The Journal of Wildlife Management; DOI: 10.1002/jwmg.712 Research Article Targeting Mitigation Efforts: The Role of Speed Limit And Road Edge Clearance for Deer–Vehicle Collisions ERLING L. MEISINGSET, Norwegian Institute for Agricultural and Environmental Research, Organic Food and Farming Division, NO-6630, Tingvoll, Norway LEIF E. LOE, Department of Ecology and Natural Resource Management, Norwegian University of Life Science, P.O. Box 5003 NO-1432, Aas, Norway ØYSTEIN BREKKUM, Norwegian Institute for Agricultural and Environmental Research, Organic Food and Farming Division, NO-6630, Tingvoll, Norway ATLE MYSTERUD,1 Centre for Ecological and Evolutionary Synthesis (CEES), Department of Biosciences, University of Oslo, P.O. Box 1066 Blindern NO-0316, Oslo, Norway ABSTRACT Deer–vehicle collisions (DVCs) have major social, economic, and animal welfare costs. Modeling of accident risk relative to environmental characteristics could enable managers to plan the layout of new roads or determine optimal location of mitigation measures to reduce the number of accidents on existing roads. We analyzed position data providing road crossings from 67 global positioning system (GPS)marked red deer (Cervus elaphus) and location data of 271 car killed deer within the home ranges of the GPSmarked red deer. We modeled collision risk as a function of speed limit, season, road characteristics, and habitat features. For a subset of the data, we tested if vegetation clearance along a highway reduced the collision frequency of red deer. The relative risk for DVCs increased with speed limit. We found a higher risk for DVCs during winter compared to the other seasons. Forest cover, distance to pasture, and terrain ruggedness substantially affected risk of DVCs. Road edge clearance reduced the frequency of DVCs, but the effect appeared in the winter season only with a decrease of 53%. Our study highlights that speed limit reduction and road edge clearance are both powerful mitigation tools to reduce the number of DVCs. Ó 2014 The Wildlife Society. KEY WORDS Cervus elaphus, collision risk model, deer–vehicle collisions, mitigation, red deer, road edge clearance, speed limit. Many cervid populations in Europe and the United States have increased dramatically in recent decades (McShea and Underwood 1997, Putman et al. 2011). At the same time, the traffic level has increased and road network has developed substantially. This has led to increased frequency of deer– vehicle collisions (DVCs), now counting several million each year around the world (Groot Bruinderink and Hazebroek 1996, Rea 2003, Mysterud 2004, Dussault et al. 2007, Langbein et al. 2011), and is a major safety issue both to humans and wildlife (Groot Bruinderink and Hazebroek 1996, Langbein et al. 2011). Direct costs related to human mortality and injury, animal welfare, and material damage are high and in many regions, loss of potential hunting objects introduces indirect costs. Since DVCs are often traumatic for animals and humans and economically costly as well, more effective mitigation measures along roads are urgently needed (Langbein et al. 2011). Analyses of Received: 17 June 2013; Accepted: 23 February 2014 1 E-mail: [email protected] Meisingset et al. Collision Risk of Red Deer temporal and spatial patterns of risk for DVCs may contribute with important information to target relevant mitigation measures for the future. The number and variety of mitigation measures to reduce DVCs have increased greatly during recent decades (Iuell et al. 2003), but many of the mitigation measures have not been evaluated scientifically (Langbein et al. 2011). Use of fences or road under- or overpasses (also in combination) can reduce DVCs substantially (Clevenger et al. 2001, Langbein et al. 2011), but these are expensive measures, difficult to implement on a larger scale, and often seriously hamper animal movement. Other mitigation measures cannot reduce DVC to the same extent, but they usually aim to reduce the number of accidents to socially and economically acceptable levels. The spatial and temporal distributions of wildlife accidents are not random (Seiler 2005, Gunson et al. 2011). Landscape and habitat features may play an important role in determining where cervids cross roads, and risk zones for DVCs are often associated with a number of factors (Dussault et al. 2006, Danks and Porter 2010, Gunson et al. 2011). Roads are often associated with habitats 1 containing nutritious forage both for browsing and grazing herbivores. Access to selected habitats can reduce road avoidance behavior of cervids (Gagnon et al. 2007) and increase the rate of road crossings (Meisingset et al. 2013). High traffic levels can reduce crossing frequency, but the level has to be very high to prevent crossings completely (Gagnon et al. 2007). A necessary step toward understanding factors causing accidents is to quantify risk (Seiler 2004, Langbein et al. 2011). Risk is more than a question of animal behavior in relation to road networks. It also involves the drivers’ behavior in terms of speed and ability to observe deer. Though speed limit is likely a key factor, only a few reports have quantified how speed limits affect the number of DVCs (Bashore et al. 1985, Seiler 2005, Danks and Porter 2010, Neumann et al. 2012). Predictive models of DVC sites are regarded as the key to mitigate accidents both at regional and local scales (Malo et al. 2004, Gunson et al. 2011). Such models can be important for identifying relevant factors for DVC risk and to assess mitigation measures at the right scale. Modeling of accident risk relative to environmental characteristics could enable managers to plan the layout of new roads or determine locations of mitigation measures. One commonly used mitigation measure is clearance of vegetation along roads to prevent collision in forest habitats (Rea 2003, Langbein et al. 2011). The idea is to increase the visibility of approaching animals and give the driver more time to react. Removal of palatable vegetation also reduces the time animals spend in the vicinity of the road and thereby decreases the number of potential crossings (Olsson 2007). This measure is commonly used to prevent moose (Alces alces)-vehicle collisions (Iuell et al. 2003, Rea 2003), but few studies have formally tested the effect of vegetation clearance along roads for reducing DVCs (Gunson et al. 2011, Langbein et al. 2011). Red deer (Cervus elaphus) is among the most widely distributed large herbivore species in Europe (Skog et al. 2009), and populations have increased in number over much of Europe depending on the management system (Milner et al. 2006). In Norway, 840 red deer were recorded killed in car accidents during the season 2007–2008, which was historically high. The main reason for the increase in number of car-killed red deer in Norway seems to be an effect of increasing deer population size and increasing traffic burden (Mysterud 2004). We developed 2 collision risk models by contrasting location data of DVCs with 1) random positions along roads, and 2) position data from global positioning system (GPS)-marked red deer within the study area in Møre & Romsdal and Sør-Trøndelag, Norway. We developed models of relative risk of collisions by incorporating information about speed limit, road standard, and habitat features. Further, we tested the effect of vegetation clearance along a highway on collision frequency of red deer within 1 part of the study area. STUDY AREA The study area covered approximately 24,000 km2 overlapping several municipalities in the counties of Møre & Romsdal and Sør-Trøndelag in central Norway (N 62.15– 63.598; E 6.82–10.698; Fig. 1; Fig. A1, available online at Figure 1. Examples of distribution of home ranges of red deer (100% minimum convex polygons; solid black lines: n ¼ 24) with major roads (gray lines) and road crossings (dark gray dots) within the study area in central Norway, 2007–2008. 2 The Journal of Wildlife Management 9999 www.onlinelibrary.wiley.com). The climate is oceanic and precipitation and temperature generally declines from coast to inland and from lowland to higher altitudes. Snow cover is normally present during winter, but depth and duration of snow cover increases from coast to inland and with altitude, and is highly variable among years (Mysterud et al. 2000). The topography is characterized by valleys and fjords, hills, and mountains, with peaks reaching more than 1,500 m above sea level. Agricultural areas are normally situated on flatter and more fertile grounds in the bottom of valleys. The study area was forested and situated in the south to middle boreal zone (Moen 1998). A more detailed description of the vegetation can be found elsewhere (Meisingset et al. 2013). Highways (termed major roads) and country roads (termed medium roads) typically followed flatter areas in valley bottoms and frequently passed through agricultural land where red deer feed (Godvik et al. 2009, Meisingset et al. 2013). Private roads (termed small roads) were located around houses, farms, and pastures, and within forested areas where they mainly were built for timber transport. Road density for major, medium, and small roads was 0.05, 0.10, and 0.22 km/km2, respectively. Density of roads was highest in agricultural and high productive forest areas (Meisingset et al. 2013). The area had no fencing along the roads to prevent DVCs. Road edges were cleared along a highway in the valley of Sunndal, in the southern part of our study area. The control area was situated in the neighboring valley Surnadal with a comparable topography, highway, traffic load, and red deer population density. These 2 valleys were typical Norwegian fjord valleys with high mountains and had a highway passing through the valley beside a river. The red deer population is located in the core areas for red deer in Norway (Mysterud et al. 2011). It has increased steadily the last 4 decades and the harvest yield has increased approximately 6-fold in the region as a whole during this period, and more than 6,300 red deer were harvested during autumn hunt in 2008. Annual harvest takes place between 10 September and 23 December. A majority of adult red deer migrate seasonally between small winter (coast; low elevation) and summer (inland; higher elevation) ranges (Bischof et al. 2012), but the proportion of migrants varies with population density within the municipality and local topography (Mysterud et al. 2011). Within our study area, 140–160 red deer were killed by car per year, at an increasing rate during the last 2 decades (Meisingset et al. 2010). METHODS Data Collection We obtained data on DVCs within the study area from the Cervid Register in Norway. We had access to data from 271 DVCs within the combined 100% minimum convex polygon home range of the GPS-marked red deer. We used 100% to be sure to include the entire area where road crossings occurred. The DVCs had exact locations (recorded with handheld GPS; accuracy 10 m) and date (ranging from Jan 2006 to Jun 2009). During summer 2008, vegetation removal was conducted along a 32-km stretch of a highway in our Meisingset et al. Collision Risk of Red Deer study area (Highway 70 in the valley of Sunndal). All vegetation up to 8 m from the road shoulder was removed to increase visibility along the road edges. In the subsequent summers, new vegetation in this clearance area was cut mechanically to prevent re-growth. Deer–vehicle collision data along this highway section were collected from the period 2003–2010, totaling 137 collisions (103 occurring before and 34 occurring after clearance). We excluded 3 months in summer 2008 when the vegetation removal was conducted, which gave 94 months of data. Data on DVCs were also collected from the 35-km control stretch (Highway 65 in Surnadal) located about 25 km from the treatment road in the neighboring municipality. In the control area, 70 DVCs occurred on the highway during the study period. We used individual GPS data from red deer to identify crossing sites. We had access to 67 adult (2 years) red deer of both sexes (45 females and 22 males) marked with GPS collars (Followit AB, Lindesberg, Sweden; Meisingset et al. 2013). We caught animals by darting on winter sites from January to April in 2007 and 2008, after a procedure approved by the national ethical board for science. Winter feeding sites for marking were spatially distributed in lowland natural wintering areas within the study area. Most animals use winter-feeding sites and we regard the possibility for a biased sampling within the deer population to be low, but in winter each deer will often be found close to the site it was captured. Our marking sites differed in relation to road densities and population densities. We selected individuals for capture based on proximity to the hidden capture personnel and position within the group (to facilitate a safe shot). Female collars recorded a position every hour, whereas male collars recorded positions every second hour apart from 10 September to 1 December when male collars recorded positions every hour. The average fix success rate was 82.5% (range 61.0–98.5%). We deleted locations obtained during the first 24 hr after marking. After a screening process that included visually plotting for removal of outliers and speed rules (deleting points that indicated a movement speed of more than 30 km/hr), 370,859 locations were available for further analysis. Deleted points composed less than 0.5% of all locations. Throughout the 2 years of data sampling, a mean of 34.8 (SD 10.66) collars were operational every month, and each GPS collar obtained positions for 11.8 (SD 5.54) months. The mean number of locations per individual was 5,454 (SD 2,784.4). Median location error for our GPS collars was 12 m (Godvik et al. 2009). The deer home ranges were spread out over a large area, varied largely in size, and had substantial overlap in the areas with the most intensive catching effort (Fig. A1). We defined deer road crossing sites where step lines intersected a road (see Meisingset et al. 2013 for a closer description). To increase precision of estimated road crossing sites, we excluded all crossings between GPS locations with a sampling interval >120 min (due to missing positions before or after crossing). We also excluded crossing sites in cases where the distance between successive positions exceeded 1 km. This resulted in 34,360 remaining crossing sites available for further analyses. The GPS-sampling interval or 3 distance between locations did not likely result in any systematic bias from excluded crossing points. We divided the roads into 3 categories based on size and expected traffic burden according to the Norwegian road authority’s classification of roads (Meisingset et al. 2013). We classified private roads as small roads because of low traffic burden (estimated cars/day: 0–100) and just 1 unpaved lane; country roads as medium roads because these were usually paved and with 1 lane in each direction and with intermediate traffic burden (estimated cars/day: 100–1,000); and highways as major roads with at least 1 lane in each direction and high traffic burden (estimated cars/day: 1,000– 4,000; www.vegvesen.no). Roadmaps were available as vector layer data at scale 1:50,000. Speed limits on all roads were available from the Norwegian road authority, which covered all registered roads within our study area. Speed limits varied from 30 to 80 km/hr. Because speed limits of 30 km/hr and 40 km/hr mainly were distributed within or close to urban areas, which is likely to influence the crossing behavior of red deer, we excluded these road segments from the analyses totaling 171 crossing sites and 2 DVC sites. Very few stretches of roads with 70 km/hr speed limits occurred within the study area; we therefore grouped 60 and 70 km/hr speed limits and treated them as 1 level. We characterized the habitat in a 100-m circular buffer zone around each crossing site. We selected the size of the buffer zone to realistically include the actual crossing site, and to be large enough to allow characterization of the topography but small enough to provide a useful measure of habitat composition in our finely fragmented study area. Within the buffer zone, we calculated the proportion of different habitat types. We derived habitat types from digital land resource maps provided by the Norwegian Forest and Landscape Institute, with scale 1:5,000 (Godvik et al. 2009, Loe et al. 2012, Meisingset et al. 2013). We classified habitat into agricultural infields (mostly pastures and meadows; termed pastures hereafter) and forest, using distance to pasture (linear distance in meters from the estimated crossing site to the closest pasture) and proportion of productive forest cover within the 100-m buffer as covariates in subsequent analyses. We calculated the same variables for 100-m buffers around DVC sites. We also quantified terrain ruggedness based on a digital elevation model with 25-m resolution. For each grid cell, we calculated a ruggedness index (Sappington et al. 2007) based on the center cell and the 8 neighboring cells (i.e., the values are based on the elevation in a 75m 75-m square). High scores represented a highly variable topography, whereas lower scores represented a smoother terrain. We processed spatial maps and data in ArcGIS 9.3 (Environmental Systems Research Institute, Inc., Redlands, CA). To calculate distance to pastures, we used the vectorediting tool, Snap Points to Lines, in Hawth’s Analysis Tools for ArcGIS (Beyer 2004). To calculate terrain ruggedness, we used Terrain Tools for ArcGIS. As an index of population density within Sunndal (location of the vegetation clearance) and Surnadal (control site), we used the annual number of harvested red deer during the 4 hunting season (obtained from the local managers in Sunndal and Surnadal municipalities). The number of shot animals gives a fairly good estimate of changing population density between years within a given area (Mysterud et al. 2007). We gathered data on snow depth (measured as max. snow depth during the specific month) from local weather stations in the 2 municipalities (stations no. 64,580 and 64,800, www. eklima.no). Statistical Analyses Predictive models of collision risk.—To test if DVC sites differed from random locations along roads and from red deer crossing sites, we fitted 2 different exponential resource selection function (RSF; Equation (1)) models with maximum likelihood estimates and nonparametric bootstrap standard errors (B ¼ 99), using the function “rsf” in the package “ResourceSelection” in R (Lele 2009, Lele et al. 2013, McDonald 2013). wðxÞ ¼ expðb1 x1 þ þ bp xp Þ ð1Þ In the first case, we fitted an RSF model with DVC (coded 1) and random locations (coded 0) as the response variable with a non-matched used-available design. In the second model, we fitted an RSF with collision site (coded 1) and deer crossing site (coded 0) as the response variable (Table A1). In the second model, each DVC site was named after the individual red deer home range(s) it was located within. Both DVC site and deer crossing site were therefore associated to an individual deer, and individual red deer home range was included as a matching factor in a matched used-available RSF. Candidate predictor variables were the same in both models (Table A2, available online at www.onlinelibrary. wiley.com) and included season, road category, distance to closest pasture, terrain ruggedness, productive forest cover, and speed limit. Seasons included spring (1 Apr–31 May), summer (1 Jun–15 Aug), autumn (16 Aug–30 Nov), and winter (1 Dec–31 Mar) and was incorporated as a 4-level factor variable. We divided road category into major and medium sized roads as a 2-level factor variable. Because no DVCs occurred at small roads, but 55.0% of red deer crossings occurred on these road segments, we excluded red deer crossing sites at small roads from the analyses. We treated distance to closest pasture as a 2-level factor variable, closer or farther than 50 m, to investigate the local effect of improved visibility in open terrain. We treated terrain ruggedness as a numeric variable; the value increased with the degree of ruggedness. We treated productive forest cover as a numeric variable from 0% to 100% cover of productive forest within the 100-m circular buffer zone. And finally, we included speed limit as a 3-level factor variable with 50, 60– 70, or 80 km/hr as levels. This type of model has been termed a collision risk model (Malo et al. 2004, Seiler 2005, Frair et al. 2008). In our case, this gives a relative probability of a site being a collision site because 2 different datasets are used to estimate the probability. We collected road crossings only for the subset of GPS-marked deer. Because only 1 GPS-marked deer was killed by a car, the collision data consisted of all reported red The Journal of Wildlife Management 9999 Meisingset et al. Collision Risk of Red Deer 0.12 0.22 þ 0.31 0.04 0.78 þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ þ 1 þ þ þ þ þ þ þ þ þ þ 1 0.65 þ þ þ þ þ þ þ þ þ þ þ þ þ þ 1 þ þ þ þ þ þ þ þ þ þ 1 þ þ þ þ þ þ þ þ þ þ 1 Model 1 2 3 4 5 6 7 8 9 10 Imp. þ þ Weight 0.202 0.182 0.097 0.076 0.068 0.046 0.041 0.036 0.035 0.032 0.00 0.22 1.47 1.95 2.19 2.96 3.18 3.43 3.51 3.66 DAIC AIC 3,952.7 3,952.9 3,954.2 3,954.7 3,954.9 3,955.7 3,955.9 3,956.1 3,956.2 3,956.4 1,966.4 1,963.5 1,958.1 1,966.3 1,963.5 1,961.8 1,969.9 1,958.1 1,967.1 1,969.2 logLik Df 10 13 19 11 14 16 8 20 11 9 þ þ þ þ þ þ þ þ þ þ 1 Road category speed limit Season road category Season speed limit Road category distance to pasture Speed Season dis- limit Season Speed limit Terrain Road category tance distance (categorical) (categorical) ruggedness (categorical) to pasture to pasture Cover of productive forest Out of the available 271 DVC sites, 83.0% (n ¼ 225) were found on major roads and 17.0% (n ¼ 46) on medium-sized roads. None were found on small roads. Out of the 34,360 identified red deer road crossing sites, 11.9% (n ¼ 4,075) were found on major roads, 33.1% (n ¼ 11,386) on medium roads, and 55.0% (n ¼ 18,899) on small roads (see descriptive statistics in Tables A1 and A2, available online at www. onlinelibrary.wiley.com). Distance to pasture (categorical) RESULTS Table 1. Degrees of freedom (Df), log likelihood (logLik) value, Akaike’s Information Criterion (AIC) values, DAIC values, and weight for the top 10 candidate models (based on AIC) for red deer–vehicle collision sites (DVC, coded 1) versus random sites along roads (coded 0) in central Norway, 2006–2009, fitted with an exponential resource selection function (RSF) model with maximum likelihood estimates. For each model (1–10), þ shows whether the specific parameter is included into the model. The total number of candidate models was 339. The relative importance of variables (Imp) is provided in the last row. deer (with 1 exception unmarked) killed by car within the combined home ranges of the GPS-marked deer. The low proportion of GPS-marked deer DVC sample is likely explained by a low number of GPS-marked individuals compared to the total population size (of mainly unmarked deer) that are subjected to DVCs. We also selected random sites along the roads within the combined deer home ranges. We tested for correlation among the explanatory variables. However, no variable was correlated by r > 0.5, so all candidate variables were subjected to model selection. We selected candidate models using the dredge function in the package “MuMIn” in R. In the first model (DVC vs. random), we included all predictor variables, as well as all 2way interactions between the categorical variables. In the second model (DVC vs. crossings), we included all the predictor variables, as well as the 2-way interaction between road category and speed limit and season and distance to pasture. We selected the final models based on Akaike’s Information Criterion (AIC). We also calculated model weights and relative importance of the variables (sum of Akaike weights of all models where the variable is included; Barton 2013) in the top-ranked candidate models. Even though we had data from both sexes, we could not analyze if 1 sex was overrepresented in DVCs because we had no precise information about sex ratio within the whole study population. In the full model, we included all DVC sites including deer of unknown sex (which substantially increase sample size). Effect of vegetation clearance.—We tested the effect of vegetation clearance on Highway 70 by using a negative binomial generalized linear model using the function glm.nb in library MASS in R (Venables and Ripley 2002). The response variable was the number of DVCs per month (n ¼ 94 months). Predictor variables included the 2-level factor variables treatment (treatment area and control area), period (before and after vegetation clearance in the treatment area), and season (winter, dark season with snow [Oct–Mar], and summer, light season and no snow [Apr–Sep]), as well as population density (numeric) and the interaction between treatment and period. We also developed a separate winter model (a subset of data from the winter season [Oct–Mar]) where we also included snow depth (numeric; measured as maximum registered snow depth each month). We ran the last model to control for the effect of snow conditions during winter, which may lead to more slippery roads and a higher concentration of deer density in the valley bottom with more road crossings as a result. We conducted all analyses in R version 2.15.2 (R Development Core Team 2011). 5 Red Deer–Vehicle Collision Sites The distribution of DVCs compared with random sites along roads differed considerably in relation to several of the predictor variables. The most supported model (Table 1) included road category, distance to pasture, productive forest cover, terrain ruggedness, and speed limit, as well as the 2way interactions between road category and speed limit, and distance to pasture and speed limit. The relative risk for DVCs increased with speed limit (Table 2). With a change in speed limit from 50 to 60–70 km/hr and from 50 to 80 km/ hr, relative risk increased by 3.9 and 8.6, respectively. The increase in relative risk from 50 to 80 km/hr was lower for medium than for major roads (interaction between road category and speed limit; Table 2). An increasing proportion of productive forest cover increased the risk for DVCs (Table 2). With a 10% increase in cover of productive forest, the relative risk increased by 1.3. We also observed an increased relative risk for DVCs in areas with more rugged terrain (relative risk is 7.4 times higher in the road locations with the lowest compared to the highest ruggedness values; Table 2). The DVC sites differed from red deer crossing sites for several of the predictor variables. The most supported model included speed limit, road category, season, distance to pasture, and productive forest cover, as well as the 2-way interactions between road category and speed limit, and season and distance to pasture (Table 3). The relative risk for DVCs increased with speed limit and was highest at 80 km/hr (Table 4, Fig. 2). An increase in speed limit from 50 to 60–70 km/hr increased the relative risk by a factor of 6.8, whereas from 50 to 80 km/hr gave an increased risk of 14.4. The relative risk for DVC was lower at medium-sized roads compared to major roads (relative risk of 0.3). As for the previous model, collision risk increased more at major roads compared to medium-sized roads when increasing the speed limit from 50 to 80 km/hr (Table 4). The relative risk decreased considerably when distance to pasture exceeded 50 m (relative risk of 0.4; Table 4, Fig. 2) and this difference occurred in all seasons. Probability of a DVC during winter was considerably greater compared to the other seasons (Table 4, Fig. 2). The increase in relative risk from autumn to winter was 4.9, whereas risk was lower in spring compared to autumn (relative risk of 0.6). An increasing level of productive forest cover increased the risk for DVCs (Table 4, Fig. 3). Relative risk increased by 1.3 for every 10% increase in cover of productive forest. Relative risk increased at a higher rate closer to pastures and in winter compared to other seasons with an increasing level of productive forest cover (Fig. 3). Effect of Vegetation Clearance The mean number of DVCs per month decreased by 31% from 1.6 (SD ¼ 2.00, n ¼ 64) before to 1.1 (SD ¼ 1.33, n ¼ 30) after vegetation clearance in the treatment area. This amounts to 6 DVCs less per year for the 32-km road stretch. In the control area, DVCs increased from 0.6 (SD ¼ 0.77, n ¼ 64) per month to 1.1 (SD ¼ 1.59, n ¼ 30) over the same period for the same length of road. The effect of vegetation clearance differed by season and reduced the number of DVCs during the winter season only (Table 5, Fig. 4), when the reduction in DVCs was as high as 53%. Population density tended to be positively related to number of DVCs. The winter model confirmed the positive effect of road edge clearance in the treatment area (Table 5), even after controlling for effects of snow cover and population density on collision risk. DISCUSSION Temporal and spatial patterns in DVCs can be used to predict the optimal locations of deer crossing structures (sites with high risk of DVC; e.g., Clevenger and Waltho 2005), the seasonal timing of efforts, and how to manipulate habitat features or road characteristics to reduce collision risk. Our study shows that both spatial and temporal (seasonal) variation occurs in collision risk by red deer, and that several factors contribute to this risk. Speed limit is important for risk of DVC. Further, road edge clearance substantially reduced DVC frequency and efforts could be targeted to the winter season. Increased speed gives the driver shorter time to respond when a deer crosses the road in front of the car, and increased distance for the vehicle to stop. Studies on collision sites of moose both in Sweden and western Maine, USA, have previously reported higher probability of moose–vehicle collisions with increasing speed limit (Seiler 2005, Danks and Porter 2010, Neumann et al. 2012). Danks and Porter Table 2. Parameter estimates and test statistics from the most parsimonious model (exponential resource selection function [RSF] model with maximum likelihood estimates and nonparametric bootstrap standard errors [B ¼ 99]) for the risk model contrasting red deer–vehicle collision (DVC) sites (coded 1) with random sites along roads (coded 0) in Møre & Romsdal and Sør-Trøndelag, Norway, 2006–2009. Estimates of factor variables are relative risk relative to a reference category, whereas estimates for numeric variables are the change in relative risk per unit of the variable (1% for productive forest cover). Parameter Road category (medium vs. major) Distance to pasture (>50 m vs. <50 m) Productive forest cover (%) Terrain ruggedness Speed limit (60 vs. 50 km/hr) Speed limit (80 vs. 50 km/hr) Road category (medium) speed limit (60 km/hr) Road category (medium) speed limit (80 km/hr) Distance to pasture (>50 m) speed limit (60 km/hr) Distance to pasture (>50 m) speed limit (80 km/hr) 6 Estimate SE Z Lower 95% CI Upper 95% CI 0.343 0.116 0.012 1.744 1.374 2.147 0.974 2.278 0.084 1.153 0.542 0.980 0.002 0.371 0.569 0.553 0.644 0.614 1.103 0.982 0.633 0.118 5.112 4.696 2.415 3.881 1.512 3.709 0.076 1.175 1.404 2.036 0.008 1.016 0.259 1.063 0.288 3.481 2.246 3.078 0.719 1.804 0.017 2.472 2.489 3.232 2.236 1.074 2.079 0.771 The Journal of Wildlife Management 9999 1 0.22 þ þ Weight 0.777 0.223 0.000 0.000 0.000 0.00 2.49 18.51 18.68 44.39 DAIC AIC 4,131.7 4,134.2 4,150.2 4,150.4 4,176.1 2,052.9 2,053.1 2,065.1 2,064.2 2,079.1 logLik Df 13 14 10 11 9 þ þ þ þ þ 1 þ þ þ þ þ þ þ 1 þ þ þ þ þ 1 1 þ þ þ þ þ 1 1 2 3 4 5 Imp. Model þ þ þ þ þ þ þ þ þ 1 Road category speed limit Season distance to pasture Road category (categorical) Terrain ruggedness Season Speed limit (categorical) (categorical) Cover of productive forest Distance to pasture (categorical) Table 3. Degrees of freedom (Df), log likelihood (logLik) value, Akaike’s Information Criterion (AIC), DAIC values, and weight for the top 5 candidate models (based on AIC) for red deer–vehicle collision sites (DVC, coded 1) versus red deer crossing sites (coded 0) in central Norway, 2006–2009, fitted with an exponential resource selection function (RSF) model with maximum likelihood estimates. For each model (1–5), þ shows whether the specific parameter is included into the model. The total number of candidate models was 76. The relative importance of variables (Imp.) is provided in the last row. Meisingset et al. Collision Risk of Red Deer (2010) found that for each 8 km/hr increase in speed, odds of moose–vehicle collisions increased by 35%, but the rate of increase was most apparent at speed limits above 72 km/hr. Also in our case, speed limit was an important factor to explain variation in relative risk of DVC. The timing of mitigation efforts is critical when considering costs of maintenance (as for road clearance) or feasibility of continuous efforts. For example, reducing speed limit may be unpopular among car drivers, but may work better if it is only for given periods of particularly high risk. The risk of DVC shows considerable temporal variation, but we found no consistent pattern among studies. Studies have reported peaks in DVC during autumn (roe deer [Capreolus capreolus]; Madsen et al. 2002), summer (moose; Danks and Porter 2010, Neumann et al. 2012), or during migration periods (moose; Dussault et al. 2007), whereas some studies have shown a peak of road crossings during rutting season and hunting activities (white-tailed deer [Odocoileus virginianus]: Allen and McCullough 1976, Sudharsan et al. 2006; moose: Lavsund and Sandegren 1991). We found higher risk for DVCs during winter compared to the other seasons. At northern latitudes, light conditions vary substantially between seasons. During autumn and winter, we often find peaks in activity pattern of deer during dark periods after dawn and before dusk (Godvik et al. 2009, Pepin et al. 2009). During winter, this often corresponds with peaks in daily traffic level. In northern summers, darkness is virtually absent. Correspondingly, the proportion of meetings between cars and deer that occur during darkness is much higher during winter than other parts of year. In addition, roads are more slippery during winter when snow is present adding to the increased risk of DVCs during this time of year. Peaks in collision rates are found in dark periods in white-tailed deer (Haikonen and Summala 2001, Sudharsan et al. 2006) and moose (Haikonen and Summala 2001, Dussault et al. 2006, Danks and Porter 2010). Also, more deer are present at low elevation valley bottoms with more roads during winter because of seasonal migration (Mysterud et al. 2011). Pastures are important deer habitats that are selected during all seasons (Godvik et al. 2009), and much of the road crossings by red deer are probably motivated to get access to pastures (Gagnon et al. 2007). Our study shows a higher risk for DVCs close to pastures even with better visibility in this open habitat type. Red deer likely take higher risks when approaching pastures and cross roads close to pastures. Proximity to pasture seems to be an important factor when evaluating DVC risks (Tappe and Enderle 2007) and has to be considered when planning mitigation measures (Gagnon et al. 2007). We also found an increased risk for DVCs with increasing cover of productive forest, whereas rugged terrain was partly supported to increase risk of DVCs (terrain ruggedness were included in the most supported model of DVC vs. random points, but not in the model of DVC vs. red deer crossing points). Both these factors probably affect the ability for drivers to detect crossing red deer in time. Accordingly, the risk of DVCs decreased with visibility in white-tailed deer (Bashore et al. 1985, Nielsen et al. 2003), 7 Table 4. Parameter estimates and test statistics from the most parsimonious model (exponential resource selection function [RSF] model with maximum likelihood estimates and nonparametric bootstrap standard errors [B ¼ 99]) risk model contrasting red deer–vehicle collision (DVC) sites (coded 1) with red deer crossing sites (coded 0) in Møre & Romsdal and Sør-Trøndelag, Norway, 2006–2009. Estimates of factor variables are relative risk relative to a reference category and estimates for numeric variables are the change in log odds per unit of the variable (1% for productive forest cover). Season (spring vs. autumn) Season (summer vs. autumn) Season (winter vs. autumn) Road category (medium vs. major) Distance to pasture (>50 m vs. <50 m) Productive forest cover (%) Speed limit (60 vs. 50 km/hr) Speed limit (80 vs. 50 km/hr) Road category (medium) speed limit (60 km/hr) Road category (medium) speed limit (80 km/hr) Season (spring) distance to pasture (>50 m) Season (summer) distance to pasture (>50 m) Season (winter) distance to pasture (>50 m) Estimate SE Z Lower 95% CI Upper 95% CI 0.552 0.062 1.579 1.259 0.885 0.025 1.918 2.665 0.126 3.223 0.759 1.227 0.617 0.177 0.245 0.166 0.419 0.220 0.003 0.385 0.360 0.607 0.463 0.291 0.559 0.304 3.123 0.251 9.510 3.002 4.020 8.040 4.979 7.404 0.208 6.963 2.610 2.196 2.028 0.898 0.419 1.254 2.081 1.317 0.019 1.163 1.959 1.063 4.130 0.189 2.323 0.021 0.205 0.542 1.905 0.437 0.454 0.031 2.673 3.370 1.316 2.315 1.329 0.132 1.213 moose (Seiler 2005), roe deer, and wild boar (Sus scrofa; Malo et al. 2004). Therefore, landscape characteristics need to be taken into account when assessing mitigation efforts. Fencing, over- and underpasses, vegetation modification of road edge, and traffic signs are the most common mitigation measures to reduce DVCs (Langbein et al. 2011). Target clearance along the roads has been increasing in recent years, in particular to reduce the number of collisions of moose and deer (Iuell et al. 2003). However, few studies have tested the effect of such measures. Our study shows a positive effect of road edge clearance, with a 53% reduction in DVCs during winter season. Interestingly, we found no effect of road edge clearance during summer. This was not caused by regrowth, because new vegetation was removed every summer in the treatment period. Two studies in Norway reported a positive Winter (<50 m) Relative risk of DVC Parameter Summer (<50 m) 1400 1400 1200 1200 1000 1000 800 800 600 600 400 400 200 200 0 0 0 15 30 45 60 75 90 0 60 75 90 1200 1200 1000 1000 800 800 600 600 400 400 200 200 0 15 30 45 60 75 90 0 15 30 45 60 75 90 Figure 3. Relative risk (SE) for red deer crossings ending in deer–vehicle collisions (DVCs) as a function of productive forest cover (%) during summer and winter seasons, at 2 distance categories form pastures (<50 m, >50 m) and at speed limits of 50 km/hr (gray lines) and 80 km/hr (black lines) in central Norway, 2006–2009. The predictions are based on the most parsimonious model and values are shown for major roads. 100 4 50 0 60−70 80 Speed limit (km/hr) Figure 2. Relative risk (SE) for red deer crossings ending in deer–vehicle collisions (DVCs) as a function of speed limit (50, 60–70, and 80 km/hr) during summer (gray) and winter (black) seasons at 2 distance categories from pastures (<50 m: solid dots, >50 m: open dots) in central Norway, 2006–2009. The predictions are based on the most parsimonious model. Predicted values are for major roads and for mean values of numeric covariates (productive forest cover ¼ 20.9%). Number of collisions / month Relative risk of DVC 45 Productive forest cover (%) 150 8 30 Summer (>50 m) 1400 0 50 15 Winter (>50 m) 1400 0 Winter <50 m Winter >50 m Summer <50 m Summer >50 m 80 km/hr 50 km/hr Treatment road Control road 4 3 3 2 2 1 1 0 0 Before After Winter Before After Summer Figure 4. Predicted number of red deer–vehicle collisions (DVCs) per month (95% CI) at the treatment road (vegetation clearance; solid dots) and the control road (open dots), before and after roadside vegetation clearance and grouped into winter and summer seasons in central Norway, 2003–2010. The Journal of Wildlife Management 9999 Table 5. Parameter estimates and test statistics for the number of red deer–vehicle collisions (DVC) per month as a function of the spatiotemporal effect of vegetation removal along Highway 70 (treatment) and other covariates in central Norway, 2003–2010. Two different negative binomial generalized linear models are fitted; first a full year model incorporating the effect of season, and second as a winter season only model. Parameter Full year model Intercept Area (treatment vs. control) Road edge clearance (before vs. after) Season (winter vs. summer) Population density index Area road edge clearance Area season Road edge clearance season Area road edge clearance season Winter model Intercept Area (treatment vs. control) Road edge clearance (before vs. after) Population density index Monthly maximum snow depth Area road edge clearance Estimate SE Z P-value 0.364 1.093 0.752 1.745 0.005 1.226 0.732 0.874 0.884 0.433 0.267 0.308 0.303 0.003 0.431 0.498 0.519 0.519 0.841 4.096 2.440 5.759 1.622 2.844 1.471 1.683 1.129 0.401 <0.001 0.015 <0.001 0.105 0.004 0.141 0.092 0.259 0.191 1.179 0.730 0.005 0.006 1.245 0.460 0.260 0.291 0.003 0.004 0.404 0.415 4.542 2.509 1.492 1.737 3.083 0.678 <0.001 0.012 0.136 0.082 0.002 effect of vegetation clearing along the railway for the number of collisions of moose (Jaren et al. 1991, Andreassen et al. 2005). After clearing the railway line, the number of collisions decreased by approximately 50%. In Sweden, an accident reduction of nearly 20% was found for moose– vehicle collisions after clearing of bushes and branches below 3 m within 20 m from the road (Lavsund and Sandegren 1991). Corresponding figures for the effects of clearing along the road are sparse for both moose and red deer in Scandinavia and elsewhere. The effect of road edge clearance is likely at least partly a car driver effect; clearance gives better visibility along roadsides and possibilities for discovering deer earlier as they are approaching the road. We show road edge clearance is an effective measure even if the traffic level is quite low (700–2,040 cars/day). Even though measures ideally should be conducted on large scale, our study demonstrates that road clearance is efficient, and probably more cost effective, when targeted at shorter road stretches with particular high DVC frequencies. MANAGEMENT IMPLICATIONS Our study shows that speed limitation within risk areas for DVCs can reduce collision risk substantially. Reducing speed limits during either the entire winter season or when snow conditions are severe may be an effective compromise if a permanent reduction in speed is difficult to implement. Targeting this measure within dense forest habitats with a variable topography in the vicinity of pastures should be most effective. Likewise, vegetation clearance could reduce DVCs, especially during winter season, and should similarly be targeted in dense forest habitats close to pastures because they are the stretches with the highest risk of DVCs. Effective clearance can thus be achieved if completed in late fall when vegetation growth has ceased. ACKNOWLEDGMENTS This study was funded by the Research Council of Norway (“Natur og næring”-program; project no. 179370/I10), the Meisingset et al. Collision Risk of Red Deer Directorate for nature management and Norwegian Institute for Agricultural and Environmental Research. Marking was mainly funded by game fund of various counties, municipalities, and red deer management units, and we are grateful to all people who have assisted in the marking of animals. Thanks to J. Cabell for carefully improving the English writing. We are grateful to the Associate Editor, and 2 anonymous referees for very valuable comments that led to substantial improvements in the manuscript. Authors have no conflict of interest related to this work. LITERATURE CITED Allen, R. E., and D. R. McCullough. 1976. Deer-car accidents in southern Michigan. Journal of Wildlife Management 40:317–325. Andreassen, H. P., H. Gundersen, and T. Storaas. 2005. The effect of scentmarking, forest clearing, and supplemental feeding on moose-train collisions. Journal of Wildlife Management 69:1125–1132. Barton, K. 2013. Package “MuMIn”: multi-model inference. <http://cran.rproject.org/> Accessed 7 May 2013. Bashore, T. L., W. M. Tzilkowski, and E. D. Bellis. 1985. Analysis of deer– vehicle collision sites in Pennsylvania. Journal of Wildlife Management 49:769–774. Beyer, H. L. 2004. Hawth’s Analysis Tools for ArcGIS. <http://www. spatialecology.com/htools> Accessed 26 Jan 2009. Bischof, R., L. E. Loe, E. L. Meisingset, B. Zimmermann, B. Van Moorter, and A. Mysterud. 2012. A migratory northern ungulate in the pursuit of spring: jumping or surfing the green-wave? American Naturalist 180:466– 474. Clevenger, A. P., B. Chruszcz, and K. E. Gunson. 2001. Highway mitigation fencing reduces wildlife-vehicle collisions. Wildlife Society Bulletin 29:646–653. Clevenger, A. P., and N. Waltho. 2005. Performance indices to identify attributes of highway crossing structures facilitating movement of large mammals. Biological Conservation 121:453–464. Danks, Z. D., and W. F. Porter. 2010. Temporal, spatial, and landscape habitat characteristics of moosevehicle collisions in Western Maine. Journal of Wildlife Management 74:1229–1241. Dussault, C., J. P. Ouellet, C. Laurian, R. Courtois, M. Poulin, and L. Breton. 2007. Moose movement rates along highways and crossing probability models. Journal of Wildlife Management 71:2338–2345. Dussault, C., M. Poulin, R. Courtois, and J. P. Ouellet. 2006. Temporal and spatial distribution of moose-vehicle accidents in the Laurentides Wildlife Reserve, Quebec, Canada. Wildlife Biology 12:415–425. 9 Frair, J. L., E. H. Merrill, H. L. Beyer, and J. M. Morales. 2008. Thresholds in landscape connectivity and mortality risks in response to growing road networks. Journal of Applied Ecology 45:1504–1513. Gagnon, J. W., T. C. Theimer, N. L. Dodd, S. Boe, and R. E. Schweinsburg. 2007. Traffic volume alters elk distribution and highway crossings in Arizona. Journal of Wildlife Management 71:2318–2323. Godvik, I. M. R., L. E. Loe, J. O. Vik, V. Veiberg, R. Langvatn, and A. Mysterud. 2009. Temporal scales, trade-offs, and functional responses in red deer habitat selection. Ecology 90:699–710. Groot Bruinderink, G. W. T. A., and E. Hazebroek. 1996. Ungulate traffic collisions in Europe. Conservation Biology 10:1059–1067. Gunson, K. E., G. Mountrakis, and L. J. Quackenbush. 2011. Spatial wildlife-vehicle collision models: a review of current work and its application to transportation mitigation projects. Journal of Environmental Management 92:1074–1082. Haikonen, H., and B. A. Summala. 2001. Deer-vehicle crashes—extensive peak at 1 hour after sunset. American Journal of Preventive Medicine 21:2009–2213. Iuell, B., G. J. Bekker, R. Cuperus, J. Dufek, G. Fry, C. Hicks, V. Hlavac, L. Keller, B. Le Maire Wndall, C. Rosell, T. Sangwine, and N. Torsløv. 2003. Wildlife and traffic: a European handbook for identifying conflicts and designing solutions. KNNV Publishers COST 341—Habitat fragmentation due to transportation infrastructure. European Co-operation in the Field of Scientific and Technical Research, Brussels, Belgium. Jaren, V., R. Andersen, M. Ulleberg, P. H. Pedersen, and B. Wiseth. 1991. Moose-train collisions—the effects of vegetation removal with a costbenefit-analysis. Alces 27:93–99. Langbein, J., R. Putman, and B. Pokorny. 2011. Traffic collisions involving deer and other ungulates in Europe and available measures for mitigation. Pages 215–259 in R. Putman, M. Apollonio, and R. Andersen, editors. Ungulate management in Europe: problems and practices. Cambridge University Press, Cambridge, United Kingdom. Lavsund, S., and F. Sandegren. 1991. Moose-vehicle relations in Sweden: a review. Alces 27:118–126. Lele, S. R. 2009. A new method for estimation of resource selection probability function. Journal of Wildlife Management 73:122–127. Lele, S. R., E. H. Merrill, J. Keim, and M. S. Boyce. 2013. Selection, use, choice and occupancy: clarifying concepts in resource selection studies. Journal of Animal Ecology 82:1183–1191. Loe, L., C. Bonenfant, E. Meisingset, and A. Mysterud. 2012. Effects of spatial scale and sample size in GPS-based species distribution models: are the best models trivial for red deer management? European Journal of Wildlife Research 58:195–203. Madsen, A. B., H. Strandgaard, and A. Prang. 2002. Factors causing traffic killings of roe deer Capreolus capreolus in Denmark. Wildlife Biology 8:55– 61. Malo, J. E., F. Suarez, and A. Dı̀ez. 2004. Can we mitigate animalvehicle accidents using predictive models? Journal of Applied Ecology 41:701– 710. McDonald, T. L. 2013. The point process use-availability or presence-only likelihood and comments on analysis. Journal of Animal Ecology 82:1174– 1182. McShea, W. J., and H. B. Underwood. 1997. The science of overabundance. Deer ecology and population management. Smithsonian Institute Press, Washington, District of Coloumbia, USA. Meisingset, E. L., Ø. Brekkum, and L. E. Loe. 2010. Hjortens habitatbruk og atferd i relasjon til vei.—En analyse av påkjørsler og posisjonsdata fra hjort. Bioforsk Rapport 83:1–34. [In Norwegian.] Meisingset, E. L., L. E. Loe, Ø Brekkum, B. Van Moorter, and A. Mysterud. 2013. Red deer habitat selection and movements in relation to roads. Journal of Wildlife Management 77:181–191. Milner, J. M., C. Bonenfant, A. Mysterud, J. M. Gaillard, S. Csanyi, and N. C. Stenseth. 2006. Temporal and spatial development of red deer harvesting in Europe: biological and cultural factors. Journal of Applied Ecology 43:721–734. 10 Moen, A. 1998. Nasjonalatlas for Norge: Vegetasjon. Statens Kartverk, Hønefoss, Norway. [In Norwegian.] Mysterud, A. 2004. Temporal variation in the number of car-killed red deer Cervus elaphus in Norway. Wildlife Biology 10:203–211. Mysterud, A., L. E. Loe, B. Zimmermann, R. Bischof, V. Veiberg, and E. L. Meisingset. 2011. Partial migration in expanding red deer populations at northern latitudes—a role for density dependence? Oikos 120:1817–1825. Mysterud, A., E. L. Meisingset, V. Veiberg, R. Langvatn, E. J. Solberg, L. E. Loe, and N. C. Stenseth. 2007. Monitoring population size of red deer: an evaluation of two types of census data from Norway. Wildlife Biology 13:285–298. Mysterud, A., N. G. Yoccoz, N. C. Stenseth, and R. Langvatn. 2000. Relationships between sex ratio, climate and density in red deer: the importance of spatial scale. Journal of Animal Ecology 69:959–974. Neumann, W., G. Ericsson, H. Dettki, N. Bunnefeld, N. S. Keuler, D. P. Helmers, and V. C. Radeloff. 2012. Difference in spatiotemporal patterns of wildlife road-crossings and wildlife-vehicle collisions. Biological Conservation 145:70–78. Nielsen, C. K., R. G. Anderson, and M. D. Grund. 2003. Landscape influences on deer-vehicle accident areas in an urban environment. Journal of Wildlife Management 67:46–51. Olsson, M. 2007. The use of highway crossings to maintain landscape connectivity for moose and roe deer. Dissertation, Karlstad University Studies, Karlstad, Sweden. Pepin, D., N. Morellet, and M. Goulard. 2009. Seasonal and daily walking activity patterns of free-ranging adult red deer (Cervus elaphus) at the individual level. European Journal of Wildlife Research 55:479–486. Putman, R., M. Apollonio, and R. Andersen. 2011. Ungulate management in Europe: problems and practices. Cambridge University Press, Cambridge, United Kingdom. R Development Core Team. 2011. R: a language and environment for statistical computing. R Development Core Team, Vienna, Austria. Rea, R. V. 2003. Modifying roadside vegetation management practices to reduce vehicular collisions with moose Alces alces. Wildlife Biology 9:81– 91. Sappington, J. M., K. M. Longshore, and D. B. Thompson. 2007. Quantifying landscape ruggedness for animal habitat analysis: a case study using bighorn sheep in the Mojave Desert. Journal of Wildlife Management 71:1419–1426. Seiler, A. 2004. Trends and spatial patterns in ungulate-vehicle collisions in Sweden. Wildlife Biology 10:301–313. Seiler, A. 2005. Predicting locations of moose–vehicle collisions in Sweden. Journal of Applied Ecology 42:371–382. Skog, A., F. E. Zachos, E. K. Rueness, P. G. D. Feulner, A. Mysterud, R. Langvatn, R. Lorenzini, S. S. Hmwe, I. Lehoczky, G. Hartl, N. C. Stenseth, and K. S. Jakobsen. 2009. Phylogeography of red deer (Cervus elaphus) in Europe. Journal of Biogeography 36:66–77. Sudharsan, K., S. J. Riley, and S. R. Winterstein. 2006. Relationship of autumn hunting season to the frequency of deer-vehicle collisions in Michigan. Journal of Wildlife Management 70:1161–1164. Tappe, P. A., and D. I. M. Enderle. 2007. Using site-level factors to model areas at high risk of deer-vehicle collisions on Arkansas highways. Pages 489–499. in L. C. Irwin D. Nelson K. P. McDermott N. C. Raleigh, editors. Proceedsings of the 2007 international conference on ecology and transportation, center for transportation and the environment. North Carolina State University, Raleigh, USA. Venables, W. N., and B. D. Ripley. 2002. Modern applied statistics with S. Springer Verlag, New York, New York, USA. Associate Editor: David Euler. SUPPORTING INFORMATION Additional supporting information may be found in the online version of this article at the publisher’s web-site. The Journal of Wildlife Management 9999
© Copyright 2026 Paperzz