4306.pdf ALCES VOL. 43, 2007 HURLEY ET AL. - SPATIAL ANALYSIS OF MVC 79 A SPATIAL ANALYSIS OF MOOSE-VEHICLE COLLISIONS IN MOUNT REVELSTOKE AND GLACIER NATIONAL PARKS, CANADA Michael V. Hurley, Eric K. Rapaport, and Chris J. Johnson University of Northern British Columbia, Natural Resources and Environmental Studies Graduate Program, 3333 University Way, Prince George, BC, Canada, V2N 4Z9 ABSTRACT: Moose (Alces alces)-vehicle collisions (MVC) can be costly ecologically by affecting population numbers, economically by vehicle damage, and socially through human injury or mortality. The purpose of this paper is to identify factors related to moose ecology, driver behaviour, and road design that are useful for predicting the spatial location of MVC on the Trans Canada Highway dis- models and used Akaike’s Information Criteria (AIC) to determine the most parsimonious model within Geographic Information System (GIS). The Receiver Operator Characteristic (ROC) discriminated subsets. A MVC probability map along the highway was created using the GIS model, providing a planning to reduce MVC risk within the parks should begin by assessing landscape-scale variables with emphasis on distance to wetland and landscape slope. This landscape-scale analysis should be predictors of moose tracks, game trails, and coniferous forest habitat. If highway planning cannot be effective in decreasing MVC, mitigation measures should include a public awareness program, speed reduction, and consideration of an alternative intercept foraging plan. ALCES VOL. 43: 79-100 (2007) Key words: driver visibility, evidence, GIS, habitat, highway design, moose, MVC, roadside vegeta- tion Wildlife-Vehicle Collisions (WVC) are a serious problem in North America (Bashore et al. 1985, Child et al. 1991, Del Frate and Spraker 1991, Oosenbrug et al. 1991, Romin Nearly 3,000 moose-vehicle collisions occur annually in North America (Child 1998) and 200 – 300 moose are killed on major Brit- ish Columbia highways each year (Child et conservative and do not take underreporting into consideration or the unknown number of mortalities on mining, logging, and rural roads. If the impacts of trains are included, this num- Columbia (Child et al. 1991). Collisions can be costly ecologically by affecting population numbers, economically by vehicle damage and lost hunting opportunities, as well as socially through human injury and mortality. for predicting areas of high MVC. Seiler (2005) stated that more detailed knowledge of occurrence of preferred moose forage (Ball and Dahlgren 2002, Seiler 2005), embank- ment of the road (Clevenger et al. 2003), and SPATIAL ANALYSIS OF MVC - HURLEY ET AL. ALCES VOL. 43, 2007 80 driver visibility (Bashore et al. 1985) would increase the predictive power of past model- ling attempts. Seiler (2005) noted how new - edge of the spatial distribution of collisions. Malo et al. (2004) suggest that WVC models should be used at both the landscape and local scales during the process of road design and implementation of mitigation measures. The purpose of this paper is to predict the spatial occurrence of moose-vehicle collisions (MVC) along the Trans Canada highway through Mount Revelstoke and Glacier Na- tional Parks along with the associated corre- lated factors. MVC rates along this stretch of 0.045 per kilometre per year (Sielecki 2004) for a total of 0.5 – 3 MVC per year within the parks. This MVC rate is relatively similar to outside of the park boundary; however, the reporting procedure within the park is more accurate for modelling purposes. The area is of high concern due to both the Trans Canada Highway and wildlife having limited movement options through narrow and high mountain passes. In addition, Parks Canada has a management objective to reduce the environmental impact of the transportation corridor, particularly on wildlife, vegetation, - tional Parks. To predict MVC and determine the re- lated process, models were developed using or using a Geographic Information System (GIS) (Finder et al. 1999, Malo et al. 2004, were included based on their contribution to ecological processes, moose biology, and driver attributes. The predictive capability of subsets. The model with the best predictive representative model to be compared among local-scale model subsets included highway design, moose evidence, roadside vegetation management, moose habitat, and driver vis- ibility. By predicting MVC locations, their reduction could be looked upon from a proac- tive perspective. By focusing on preventative measures as opposed to relying on mitigation measures, the implementation is not as costly, ecologically, economically, or socially. STUDY AREA The study site was restricted to the Trans Canada Highway dissecting Glacier and Mount Revelstoke National Parks within the Rocky Mountain Highway district in South-Eastern British Columbia, Canada (Fig. 1). Rugged, have resulted in limited transportation corridor operation of this segment of the Trans Canada Highway therefore faces numerous challenges weather, slope, rock instability, and collisions Parks Canada is responsible for the planning, Fig. 1. Regional setting of Glacier and Mount Revelstoke National Parks ALCES VOL. 43, 2007 HURLEY ET AL. - SPATIAL ANALYSIS OF MVC 81 construction, and operation of the highway within the National Park Boundaries. Glacier and Mount Revelstoke National Parks encompass 3 biogeoclimatic zones, the Interior Cedar Hemlock (ICH), Englemann Spruce-Subalpine Fir (ESSF), and the Alpine Tundra Zone (AT). The ICH is primarily comprised of old-growth cedar (Thuja plicata) and mountain hemlock (Tsuga mertensiana). In the ESSF, the lower subalpine forests are dominated by Englemann spruce (Picea en- gelmannii Abies lasiocarpa), and mountain hemlock. Mean annual pre- cipitation is 700-3,000 mm, most of which (70 – 80%) falls as snow (Meidinger and Polar 1991). METHODS Data Collection MVC data were contributed by Mount Revelstoke and Glacier National Parks (John Flaa, personal communication, Parks Canada). of each MVC was recorded by park wardens by either marking the collision on a map or by recording the collision location using a Global Positioning System (GPS) unit. An assump- tion was made that the reporting system has The primary reporting method transformation was from map marking to GPS use in the year 2000, representing 80% and 20% of MVC locations using each respective method. The UTM co-ordinates were recorded in a database along with date of kill, hour of kill, and infor- mation regarding the number and species of wildlife. The UTM coordinates for each MVC were plotted onto the highway layer within the study area using ArcGIS (ESRI 2005). The study encompassed a spatial analysis generated reference points so that logistic regression could be used to contrast high- way points with and without MVC (Fig. 2). Reference points were created by randomly generating numbers that represented distances along the highway. Road distances started at 0 km from the southern entrance of Mount the Northern entrance of Glacier National Park. Random reference points that shared the coordinates with a snow shed were not included. Changes in land cover due to natural or human disturbance over time were assessed using Parks Canada stand origin data. This - lations could be studied between independent variable data collected in one season with MVC data spanning nearly 4 decades. Both coniferous and deciduous cover has regener- ated since the right of way was cleared for on the assumption does not warrant concern after highway construction. Since highway natural disturbances have occurred within the 500 m highway buffer area since highway construction. Fig. 2. Topographical Slope Classes assessed at each point. The thick lines represent the high- way and the thin lines represent the adjacent SPATIAL ANALYSIS OF MVC - HURLEY ET AL. ALCES VOL. 43, 2007 82 Depending on accuracy and availability of spatial data, each variable was either mea- (landscape scale). We chose variables based - we used past studies and our knowledge of the study area to select potentially relevant Landscape-scale Variable Analysis (GIS) We used a GIS to measure 15 landscape- scale variables (500 m radius) (Table 1). All continuous variables were averaged within the 500 m radius buffer centered on each collision and random point. A 500 m buffer around each location represented the road-effect zone effect zone is the area that encompasses the majority of ecological effects resulting from road construction and use and is typically the focus of planning and mitigation (Forman 1999). A minimum of 500 m was kept be- tween random reference points upon creation in order to ensure independence. The 500 m radii represented the area over which collision attributes were sampled using a GIS at the landscape-scale. We used British Columbia Provincial Government Terrain Resource Information Management (TRIM) spatial data in GIS to represent highway segments, elevation, slope, and aspect. All TRIM data had a scale of 1:20,000 with a resolution of 25 m by 25 m cell size. Topographical criteria were included due to the inherent nature of moose migration from hills to valleys during the winter (Gun- dersen et al. 1998, Hundertmark 1998). Thus, measures of slope and aspect were included in an effort to gain insight into the effects of moose movement on MVC. The distance to water bodies and wetland were measured due to the fact that moose seek Variable Unit Aspect (GIS) Mean aspect within 500 m buffer degrees Built (GIS) Distance to the nearest human development m Crossroad m Elevation (GIS) Elevation above sea level generated using a digital elevation model m Forest Edge m Hiking (GIS) Distance to the nearest hiking trail m High Use Habitat (GIS) Area of high moose habitat within 500 m buffer as per Parks Canada data m2 Land Cover (GIS) Dominant land cover type within 500 m buffer Shrub/ Coniferous/ Lines (GIS) Distance to the nearest communication line m Rail (GIS) Distance to the nearest railway line m Risk Sign (GIS) Distance to nearest wildlife-risk sign m Slope (GIS) Mean slope within 500 m buffer degrees Water (GIS) Distance to the nearest water body boundary m Water Int m Wetland (GIS) Distance to the nearest wetland boundary m Table 1. Landscape-scale variables measured at each MVC site and reference point to model the fac- tors that determine moose-vehicle collision locations within Mount Revelstoke and Glacier National ALCES VOL. 43, 2007 HURLEY ET AL. - SPATIAL ANALYSIS OF MVC 83 (Peek 1998). The distance to water and wetland TRIM data. We measured the presence/ab- sence of high use habitat at each collision and reference point to determine the relationship of MVC with critical habitat range. The domi- nant land cover type was determined within a 500 m buffer to further assess habitat-related attributes and also potential effects on driver data within Mount Revelstoke and Glacier National Parks were based on Parks Canada scale of 1:50,000 (Achuff et al. 1984). We used GIS to record the distance from each MVC to rail lines, power lines, hiking trails, and built areas. The distance to rail and power lines was based on Parks Canada spa- tial data while the distance to built areas was Rail lines are plowed in the winter, providing a potential movement corridor. In addition, the vegetation clearance within rail line and power line corridors creates the potential for the presence of early seral forage. The human development affects the occurrence of MVC by means of habitat alteration, human activity, and potential predator avoidance (Malo et al. 2004, Seiler 2005). Hiking trails potential for increased movement, predation, and effect of human use on moose distribution. found moose to be more vulnerable to wolves at sites closer to trails and streams. distance of each MVC location from the near- est wildlife risk sign and highway curvature. We used the distance to wildlife risk sign criteria to assess the role of driver awareness on MVC. The distance to highway curvature was analyzed to assess driver visibility at a landscape-scale. Spatial representation of GIS model — Using landscape-scale GIS data, we developed a model with the structure: 0 1 1 k k) ————————————— 0 1 1 k k) where Y is the predicted probability of a MVC k k (Manly et al. 1993). The predictive MVC probability surface was created using Local-scale Variable Analysis From June to August 2005, we collected data for local-scale analyses. We used a GPS to locate each MVC and random reference sites scale variables (Table 2). Variables ranged from habitat related to driver and highway attributes, each contributing to one of the 5 local-scale model subsets. Habitat — At each site, habitat character- istics were measured using a variety of meth- Fig. 3. Probability surface showing the likelihood of MVC for Mount Revelstoke and Glacier National Parks using the GIS model. SPATIAL ANALYSIS OF MVC - HURLEY ET AL. ALCES VOL. 43, 2007 84 Variable Unit Ang 5 m Mean distance at which an observer standing 5 m from the pavement edge could no longer see passing vehicles taken from each direction on both sides of the highway m Ang 10 m Mean distance at which an observer standing 10 m from the pavement edge could no longer see passing vehicles taken from each direction on both sides of the highway m Browse Presence of browse within 100 m transect P/A Browse (Roadside) Presence of browse within 25 m transect P/A Corridor Width Width of highway corridor clearance including pavement m Dist Cover Mean distance to vegetative cover (trees and shrubs >1 m high) taken from both sides of the road m Ditch Presence of ditch adjacent highway P/A Ecotone Presence of an ecotone P/A Game Trail Absent/Low/High A/L/H Habitat Class (MF)/Coniferous Forest(CF)/Wetland(W)/Shrub(S) OFM/CF/W/S Inline Mean distance at which an observer standing at the pavement edge could no longer see passing vehicles taken from each direction on both sides of the highway m Jersey Barrier Presence of jersey barrier P/A Median Presence of median P/A Passing Lane Presence of a passing lane P/A Pellets Presence of pellets within 100 m transect P/A Pellets (Roadside) Presence of pellets within 25 m transect P/A Roadside Age Class Highest age of shrub within 25 m transect (1-3 yrs) (7-10 yrs) Roadside Vegetation Type of vegetation species within 25 m transect P/A Slope (0-5 m) Mean slope of the land 0-5 m perpendicular to the pavement edge taken from both sides of the road degrees Slope (5-10 m) Mean slope of the land 5-10 m perpendicular to the pavement edge taken from both sides of the road degrees Slope (10-30 m) Mean slope of the land 10-30 m perpendicular to the pavement edge taken from both sides of the road degrees Speed Mean recorded speed of passing vehicle km/h Topo Terrain slope category Tracks Presence of tracks within 100 m transect P/A Tracks (Roadside) Presence of moose tracks within 25 m transect P/A Table 2. Local-scale variables measured at each site and reference point to model the factors that de- termine moose-vehicle collision locations. ALCES VOL. 43, 2007 HURLEY ET AL. - SPATIAL ANALYSIS OF MVC 85 attractants of moose to highway corridors. We placed 25 m transects perpendicular to the highway and measured plant species presence and age at 5 m intervals within 4 m2 determine the most recent year of roadside clearing. The highest age of a shrub within the 25 m transect was used as an indicator of time since the roadside was cleared. We also recorded evidence of browsing, moose tracks, Some roadside vegetation species were grouped into families due to their low oc- currence. Western mountain ash (Sorbus scopulina) and saskatoon berry (Amelanchier alnifolia) were grouped into the rose family. Narrow-leaved hawkweed (Hieracium um- bellatum), common dandelion (Taraxacum ), pearly everlasting (Anaphalis margaritacea), yarrow (Achillea millefolium), Leucanthemum vulgare) were that were rarely present (1 – 3 occurrences) and could not be grouped into a family were roadside vegetation was modelled at the spe- cies level (Table 3). We placed a 100 m transect perpendicular and assess the roadway for presence of moose. Coniferous Forest (CF), Wetland (W), or Shrub (S). We recorded the dominant land cover class at 10 m intervals on the transect. Evidence of moose included wildlife trails, pellets, tracks, or browse. If the highway bisected two habitat types, this ecotone was noted. Ecotone was used as a variable to investigate any habitat edge effect that could potentially be correlated with MVC. The distance to the nearest forest edge perpendicular to the road was measured The distance to crossroads and water bodies intersecting the road were also measured in the same manner. The distance to crossroads was tested to determine whether intersections opportunity of a collision. Human and wildlife movement — We recorded a number of highway attributes that the ability of drivers to avoid a MVC. We used an inclinometer to measure the slope immedi- ate to the roadbed (0 – 5 m), the verge (5 – 10 m), and the adjacent land (10 – 30 m). We and topographic measurements tested whether embankments had positive or negative rela- tionships with moose-vehicle collisions. Driver visibility — Driver visibility was measured as the shortest distance to the point at which a car becomes out of sight of an observer from 3 different locations adjacent the highway. Field visibility variables mea- moose on the right-of-way. Since it could Species Modelling Name Common Horsetail (Equisetum arvense) HORSETAIL Grass GRASS Willow (Salix sp.) WILLOW Red-Osier Dogwood (Cornus stolonifera) DOGWOOD Sitka Alder (Alnus crispa) ALDER Western Red Cedar (Thuja plicata) CEDAR Spruce (Picea sp.) SPRUCE Thimbleberry THIMBLEBERRY Common Red Paintbrush (Castilleja miniata) PAINTBRUSH Black Twinberry (Lonicera involucrata) TWINBERRY Spreading Dogbane (Apocynum androsaemifolium) DOGBANE Lupine (Lupinus sp.) LUPINE Aspen (Populus tremuloides) ASPEN Table 3. Roadside vegetation species present within SPATIAL ANALYSIS OF MVC - HURLEY ET AL. ALCES VOL. 43, 2007 not be determined from what side or which direction a vehicle struck an animal, 4 vis- ibility measurements were taken at each site, 2 facing each direction, on each side of the highway. One in-line (from road edge) and 2 angular measurements were measured (5 m and 10 m from the road edge). Recognising that trucks were more visible at greater distances than cars or motorcycles, visibility distances were always measured using trucks. The mean distance to vegetative cover (trees and shrubs > 1 m high) was measured on both sides of the road to determine driver visibility. The cor- ridor width was the total area cleared for the highway including a combination of roadside clearance on both sides of the highway and the highway pavement width. The presence/absence of roadside ditches was re- and animal movement. The presence/absence of jersey barriers, passing lanes, and medians resulting from highway design and construc- tion. The average speed limit was read by means of a Bushnell Radar Gun. Highway speed was recorded as the mean of 20 vehicles (10 vehicles going in each direction). Actual vehicle speed was recorded as opposed to speed limit due to the inherent nature of vehicles not included in model development due to the absence of variability within the study area. All distances were measured using a range - ence/absence and continuous/discontinuous variables were estimated visually. Data Analysis Due to the binary nature of the dependent variable (0 = reference, 1 = collision), and the inclusion of categorical independent variables, the data were analyzed using bivariate logistic regression. The variables were grouped into and Akaike’s Information Criteria (AIC) was used to determine the most parsimonious model within each subset. The use of model selection criteria enabled inference to be drawn from several models simultaneously, so that a ‘best set’ of similarly supported models could be chosen (Johnson and Omland 2004). We - isolated, understood, and adapted to mitigation strategies. Five subsets modelled local-scale/ - ined GIS landscape-scale hypotheses. The that affected the driver visibility of moose. The second subset included the variables that indicated the evidence of moose in the terrain perpendicular to the highway. Highway design was assessed in the third subset. The fourth and age in order to relate MVC to roadside local-scale subset tested moose habitat features completed among the best AIC local-scale in order to identify the most parsimonious model overall. This round of AIC did not include the landscape-scale GIS models in its comparison due to the difference in scale relative to the 5 local-scale models. variables grouped into common hypothesized subsets, 2 combination models were developed to help further reveal the MVC phenomena. We recognise that these interaction models were not initial hypotheses, but arose as model subsets. Variables chosen for interac- tions included those that previously showed and Lemeshow 2000). To reduce multicollinearity among the modelled variables (Zar 1998), correlation screening was completed prior to model de- ALCES VOL. 43, 2007 HURLEY ET AL. - SPATIAL ANALYSIS OF MVC 87 which compared each variable combination, and removed those that were highly correlated (r > 0.75) (Seiler 2005). In the GIS model subset, the distance to communication lines was omitted from further analysis as it was highly correlated (Pearson correlation coef- showed a lower correlation with MVC points - posed to 0.49). In the driver visibility model subset, angular visibility 5 m was eliminated as it was highly correlated with inline vis- ibility. Angular visibility 5 m was chosen to be eliminated as opposed to inline as angular visibility 5 m is measured in between inline and angular visibility 10 m thus providing a larger range of measurements. In addition, inline is also taken from the road edge closer to where a collision occurs. Also in the driver visibility model subset and the highway design model subset, slope (5 – 10 m) was highly correlated with slope (0 – 5 m). To provide a greater range of slope measurements, slope (5 – 10 m) was eliminated as it is intermediate to the other two slope measurements (0 – 5 m and 10 – 30 m). contribution that a unit increase in the inde- pendent variable made to the outcome prob- of the individual independent variables. We variables on the collision probability. Each topographic and distance variable was modelled as a simple linear and then a for further model comparisons if the more relative to the simple linear form. For the GIS further modelling for the 3 topographic vari- ables of elevation, slope, and aspect. For the for inline visibility and angular visibility at 10 m were included for further modelling. We used the change in deviance to assess - amine high-leverage points which may have The 3 points with the highest leverage were investigated to determine the location in the parks, and the corresponding change in coef- both statistical and biological consideration, the points remained in the model as 95% of the cases were within +/- 2 (Menard 2001). Autocorrelation had to be corrected, as (Neilsen et al. 2002). Autocorrelation was assessed using PASSaGE by calculating the Moran’s I using the unstandardized model residuals and distance between points. Ro- bust standard errors were estimated using the Huber/White sandwich estimator in the program STATA (2002) to correct for auto- Huber/White sandwich estimator is robust decreased the potential for type I errors by levels (Lennon 2000). Model Validation The Receiver Operating Characteristic (ROC) was used to determine the degree of that is independent of probability cut-off levels (Boyce et al. 2002). ROC validation was de- veloped using independent data not included during model creation. Twenty percent of validation. To represent the variance associ- ated with the process of choosing validation data, we repeated the ROC procedure 5 times. Each iteration used a different set of randomly selected collision and reference points. This validation procedure was followed for each SPATIAL ANALYSIS OF MVC - HURLEY ET AL. ALCES VOL. 43, 2007 88 RESULTS AIC Model Comparison The use of AIC in model comparison showed selection uncertainty being within models in certain subsets meaning a small difference in performance. This development be interchanged as the model of choice were - eses often only differing in one variable to assess whether that certain factor is of critical importance to the susceptibility of MVC. Driver visibility model subset — Of the 10 driver visibility candidate models, the - vided support as the most parsimonious with an AICw included the variables of vehicle speed, cor- ridor width, and presence/absence of passing lanes. Adding variables of roadside slope or visibility distance to this model did not con- tribute to the AICw (AICw = 0.283 and 0.224, respectively). The AICw for the additional - the odds of MVC (Table 5). Corridor width Visibility models; MVC were more likely with increasing corridor widths. GIS model subset — The Topographic models within the GIS subset (AICw = 0.537) variables included slope, aspect, and eleva- tion while water bodies included lakes, rivers, and Wetland model hypothesis resulted in w (AICw = 0.299) while w w such as the Human Built model using variables of hiking trails, distance to rail, and distance to built area. - ence was found with MVC being correlated to to MVC in the GIS/Driver Visibility model but not the GIS model alone included eleva- tion and aspect. The distance to wetland had closer to wetland. The GIS model produced Roadside vegetation model subset — Of the Roadside Vegetation Models, the Forage Species hypothesis had the greatest AICw, al- though the weight was only 0.504, suggesting Hypothesis/Model Variables -2LL AIC AICw SPEED + PASSING LANE + CORRIDORWIDTH 4 132.7 140.71 0.44 Adjacent Roadside Slope and SLOPE(0-5M) + SLOPE(10-30 M) + PASSING LANE + SPEED + CORRIDORWIDTH 129.53 141.53 0.28 and Visibility INLINE + INLINE2 + ANG10 + ANG102 + PASSING LANE + SPEED + CORRIDORWIDTH 8 125.93 141.93 0.22 Table 4. Results of driver visibility AIC candidate model selection within Mount Revelstoke and Glacier National Parks. ALCES VOL. 43, 2007 HURLEY ET AL. - SPATIAL ANALYSIS OF MVC 89 considerable uncertainty in model selection (Table 8). Variables included in this model as reported in the literature. Shrub Age alone or when combined with Forage Species did model hypotheses were not included in Table w. One of based on non-forage species with an AICw of 0.041. Within the Roadside Vegetation model, the presence of grasses was positively correlated to MVC sites, while the presence of a kill (Table 9). Moose habitat model subset — The Land Cover Type hypothesized model was the most parsimonious of the Moose Habitat candidate models (AICw = 0.479) (Table 10). The addition of the distance to water inter- Variable S.E. (Robust) W P (Robust) Speed* 0.05 10.05 0.00 Corridor Width* 0.05 0.02 0.03 Passing -0.13 0.49 0.07 0.80 Constant 4.7 12.09 0 Table 5. Logistic regression analysis results for the best driver visibility AIC model. *P Hypothesis/Model Variables -2LL AIC AICw Water Bodies ELEVATION + ELEVATION2 + SLOPE + SLOPE2 + ASPECT ASPECT2 + WETLAND + WATER 9 44.08 0.54 Wetland WETLAND + ELEVATION + ELEVATION2 + SLOPE + SLOPE2 51.35 0.3 Moose Movement ELEVATION + ELEVATION2 + SLOPE +SLOPE2 + ASPECT + ASPECT2 + RAIL 9 Parks. Variable S.E. (Robust) W P (Robust) Wetland* -0.00 0.00 9.80 Slope* -1.05 0.30 7.44 Slope2* 0.02 0.01 0.00 Aspect2* 4 4 4.31 0.04 Elev2 5 3.309 0.058 Aspect -0.085 0.0479 Elev 0.055 0.034 2.475 0.112 Water 0.001 0.003 0.135 0.707 Constant 0.958 0.004 0.953 Table 7. Logistic regression analysis results for the best GIS AIC model. *P SPATIAL ANALYSIS OF MVC - HURLEY ET AL. ALCES VOL. 43, 2007 90 sections to this Land Cover model decreased the AICw (AICw = 0.441). The remainder of the candidate hypotheses all had AICw under alone resulted in an AICw of 0.004. Conifer- - ence on the odds of a MVC within the Moose Habitat model (Table 11). Moose evidence model subset — The AICw was 0.529 for the Trails and Transect Evidence hypothesized model, providing sup- port as the most parsimonious of the Moose Evidence candidate models (Table 12). This model included moose evidence within the 100 m transect as well as the presence/ab- sence of game trails. The candidate models with only Trails (AICw = 0) or only Transect evidence (AICw = 0.048) performed poorly on their own and were not included in Table 12. The inclusion of roadside tracks, browse, of the best model (AICw = 0.315) nor were the roadside variables effective predictors on their own (AICw = 0). Evidence of moose was positively correlated with MVC sites with the presence of tracks being the most important, followed by the presence of game trails (Table 13). This best AIC moose evidence model of Trails and Transect Evidence correctly clas- Highway design model subset — The comparison of the 9 Highway Design candi- date models resulted in the Highway Corridor (AICw = 0.553) (Table 14). The Full Model, which included the additional variable of dis- tance to crossroad, was no more parsimonious Hypothesis/Model Variables -2LL AIC AICw Forage Species WILLOW + DOGWOOD + ALDER + CEDAR + ASPEN + HORSETAIL + GRASS + SPRUCE + ROSE 10 139.09 159.09 0.5 Forage Species and Shrub Age ROADSIDE AGECLASS + WILLOW + DOGWOOD + ALDER + CEDAR + ASPEN + HORSETAIL + GRASS + SPRUCE + ROSE 11 138.48 0.25 Shrub Age ROADSIDE AGECLASS 2 157.4 0.17 Table 8. Results of roadside vegetation AIC candidate model selection within Mount Revelstoke and Glacier National Parks. Variable S.E. (Robust) W P (Robust) Grass* 1.08 0.52 5.19 0.04 Alder -0.98 0.52 4.05 Spruce 1.15 3.70 0.07 Horsetail -0.88 0.47 Dogwood 0.22 Willow 0.73 1.41 0.23 Rose -0.31 0.54 0.41 0.57 Cedar 1.03 0.37 0.59 Aspen -0.14 0.48 0.08 0.78 Constant -0.91 0.14 Table 9. Logistic regression analysis results for the best roadside vegetation AIC model. *P ALCES VOL. 43, 2007 HURLEY ET AL. - SPATIAL ANALYSIS OF MVC 91 Hypothesis/Model Variables -2LL AIC AICw Land Cover Type CF + OFM + WETLAND + SHRUB 5 141.53 151.53 0.48 and Land Cover Type OFM + CF + SHRUB + WETLAND + WATERINT 0.44 Full Model ECOTONE + FORESTEDGE + WATERINT + CF + OFM + WETLAND + SHRUB 8 Table 10. Results of moose habitat AIC candidate model selection within Mount Revelstoke and Glacier National Parks. Variable S.E. (Robust) W P (Robust) Coniferous forest* 0.04 0.01 0.00 Shrub -0.04 0.02 3.42 0.05 Wetland 0.04 0.03 2.71 0.18 0.00 0.01 0.05 0.80 Constant -0.97 0.81 1.22 0.23 Table 11. Logistic regression analysis results for the best moose habitat AIC model. *P Hypothesis/Model Variables -2LL AIC AICw Trails and Transect Evidence PELLETS 5 94.43 104.43 0.53 Full Model BROWSEROAD 7 91.4 105.42 0.32 Roadside Evidence and Transect Evidence 95.73 107.73 0.1 Table 12. Results of moose evidence AIC candidate model selection within Mount Revelstoke and Glacier National Parks. Variable S.E. (Robust) W P (Robust) Tracks* 1.89 0.00 Pellets 2.47 5.53 0.13 Trail Trail(high)* 1.33 2.04 0.02 Trail(low) 0.21 0.04 0.88 Browse 0.59 1.95 0.11 Constant -2.93 0.50 4.91 Table 13. Logistic regression analysis results for the best moose evidence AIC model. *P SPATIAL ANALYSIS OF MVC - HURLEY ET AL. ALCES VOL. 43, 2007 92 (AICw = 0.215). The hypothesis that variables associated with moose movement resulted in a model with a lower AICw (AICw = 0.144). The additional highway design hypotheses modelling smaller variable groupings were all under AICw of 0.1 and not included in the model under an AICw of 0.1 was using the variables of topographic class, slope, and pres- ence of ditches. Corridor width displayed a and the Highway Design models (Tables 5 and 15, respectively). In each model, MVC were more likely with increasing corridor widths. The Highway Design model showed the poor- est performance among the model subsets with Interaction models — combined GIS and driver visibility models to Table 14. Results of highway design AIC candidate model selection within Mount Revelstoke and Glacier National Parks. Hypothesis/Model Variables -2LL AIC AICw Highway Corridor Engineering TOPO + SLOPE(0-5 M) + SLOPE(10-30M) + MEDIAN + JERSEY + PASSING LANE + CORRIDORWIDTH + DITCH 9 130.89 148.89 0.58 Full Model TOPO + DITCH + SLOPE(0-5 M) + SLOPE(10- 30M) + MEDIAN + JERSEY + PASSING LANE + CROSSROAD + CORRIDORWIDTH 10 130.8 150.84 0.22 Moose Movement TOPO + SLOPE(0-5 M) + SLOPE(10- 30 M) + CROSSROAD + JERSEY + CORRIDORWIDTH + DITCH 8 135.71 151.71 0.14 Table 15. Logistic regression analysis results for the best highway design AIC model. *P Variable S.E. (Robust) W P (Robust) Corridor width* 0.03 10.05 0.03 Passing lane 0.75 0.49 2.07 0.13 Slope (0-5 m) -0.03 0.02 1.85 0.13 Median 1.02 1.50 Ditch -0.29 0.53 0.34 0.58 Jersey barrier 0.20 0.47 0.14 Slope (10-30m) 0.00 0.01 0.01 0.93 Topo 8.87 Topo(2b) 0.57 1.24 0.51 Topo(3a) -1.35 0.87 -1.59 0.12 Topo(3b) -1.57 0.90 -1.7 0.08 Topo(3c) 0.91 1.31 0.81 0.49 Topo(4) 0.35 0.91 0.41 0.70 Topo(5a) -0.42 0.89 -0.44 Topo(5b) -0.41 0.98 -0.44 -1.21 1.18 -1.03 0.30 Constant -2.24 1.37 -1.94 0.10 ALCES VOL. 43, 2007 HURLEY ET AL. - SPATIAL ANALYSIS OF MVC 93 corridor width, wetland-speed, and wetland- corridor width were included as interactions. In the GIS/Driver Visibility interaction model, with greater speeds. When GIS was combined with Driver Visibility, the interaction model was lower than GIS alone, yet still impressive, correctly classifying 92.4% of points. The second combination model included variables from the moose habitat and driver visibility models. Interaction terms consisted of co- niferous forest with both speed and highway corridor width (Table 17). No factors were Habitat/Driver Visibility interaction model. When Driver Visibility was combined with moose habitat, the interaction model had a higher ROC score than the Driver Visibility model alone, yet was still poor; only correctly ROC Validation - nation ability as reasonable and rates higher than 90% as very good discrimination because the sensitivity rate is high relative to the false positive rate. Using this 70% as a minimum threshold, the acceptable models after ROC validation in descending order include GIS, GIS + Driver Visibility, Moose Evidence, and Moose Habitat. Highway Design, Roadside Variable S.E. (Robust) W P (Robust) -0.02 0.01 0.02 Elev2 -5 -5 5.513 0.050 Aspect2 -4 -4 5.343 0.008 Elev 0.138 0.077 5.073 0.072 Slope2 0.023 0.012 4.098 0.053 Aspect -0.129 0.053 3.920 0.015 0.008 0.004 3.537 0.053 -5 2.825 -5 0.381 Water -0.002 0.003 0.127 Passing -0.047 0.002 Constant -27.371 0.345 *P Variable S.E. (Robust) W P (Robust) Wetland 0.044 0.028 3.395 0.124 Shrub 0.022 2.333 0.102 0.001 2.215 0.117 Passing -0.511 0.438 1.327 0.243 0.879 0.307 0.009 0.014 0.345 0.520 Constant -1.254 0.900 1.573 0.050 Table 17. Logistic regression analysis results for the best moose habitat/driver visibility interaction AIC model. *P SPATIAL ANALYSIS OF MVC - HURLEY ET AL. ALCES VOL. 43, 2007 94 Vegetation, Driver Visibility and the Moose Habitat/Driver Visibility models were below test among the best local-scale model from each of the 5 subsets strongly supported the Moose Evidence model as the most parsimoni- ous (AICw = 1.0), adding further support to its MVC predictive model. DISCUSSION Model Performance Although ROC scores for the GIS, GIS and Driver Visibility Interaction, Moose Evi- reasonably high discrimination, results should be interpreted with caution. As the study area is within a National Park, the land processes outside park boundaries, such as forestry, close to the park entrance over those near the centre of the park. As the study area is situ- ated within both a high mountain pass and a protected area, the transportation challenges model results should therefore not be directly to be used elsewhere, the structure could be - priately adapted to the location and species. We assumed that land-use remained constant over the reporting period, although minor changes were most likely inevitable despite the National Parks having been managed in a relatively constant ecological state. Additional caution should be used when interpreting these models as not all of the collisions that have occurred in the past were reported. The total number of collisions in- volving motor vehicles and large animals in Canada has generally been underestimated by of these reporting discrepancies include the unknown taking of carcasses before highway contractors are alerted, carcasses falling out of sight or animals moving away to die at unknown locations. In addition, drivers may report the collision to another jurisdiction or fail to report a minor collision, instead paying for the damages privately (Sielecki 2004). variation present among the models can be factors. The models were developed using the - sures previously shown to have successfully et al. 1999, Clevenger et al. 2003, Malo et al. 2004). Notwithstanding the inclusion of ad- be due to one simple factor such as weather, driver alertness, or moose behaviour. There is a possibility that the inclusion of Mount Revelstoke MVC in the overall model af- distance gap between parks. The two parks Model ROC Validation S.E. AICw GIS 0.035 n/a GIS + Driver Visibility 92.4% n/a Moose Evidence 1.0 Moose Habitat 70.2% 0.115 0 Moose Habitat + Driver Visibility 0.117 0 Driver Visibility 0.12 0 Roadside Vegetation 59.2% 0.123 0 Highway Design 0 Table 18. Model ROC Validation results on the best AIC model from each subset. ALCES VOL. 43, 2007 HURLEY ET AL. - SPATIAL ANALYSIS OF MVC 95 do, however, share ecosystem characteristics and are managed under one division of Parks Canada. In addition, the spatial error in re- porting MVC locations may have affected measurements based on the assumption that locations were accurate. Interpretation of Contributing Factors relationship with MVC in the Driver Visibility model. Higher speeds leading to a greater - vides support to the literature, although Seiler (2005) and Malo et al. (2004) modelled speed limit as opposed to actual radar speed. The to MVC in both the Driver Visibility model and the Highway Design model, although these 2 models were poor predictors overall. MVC sites were found at highway locations with greater corridor width than reference sites. Clevenger and Waltho (2000) found that wildlife use of highway passages was posi- tively correlated with road width. Improved visibility due to greater vegetation clearance may not have displayed importance as the bulk of accidents in the 2 parks occurred at correlation of MVC with distance to road curve, inline visibility, and angular visibility. to a decrease in vehicle speed while Joyce and Mahoney (2001) found more MVC at night due to increased moose activity. Furthermore, roadside brushing likely augments the risk of collision by maintaining early seral vegeta- tion, which attracts wildlife to the highway (Child et al. 1991, Rea 2003). Other studies have provided support for animals preferring to cross highways that are closer to vegeta- tion cover (Jaren et al. 1991, Clevenger et al. 2003, Malo et al. 2004, Seiler 2005). These contradicting theories of increased visibility and increased moose attraction may have led to the poor predictive abilities of the Driver Visibility and Highway Design models. The positive correlation between a wider highway corridor width and MVC may simply be a function of the highway being reduced to narrow widths along steeper sections. Both corridor width and speed no longer showed Seiler (2005) where the distance between correlated to MVC; however, if vehicle speed was weakened. Coniferous forest as a single variable in the Habitat model was, however, to be an important habitat type, with moose use ranging from 31 – 49% use per season in - cant contributor to the Habitat model. Perry forest to be of slightly less important moose addition, moose avoid wolves by spacing out and Pletscher 2000). model was observed in the slope variable which to MVC in previous studies (Gunson et al. et al. (2003) found that mammals were more likely to cross when the highway was level with the adjacent terrain. Where the two national parks are within the Selkirk Mountain range, valley corridors and limited gentle sloping landscapes. Snow accumulation is less in the valley bottoms, providing important ungulate habitat in the late autumn, winter, and early - ged mountain terrain forces both wildlife and human movement through the valley passes SPATIAL ANALYSIS OF MVC - HURLEY ET AL. ALCES VOL. 43, 2007 MVC within Mount Revelstoke and Glacier National Parks have occurred in winter months providing support for this theory. correlation to MVC within the GIS model whereas the distance to water did not. Moose (Peek 1998). The distance from water to MVC locations may not show a correlation simply due to the general fact that there are lakes and rivers dispersed throughout the parks and not in one particular area. The poor prediction ability of the road- side vegetation model may be attributed to a relatively homogeneous highway corridor throughout the 2 parks. Moose are browsing specialists with 90% of average diets being shrubs and trees (Perry 1999). Many of the preferred shrub species for moose were rela- tively common at both MVC and reference locations. The presence of grass was the Vegetation model and this may have been due to the overall scarcity of grasses in the steeper, higher elevation reference point lo- cations, instead being more prevalent within most likely did not contribute to the AICw in Roadside Vegetation candidate models due to the majority of roadside shrubs being the entire park. Moose tracks and high-use game trails 100 m transect perpendicular to the highway on can be a simple indicator of MVC locations. Roadside moose evidence was not included in w was not improved after inclusion in the full model or on its own. Roadside evidence may not have improved the model due to the presence of roadside browsing at the majority of both MVC locations (89%) and reference points (75%). Scale-dependent Factors step in assessing contributing variables within This landscape-scale/GIS approach shows - may have shown less predictive ability than the landscape-scale model, but were nevertheless and revealing factors important at both scales of analysis. For this reason, we created the GIS and Driver Visibility interaction model; although, the ROC score for this interaction model was no higher than that of the GIS model on its own. Although the Moose Habitat and Moose Evidence models suggested that habitat was a strong predictor of MVC, the distance to high use habitat and land cover variables in the GIS model subset were not present in the difference in predictability between the dif- ferent models seems to be a scale-dependant issue where local effects within 100 m such as forest type and moose evidence are more pro- direct habitat variables. Often, availability of the actual use of the habitat is restricted to one scale (Johnson et al. 2002). In addition, use habitat or land cover type might not have been selected for by moose and if so it may be so only at certain times of the year, thus introducing a temporal aspect to the model. Joyce and Mahoney (2001) suggest that MVC occur in areas of low and high moose density. 1937). Predictions from an anthropogenic concept states that animals have programmed ALCES VOL. 43, 2007 HURLEY ET AL. - SPATIAL ANALYSIS OF MVC 97 neurohormonal cues in how the environment is interpreted which can be species, gender, social, or season dependant (Bubenik 1998). The models were created using variables stemming from an anthropogenic perspective, however, human impressions on where moose should live do not ultimately determine where a moose will be. An opposite scale-related phenomenon may have occurred within the Highway Design model where the poor predictive ability may be attributed to the local-scale variables being overshadowed by landscape-scale factors. The class variable may not have been large enough topographic factors as seen in the GIS model. Linear landscape elements such as riparian corridors, ditches, steep slopes, and ridges may funnel animals alongside or across the roadway and thereby increase the risk of collisions (Malo et al. 2004, Seiler 2005). The importance of highway corridor width decreased when combined with landscape- scale factors of slope and wetland in the GIS/ Driver Visibility interaction model. The speed and slope interaction variable did, however, different scales were combined, suggesting MVC are correlated to locations with higher vehicle speeds and lower slope values. Management Implications GIS is a powerful tool in the initial where local-scale mitigation measures are needed. If the need for local-scale analysis should be modelled due to their reasonably high predictive abilities. Attention should be focused on highway segments close to wet- corridors, presence of coniferous forest, moose evidence, and at higher vehicle speeds. Improved road planning is the primary practice that should be regarded as the means to reduce the ecological effects that transport infrastructure impose. This study has helped observe some of the underlying processes that contribute to MVC within the parks. The Trans Canada Highway in Mount Revelstoke and Glacier National Parks, is a well established transportation route and mitigation measures will be the only option unless road altera- tion or new construction occurs. Although the processes within the predictive models are best suited for highway planning, the knowledge can be used as a basis for mitiga- tion decisions. An effective and acceptable countermeasure should reduce animal-vehicle interactions while still allowing for necessary animal behaviour and movements (Bashore et al. 1985). Suggested measures include reduc- tions in vehicle speed and intercept foraging. additional enforcement which can be costly. Intercept foraging involves the development of alternative feeding sites away from the transportation corridor (Schwartz and Bartley 1991). Wood and Wolfe (1988) determined that intercept foraging was an effective short- however, they cautioned that wildlife may become dependant on the supplemental food resulting in the attraction of additional wildlife. A fencing and wildlife underpass combination could be effective along the highway adjacent the Beaver River. Whenever possible, these a public awareness program such as the Wild- life Collision Prevention Program in British Columbia. Complete reliance should not be put into educational programs to enhance public awareness about WVC as their success has not yet proven effective (Romin and Bis- be a starting point. The models presented here may provide useful tools for road planners, but effective SPATIAL ANALYSIS OF MVC - HURLEY ET AL. ALCES VOL. 43, 2007 98 concrete approach that includes consideration of the landscape outside of park boundaries and more in-depth knowledge of the local work would be to investigate actual moose movement in the study area using telemetry data to map key crossing points. These data in combination with the collision points and modelling could provide invaluable informa- in the national parks. ACKNOWLEDGEMENTS This study could not have been conducted moral support of Maya Dougherty. Mount Revelstoke and Glacier senior park warden John Flaa went above and beyond his call of duty in providing spatial and collision data, not to mention the generous in kind support theoretical support by our companions Tony - improved the manuscript. GIS support from Scott Emmons and Ping Bai facilitated the analysis process. REFERENCES ACHUFF, P. L., W. D. HOLLAND, G. M. COEN, and K. VAN TIGHEM. 1984. Ecological Land Classification of Mount Revelstoke and Glacier National Parks, British Col- umbia. Volume 1 Integrated Resource Description. Alberta Institute of Pedology. Publication Number M-84-11. Edmonton, Alberta, Canada. BALL, J. P., and J. DAHLGREN. 2002. Brows- ing damage on pine (Pinus sylvestris and P. contorta) by a migrating moose (Alces alces) population in winter: Relation to habitat composition and road barriers. Scandinavian Journal of Forest Research 17:427-435. BASHORE, T. L., W. M. TZILKOWSKI, and E. D. BELLIS. 1985. Analysis of deer-vehicle collision sites in Pennsylvania. Journal of BIFULCO, R., and H. F. LADD Choice, Racial Segregation and Test-Score Gaps. Proceedings from the Annual Meet- ing of Allied Social Science Associations. Boston, Massachusetts, USA. BOYCE, M. S., P. R. VERNIER, S. E. NIELSEN, and F. K. A. SCHMIEGELOW. 2002. Evaluating resource selection functions. Ecological Modelling 157:281-300. BUBENIK, A. B. 1998. Behavior. Pages 173-221 in A. W. Franzmann and C. C. Schwartz, editors. Ecology and Man- agement of the North American Moose. Smithsonian Institution Press, Washing- ton, D.C., USA. CHILD, K. N. 1998. Incidental mortality. Pages 275–285 in A. W. Franzmann and C. C. Schwartz, editors. Ecology and Man- agement of the North American Moose. Smithsonian Institution Press, Washing- ton, D.C., USA. _____, S. P. BARRY, and D. A. AITKEN. 1991. Moose mortality on highways and railways in British Columbia. Alces 27:41-49. CLEVENGER, A. P., B. CHRUSZCZ, and K. E. GUNSON. 2003. Spatial patterns and fac- tors influencing small vertebrate fauna road-kill aggregations. Biological Con- servation 109: _____, and N. WALTHO. 2000. Factors in- fluencing the effectiveness of wildlife underpasses in Banff national park, Alberta, Canada. Conservation Biology DAMAS, and SMITH. 1982. Wildlife mortality in Transportation Corridors in Canada’s National Parks, Volume I and II. Parks Canada, Ottawa, Ontario, Canada. DEL FRATE, G. G., and T. H. SPRAKER. 1991. Moose vehicle interactions and an asso- ciated public awareness program on the ALCES VOL. 43, 2007 HURLEY ET AL. - SPATIAL ANALYSIS OF MVC 99 (ESRI) ENVIRONMENTAL SYSTEMS RESEARCH INSTITUTE. 2005. ArcMap GIS version 9.1. ESRI Inc., Redlands, California, USA. FINDER, R. A., J. L. ROSEBERRY, and A. WOOLF. 1999. Site and landscape conditions at white-tailed deer/vehicle collision loca- tions in Illinois. Landscape and Urban Planning 44:77-85. FORMAN, R. T. T. 1999. Horizontal processes, roads, suburbs, societal objectives, and landscape ecology. Pages 35-53 in J.M. - scape Ecological Analysis: Issues and Ap- plications. Springer-Verlag Incorporated, New York, New York, USA. _____, and L. E. ALEXANDER. 1998. Roads and their major ecological effects. An- nual Review of Ecology and Systematics 29:207-231. GUNDERSON, H., H. P. ANDREASSEN, and T. STORAAS. 1998. Spatial and temporal correlates to Norwegian moose-train col- lisions. Alces 34:385-394. GUNSON, K. E., B. CHRUSZCZ, and A. P. CLEV- ENGER - scape and highway influence ungulate vehicle collisions in the watersheds of the central Canadian Rocky Mountains: A fine-scale perspective? Proceedings from the International Conference on Ecology and Transportation 2005. San Diego, California, USA. HOSMER, D. W., and S. LEMESHOW. 2000. Ap- plied Logistic Regression. Second Edi- tion. John Wiley and Sons Incorporated, New York, New York, USA. HUBER, P. J likelihood estimates under nonstandard conditions. Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability 1:221–223. University of California Press, Berkeley, California, USA. HUNDERTMARK, K. J. 1998. Home range, dis- persal and migration. Pages 275-285 in A. W. Franzmann and C. C. Schwartz, editors. Ecology and Management of the North American Moose. Smithsonian Institution Press, Washington, D.C.,USA. JAREN, V., R. ANDERSEN, M. ULLEBERG, P. H. PEDERSEN, and B. WISETH. 1991. Moose- train collisions: the effects of vegetation removal with a cost-benefit analysis. Alces 27:93-99. JOHNSON, C. J., K. L. PARKER, D. C. HEARD, and M. P. GILLINGHAM. 2002. Movement parameters of ungulates and scale-specific responses to the environment. Journal of Animal Ecology 71:225–235. JOHNSON, J. B., and K. S. OMLAND. 2004. Model selection in ecology and evolu- tion. Trends in Ecology and Evolution 19:101-108. JOYCE, T. L., and S. P. MAHONEY. 2001. Spa- tial and temporal distribution of moose- vehicle collisions in Newfoundland. Wildlife Society Bulletin 29:281-291. KUNKEL, K. E., and D. H. PLETSCHER. 2000. Habitat factors affecting vulnerability of moose to predation by wolves in southeast- ern British Columbia. Canadian Journal of Zoology 78:150–157. LEBLANC, Y., F. BOLDUC, and D. MARTEL Upgrading a 144 km section of highway in prime moose habitat: where, why, and how to reduce moose-vehicle collisions. Proceedings from the International Con- ference on Ecology and Transportation 2005. San Diego, California, USA. LENNON, J. J. 1999. Resource selection func- tions: taking space seriously. Trends in Ecology and Evolution 14:399–400. MALO, J. E., F. SUAREZ, and A. DIEZ. 2004. Can we mitigate animal-vehicle accidents using predictive models? Journal of Ap- plied Ecology 41:701-710. MANLY, B. F. J., L. L. MCDONALD, and D. L. THOMAS. 1993. Resource Selection by Animals: Statistical Design and Analy- sis for Field Studies. Chapman & Hall, MEIDINGER, D., and J. POLAR. 1991. Special SPATIAL ANALYSIS OF MVC - HURLEY ET AL. ALCES VOL. 43, 2007 100 Report February 1991. Research Branch, Victoria, British Columbia, Canada. MENARD, S. 2001. Applied Logistic Regres- Applied Logistic Regres- sion Analysis. Sage Publishing, Thousand Oaks, California, USA. NEILSEN, S. E., M. S. BOYCE, G. B. STENHOUSE, and R. H. M. MUNRO. 2002. Modeling grizzly bear habitats in the Yellowhead ecosystem of Alberta: taking autocorrela- OOSENBRUG, S. M., E. W. MERCER, and S. H. FERGUSON. 1991. Moose-vehicle colli- sions in Newfoundland - Management considerations for the 1990’s. Alces 27:220-225. PEEK, J. M. 1998. Habitat relationships. Pages 275-285 in A. W. Franzmann and C. C. Schwartz, editors. Ecology and Management of the North American Moose. Smithsonian Institution Press, Washington, D.C., USA. PERRY, J., editor. 1999. Moose, Mule Deer Research Partnership. File Report 99-5. REA, R. V. 2003. Modifying roadside vege- tation management practices to reduce vehicular collisions with moose Alces alces. Wildlife Biology 9:81-91. ROMIN, L. A., and J. A. BISSONETTE Deer-vehicle collisions: status of state monitoring activities and mitigation measures. Wildlife Society Bulletin SCHWARTZ, C. C., and B. BARTLEY. 1991. Reducing incidental moose mortality: considerations for management. Alces 27:227-231. SEILER, A. 2005. Predicting locations of moose–vehicle collisions in Sweden. Jour- nal of Applied Ecology 42:371–382. SIELECKI, L. 2000. Wildlife Accident Re- porting System Annual Report. British Columbia Ministry of Transportation, Victoria, British Columbia, Canada. SWETS, J. A. 1988. Measuring the accu- racy of diagnostic systems. Science 240:1285–1293. TABACHNICK, S. C., and L. S. FIDELL. Using Multivariate Statistics, Third Edi- tion. Harper Collins College Publishers, New York, New York, USA. VON UEXKULL, J. 1921. Umwelt und Innen- welt der Tiere. Second edition. Springer Verlag, Berlin. _____. 1937. Umweltforschubg. Zeitschrift für Tierpsychologie. 1:33-34. WHITE, H. 1980. A heteroskedasticity- and a direct test for heteroskedasticity. Econometrica 48:817-830. WOOD, P., and M. L. WOLFE. 1988. Intercept feeding as a means of reducing deer- vehicle collisions. Wildlife Society Bul- WOODS, J. G., and R. H. MUNROE rails and the environment: wildlife at the intersection in Canada’s Western Moun- tains. Proceedings from the Transporta- tion Related Wildlife Mortality Seminar, Orlando, Florida. USA. ZAR, J. H. 1998. Biostatistical Analysis. International editions. Prentice-Hall, New Jersey, New York, USA.