Results
SCAT-BASED PREDATOR RESOURCE SELECTION FUNCTIONS
We surveyed 183 km in 2013, 652 km in 2014, and 405 km in 2015 for scats recording all predators, and an additional 82 km in 2016 for cougar scats only, for a total of 1,322 km over 125 transects. We used detections of 373 bear, 42 cougar, 223 coyote, and 470 wolf scats to determine RSFs. In all species, top RSFs were better supported than both the respective null and full candidate models with all 13 variables (Appendix S6, Supporting Information). Bears selected against conifer forests (β = -0.71, [95% C.I. -1.23, -0.19]), motorized trails (β = 0.86, [95% C.I. 0.41, 1.31]), and roads (β = 0.00005, [95% C.I. 0.00003, 0.00007]), and for areas with cutblocks (β = 0.84, [95% C.I. 0.23, 1.45]), high NDVI (β = 0.0002, [95% C.I. 0.00007, 0.00033]), steep slopes (β = 0.02, [95% C.I. 0.019, 0.039]), and non-motorized trails, particularly when farther from areas with motorized trails (β = 0.00005, [95% C.I. 0.00001, 0.00009]; Table 1). Wolf scats most likely occurred near waterways (β = -0.0001, [95% C.I. -0.00015, -0.00005]), on gentler slopes (β = -0.04, [95% C.I. -0.06, -0.02]) and non-motorized trails (β = 1.29, [95% C.I0.99, 1.59]), but farther from motorized trails (β = 0.00005, [95% C.I. 0.00004, -0.00006]; Table 1).
Equally supported models for coyotes included areas with gentler slopes (β = -0.05, [95% C.I. -0.08, -0.02]) and vehicle-restricted trails (β = 1.62, [95% C.I. 1.27, 1.97]), and areas farther from vehicle-permitted trails (β = 0.00006, [95% C.I. 0.00004, 0.00008]; Table 1); we retained proportion of shrub cover (β = 2.63, [95% C.I. 0.21, 5.05]) because its confidence limits did not overlap zero. Cougar scats were more likely in areas with less conifer forest cover (β = -1.92, [95% C.I. -3.38, -0.46]) and high edge density (β = 8.39, [95% C.I. 1.12, 15.56]) because the confidence limits of only these two variables did not overlap zero.
Based on model predictions, both wolf and bear scats were more likely found in the western portion of the study area along river drainages than in the eastern portion of the study area. Around the Ya Ha Tinda elk winter range, likelihood of scats was high for wolves but low for bears (Figs 2a, 2d). Cougar scats increased from the west (low) to east (high), corresponding with more forest edge in the forest-managed lands. Probability of coyote scat occurrence was fairly consistent across the study area (Figs 2b, c).
ELK PRESENCE IN PREDATOR SCATS
We analyzed the contents of 476 scats (130 bear, 33 cougar, 114 coyote, and 199 wolf); 226 were analyzed via macroscopic analysis and the remainder via DNA analysis. Elk was equally found in coyote (36% of scats), wolf (38%), and cougar (46%) scats collected from across the study area (Wilcoxon Rank Sum test; all pairwise P ≥ 0.27). Bear scats contained elk less frequently compared to the three other predators (19%, all pairwise P < 0.001). We did not use herbaceous forage biomass and elk resource use (RUF) in the same models for predicting elk in a scat, nor did we use slope, elevation, and ruggedness in the same models because they were highly correlated (r > 0.60)
The best-supported, predator-specific models predicting elk presence in a scat included most consistently the positive effect of herbaceous biomass, except for cougar; inclusion of other variables depended on the predator species (Table 2). For bears, the top model included the positive effect of herbaceous biomass (β = 0.061, [95% C.I. 0.03, 0.10]), and negative effect of open cover type (β = -4.83, [95% C.I. -8.58, -1.63], Table 2). For wolves, we selected the model including only herbaceous biomass (β = 0.21, [95% C.I. 0.16, 0.27]), terrain ruggedness (β = 0.84, [95% C.I. 0.53, 1.19]), and percent of area covered by deciduous forest (β = -36.29, [95% C.I. -56.85, -18.25]; Table 2) because the confidence limits of the beta coefficient for burns included zero (Appendix S7, Supporting Information). We selected the model for elk presence in coyote scat that included the positive effects of herbaceous biomass (β = 0.048, [95% C.I. 0.030, 0.070]) and distance to water (β = 0.00032, [95% C.I. 0.000060, 0.00058]) and the negative effect of road density (β = -0.88, [95% C.I. -1.83, -0.13]; Table 2). The most uncertainty was found in the models of elk presence in cougar scats largely because of low sample size (Appendix S7, Supporting Information). Based on parsimony, we selected the model with only forest edge density (β = 1.25, [95% C.I. 0.31, 2.49]; Table 2).
The western portion of the study area (i.e. , BNP) had low relative probability of elk being in all predator scats. Predictions from the above models indicated elk were most likely found in cougar scats in the eastern part of the study area and wolf and coyote scats at Ya Ha Tinda (Figs 2-3, Table 3).
SCAT-BASED PREDATION RISK TO ELK
In combining predator distribution and scat contents, the predation risk to elk was highest from wolves and coyotes at the Ya Ha Tinda (Table 3) where resident elk summer. Risk from bears was highest for areas west of Ya Ha Tinda and at the Ya Ha Tinda. In contrast, risk from cougars was widespread and relatively high only east of the Ya Ha Tinda (Table 3). Highest total predation risk to elk from all predators occurred for the YHT resident elk (Table 3). Rank correlation of the total risk (all predators combined; Fig. 4) corresponded well with predation risk inferred from kill sites (rs = 0.98, P< 0.0001; Appendix S8, Supporting Information).
Discussion
Because different factors may influence where prey are encountered and killed (Hebblewhite, Merrill, & McDonald, 2005; Kauffman et al., 2007), predictions of spatial risk based only on predator distribution may not be sufficient to assess predation risk in terms of an actual mortality event but this distinction is rarely acknowledged (Moll et al., 2017). At the same time, knowing the likelihood of predators being present may be sufficient to address questions regarding prey responses and frequency of anti-predator behaviours (Robinson & Merrill, 2013).Indeed, Prugh et al., (2019) argued for even further expansion of the stages of predation to include engaging, attacking, and surviving given an encounter to adequately understand and address questions of predation risk. However, quantifying these stages is challenging and likely requires time-intensive, observational studies (Wikenros et al., 2009; Tallian et al., 2017), or detailed fine-scale movements of both predators and prey (Basille et al., 2015; Greggor et al., 2016). Where risk of mortality is desired, the most common approach has been to examine kill-sites. Kill-site data based on annual winter mortality surveys or marked animals require extended periods and expense to accumulate sufficient samples, giving an approach based on scats a distinct advantage in assessing predation risk early in a study. It is also advantageous in terms of being non-invasive and cost-efficient (Wasser et al., 2004; Orkin et al., 2016), particularly a multi-predator community because dogs can simultaneously find scats from multiple predators. At the same time, reliability of using fecal material to estimate predator distribution has been questioned because of false species identification and sampling biases (Morin et al., 2016; DeMatteo et al., 2018). We found distinguishing between scats of bear species was more difficult than anticipated and warrants DNA analysis to verify species where the uncertainty needs to be eliminated. Biases in detecting scats, can be addressed with refined training and handling of dogs. For example, in the blind trials we performed, dogs had a high (> 90%) detection rate of scats (Spilker, 2019). Potential biases in sampling designs, such as collecting scats only on trails can be problematic (Steenweg et al., 2015), but we remedied this by locating transects on and off trails. Nevertheless, features such as steep cliffs and rugged terrain can constrain sampling, whereas this may not be the case if large carcasses can be observed during aerial flights.
A second issue for using scats for sampling the distribution of any species relates to whether deposition reflects where they spend time, and in the case of predators, where they make an actual kill. We present evidence for scat location and contents corresponding to these two components of predation because the factors we found associated with scat locations and contents (i.e. elk) were similar to those reported by others for both the distribution of these predators and for elk kill sites. For example bear scats were associated with high forage quality (NDVI) and quantity (cutblocks), similar to models for grizzly bears where bears were associated with greenness and open canopy cover (Nielsen et al., 2002; Apps et al., 2004). Wolf scats were associated with flat areas (Hebblewhite & Merrill, 2007) and cougar scats were associated with areas of high forest edge (Atwood et al., 2009; Elbroch et al., 2013). Indeed, we found that for wide-ranging species like wolves and grizzly bears, predictions from the scat-based RSFs corresponded well with predictions from telemetry-based RSFs in the study area (Appendix S2, Supporting Information). The ability of our scat-based approach to make predictions similar to the kill-site analysis also is not surprising because the approaches include the same components of predation, i.e. , a risky place is one where there is a high probability of encountering and being killed by a predator. For example, elk select for areas with abundant forage biomass in summer (Hebblewhite & Merrill, 2009; Smolko 2014), which is where we found a higher probability of elk being present in the scat of all species; we also found elk being in wolf scats associated with rugged terrain, which is where Torretta et al., (2018) also reported wolf kill sites, suggesting ruggedness reduced agility to navigate in steep terrain.
A major consideration in developing and applying estimates of spatial risk is to appropriately match the approach to the spatial and temporal scales for the processes and questions addressed (Moll et al., 2017; Cusack et al., 2019; Prugh et al., 2019). For example, experimental approaches such as giving-up densities (Altendorf et al., 2001) or interactions caught on remote cameras (Hernández et al., 2005) may be most appropriate to make fine-scale, behavioral or site-specific inferences of predation risk, whereas simultaneously monitoring sequential movements of predators and prey at short temporal scales may lead to understanding how certain evasion tactics are successful in only limited situations. Here, we show that a scat-based approach to predicting predation risk for elk corresponds well with outputs derived from modeling kill sites, but submit the approach is most appropriate when directed at answering questions at broad spatial and temporal scales (Orkin et al., 2016). For example, because it was not feasible to collect sufficient scats except over a 12-week summer season, we gained little insight into the seasonal dynamics of predation risk. At the same time, the broad-scale spatial patterns in predation risk for elk we found are consistent with our demographic understanding of predation rates on both adult and calf elk in this area ( Hebblewhite et al., 2018; Berg, 2019). Strengthening the link between risk of predation from the perspective of the prey, as represented in the above metrics, and kill or predation rates is a key next step for addressing questions of predator-prey dynamics.
ConclusionWe illustrate a new approach for estimating predation risk to prey based on distribution and contents of predator scats using scat dogs, and found it corresponds well with the results of other approaches. It has the advantage of being able to distinguish key components of predation, such as where prey may encounter predators and where they are killed. It can be used to sample broad areas over a relatively short time frame to get a snapshot of spatial predation risk, which lends itself to repeat sampling for detecting changes in spatial risk in the same area over time. As with other methods, appropriate sampling design and reducing uncertainty with observer training (e.g., dogs and handlers) and auxiliary data such as DNA to confirm species identification will be key considerations.