Barton springs salamander response to environmental

variables in Austin, TX
Audrey Archer & Maggie Ernest
October 7, 2014

Introduction
Background

Barton Springs Salamander (Eurycea sosorum)

are known to exist in only four spring outflows
collectively known as Barton Springs in
Austin, Texas. The largest spring, Main
Barton Spring, is the water source for Barton
Spring Pool, a popular public swimming spot.
The other springs are significantly smaller,
with both Eliza and Old Mill Springs having
man made structures around them to contain
and direct their flow. Upper Barton Springs is
located less than half a mile away from the
main spring, yet has a significantly different
chemistry and frequently goes dry (Eckhardt
2009). Urban runoff, increased development
in the Barton Creek watershed, toxic chemical
spills, and certain pool maintenance practices
have threatened Barton Springs Salamander
populations to the point that it was listed as a
federally endangered species in 1997 (Texas
Parks & Wildlife Department (TPWD); U.S.
Fish & Wildlife Services (USFWS)).
In order to maintain the citys permit from
USFWS to operate Barton Springs as a
recreation area, the Watershed Protection
Department was tasked with updating the
1998 Habitat Conservation Plan (HCP) for the
species (COA Watershed Department 2013).
In doing so, the department has conducted
intensive monthly surveys at each of the four
springs since 1998; the data collected from
these surveys was the data source for our
analyses. A preliminary look at the data
(Figure 1) from these surveys shows that
population abundance rises and falls similarly,
but at largely different magnitudes between
sites.

Figure 1. Surveys of Salamander abundance began

as early as 1993, however standardized protocols did
not exist until 2003; this is reflected by the sharp
increase in recorded abundance after this time.
Objectives

The principle objective of this study is to

model salamander abundance in relation to
recorded environmental variables. As a
secondary objective of the study, we determine
the relative importance of each of the
significant variables in the model in terms of
the magnitude of their effect. This is
important from a management perspective
because often, time, money, and labor are
constrained, therefore, identifying the
predictor variable with the greatest influence
could target management priorities most
efficiently.

Methods
Data Preparation

We only considered data for Eliza and Barton

Springs that were collected between 2003 and
2013, because these two sites had the greatest

salamander abundance and variation over

time and because survey protocols were not
standardized until 2003. The HCP details the
protocol
for
determining
abundance:
observers oriented in a line perpendicular to
the current, moving in concert from
downstream to upstream, removing all loose
substrate and replacing it behind the line. This
creates a moving, 6 to 10-inch strip of
coverless habitat that the salamanders are
reticent to cross. Each salamander crossing
coverless habitat upstream to downstream was
added to the cumulative number; any
salamander returning to upstream habitat
from behind the line was subtracted (161).
Unfortunately, we were unable to find any
information on how long these transects were.

determining salamander abundance. As

described previously, different numbers of
transects were sampled, but the length of these
transects is unknown. The overall procedure
for our analysis was to reduce the full model in
a stepwise fashion until we identify the
minimum adequate model. Subsequently, we
standardized the estimated coefficients by
their z-scores so that we can compare and
determine which predictor variables have the
greatest influence on salamander abundance.

Results
Exploratory Data Analysis

Before running the regression, we conducted

exploratory data analysis. In general, we see
that salamander abundance is positively
skewed, but when log transformed appears to
have a normal distribution (Figure 2). This
reinforces our assertion that Poisson regression
is the most appropriate analysis for this data
set. It is also evident that Eliza and Barton
Springs have very different totals for
salamander abundance (Figure 3). We also see
that Barton Spring is much deeper and has
more boulders than Eliza Spring. On the
other hand, we find that Eliza Spring has
more concrete and silt than Barton Spring
(Table 1).

With regards to the many environmental and

site parameters recorded, we only selected
those that are known to be most important to
salamander habitat and were also consistently
recorded. The parameters meeting these
conditions were water depth (in), water
temperature (C), dissolved oxygen (mg/L),
carbon dioxide (mg/L), and percent benthic
ground cover, including silt/sediment (<
1/16mm), sand ( 1/16 - 2 mm), gravel/pebble
(2 - 64 mm), cobble (64 - 256 mm) boulder (>
256 mm), and concrete. As anticipated, the
ground cover classes were highly correlated
with one another (having correlation
coefficients greater than 0.70), so we did not
include these in the model.
Statistical Analyses

Because our response variable, salamander

abundance, was count data with a strong
positive skew, we employed a Poisson
regression. The null hypothesis for the model
was that none of these predictor variables
would have a significant effect on salamander
abundance; our alternative hypothesis was
that one or more of the predictor variables
would have an effect on salamander
abundance. We included an offset in order to
account for different levels of effort in

Figure 2. Histogram of salamander abundance

revealing a positive skew of the raw data, which is
approximately normalized through a log
transformation.

variables have no effect on salamander

abundance.
Table 2. Coefficient estimate and p-values for the
minimum adequate model (Salamander ~ Site + CO2 +
Silt + Boulder + Concrete). All predictor variables are
statistically significant. Site and boulder are both
positively related to salamander abundance whereas CO2,
silt, and concrete all have a negative relationship with
salamander abundance.
Figure 3. Boxplots of salamander abundance for both
Eliza and Barton Springs. This suggests that their
relative abundances are different from each other.
Table 1. Mean values and standard
errors for variables of Eliza and Barton Springs. The
values in bold are greater than the other spring

By
back-transforming
the
estimated
coefficients, we were able to determine the
effect changes in these variables have on
salamander abundance. The effect of site is by
far the greatest: compared to Barton Spring,
Eliza Spring has 1571% more salamanders.
Regarding ground cover, there salamander
abundance increases by 35% with each
additional percent boulder, decreases by 5%
for concrete, and decreases by 1% for silt.
There is also a 1% decrease in abundances for
each additional unit increase in carbon
dioxide levels. The 95% confidence interval
for boulder is much larger compared to the
other predictors (Figure 4), having a much
more variable effect on salamander
abundance.

We considered each of the following variables

as main effects as our initial Poisson
regression: site (as a factor), carbon dioxide,
dissolved oxygen, water depth, water
temperature, silt, gravel, boulder, and
concrete.
This
model
was
highly
overdispersed, with a deviance value of 112.3;
this violates the assumption that the mean is
equal to the variance. To account for such
high dispersion, we updated the model with a
quasipoisson family type. We proceeded with
a stepwise reduction process, removing
insignificant variables (p > 0.05), until we
reached the minimum adequate model. The
resulting model included site, carbon dioxide,
silt, boulder, and concrete as predictor
variables for salamander abundance (Table 2).
Therefore, we rejected our null that these

In

order to address our second objective - to

make comparisons possible in order to identify
which predictor variables have the greatest
effect on salamander abundance - we
standardized all of the predictor variable
coefficients by their z-score (Table 3). We
back-transformed these coefficients as well, in
order to interpret them more readily. These
results showed site to be the most important
factor, which was unsurprising given the

differences in magnitude of abundance for

each of the sites. After site, percent boulder
cover was the most important factor, having
nearly two times the effect of silt and carbon
dioxide.

well and the leverage plot does not show that

any observations having undue leverage over
the data. Notably, data point 52 sticks out in
each of these diagnostic plots; however, it does
not appear to be an error or contain any
information that is out of the ordinary for the
data set, so we chose to keep the data point in
the model. Overall, we found that the model
is significantly different from the null model
with a change in deviance of 479 on 5 degrees
of freedom (p-value = 0.05). Despite this, we
find that our model does not fit the observed
data well, with a deviance of 12366 on 109
degrees
of
freedom
(p-value
~
0). Consequently, while the model is better
than the null, it is not worthy of predicting
salamander abundance for Eliza and Barton
Springs.

Figure 4. Coefficient plot (log scale) for predictor

variables of minimum adequate model (except site).
Confidence interval is based on one standard deviation

Table 3. Coefficients and p-values for each

predictor variable after standardization using zscores. These reveal relative magnitude of effect of
each predictor on salamander abundance. Site is the
most important factor influencing abundance,
followed by boulder.

Figure 6. Diagnostic plots for the

minimum adequate model.

Discussion
While the model we produced does not
accurately predict salamander abundance, it
does shed light on the importance of benthic
ground cover; specifically, the amount of
larger rocky substrates relative to smaller
particles. In reviewing the HCP, this is likely
due to the vast majority of salamanders
[being] found in interstitial areas beneath

Diagnostic plots reveal that our model fits

moderately well with our observed data
(Figure 6). We see from our residual plots that
there is a slight increase in variance with
larger values. Additionally, the QQ plot fits
4

rocks in flowing water...allowing these spaces

to serve as protection from aquatic and
terrestrial predators (165). We would expect
this microhabitat preference to show up in
levels of cobble rather than boulder, however
we removed cobble due to its strong
correlation with other variables. Because of
this, it is advisable to consider the relationship
between cobble and abundance before making
the decision to increase the number of
boulders at a site with the expectation of
creating more suitable salamander habitat.
This advice is emphasized considering the
abundance of salamanders at Eliza Spring,
where essentially no boulders exist.

in favor of monitoring CO2 levels over DO

levels, if such a decision must be made.
One glaring result of this analysis is the hugely
positive relationship between abundance in
Eliza Spring when contrasted against Barton
Spring. While we expected to see a significant
difference between the two sites, the statistical
results did not offer any logical reasoning for
the disproportionate abundances as we had
hoped. We again attribute the missing
explanatory variable, flow regime, to be a
possible explanation for the difference: Eliza
Spring is more typical of the shallow, flowing
streams in which the majority of salamanders
of related species are found (Watershed
Protection Department 2013). This contrasts
the deep water of the Main Spring, which is
maintained for swimming. That being said,
water depth did not play a significant role in
our model. A visual comparison of the two
sites, with Barton Spring being a clear blue
and Eliza more green, begs the question of
whether or not different nutrient levels are
what is driving the difference in salamander
abundance. It is also possible that the presence
of human activity in Barton Spring compared
to Eliza Spring (which is closed to the public),
has a negative effect on salamander
abundance. This is a difficult variable test
without conducting experiments that control
for human activity.

The excess variance in our model, which

informed us of the models poor fit of the
predicted values to observed, is likely
attributable to the omission of key
environmental variables that are important to
Barton Springs salamanders. One such
variable is water flow, which was
unfortunately only recorded from 2008 and
2011 and even then was not recorded
regularly (and thus not included in our
analysis). According to the HCP, flowing
water prevents accumulation of excess
sediment, which creates accessible hiding
spaces for the salamanders. This may explain
why there is a negative (though slight)
relationship between silt/sediment cover and
abundance. Flowing water also provides
constantly renewing dissolved oxygen, which is
essential for salamander metabolism and
respiration. Considering this, we were
surprised to find no significant relationship
between dissolved oxygen and abundance.
This may be attributable to the fact that the
species evolved in a karst groundwater system
where oxygen concentration is naturally
undersaturated, so is sufficiently adapted to
varying levels of dissolved oxygen (HCP 2013).
Carbon dioxide levels, however, did have a
significant
negative
relationship
with
salamander abundance; this provides evidence

Resources
COA Watershed Protection Department. Eliza
Spring Daylighting. Retrieved on Oct 6, 2014.

COA Watershed Protection Department (2013).

Habitat Conservation Plan for E. sosorum and
E. waterlooensis.
Texas Parks & Wildlife Division (TPWD). Barton
Springs Salamander Retrieved on Oct 6, 2014.
U.S. Fish & Wildlife Services (USFWS) Species
Profile: Barton Springs Salamander Retrieved on
Oct 6, 2014.