Deer resource selection analysis
We defined “used” locations as those where we received a successful
location fix from any of our collared deer across the study period, for
a total of 3,924 used locations. Available locations were randomly
generated across the District of Oak Bay, bounded by the area for which
we searched and captured deer. We generated available points in a 3:1
ratio, the number needed to achieve an asymptotic distribution of
variable means (Gerber and Northrup 2020, Northrup, Hooten, Anderson Jr
and Wittemyer 2013). For each used and available location, we calculated
mean vegetation greenness (NDVI) and the percent area of tree cover
within a 50-m radius buffer, a size selected to minimize error
associated with GIS data resolution while also representing small-scale
resource choice. We also calculated the percent area of three small,
medium, and large residential lots, as well as road density
(km/km2). Parks and golf courses were poorly
represented inside buffer areas, so we measured proximity to these
features.
We evaluated β coefficients from a single global model containing all
selected landscape features. We chose not to do model selection as our
goal was not to find the most parsimonious (reduced) model with a small
subset of component variables, but rather to ascertain the selection
strength of multiple variables (Burnham and Anderson 2002). We used a
logistic regression in a generalized linear model (GLM with binomial
errors and a logit link) with used locations (1’s) and randomly selected
available locations (0’s) regressed against landscape covariates. We
examined the second order of selection (Johnson 1980) which examines use
within a group of animals: in this case the population of Oak Bay.
Third-order selection – selection of resources by individuals within
their home-range – is a useful analysis when considering
individual-specific behaviours or when resources are not available to
all members of a population (Manly, McDonald, Thomas, McDonald and
Erickson 2007), but given our small study area and the movement of
individuals across that area (making all resources accessible), the
second-order analysis was of interest here. We used k-fold
cross-validation (Roberts, Bahn, Ciuti, Boyce, Elith, Guillera-Arroita,
Hauenstein, Lahoz-Monfort, Schröder, Thuiller, Warton, Wintle, Hartig
and Dormann 2017) to examine model fit and calculated odds ratios (OR)
from β coefficients as eβ.