We investigated a number of factors hypothesized to influence the selection of crossing events (including both the crossing location and time) by barren-ground caribou on the Tibbitt to Contwoyto winter road network. We used logistic regression to contrast the variables associated with observed crossing events and a set of randomly identified available crossing events. We hypothesized that the road-crossing decisions of caribou would be influenced by human disturbance in the form of traffic and hunting pressure as well as factors known to influence caribou movement, such as weather, time of year, and previous movement by the individual animal.
Modelling Approach
The modelling approach paralleled methods used in resource selection analyses (Boyce et al., 2002), where observed events were contrasted against ecologically realistic opportunities (i.e. available movements or locations). These observed and available crossings of an industrial winter road were derived from GPS collar data for caribou in the Bathurst, Bluenose East and Beverly/Ahiak herds. Using this method, we quantified relationships between the observed road crossings of caribou and a range of human and ecological disturbance factors.
Crossing Data
We used GPS-collar data to infer the movements and crossing events of monitored caribou. These data were provided by the Government of the Northwest Territories (GNWT) (Adamczewski & Boulanger, 2016; Gunn et al., 2013). Most of the GPS collars (Telonics and Iridium; model TGW-4577-4s) were programmed to gather locations 3 times per day, with some collars programmed to collect 24 locations per day when in proximity to the Tibbitt to Contwoyto Winter Road and the associated diamond mines. We used only the data from these higher frequency collars to estimate the time and location of crossing events. As records for industrial traffic were only available for 2018–2020, we limited the analysis of crossings to that period.
For each observed crossing event, we generated 5 random locations and times where caribou could have crossed the road (Northrup et al., 2013), termed the available crossing events. We used the available crossings to represent the range of ecologically realistic crossing opportunities for collared caribou during the study period. We generated the available crossings from GPS-collar locations of caribou near the road during the study period. The geographic origin of an available crossing was selected randomly from an observed GPS-collar caribou location that was within 5.2 km of the road, which was the 97.5th percentile of straight-line distances to the winter road from caribou locations that were recorded immediately before an observed road crossing event. The location for each available crossing was set as the point where a straight-line from the randomly selected location intersected the nearest segment of the winter road.
The definition of resource availability can influence the strength of selection or avoidance of a resource (C. J. Johnson & Gillingham, 2008). We chose an availability threshold (5.2 km) that was representative of the distribution of caribou that chose to cross the road and represented plausible crossing events for comparison with the observed events. The dates and times for the available crossing events were taken from the timestamp of the crossing locations used to generate them.
Ecological And Disturbance Covariates
We identified factors that were hypothesized to affect the decisions of caribou when crossing the Tibbitt to Contwoyto Winter Road. These included environmental and disturbance variables that were previously reported to influence the distribution, movements, and behaviour of caribou (Table 1).
Table 1
Description, mechanism, and predicted effect (positive or negative) of covariates hypothesized to influence the crossing choice of barren-ground caribou on the Tibbitt to Contwoyto Winter Road in the central NWT, Canada.
Covariate | Type | Description |
Open road | Binary | Open period of the winter road for that year, from official open/close dates. |
Active road | Binary | Road in use when caribou approached or crossed the road. Represented by any traffic on the winter road at the crossing location for up to 3 days prior to the crossing event. |
Traffic count | Continuous | Index of traffic activity during the crossing event. Represented by an estimated count of vehicles 8 hrs prior to and 2 hrs after the crossing event. |
In no-hunting zone | Binary | Crossing occurred within the Bathurst mobile no-hunting zone. |
Windchill | Continuous | Measure of thermal stress derived from temperature and wind speed. |
Directionality coefficient | Continuous | Average persistence in directionality of caribou movements over the previous 7 movement steps. |
Day of year | Continuous | Number of days since January 1st as an index of onset of migration. |
Year | Categorical | Year of observation. |
Individual ID | Random | Random effect to control for multiple locations for individual caribou. |
We used right-of-way data for the Tibbitt to Contwoyto Winter Road from the GNWT (J. Williams, Government of the Northwest Territories, personal communication), Mackenzie Valley Land and Water Board (Mackenzie Valley Land and Water Board, 2020), and De Beers Canada (Ryan Marshall, De Beers Canada, personal communication), along with a portion digitized from satellite imagery (NASA SENTINEL-2 dataset) to develop a shapefile denoting an estimated annual alignment of the road. Apart from adjustments in alignment to bypass ice ridges and overflow, the location of the road was similar year-to-year.
We used dispatch records to estimate the volume and timing of industrial traffic (Rudolph Swanepoel, Det’on Cho Logistics, personal communication). These records were logged in a computer database by dispatchers, and included the date, time, and location for the start and end of each commercial vehicle’s trip along the road. These data captured most of the traffic on the winter road, predominantly large trucks moving material to and from the mines. We estimated hourly traffic volume at 2-km intervals along the road. Vehicles were assumed to take the shortest path at a constant speed.
We used the estimated traffic volumes to establish when vehicle use occurred on the winter road during each study year. In a review of satellite imagery (NASA SENTINEL-2 dataset) we confirmed that the roadbed was present beyond the ± 2-week time window of traffic activity. We only kept observed or available crossing events from caribou that were within a ± 2-week time window of estimated traffic presence.
We used the boundaries of the Bathurst mobile no-hunting zone to estimate the presence or absence of hunting activity. The Bathurst mobile no-hunting zone was established by the GNWT and Indigenous governments as a joint management action in response to the decline of the Bathurst barren-ground caribou herd (Government of the Northwest Territories, 2019). The zone represented a minimum convex polygon formed from the locations of GPS-collared caribou, with each location including a 10-km spatial buffer (D. Cluff, Government of the Northwest Territories, personal communication). The boundaries of the mobile zone were adjusted periodically to represent the shifting distribution of the Bathurst herd. The open and close dates for each no-hunting zone were provided by the North Slave office of the Department of Environment and Natural Resources, GNWT.
We used NASA’s Modern Era Retrospective Analysis for Research and Applications (MERRA-2) dataset to estimate wind speed and temperature (Gelaro et al., 2017). From these data, we modelled the windchill at the crossing site (\({T}_{wc}=13.12+0.6215{T}_{a}-11.37{v}^{0.16}+0.3965{T}_{a}{v}^{0.16}\); \({T}_{wc}\): wind chill index (kg*cal/m2/h), \({T}_{a}\): air temperature (°C), \(v\): wind speed (km/h). The windchill equation was developed for humans, not Rangifer (Osczevski & Bluestein, 2005). Those values, however, do provide a measure of relative change in the combined thermal effect of air temperature and wind speed across the winter. We used the lowest calculated windchill over the 3 days prior to a crossing as an index of thermoregulatory stress experienced by caribou (Leclerc et al., 2019).
We summarized the directional persistence of individual caribou prior to the observed or available crossing event. Directionality was calculated as the relative mean resultant distance of previous steps. We calculated this metric from the 7 steps prior to the crossing event by setting each step to a distance of one, then summing the vector distance (i.e., resulting straight line distance) and dividing by the number of steps. For the step direction we used the relative turning angles (i.e., the angle between successive steps). The calculated directionality value indicated the degree of tortuosity of movement, from which the type of movement the animal was engaged in can be inferred. Values close to one, generated from consecutive steps in the same direction, indicated directional persistence in travel. Values close to zero, generated from high angle turns, indicated low directional persistence, possibly associated with foraging behaviour (Benhamou, 2004).
Development And Assessment Of Statistical Models
We used logistic regression to contrast observed and available crossings of the Tibbitt to Contwoyto Winter Road. We developed a set of a priori model hypotheses to test the factors that may have influenced the crossing choices of monitored caribou (Table 2). Models were developed according to reported behaviour of barren-ground and migratory caribou near industrial features and roads as well as our observations of the population during an accompanying study of the behavioral responses of caribou to vehicle traffic (Smith, 2022). We used the Akaike Information Criterion for small sample sizes (AICc; Hurvich and Tsai 1989) to identify the most parsimonious model of the set (D. R. Anderson et al., 2002; Burnham & Anderson, 2002). The top-ranked model had the greatest degree of fit to the data with the fewest parameters. We reported the difference in AICc values for each model relative to the most parsimonious model (ΔAICc). We considered all models with a ΔAICc of < 2.0 as informative.
The AICc provides a relative measure of model fit, but not predictive accuracy. We used the area under the curve (AUC) of the receiver operating characteristic (ROC) curve to measure the predictive accuracy of each model (Boyce et al., 2002). We used a leave-one-out cross-validation (LOOCV) to generate independent predicted probabilities for each observed and available crossing event (Fielding & Bell, 1997; Pearce & Ferrier, 2000). Using that procedure, each predicted value was generated after sequentially withholding the respective crossing event from model estimation. Area under the curve values between 0.7 and 1 were considered to have good predictive ability, and models with an AUC of approximately 0.5 were assessed as having no predictive capacity (Boyce et al., 2002; Fielding & Bell, 1997).
We used the variance inflation factor (VIF) to test models for multicollinearity. No parameters had a VIF > 10, which is a common threshold to determine problematic multicollinearity (Vittinghoff et al., 2012). We used 95% confidence intervals to assess the strength of model covariates. Confidence intervals that included a value of zero were assumed to have some combination of small effect size or low precision. We used a random effect for each collared caribou to account for repeated crossings by individuals. All data generation and modelling were conducted using R (R Core Team, 2021) and ArcGIS (ESRI, 2020). Logistic regression was performed using the glmmTMB package for R (Brooks et al., 2017). Mapping was conducted using the QGIS software (QGIS Development Team, 2022).
Table 2
Candidate logistic regression models designed to test a range of factors hypothesized to influence the crossing choice of barren-ground caribou on the Tibbitt to Contwoyto Winter Road in the central NWT, Canada.
Model group | Hypothesis | Model Structurea |
Traffic | Presence of the active winter road or the level of traffic on the winter road reduced crossing events by caribou. | Open road + year Active road + year Traffic count + year |
Hunting | Presence of hunters on the winter road reduced crossing events of caribou. | In no-hunting zone + year |
Movement | Crossing of the winter road was influenced by ecological factors that influenced the seasonal movement behaviour of caribou. | Day of year + year Windchill + year Directionality coefficient (relative) + year Windchill + directionality coefficient (relative) + year |
Global model | Additive effect of all model covariates is important. | Traffic count + day of year + directionality coefficient (relative) + in no-hunting zone + windchill + year |
Null model | Model covariates are not important for explaining crossing decisions of caribou. | 1 |
aIndividual ID of collared caribou was included as a random effect in all models.