Bed site selection by a subordinate predator: an example with the cougar (Puma concolor) in the Greater Yellowstone Ecosystem

Ethics statement

Our capture protocols for cougars, as outlined in Elbroch et al. (2013), adhered to the guidelines outlined by the American Society of Mammalogists (Sikes & Gannon, 2011), and were approved by two independent Institutional Animal Care and Use Committees (IACUC): the Jackson IACUC (Protocol 027-10EGDBS-060210) and National Park Service IACUC (IMR_GRTE_Elbroch_Cougar_2013-2015). Every effort to ameliorate suffering of cougars was made, and no cougars were killed or sacrificed during capture events. Our study was carried out on the Bridger-Teton National Forest (United States Forest Service, USFS Authorization ID JAC760804), Grand Teton National Park (NPS Permit GRTE-2012-SCI-0067), and the National Elk Refuge (USFW permit NER12), with permission to handle cougars granted by the Wyoming Game and Fish Department (Permit 297).

Study area

Our study area encompassed approximately 2,300 km² of the southern Greater Yellowstone Ecosystem in Teton County, Wyoming, and included portions of the Bridger-Teton National Forest, Grand Teton National Park, and the National Elk Refuge (Fig. 1).

Elevations in the study area ranged from 1,800 m in the valleys to >3,600 m in the mountains. The area was characterized by short summers and long winters with frequent snowstorms. Average summer temperatures were 6.9 °C, and average winter temperatures were −7.2 °C (Gros Ventre SNOTEL weather station). Precipitation occurred mostly as snow, and maximum snow depths ranged from 100 cm at lower elevations to >245 cm at intermediate and higher elevations (2,000 m+). Habitats included foothill grasslands, big sagebrush (Artemisia tridentate) dominated shrub-steppe, Douglas-fir (Pseudotsuga menziesii) forests, aspen (Populus tremuloides) forests, and higher elevation coniferous forests, composed of lodge pole pine (Pinus contorta), subalpine fir (Abies lasiocarpa), Engelmann spruce (Picea engelmannii), and white bark pine (Pinus albicaulis). Riparian corridors were dominated by cottonwood (Populus ungustifolia, Populus balsamifera, and Populus trichocarpa) and willow (Salix spp.) communities (Marston & Anderson, 1991).

Other carnivores in the study system included grizzly bears, American black bears, wolves, coyotes (Canis latrans), and red foxes (Vulpes vulpes). Ungulate prey included elk (Cervus elaphus), mule deer (Odocoileus hemionus), white-tailed deer (O. virginianus), Shiras moose (Alces alces shirasi), bighorn sheep (Ovis canadensis), and North American pronghorn (Antilocapra americana).

Cougar capture, collar programming, and bed site identification

Each winter, cougars were located, immobilized, and fitted with satellite GPS collars (Lotek Wireless, Inc.; Newmarket, Ontario, Canada; or Vectronic Aerospace GmbH., Berlin, Germany). We used trailing hounds to force cougars to retreat to a location where we could safely approach them. Cougars were immobilized with ketamine (4.0 mg/kg) and medetomidine (0.07 mg/kg), and their temperature, heart rate, and respiration were monitored at five-minute intervals while they were processed, sampled, and fitted with a collar. Once animals were completely processed, the effects of the capture drugs were reversed with Atipamezole (0.375 mg/kg), and cougars departed capture sites on their own.

We programmed collars to acquire location data between 12 and 24 times per day and received and uploaded data to Google Earth daily. We identified GPS clusters visually in Google Earth, which we defined as any ≥2 subsequent locations ≥4 h apart, within 150 m of each other, and occurring within two weeks of each other. Because of the marked temperature differences between winter and summer and the migratory behavior of cougar prey, we analyzed data for two seasons. Following Elbroch et al. (2013), we defined seasons based on well-established elk migrations: summer, defined as June 1–November 30, and winter, defined as December 1–May 31.

We visited and examined clusters in the field, and 98% of all investigations were performed by CyberTracker-certified observers (Evans et al., 2009; Elbroch et al., 2011), ensuring expertise and consistent field effort. Bed sites were identified as a circular depression in the vegetation or snow containing identifiable cougar hair. When consuming prey, cougars often bedded within the immediate vicinity of their kills. So, to ensure that our examination of bed site selection was distinct from kill site selection, we only included beds that were unassociated with known kills. Therefore our analysis included only those beds selected by cougars in between time periods associated with handling prey, and more than 500 m from confirmed kill sites.

Second-order selection: comparing bed site attributes to landscape attributes

Following methods for resource selection functions (Boyce, 2006), we employed logistic regression with a binomial distribution (logit link function) to compare bed sites with random locations on the landscape; for each bed site, we sampled five random points from within the minimum convex polygon (MCP) bounding the location data of all our study animals during the study period, which we defined as our “study area.” Bed sites were mapped in ArcGIS 10.1 (ESRI, Redlands, CA, USA), and random points were generated using the Geospatial Modeling Environment (GME, Beyer, 2009–2012). Each bed site or random location was then assigned the following attributes: habitat type (n = 5), distance to forest edge, slope (30 m resolution), terrain ruggedness (a 3-dimensional vector ruggedness measure termed VRM (Sappington, Longshore & Thompson, 2007)), elevation, and aspect (categorized as north, south, east, and west).

We reclassified 87 land cover classes described in a Gap Analysis Program (gapanalysis.usgs.gov/gaplandcover) at 30 m resolution into five general habitat classes by lumping biologically similar cover classes together: (1) grasslands, meadow, or barren; (2) riparian and water bodies; (3) sagebrush and shrub-steppe; (4) forest; and (5) disturbed, agricultural, and urban. Aspect and VRM were derived from the digital-elevation model (http://datagateway.nrcs.usda.gov/) following the method of Sappington, Longshore & Thompson (2007). Forest edge was created following the methods of Elbroch et al. (2015b) by drawing a perimeter around each forested section.

To begin, we devised a list of bed site attributes that would support the thermoregulation hypothesis versus the predator avoidance hypothesis, from which we developed a set of a priori models to test against each other. We believed the following covariates influenced thermoregulatory properties of bed sites in both winter and summer: distance to forest edge, habitat type, aspect, and elevation. We believed that proximity to forest edges, aspect, and habitat class would modulate solar radiation and exposure to the sun or cold winds. Higher elevations in our study area were correlated with increased snowfall and lower temperatures, and thus we expected that in the winter, cougars would choose bed sites at lower elevations.

We believed the following covariates would influence predation risk by competitors at bed sites in both seasons: habitat class, proximity to forest edge, slope, and VRM. To mitigate predation risk, we expected cougars to select for more structurally complex habitat types and against open vegetation classes where cursorial wolves would have the advantage. We also expected beds to be closer to forest edges, to increase the likelihood of detecting approaching competitors. For example, research by Wam, Eldegard & Hjeljord (2012) found that wolves often bedded on “overlooking sites,” likely to facilitate the detection of conspecific intruders. We also believed that steeper slopes and more rugged terrain (high VRM) would increase bed site inaccessibility and/or facilitate escape from a potential predator.

Prior to any statistical analyses, we employed a correlation matrix to evaluate collinearity (|r| > 0.50) among predictor variables. Predictor variables were not correlated (all |r| < 0.50), so we included them all in our analyses. Cougar ID was included as a random intercept to account for variation among individuals. We then created candidate models for each season and hypothesis (winter thermoregulation, winter predator avoidance, summer thermoregulation, and summer predator avoidance) from all possible combinations of their distinctive, biologically-relevant predictor variables. We calculated Akaike’s Information Criterion adjusted for small sample size (AICc), ΔAICc, and Akaike weights (wi) for each model in each model set, and considered the top model and any subsequent model differing by <2 AIC_c units to have produced substantial empirical support for explaining variation in the data; redundant covariates in top models (e.g., nesting) were considered uninformative (Arnold, 2010), and when it occurred, we selected the simplest top model.

As models were constructed and analyzed in the same way across hypotheses and model sets, we also compared performance characteristics of our top thermoregulation and predator avoidance models, to see which best fit the data and which hypothesis garnered the most support from our data. Further, we ran an additional post-hoc analysis for each season, composed of a single model containing the significant parameters from each of the top thermoregulation and predation avoidance models, to determine if a combination of the two hypotheses performed better than either one alone (based upon AIC parameters).

Fourth-order selection: microsite characteristics of bed sites

We calculated seasonal 95% fixed-kernel home ranges for each marked, resident adult cougar. Kernel density estimates (Worton, 1989; Kie et al., 2010) and isopleths were quantified in the Geospatial Modeling Environment (GME, Beyer, 2009–2012); for cougars sampled over multiple years, we calculated annual home ranges and then averaged parameters.

We collected microsite attributes at verified bed sites and 50 random points in each cougar’s seasonal home range (e.g., 50 points in their winter home range, and 50 in their summer home range). Following the methods of Elbroch et al. (2015b), we gathered the following microsite data: canopy cover, concealment, and habitat characteristics. Canopy cover was measured in each cardinal direction from the center of each location or bed with a convex spherical crown densiometer (Forestry Supplier, Kackson, MS, USA). Concealment was measured with a subdivided concealment board (Noon, 1981) measuring 1 m tall and 50 cm wide (Elbroch et al., 2015b). The concealment board was held at the center of each site, and we recorded the percent of the concealment board obscured by natural features when viewed from 10 m away in each cardinal direction. We also noted whether each site occurred on, under, or within one meter of a prominent physical feature such as a tree, cave, cliff band, boulder, or log jam. In comparison to our landscape-level analyses that employed a habitat class layer in ArcGIS sampled at a 30 m resolution, we also collected habitat and topography data at 10 m in each cardinal direction from the bed site. We defined the microhabitat of bed sites in the field as one of seven habitat types (forest, forest edge/transitional habitat, meadow, sagebrush, riparian, and a “rugged” barren habitat type consisting of cliff bands and talus fields) and the topography of each site as either a bench, cliff band, drainage, ridgeline, sloping hillside, or flat. Slope was derived in ArcGIS from the digital-elevation model (http://datagateway.nrcs.usda.gov/) at a resolution of 15 m. Finally, we also recorded distance to what we termed “escape terrain.” Initial research by Akenson, Henjum & Craddock (1996), Akenson et al. (2003) and preliminary field examinations in our study area demonstrated that many cougar beds were in or very near to rugged, structurally complex landscape features such as talus fields, cliff bands, and areas of downed woody debris. To test if they were selecting for proximity to these features, we used a range finder (Bushnell) to measure the distance from a bed to the nearest escape terrain feature, up to 200 m.

As with our landscape analysis, we first determined covariates that supported the thermoregulatory versus predator avoidance hypotheses. In both the winter and the summer, we included the following variables in our microsite selection models to support the thermoregulatory hypothesis: canopy cover, habitat class, and the presence of a physical feature such as a tree, cave, or boulder. As has been shown in cougar dens (e.g., Bleich et al., 1996), such covariates may provide protection from the elements, therefore assisting in body temperature regulation.

For both winter and summer, we included the following covariates to support the predator avoidance hypothesis of bed selection at the microsite level: concealment, slope, topography, habitat class, proximity to escape terrain, and the presence of a feature such as a tree or boulder. Vegetative concealment hides cougars from potential enemies, and proximity to escape terrain, structured habitats, steeper slopes, and physical features facilitate escape should a cougar be discovered in its bed.

We repeated the approach described above for our landscape-level analyses. We employed a correlation matrix to evaluate collinearity (|r| > 0.5) among continuous predictor variables and chi-square test of independence to test for collinearity between categorical predictors, but no predictor variables were correlated (all |r| < 0.50). Then, for each hypothesis and each season, we employed generalized linear mixed models with a binomial distribution (logit link function) and a random effect (intercept) to account for variation among individual cougars. For each hypothesis, we created a candidate model set derived from all possible combinations of the biologically-relevant covariates. We then calculated and compared AICc values, ΔAICc, and Akaike weights (wi) to determine whether the best predator avoidance model, the best thermoregulatory model, or the post-hoc combination model containing the significant parameters of both hypotheses best explained bed site selection at the microsite level.

Second-order selection: comparing bed site attributes to landscape attributes

From 23 December 2012 to 19 January 2016 we visited 1,718 clusters. We documented 599 beds, 754 kills, and 366 sites where we did not find anything. The 599 beds were from nine different cougars (three males, six females). Of those, 312 were wintertime beds, and 287 were summertime beds.

In the wintertime, cougar bed selection supported both the thermoregulation and predator avoidance hypotheses (Table 1). The results from the post-hoc combination model showed that cougars selected beds in the winter that were at lower elevations (p < 0.001; β =  − 0.006), on steeper slopes (p < 0.001; β = 0.097), and closer to forest edges (p < 0.001; β =  − 0.008; mean distance: 57.29 m ± 54.99 m). They selected against sagebrush (p < 0.001; β =  − 1.299), meadow (p < 0.001; β =  − 1.355), and riparian habitat types (p = 0.005; β =  − 1.134). Though cougars did exhibit selection for eastern (p < 0.001; β = 1.375) and western aspects (p = 0.033; β = 0.514), selection was strongest for southern aspects (p < 0.001; β = 1.655).

In the summertime, cougar bed site attributes strongly supported the predator avoidance hypothesis over the thermoregulation hypothesis (Table 1). We did not run an additional post-hoc combination model as the significant variables from the top thermoregulation model were already included within the top predator avoidance model. The top predator avoidance model found that cougars selected bed sites closer to forest edges (p < 0.001; β =  − 0.003; mean distance: 75.93 m ± 95.12 m) and on steeper slopes (p < 0.001; β = 0.077). They also selected against sagebrush (p < 0.001; β =  − 0.775) and meadow habitat types (p < 0.001; β =  − 1.333).

Fourth-order selection: microsite characteristics of den sites

We assessed microsite attributes at 60 summertime beds (from six cougars) and 80 wintertime beds (from eight cougars). In the winter, bed site attributes supported the thermoregulation hypothesis over the predator avoidance hypothesis (Table 2). Based upon our top model, cougars disproportionately selected bed sites with high canopy cover (p < 0.001; β = 0.057; mean percent cover: 87.4% ± 22.9%) and in “rugged” barren habitats (p = 0.007; β = 3.048) characterized by cliff bands and talus fields. Selection for rugged habitat classes likely also mitigated predation risk. Though the presence of a feature such as tree or boulder was included in our top model (Table 2), the parameter was not statistically significant (p = 0.999).

In the summer, cougar bed selection supported both the thermoregulation and predator avoidance hypotheses (Table 2). The results from the combination model showed that cougars selected bed sites with higher concealment (p = 0.002; β = 0.030; mean percent concealment: 82.2% ± 20.6%), higher canopy cover (p < 0.001; β = 0.053; mean percent canopy cover: 91.7% ± 17.2%), on steeper slopes (p = 0.004; β = 1.816), on benches (p = 0.002; β = 2.340), and in cliff band (p = 0.015; β = 2.440) topography.