Patuxent Wildlife Research Center 

Models for Managing Habitat of a
Swainson’s Warbler Breeding Population, Bond Swamp National Wildlife
Refuge, Georgia


Ocmulgee River near Macon, Georgia 
Field data collection—We established 21 transects totaling 18.4
km in AprilJune 2001 for Swainson’s Warblers searches at BSNWR .
Swainson’s Warbler territories were located also by searching the
entire refuge in addition to conducting distance sampling and playback
surveys for territorial Swainson’s Warblers along the transects.
From late April through midJune 2001, we collected data by
following singing males as they moved through their territories,
maintaining a distance of at least 10 m from these individuals.
During our intensive territory observations, we recorded GPS
coordinates for each location at which the individual bird was seen or
heard. Each individual was
followed for at least 12 hr. We
obtained an average of 20 locations per individual.
Counting dense cane stems at BSNWR (3.5 m x 97.5 m transect) in a 50 x 50 m plot of a Swainson's Warbler territory. 
We produced a simple map (10 x 10 m grid) from GPS locations of each SWWA territory. These maps identified a 50 x 50 m (0.25 ha) area used most intensively by each individual during our territory observations. We then established the 50 x 50 m plot above for measurement of habitat variables in the corresponding area of each territory.
In addition, we established 50 x 50 m (0.25 ha) random unoccupied plots for habitat measurements along our transects for the Swainson’s Warbler surveys. We measured the following habitat variables in all Swainson’s Warbler territory plots and random plots between 20 June and 1 August 2001: average cane height (m) in plot, to nearest 0.25 m, based on mean of 2 observers; total cane area (m2) in plot, mean from 2 observers; number of cane stems on 2 perpendicular 3.5 x 50m belt transects (stem density); number of shrub, vine, and small tree stems (≤2.5 cm dbh) on same transects above (stem density); number of samples with dead ground cover (100 point samples of dead ground cover, every m on transects); number of samples with no dead ground cover (ibid.); number of samples with water ground cover (ibid.); number of samples with broadleaved forb ground cover (100 point samples for live ground cover, every m on transects); number of samples with grass ground cover (ibid.); mean leaf litter depth from 20 samples, taken every 5 m along transects; % canopy cover (number of point samples with canopy cover present, out of 100 samples; diameter at breast height (dbh in cm) of tree in plot with maximum dbh; distance from center of plot (m) largest tree (dbh); number of tree falls in plot; number of canopy tree openings in plot; basal area (x 10 = sq. ft./acre); pH water of plot soil sample; pH salt (CaCl2) of plot soil sample; % silt in plot soil sample; % sand in plot soil sample; % clay in plot soil sample; and electrical conductivity (deciSiemens/m [dS/m]) of plot soil sample.
We used SAS Version 8 for statistical analysis of habitat plot data. We analyzed data for 56 Swainson’s Warbler territory plots (SWWA) (50 x 50 m, n = 56) and 110 randomly selected and unoccupied (RAND) plots. We developed a set of alternative multivariate habitat models for Swainson’s Warbler territories at BSNWR using stepwise binary logistic regression. We avoided potential multicollinearity by preparing a Pearson correlation matrix prior to logistic regression analysis and eliminated 4 variables that were highly correlated (r^{2} > 0.49) with others in our data set. We then ran several different logistic regression analyses with the remaining 18 variables, using varying alphalevels for entry and retention of variables in each model (alpha = 0.05, 0.10, and 0.15, respectively).
We developed 3
habitat models using logistic regression, with increasingly permissive
alphalevels for entry and retention of variables in each model.
The most parsimonious model included 4 variables: cane stem
density, shrub+vine+sapling stem density, mean litter depth, and ground
coverage by water (alpha = 0.05 for both entry and retention; AIC = 80.6).
Other models contained more variables (611) and described the
habitat in more detail, both biologically and intuitively.
Not surprisingly, predictive capability improved with the addition
of more variables to the model (i.e., with increasingly "levels), as
estimated by several rank correlation indices calculated for the three
models (SAS Institute Inc. 1990). Rank
correlations for assessing predictive ability of the models by GoodmanKruskal
Gamma values were 0.941, 0.950, and 0.981 (out of 1.00) for the 4, 6,
and 11variable models, respectively.
Percent concordance of the predicted probabilities and observed
responses were high for all models (97.0, 97.5, and 99.0 respectively).
A manager's 4variable
habitat model for Swainson's Warbler territories, BSNWR:
SWWA = 8.8304 + 0.0109 (number of cane stems) + 0.0158 (number of shrub stems) + 0.0945 (mean depth of litter)  0.4107 (% ground cover by water)* Chisquare = 141.67, df = 4, P <
0.0001 

Bird’s eye view of ideal foraging habitat at litter layer and in a cane patch at BSNWR. It’s open for walking SWWAs to move and turn leaves in search of insects and spiders. Litter depth and cane stems are important indicators of SWWA territories. 
Our
analysis confirms that cane is an important component of Swainson’s
Warbler breeding habitat at BSNWR. Not
only were mean cane area, cane height, and cane stem density much greater
in territories than in random plots, but cane stem density was the most
significant variable entered into each of our 3 logistic regression
models. The greater
shrub stem density we found in territories, compared with unoccupied
random plots, is of interest with regard to Graves’s (2001) work in
Virginia. Shrub stem density
was the second variable entered in our logistic regression models,
demonstrating that density of noncane shrublevel vegetation may be
nearly as important as the canerelated variables in separating
Swainson’s Warbler territory plots from unoccupied plots at BSNWR.
While Swainson’s Warblers may preferentially establish breeding
territories within cane patches in areas where cane is abundant, overall
structure of the shrub layer (including both cane, if present, and other
shrublayer species) may be a more important factor in determining
suitable breeding habitat. The
shrub stem count in our study area also included vines such as greenbriar.
In Graves’s (2001) logistic regression model, Greenbriar
abundance was one of the most important variables for predicting
Swainson’s Warbler presence.
Leaf litter depth was the third most important variable in our
models and was greater in territories than in random plots.
Conversely, unoccupied plots had a greater percentage of areas with
no leaf litter and/or the presence of standing water; the latter variable
was the fourth most important variable in the habitat models, and was
negatively correlated with Swainson’s Warbler territories. As in our
study, Graves (2001) also found the presence of standing water to be
negatively correlated with Swainson’s Warbler presence in Virginia.
The mean difference in litter depth between our Swainson’s
Warbler territory plots and unoccupied plots was 50% greater in
territories. This difference in litter depth may be important, in terms of
food availability and foraging efficiency for the Swainson’s Warbler.
^{3} Email: joe_meyers@usgs.gov
U.S. Department of the Interior
 U.S.
Geological Survey Patuxent Wildlife Research Center, Laurel, MD, USA 207084038 Contact: Joe Meyers, email: joe_meyers@usgs.gov Last modified: 01/08/2003 USGS Privacy Statement 