270 
THE WILSON JOURNAL OF ORNITHOLOGY • Vol. 123, No. 2. June 2011 
cycling for all variables. All autocorrelation 
functions were evaluated at the a = 0.05 
significance level. We developed 17 regression 
models a priori, based upon previously published 
descriptions of predator-prey-nest success dynam¬ 
ics under the bird-lemming hypothesis (Summers 
1986. Bety et al. 2002). We fit models using 
maximum-likelihood, ranked competing models 
using Akaike’s Information Criterion corrected 
for small sample sizes (AIC<; Burnham and 
Anderson 2002), and compared among models 
using Akaike weights (w,-); interpreted as the 
probability the i a ‘ model was the best, given the 
data, and the set of models evaluated. We 
calculated the difference between the top model 
(lowest AIC t .) and all other models (Aj.) and 
considered models ^2 A / to be in the top model 
set. We estimated a pseudo-/? 2 as 1 - (the 
deviance of the f h model/the deviance of the 
intercept-only model) to quantify the amount of 
variation in the data accounted for by our best 
model (Cameron and Windmeijer 1996). 
Peak lemming years (LP) are predicted to be 
positively correlated with nest success (Table 1) 
and fox den occupancy tinder the bird-lemming 
hypothesis. However, trough lemming (LT) years, 
1 year following a peak lemming year, are predicted 
to be negatively correlated with nest success in the 
current year (Table 1). We predicted nest density 
(D) would be negatively correlated with nest 
success (Table 1) if high nest density increased 
fox den occupancy (FD) and predation pressure. 
However, nest density and nest success are also 
likely influenced by annual fluctuations in spring 
phenology. We predicted that median hatch date 
(HATCH; standardized to 1 Jun = 1) should be 
negatively correlated with nest success (Table I). 
We considered three models that included a 
categorical variable (T) indicating whether fox 
removal occurred (1994-1997) to evaluate the 
effect fox removal had on nest success and also 
indirectly the influence of arctic fox abundance. 
Fox removal was used as part of an experiment to 
manipulate tox predation pressure on nesting 
Canada Geese and consisted of a concerted effort 
to trap or shoot all fox on the study area beginning 
in late April and continuing through the median 
hatch date for Canada Geese (Walter 1996, 1999 ). 
We predicted arctic fox removal would reduce the 
number ot nest predators on the study area 
resulting in higher nest success. 
We used analysis of variance (ANOVA) and 
Spearman rank correlation tests, post hoc, to 
evaluate assumed associations among covariates 
under the bird-lemming hypothesis (Hollander 
and Wolfe 1999). Data summaries are presented 
as mean ± SD. We used Program R, Version 2.11 
(R Development Core Team 2010) for all statis¬ 
tical analyses. 
Our measure of arctic fox den occupancy was 
based on ^10 surveyed dens in all but 1 year 
however, 95% CIs for the proportion of occupied 
fox dens were large. We conducted a simulation 
to quantify the influence of uncertainty in our fox 
den occupancy covariate estimates in each year on 
our model selection results and fox den coefficient 
values. We selected 200 random samples with 
replacement from the fox den data for each year, 
1993-2004. The sample size in each simulated 
year was equal to the sample size observed for 
that year, and resulted in a vector of Is (occupied 
den) and 0s (unoccupied den). The proportion of 
occupied fox dens was calculated for each year for 
that iteration of the fox den data. These simulated 
fox den covariate data for each year were included 
in an analysis of all 17 a priori models and the 
models were ranked based on A,. This simulation 
produced 800 A, and 800 fox den coefficient (pm' 
estimates; 200 of each for every model (n - 4) 
where it occurred. We calculated the proportion of 
200 iterations where the FD covariate was 
included in the top model (A/ = 0) or top model 
set (A, 2) as a measure of model selection 
uncertainty caused by high variance in FD 
covariate estimates. We also considered the 20th 
and 780th ranked values of all simulated coeffi 
cient estimates for FD to represent the 95% Cl ot 
this parameter. We concluded the coefficient 
estimates were not sensitive to uncertainty in the 
covariate data used in the analysis if the 95% Cl 
did not overlap 0. 
RESULTS 
Nest Success and Covariates .—Estimated prob¬ 
ability of Canada Goose nest success, derived from, 
on average, 165 ± 64 nests per year (min = 24. 
max = 230). ranged from 0.01 to 0.85 with a mean 
of 0.48 ± 0.27 between 1993 and 2004. Estimates 
of nest success exhibited substantial annual vaiia 
tion (Fig. 2A) although autocorrelation function 
indicated no statistically significant correlations in 
nest success among years (Fig. 3A). 
Nest density at Nestor One ranged from 1.23 10 
12.22 nests/100 ha of wetland nesting habitat with 
a mean of 7.92 ± 2.95 between 1993 and 200* 
Nest density had low variation and was declining 
