Spatially explicit capture-recapture
We used a spatially explicit capture-recapture (SECR) framework to
estimate bobcat density (Efford and Fewster 2013). Because we performed
the sampling on pre-determined transects revisited three times during
the study area, we created “detectors” by splitting the study areas
into 1 km x 1 km grid cells. Only grid cells that overlapped the
transects were retained and we defined ‘detectors’ as the centroid of
each grid cell. We then assigned all bobcat-identified scats collected
on the transect/s in a given cell the unique code of that cell (or
detector) (Royle et al. 2014). Bobcats can move large distances (several
km in a day) and have large home ranges averaging 15.83 to 39.70
km2 (Ferguson et al. 2009); the distance between the
center of each cell and locations of scats were therefore negligible
from a bobcat movement and space use perspective and assigning the scats
location to the cell centroid facilitated the development of capture
history data and data analysis.
The following modeling framework and workflow used package secr(Efford 2022) implemented in the program R (R Core Team 2022). We used
ArcGIS (ESRI, Redlands CA) to create the habitat mask used as an
effective sampling area in our analysis. To model detection, we
calculated the sigma (σ) model parameter using a root pooled variance
function as a measure of 2D dispersion of the centroids, pooled over
individuals (Efford 2022). We found that a buffer width of 5 × σ around
our detector array reduced the probability of capturing a bobcat outside
this buffer to zero and increasing buffer width beyond this value had no
discernable effect on the estimated density (Figure 2). This area is
thus typically used as the effective sampling area in spatial
capture-recapture models (Borchers and Efford 2008). The value of σ was
1230 m. To investigate differences between the 2 study areas, AEP and
Vinton-Zaleski, we built a habitat mask by creating buffers around the
detectors equal to 5 × σ (6152 m) in ArcGIS; the resulting mask had two
different polygons, corresponding to the two study areas, and they had
an area of 543 km2 (AEP) and 580 km2(Vinton-Zaleski).
We tested several detection functions and selected a ‘cumulative
lognormal ’ detection function to use in subsequent analyses, as this
function performed better than other detection functions based on Akaike
Information Criterion corrected for small sample size (AICc) (Akaike
1998) comparisons of SECR models fit with half-normal, compound
half-normal, cumulative gamma, and cumulative lognormal (Table 3). We
also compared several predictor variables for detection including length
of transect per grid cell (t_length ), and various habitat
variables (proportion of developed, forest, open, and wetland habitat)
against a constant detection (null) model. We found that the constant
detection model performed the best, but several other models were
<2 ΔAICc from this model (Table 4). The model that included
detection as a function of the length of transect per grid cell failed
some variance calculations and thus was not included in model
comparison.
We fit a SECR state (observation) model using a spatial Poisson process
for animal activity centers (Borchers and Efford 2008) and included a
categorical predictor (study area: AEP or Vinton- Zaleski), as we
expected differences in density between the two areas based on
preliminary studies (Prange and Rose 2020; Popescu et al. 2021). We
compared this model to a constant density model (null) using AICc.
Lastly, because the data were collected within a single year (July 2018
to April 2019), it included a single birth pulse and each survey was
conducted over the course of several months, we did not investigate
potential differences in density between the three surveys. Instead, we
quantified the overall bobcat density and abundance during the study
period and differences between the two focal areas.