Thompson et al • LANDSCAPE EFFECTS ON CERULEAN WARBLER ABUNDANCE 723 
in numbers of birds. We calculated the number of 
singing Cerulean Warblers in each segment. We 
created maps of selected land cover types for the 
two different buffer sizes around each river from 
Version 7-21-2000 of the National Land Cover 
Data (NLCD: http://www.mrlc.gov/). We com¬ 
pared 2000 and 2006 versions of the NLCD since 
surveys spanned 1999-2006 and there was <1% 
decline in forest cover. Thus, wc used the 2000 
NLCD for all analyses to avoid having to address 
potential compatibility issues between classifica¬ 
tions. We mapped upland forest as all forested 
upland classes from the NLCD, bottomland forest 
as the woody wetland class from the NLCD. 
developed land as all developed classes from the 
NLCD. and forest as bottomland and upland forest 
combined. We intersected each buffered river 
segment with the land cover map in a geographic 
information system and calculated percent cover¬ 
age of each land cover. 
Statistical Analysis.—We used an information- 
theoretic approach (Burnham and Anderson 2002) 
to evaluate our three hypotheses concerning factors 
affecting Cerulean Warbler abundance. We con¬ 
structed a set of three candidate models represent¬ 
ing our three hypotheses plus a null model 
consisting of only an intercept term representing 
constant abundance. We included percent cover of 
upland forest, bottomland forest, and developed 
land in the 250-m buffer as fixed effects in the 
model for hypothesis # I (local effects). We 
included percent cover of upland forest, bottom¬ 
land forest, and developed land in the 10-km buffer 
for hypothesis # 2 (landscape effects). We included 
percent cover of bottomland forest and developed 
land in the 250-m buffer and percent forest in the 
10-km buffer for hypothesis # 3 (local and 
landscape effects). We combined forest types in 
the 10-km buffer to reduce cross scale correlation 
with bottomland forest in the 250-m buffer: 
tolerance values were >0.48 for all variables in 
the model indicating no substantial multicollinear- 
ity (Allison 1999). We used percent forest cover in 
a 10-km buffer as a metric of habitat availability 
and fragmentation (sen.su Robinson et al. 1995) 
because other fragmentation statistics are highly 
correlated with percent forest cover in Midwestern 
landscapes (Robinson et al. 1995, Thompson et al. 
2002). Percent forest cover best explains variation 
in nest predation (Lloyd et al. 2005), and is a strong 
predictor of Brown-headed Cowbird (Mnlothrus 
uter) abundance and parasitism (Donovan et al. 
2000, Chace et al. 2005, Lloyd et al. 2005). 
We compared support for the models by 
ranking models from most to least supported 
using Akaikc's Information Criterion for small 
sample sizes tAKV; Burnham and Anderson 
2002). We evaluated the goodness-of-fit of the 
selected model using a /.-fold cross validation 
procedure (Boyce et al. 2002). We sequentially 
removed 15 randomly-selected observations with¬ 
out replacement and evaluated how well predic¬ 
tions from a model fit to the remaining observa¬ 
tions, compared to observed values for the 15 
observations, eight times and calculated the 
Pearson correlations between observed and pre¬ 
dicted values We plotted predicted counts of 
Cerulean Warblers for ~1() values across the 
range of each supported covariate that had 
biologically meaningful effects while holding 
other covariates at their mean. 
We fit a generalized Poisson model with 
random intercepts by maximum likelihood (Proc 
GLIMIX; SAS Institute Inc. Cary, NC. USA). We 
initially fit both a standard Poisson and general¬ 
ized Poisson model (Joe and Zhu 2005) to our 
global model and. since the generalized Poisson 
had a lower AlC r and overdisperion parameter 
(c*), we used it for all candidate models. We 
specified rivers as the subject for the random 
effect which allowed the intercept to vary among 
rivers. A river was surveyed in 1 year by the same 
observer and in a narrow range of dates; this 
model allowed us to accommodate year, observer, 
and date effects on detection probability. This 
model also acknowledges the likely correlated 
abundances of Cerulean Warblers among seg¬ 
ments on the same river. Counts estimated by this 
model are an index of relative abundance, but the 
random intercepts can account for difference in 
detection among rivers (and hence observers) 
when estimating the fixed effects. We acknowl¬ 
edge the desirability and benefits of methods that 
directly consider the probability of detection 
(Rosenslock et al. 2002); how ever, survey designs 
that cannot estimate detection probability may 
still provide useful indices of abundance (Johnson 
2008). 
RESULTS 
We conducted surveys along 16 rivers from 
1999 to 2006 that we subsequently divided into 
123 5-km segments. We detected 576 singing 
male Cerulean Warblers with an average of 4.7 
singing males per 5-km segment. Land cover 
varied among rivers ranging from 30.7% forest in 
