2.2 | Camera trap survey
We deployed remotely-triggered camera traps (Reconyx© PC 850, 850C, 900, 900C) throughout each site with camera placement and sampling design proportional to study area size (Table 1). Our study uses data from four surveys at DMP (2017, 2018, 2019, 2020), two surveys at SNWR (2016, 2018), two surveys at UMBS (2015, 2016), and four surveys at HMC (2016, 2017, 2018, and 2019). We captured the heterogeneity of habitat and other environmental features to ensure ecological representation in the micro-site selection of camera traps. Camera traps were affixed to trees > 0.5m diameter and placed 0.5-1.0 m off the ground. Site-specific placement of camera traps was determined by signs of animal activity such as game trails and scat. Camera trap settings included: high sensitivity, one-second lapse between three pictures in a trigger, and a 15-second quiet period between triggers. Camera traps were not baited.
Image identifications were initially crowd-sourced and filtered for carnivores using a public-science program called Michigan ZoomINin combination with a consensus algorithm and expert validation (Gadsden et al. 2021). Carnivore species identifications were later sorted and confirmed by at least two independent researchers in the Applied Wildlife Ecology Lab.
2.3 | Temporal activityTime stamps associated with the camera trap images were used to conduct temporal analyses. Prior to all analyses, a 30-minute quiet period was introduced for every species to account for pseudoreplication, given the tendency of some animals to remain in front of the camera trap and trigger it multiple times. Since surveys were conducted across different times of the year, we scaled times to sunrise and sunset times using thesunTimes function in the ‘circular’ package in R (Ridout and Linkie 2009). 2.31 | Variation between sites
We first compiled all raccoon triggers from each survey within a site to have an aggregate across years of overall raccoon temporal activity at each site. We then compared raccoon temporal activity between sites using the Mardia-Watson-Wheeler (MWW) test, which is a nonparametric test of differences in the angular means between samples of circular data using the ‘circular’ package in R (version 4.1.0). When the W value is high it results in a significantp value (p < 0.05), which we conclude to mean that the compared temporal activities are unique.
2.3.2 | Seasonal and yearly variation
Our multi-site camera study allowed us to compare differences in raccoon temporal activity based on landscape level differences along an urban-rural gradient. Comparing between seasons can confound inferences from the analyses, due to different seasons potentially resulting in different detection rates (Marcus Rowcliffe et al. 2011). While we did not have identical seasonal coverage for every site, the multiple surveys at every site resulted in coverage for the entire year at every site with the exception of UMBS (Figure S1). To determine if there was consistency at sites regardless of season and year, we compared raccoon activity between each survey within each site, and then looked for broader patterns across sites.
2.3.3 | Coyotes on raccoon temporal activity
For each survey, we used a kernel density estimation for the independent coyote triggers and designated the cameras that fell within the top quantile of as ‘HIGH’ coyote intensity of use zones in ArcGIS Pro (version 2.3.1). We used this rather than a fixed cutoff value of expected detection rate because our sites spanned the entirety of the urban-rural gradient and expected detection rates for coyote vary depending on the composition of a site (Magle et al. 2014). Coyote triggers were checked for spatial independence using Moran’s I prior to kernel density estimation. We compared raccoon temporal activity between the high coyote cameras and the rest of the site using the MWW test. For additional evidence that temporal shifts by raccoons were due to avoidance of coyotes, we then compared the overlap between coyote and raccoon time use in the two raccoon test groups from the MWW test. To do this, we calculated an overlap (Δ) coefficient of temporal activity for coyotes and raccoons within each group (‘HIGH’ and ‘LOW’ coyote intensity of use) along with 95% confidence intervals generated from 10,000 parametric bootstraps of the temporal distribution models. Δ values range from 0 to 1, with 0 indicating completely distinct and non-overlapping temporal activity between comparison groups, and 1 indicating complete overlap. Δ1 was used for comparisons when one of the sample groups had less than 50 triggers; otherwise Δ4 was used to estimate temporal overlap (Ridout and Linkie 2009). Finally, the activity distributions were visually assessed to determine qualitative characteristics of shifts (e.g. raccoons shifting towards increased nocturnality in high coyote zones).