chemical oxygen demand; NOx nitrate and nitrite.
ϭ total suspended solids; BOD ϭ biochemical
ϭ
ϭ