Anthropogenic impacts on wildlife behavior arise both from the immediate presence of people, which induces fear responses in many species, and the human footprint on the landscape (i.e., development), which affects animal movement and habitat use. Where both disturbance types co-occur, disentangling their impacts remains a challenge. Disturbance effects may interact such that species respond to increased human footprint by either reducing (habituation) or increasing (sensitization) avoidance of human presence. We experimentally manipulated perceived human presence, using playbacks of people talking, across a gradient of human footprint in California’s Santa Cruz Mountains and used camera traps to quantify the behavioral responses of bobcats (Lynx rufus), striped skunks (Mephitis mephitis), and Virginia opossums (Didelphis virginiana), mesopredators known to exhibit strong fear response to human presence, but which vary in their use of developed areas. Bobcats and skunks reduced activity in response to human playbacks but showed no change in responsiveness to playbacks across the development gradient, suggesting that these species are similarly fearful of humans at all development levels tested. Opossums exhibited a significant interaction between development and playback treatment such that reductions in activity level in response to human playbacks were strongest at higher levels of building density. Our results indicate that, rather than habituating to human presence some mesopredators retain a strong fear of humans or become more fearful when inhabiting more developed areas. We suggest that consistently high responsiveness to immediate human presence may benefit mesopredators living in human-dominated landscapes by mitigating the risk of anthropogenic mortality. Keywords: sensitization, habituation, ecology of fear, human impact, development gradient, playback experiment 

We ran a playback experiment at thirteen sites throughout the Santa Cruz Mountains between October 7th and December 4th of 2019. We chose playback sites of differing levels of human footprint ranging from remote protected areas to suburban development, whereby building density ranged from 0 to 391 buildings within a 500 m radius of the site.

We manipulated human presence following the protocol similar to that described by Suraci et al. (2019a). At each site, we deployed a single battery-powered speaker broadcasting either human or control vocalizations. Human playbacks consisted of a single female or male voice reading passages or responding to interview questions. As a control we used Pacific treefrog (Pseudacris regilla) vocalizations since this species is found throughout the Santa Cruz Mountains, does not pose a threat or represent a food source to any of our focal species, and can commonly be heard during day and night (Smith et al. 2017; Suraci et al.2019a,b). We used ten exemplars of each playback type, ranging in duration between 24 and 229 s, which were played in random order. Playbacks were broadcast at a consistent volume of 75 dB. The speakers were set to be silent 60 percent of the day and broadcast 40 percent of the day. Each experimental replicate ran simultaneously for eight weeks at each site. We utilized a repeated-measures design such that each site received either human or control playbacks for four consecutive weeks followed by the opposite treatment for a subsequent four weeks (Suraci et al. 2016, 2019a). We randomly selected half of the sites to begin with the control treatment and half to begin with the human treatment. Over the course of the experiment wildlife within hearing range of the speaker were exposed to either control or experimental playbacks for 9.6 hours per day for 56 continuous days. Each site was checked once a week to ensure that playback equipment was functioning properly. 

At each playback site we deployed a motion-sensitive wildlife camera (Bushnell Trophy Cam; Bushnell Corp., Overland Park, KS, USA). The cameras were programmed to take a burst of three photographs when triggered by motion with a one-minute delay between bursts. The cameras were active for the full eight weeks of the experiment. We placed a scent lure (perforated sardine tin) and food bait (boiled chicken egg) at each site (Fig. 1) to increase the chance that an individual of any species would investigate the site and be captured on camera (Suraci et al. 2019a). Lures and baits were replaced weekly. All playback images were scored for the presence of mesopredators by at least two trained individuals, with species assignment based on consensus between independent scorers. We defined an independent predator detection as an image or group of images of a particular species that was separated from another detection of the same species on the same camera by at least thirty minutes.  

To test the effects of human presence and human footprint on mesopredator activity levels, we quantified the number of detections per week of each species at each camera site (Moll et al. 2018; Suraci et al. 2019a). Because experiments were run for eight weeks, we derived eight activity level estimates for each camera site (four during the human treatment and four during the control treatment), with the exception of a single site for which camera failure resulted in only seven weeks of data (four human and three control). A detection event is the result of two processes: (i) whether or not a species is present in the vicinity of a camera site and thus available to be sampled (a binary process); and (ii) the species activity level at a camera site if present, which determines the number of times the species is detected (a Poisson process) (Moll et al. 2018; Suraci et al. 2019a). We therefore fit zero-inflated Poisson (ZIP) (Zuur et al. 2009) models to predator detection data, which can capture both processes through (i) a binomial component modeling whether or not each detection estimate is a zero and (ii) a Poisson component modeling the number of detections per week (a proxy for activity level). In all models, we fit random intercepts for camera site on both the binomial and Poisson submodels to account for repeated measurements of predator activity level at each site. For each focal species (striped skunks, bobcats, and opossums), we fit a full ZIP model consisting of (i) a building density covariate on the binomial (i.e., zero or nonzero) component to control for the potential influence of development on whether a given mesopredator was available to be sampled, and (ii) covariates for playback treatment, building density and their interaction on the Poisson component (i.e., number of detections per week). For each species, we also fit a reduced model with the playback treatment x building density interaction removed. For species for which the playback treatment x building density interaction term was not significant in the full model (i.e., 95% credible intervals crossed zero), we interpret the results from the no-interaction model. All models were fit in a Bayesian framework using the Stan programming language called through R via the rstan package (Stan Development Team, 2020). For each model, we ran 4000 iterations of three Hamiltonian Monte Carlo chains retaining 1000 samples from the posterior distribution of each chain. We used vague priors for all variables and random starting points for all chains. We checked model convergence by visually inspecting trace plots and confirming that the Gelman-Rubin statistic (“R-hat”) was < 1 for all parameters. We tested model fit to the data using Bayesian P-values, which compare statistics calculated from model-generated data with those calculated from observed data. Here, we calculated Bayesian p-values for the mean and skew of model-generated and observed data (Hooten and Hobbs 2015). 

The attached dataset will support the replication of all analyses in the above paper. The attached ReadMe file defines the column headings of our data set. The number of detections of each species per week for site HPB13 during week 4 of the control (Pacific tree frog) is missing due to an SD card handling error.