Wildlife that share habitats with humans with limited options for spatial avoidance must either tolerate frequent human encounters or concentrate their activity on those periods with the least risk of encountering people. Based on 5,259 camera trap images of adult wolves from eight territories, we analyzed the extent to which diel activity patterns in a highly cultivated landscape with extensive public access (Denmark) could be explained by diel variation in darkness, human activity, and prey (deer) activity. A resource selection function that contrasted every camera observation (use) with 24 alternative hourly observations from the same day (availability), revealed that diel activity correlated with all three factors simultaneously with human activity having the strongest effect (negative), followed by darkness (positive) and deer activity (positive). A model incorporating these three effects had lower parsimony and classified use and availability observations just as well as a ‘circadian’ model that smoothed the use-availability ratio as a function of time of the day. Most of the selection for darkness was explained by variation in human activity, supporting the notion that nocturnality (proportion of observations registered at night vs. day at the equinox) is a proxy for temporal human avoidance. Contrary to our expectations, wolves were no more nocturnal in territories with unrestricted public access than in territories where public access was restricted to roads, possibly because wolves in all territories had few possibilities to walk more than a few hundred meters without crossing roads. Overall, Danish wolf packs were 6.5 (95% CI: 4.6-9.6) times more active at night than at daylight, which makes them amongst the most nocturnally active wolves reported so far. These results confirm the prediction that wolves in habitats with limited options for spatial human avoidance, invest more in temporal avoidance.

Population monitoring and data collection

Since 2017, the Natural History Museum Aarhus and Aarhus University have monitored all wolves in Denmark for the Danish Environmental Protection Agency. The occurrence and turnover of individuals are registered from genetic markers obtained from scat, hair, saliva, or urine samples collected by systematic patrolling of forest roads and by snow tracking (active monitoring) as well as saliva samples from livestock kills obtained by the Danish Nature Agency.

A territory was defined as the area patrolled by a single wolf, pair, or pack for a minimum of six months. The core areas and approximate territory extensions were estimated from the distribution of wolf signs (scats, tracks, kills, photos, etc.) within the landscape. With permission from the landowners, we placed wildlife cameras in places known (from the appearance of footprints, scats, or other signs) or suspected (leading lines in the landscape which from experience are known to be used by wolves when commuting, e.g. forest roads) to be used by the wolves within the territories. At locations with public access, visitors were informed about the presence of the cameras through signs containing project information and our contact details. We used cameras with fast trigger times able to record fast-moving species, that recorded videos and/or multiple pictures. Cameras were usually visited every two to six weeks, checking battery levels and changing memory cards. Where possible, the wolves on the images were identified to age defined as pup (born same calendar year) or adult (not a pup, hence all grown-up wolves observed January-June were coded as adults) and coded in the database. If multiple wolves on the same photo or video sequence were identified as different ages or individuals, they were registered as different records in the database. Prior to the analyses, such doublets or triplets were removed, so only one unique camera observation entered the analysis as an observation unit.

As the cameras were placed to maximize the number of wolf observations, sampling effort was concentrated in the central parts of the territories where wolf sign concentrations were highest. Cameras aimed at recording wolves were usually placed along trails and forest roads used by wolves when traversing their territories and at places with a high density of scats and footprints, that indicated frequent use by wolves at a given time. In a subset of the territories we also had cameras placed in the terrain, optimized to register all large and medium-sized mammal species. For wolf population monitoring purposes, observations from both types of surveys were entered into the Danish National Database of Wolf Observations. The effort expended in terms of camera days was not registered in this database. Observations of general wildlife were logged in a separate database, which also included information on the effort expended in terms of the number of camera days (26,210 in total). Due to resource constraints, this database only contained a subset of the total number of camera observations available in the raw data. The wolf data used for this analysis were therefore drawn from the first database. The number of different camera locations, resulting in wolf observations varied from 26 to 198 per territory (median: 51) and the total area covered (100% minimum convex polygon) by cameras delivering wolf data for the analysis, ranging from 6.5 to 79.3 (median: 21.3) km² per territory.

Selection of observations for analyses

We selected wolf camera trap data separated by a minimum of 5 minutes from eight independent territories from six areas (one territorial area was occupied by three different constellations of individuals during different time periods. This selection resulted in 5,259 camera observations of adults, 1,814 observations only showing pups (representing five litters from four territories), and 158 observations where the age could not be determined (excluded from the analyses). Of the 5,259 observations of adult wolves, 3,280 (62%) originated from cameras for monitoring wolves, 1,257 (24%) from cameras for monitoring general wildlife, and 722 (14%) from cameras where the initial purpose had not been recorded. As any dependence between observations within territories was accounted for in the statistical analyses by stating territory as random effect (see below), we decided to use the full data set rather than a reduced data set based on observations separated by 30 minutes, as often recommended to avoid serial dependence of observations. Under all circumstances, increasing the minimum sampling interval from 5 to 30 minutes only reduced the data set by 250 observations, and did not change the outcome of any of the statistical analyses.

Digitized data on wildlife and human activity was available from three of the six territory areas (five of eight territories. As ungulates, especially cervids (Cervidae, ‘deer’ hereafter) constitute the main and most selected prey type of wolves in Central Europe, we used 11,315 camera observations of deer (species composition: red deer Cervus elaphus: 60%, roe deer Capreolus capreolus: 31%, fallow deer Dama dama: 4%, unidentified deer species: 5%) to represent diel activity of prey. The 15,017 observations of humans were divided between 48% pedestrians, 16% bicyclists, 31% motorized vehicles, and 5% horse riders.

Seasonal definition

To account for seasonal variation, we divided the year into quartiles: November-January (mean day length at 56°N: 8.47 hours; range: 7.92-9.68 hours), February-April (11.97; 9.22-14.77), May-July (16.01; 14.83-16.55) and August-October (12.57; 9.73-15.32). These not only contrasted the two three-month periods with the shortest and longest daylengths, but also provided a good division of the ecological and reproductive annual cycle for wolf packs that have young offspring in May-July, mobile offspring (frequenting rendezvous sites) in August-October, increasingly independent offspring through November-January, and a pre-parturition period from February-April, when last year’s offspring have attained full independence.

Statistical analysis of diel activity patterns of wolves, deer, and humans

For each three-month period, we quantified the general variation in the diel activity of juvenile and adult wolves (W hereafter), deer (D), and humans (H), by modelling the relative frequency of observations per one-hour-interval from midnight to midnight (0: 00:00-00:59, 1: 01:00-01:59, etc.). We used the R package ‘mgcv’ v. 1.8 to fit generalized additive models (GAM) with beta distribution and logit-link, an adaptive cyclical cubic smoothing spline for time, and territory ID as random effect. Models were visually validated by plotting standardized model residuals against fitted values.

As data for human and deer activity was not available for all areas, extrapolation was necessary. As the season-specific diel activity curves for humans were highly correlated between the different study areas (Supporting information), we produced one season-specific diel activity function based on all data pooled. Among the deer, diel activity correlated less between fenced and unfenced areas. As we know from GPS-data that red deer in fenced areas move shorter hourly distances around dusk and dawn than red deer in unfenced nature areas (P. Sunde and R.M. Mortensen, unpublished), we created one activity distribution for deer based on data from all three areas (“Deer-Total”, abbreviated to “DT”), and one differentiated between fenced and unfenced areas (“Deer-Local”, abbreviated to “DL”).

To quantify the extent to which diel activity levels of adult wolves, deer, and humans were associated with light conditions and correlated internally throughout the year, we created a correlation matrix comprised of 24 * 365 = 8,760 hourly time observations, covering the entire year (1 January-31 December). For each hourly time observation, we assigned light conditions (categorical variable: night [sun angle < 0°, coded as 1] vs. day [sun angle ³ 0, coded as 0], “ND”) and the relative diel activity of W, H, DT, and DL (separated between fenced and unfenced areas) as predicted for each of the four seasons by the GAM-models.

Statistical analysis of predictors of timescape selection

We analysed the extent to which wolves selected to be active as a function of ND, H, and DT/DL by means of a Resource Selection Function (RSF) based on use-availability. For each wolf observation (used) we created 24 pseudo-observations, representing every hour for the same date and geographical location of the wildlife camera image.

We used mixed-effect binary logistic regression (response variable: used vs. unused) with territory ID as random intercept. As models with camera type as random effect resulted in similar results as models without camera type, we did not include camera type as random effect. We evaluated the extent to which a fixed effect (ND, DT or DL, H) predicted variation in diel activity from the magnitude of its selection coefficient in the logistic regression equation and considered it important if deviating significantly (p < 0.05) from 0. To establish the raw effects, we first ran models containing each of the three types of fixed effects in isolation (four different models in total as deer activity was represented by two alternative variables). To establish the extent to which the magnitude of a given fixed effect remained (or was reduced) when controlling for one or both two other competing fixed effects, we furthermore ran models representing all possible combinations of main effects of ND, H, and DT/DL. If a fixed effect remained statistically significant when controlling for the other two fixed effects, we took it as an indication that it contributed genuinely to explaining the variation in diel activity. We used the R package ‘mgcv’ v. 1.8 to fit generalized mixed models (GLM) with binomial distribution, logit-link, and territory ID as random effects. Models were visually validated by plotting standardized model residuals against fitted values.

To explore the predictive power of H, ND, and DT/DL (in isolation as well as in combination), we calculated Somers’ D, which is {Institute, 2013 #568}a nonparametric index of a model’s ability to correctly classify the dependent variable, derived as D = 2 (AUC – 0.5) where AUC is the area under the model’s receiver operation curve. If D = 1, all observations are correctly classified by the model, whereas D = 0 indicates a non-informative model. To compare the predictive effects of the three fixed-effects variables with a model based on mere smoothing of the basic circadian pattern, we contrasted the D-values of the models based on different combinations of the three fixed effects with those of a mixed-effects logistic GAM model where activity was smoothed as a function of time of day. The GAM model was fit with an adaptive cyclical cubic smoothing spline for time, and territory ID as a random effect. If Somers’ D for a fixed effects model approached the D-value of the GAM model, we took that as an indication that the fixed effects model was able to capture most of the diel activity variation. For all models considered, we calculated AICc. We considered a difference in AICc-values (DAICc) > 7 as substantial evidence for the model with the lowest AICc-value to have the most support in data.

Analysis of variation in nocturnality

Using nocturnality as a proxy for temporal human avoidance, we modelled the conditional probability that a wolf image was registered at night (sun angle < 0°; p_n) as opposed to at day (sun angle ³ 0°) as a logistic regression function, where we included the logit-transformed night length (logit [NL], the logit-transformed proportion of the day the sun is below the horizon) as a nuisance variable to adjust for variable night length. From the logistic regression equation, we derived conditional nocturnality (N) as the difference on a logit-scale between the predicted probability that an observation under given circumstances would be nocturnal, relative to what should be expected from the length of the night (N = logit [p_n] – logit [NL]). From simple algebra, it follows that the activity selection ratio between night and day (SR_N:D: how many more times wolves were observed per time unit at night as opposed to daytime) can be derived as R_N:D = exp [N].  At the equinox (where NL = 0.5, hence logit [NL] = 0), this expression could be simplified to N_{(NL = 0.5)} = logit [p_n] = ln (p_n / [1-p_n]) Û R_N:D = p_n / (1- p_n). As an example, if a sample of study subjects was observed equally often at day and night, standardized at the equinox, p_n[NL=0.5] = 0.5 and SR_N:D = 0.5 / (1-0.5) = 1. In comparison, situations where 80% and 90% of the observations at the equinox were nocturnal corresponded to wolves being 4 (0.8/0.2) and 9 (0.9/0.1) times more active during darkness than in daylight, respectively.

The analysis was run as a mixed effect logistic regression function with territory ID as random intercept, using the GLIMMIX procedure in SAS 9.4 with a logit link function, binomial error distribution, using Satterthwaite’s approximation to estimate degrees of freedom. As fixed effects, we included logit [NL], social status (single wolf, pair, or pack), human access (restricted or free), and camera purpose (wolf monitoring, general wildlife monitoring, not registered). As models with camera type as random effect resulted in similar results as models without camera type, we did not include camera type as random effect.