To manage predation risk, prey navigate a dynamic landscape of fear, or spatiotemporal variation in risk perception, reflecting predator distributions, traits, and activity cycles. Prey may seek to reduce risk across this landscape by using habitat at times and in places when predators are less active. In multi-predator landscapes, avoiding one predator could increase vulnerability to another, making the landscape of fear difficult to predict and navigate. Additionally, humans may shape interactions between predators and prey, and induce new sources of risk. Humans can function as a shield, providing a refuge for prey from human-averse carnivores, and as a predator, causing mortality through hunting and vehicle collisions and eliciting a fear response that can exceed that of carnivores. We used telemetry data collected between 2017 and 2021 from 63 GPS-collared elk (Cervus canadensis), 42 cougars (Puma concolor) and 16 wolves (Canis lupus) to examine how elk habitat selection changed in relation to carnivores and humans in northeastern Washington, USA. Using step-selection functions, we evaluated elk habitat use in relation to cougars, wolves, and humans, diel period (daytime vs. nighttime), season (summer calving season vs. fall hunting season), and habitat structure (open vs. closed habitat). The diel cycle was critical to understanding elk movement, allowing elk to reduce encounters with predators where and when they would be the largest threat. Elk strongly avoided cougars at night but had a near neutral response to cougars in the day, whereas elk avoided wolves at all times of day. Elk generally used more open habitat where cougars and wolves were most active, rather than altering use of habitat structure depending on the predator species. Elk avoided humans during the day and ~ 80% of adult female mortality was human caused, suggesting that humans functioned as a “super predator” in this system. Simultaneously, elk leveraged the human shield against wolves but not cougars at night, and no elk were confirmed to have been killed by wolves. Our results add to the mounting evidence that humans profoundly affect predator-prey interactions, highlighting the importance of studying these dynamics in anthropogenic areas.

To evaluate elk movement we captured, collared, and monitored adult female and neonatal elk for 53 months (January 2017 to June 2021). Adult elk were fit with global positioning system (GPS) radio-collars (Model Survey, Vectronic Aerospace, Berlin, Germany) that recorded a fix every 4 hours and equipped with mortality sensors that sent emails and SMS notifications after 9 hours of inactivity. Neonatal elk 0-10-days old were fit with expandable GPS collars (Model Survey, Vectronic Aerospace, Berlin, Germany) that transmitted 1 fix daily or very high frequency (VHF) tracking collars (Models M2230B and M4210, Advance Telemetry Systems, Isanti, Michigan, USA). Calf collars signaled a mortality after 8-hours of inactivity, and calves were monitored remotely (if GPS collared) or with radio-telemetry (if VHF collared) daily from capture to the end of summer (31 August), twice per week through the fall (September – December).

We captured cougars using trained dogs and baited cage traps and fit them with GPS collars (Model Vertex Lite, Vectronic Aerospace). Wolves were captured with padded leg-hold traps and by aerial darting, and fit with GPS collars (Models Vertex Lite and GPS Plus, Vectronic Aerospace, and Model TGW, Telonics Inc., Mesa, Arizona, USA). Cougar collars were programmed to record a fix every 4 hours and wolf collars were programmed to collect locations every 4-12 hours. Elk and cougar capture and handling followed protocols approved by the University of Washington Institutional Animal Care and Use Committee (IACUC Protocol #4226-01). Wolves were captured as part of existing management and conservation activities by the Washington Department of Fish and Wildlife and the Spokane Tribe of Indians in accordance with their agency-approved capture and handling protocols and the guidelines of the American Society of Mammologists.

We used step selection functions to examine how elk navigated the landscape of fear. We created separate population-level models for elk in the summer (June – August) and in the primary hunting season in the fall, hereafter “fall” (September – December). We excluded locations associated with migration and removed individual elk with fewer than 200 fixes (n = 3) from all datasets (all others had > 700 fixes). To determine if selection reflected maternal behavior to mitigate calf predation risk, we also created and compared summer step selection functions for cow elk known to have a calf (collared calves that survived through summer) to those known not to have a calf (either the cow was not pregnant at capture the prior winter, or the calf died before reaching 10-days old). For the step selection functions, we generated 20 random steps per taken step with the turning angle drawn from a von Mises distribution and step-length drawn from a gamma distribution using the amt package in program R version 4.2.3.

Seasonal models contained the following covariates: cougar and wolf use indices, elevation, slope, percent open habitat, and the human footprint. We took different approaches to describe cougar and wolf use to account for differences in their density and distribution. The area is fully occupied by cougars, so we used resource selection functions (RSFs) as an index of cougar risk to elk. Cougar RSF values ranged from 0-1 and incorporated data from 42 cougars. Because we collared wolves from all wolf packs in the study area and movement from individual wolves within a pack tends to reliably describe pack level use, we effectively had a census of wolf packs. Additionally, track and camera surveys indicated that transient wolves outside of known wolf pack territories were rare. Therefore, to describe wolf pack presence while accounting for areas outside of pack territories, we created Localized Density Distributions (LDDs) for summer and fall to use as an index of wolf risk to elk. Wolf LDDs were scaled from 0-1 to align them with the cougar index, and they incorporated data from 16 wolves.

We sourced elevation and slope layers from the Shuttle Radar Topography Mission (SRTM) Digital Elevation Model. We created a continuous covariate describing the proportion of open terrain within a 250 m moving window at 30 m resolution, selecting a 250 m buffer to reflect the median step-length taken by elk (summer: median = 243 m, fall: median = 255 m). We defined open terrain as areas classified agriculture, mesic grass, xeric grass, and xeric shrub in annual layers produced by the Cascadia Partner form (TerrAdapt:Cascadia; https://www.cascadiapartnerforum.org/terradapt). To describe human presence and associated landscape impacts, we used the TerrAdapt:Cascadia annual human footprint index, which represented the degree to which a landscape was used and altered by humans. Values ranged from 0 (unimpacted areas) to 1 (urban areas). Intermediate values described timber harvest (~ 0.3-0.5), agricultural and rural development (~ 0.5-0.7), and residential areas (> ~ 0.8-0.9). We classified the end of each step as day (between sunrise and sunset) or night (after sunset and before sunrise) using the R package suncalc. All covariates were mapped at 30 m resolution. Elevation and slope were standardized such that the mean = 0 and standard deviation = 1, whereas cougar, wolf, human footprint, and % open habitat layers ranged from 0-1.