Habitat Model
To control for repeated captures of animals (e.g., a single
moose being photographed repeatedly), we first binned moose detections
daily, where each site was assigned a positive detection if any moose
were captured that day or a negative detection (absence) if no moose
were detected that day. We then binned data monthly to estimate moose
site use, controlling for the number of days each camera was functional.
The number of days moose were present and absent at each site was
combined to generate the response variable. In this approach, we assumed
that if a moose was not detected at a site within the years of sampling,
we could reliably state it did not occur there – rather than assuming
false absence as in an occupancy framework
( MacKenzie et al., 2003).
Moose habitat use was examined using binomial family generalized
linear models (GLM). Our predictor variables for each model included the
30 land cover, human features, and burn area variables detailed above.
We summarize the area of each response variable by generating twenty
buffers for each camera site, ranging from 250 m to 5000 m in diameter
at 250-m intervals. Our GLM models were run across all sites using
bidirectional stepAIC model selection
( Zhang, 2016). A top model and
scale were then determined for all data and for each age or sex category
examined. After determining a top model, we examined predictor variance
inflation factors (VIFs; ensuring each was < 4), component
plus residual plots (to check for missing polynomial
relationships( Fox et al., 2012)),
and residual versus leverage plots for top models. The estimate and
standard error were then assessed for each variable in each model.