To assess evidence for species-specific responses to snow cover (date of
50% snow cover each year) and temperature (accumulated thaw-degree days
from 1 January–10 June each year), we estimated slope parameters of
least-squares linear regression models. We first fit models with unique
slopes for each species (i.e., models with full interaction between
species and environmental variables), with arrival date, duration of the
pre-lay period, date of nest initiation, clutch size, and nesting effort
(the annual number of nests of each species enumerated on known-area
plots) as response variables. If none of the species-specific slopes in
these models differed from 0 (i.e., no evidence of a species-specific
response to snow cover or temperature), we then fit more parsimonious
models with a common slope among species to further assess support for a
response to the two environmental variables. Next, we compared these
four response variables among the four species using standard anova
techniques and conducted Tukey’s HSD tests for post-hoc examination of
differences between species. For these assessments, we considered
results to be biologically meaningful at α = 0.05.
To assess factors that influenced the growth of brant and snow goose
goslings, we fit linear models to estimate body mass as a function of
gosling age and sex. Because we weighed goslings only once, we did not
estimate a growth curve for the period from hatch until fledge, but
instead modeled the growth of both goose species over the span of ages
represented in our samples (see Hupp et al., 2017). Although the mass
gain of goslings from both species is non-linear overall (Ankneyet al.,
1991), gosling growth rates are well approximated by a linear fit for
the period shortly prior to fledging over which we recaptured goslings
(Cooch et al., 1991). For semipalmated sandpiper and longspur chicks, we
log10-transformed both mass and age to control for
inherent patterns of unequal variance from hatch until fledging (NB:
goslings were captured only during the linear phase of their growth
cycle such that variances were equal across ages, making log
transformation unnecessary). Because we uniquely marked goslings and
semipalmated sandpiper chicks, we modeled the growth of individuals of
these species, but used the average brood mass per nest visit as our
response variable for longspurs. Finally, to focus on post-hatch factors
affecting chick growth, we excluded mass measurements collected on the
day of hatch for longspurs. For semipalmated sandpipers, we similarly
censored values collected prior to two-days age because chicks of
Arctic-breeding shorebirds rely primarily on internal yolk reserves to
fuel growth after hatch (Norton, 1973; Schekkerman et al., 1998).
We compared similar environmental metrics across all four chick-growth
model sets (Table 2b; see Supporting Information S1.2), and fit
linear-mixed effects models using least-squares regression to each
species independently using the package ‘lme4’ (Bates et al., 2015), and
standardized all four environmental variables to minimize collinearity
and facilitate model interpretation. We fit an intercept-only (null)
model in all model sets, and otherwise included chick age (semipalmated
sandpipers and longspurs) or chick age and sex (brant and snow geese) as
unstandardized covariates in all models. We combined these covariates
along with additive combinations of the aforementioned environmental
variables to create an all-subsets model set of 17 models for each
species. We fit mixed-effects models in order to control for multiple
observations of individuals from the same nest (all species) and repeat
measures of individuals across time (semipalmated sandpipers). We
employed multimodel comparisons to rank the support of each model based
on Akaike’s information criterion adjusted for small sample size
(AICc ) and averaged model results for each
species in proportion to Akaike weights wifollowing the approaches of Burnham and Anderson (2002). We calculated
the conditional and marginal R 2 of each model
using the R package ‘piecewiseSEM’ (Lefcheck, 2016) to assess objective
model performance (Nakagawa & Schielzeth, 2013), and performed
multimodel comparisons and model averaging using the R package
‘AICcmodavg’ (Mazerolle, 2019). We considered predictor variables with
95% confidence intervals that did not overlap zero to be biologically
meaningful, and generated model-averaged predictions for each species
using contrasting values of these biologically meaningful predictor
variables to visualize chick mass under varying conditions.
Specifically, we generated predictions representing growth under what we
term optimal (i.e., 75th-quartile values for predictor
variables with positive parameter estimates,
25th-quartile values for predictor variables with
negative parameter estimates) and sub-optimal (i.e.,
25th-quartile values for positive parameters,
75th-quartile values for negative parameters)
conditions. For environmental variables whose model-averaged parameter
estimates overlapped zero (i.e., uninformative predictors), we set these
predictor values at mean levels in the model-averaging process. All
analyses were performed in R (R Core Team, 2021), and values represent
mean ± SD unless otherwise noted.