Cover Image

Not for Sale



View/Hide Left Panel
Click for next page ( 58


The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement



Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.

OCR for page 57
57 case the likelihood of collisions for each road section would analyses220, and Microsoft Excel and ArcView GIS 3.377 for all show a Poisson distribution.32 For each of the four water- other analyses. sheds, the research team classed the highways into segments 100 m long and recorded presence (1) or absence (0) of the observed UVC points in each segment. A Kolmogorov- Findings and Results Smirnov one-sample test was used to determine whether the Summary of UngulateVehicle Collision Data empirical distribution differed from a Poisson distribution. Also a 2 test based on overall highway length was used to A total of 546 UVC observations were recorded between determine if an obvious UVC aggregation was significant August 1997 and November 2003 on all highways in the study along the cleared section or low valley bottom of Kootenay area. Deer (mule deer, white-tailed deer, and unidentified Highway 93 South. Finally, the research team determined the deer) were most frequently involved in collisions and com- aggregation of UVCs along each highway (i.e., whether kills posed 58% of the kills, followed by elk (27%), moose (7%) were evenly spread or clumped) by determining the percent- bighorn sheep (3%) and other ungulates, including moun- age of mile-markers associated with a UVC location. tain goats and unidentified species (5%). Univariate analyses were used to identify which of the The majority of UVCs occurred on the TCH east of Banff continuous variables (unpaired t-tests) and categorical vari- National Park in the province of Alberta (46%), followed by ables (2 contingency tests) differed significantly (P < 0.05) Highway 93 South in Kootenay National Park (22%), High- between high- and low-kill sites within the spatially accurate way 40 in Kananaskis Country (12%), the TCH in Yoho Na- and mile-marker datasets. The significance of each differen- tional Park (10%), and the TCH in Banff National Park tiated class within the categorical variables was evaluated (10%). Calculating the average number of kills per mile for using Bailey's confidence intervals.48 each highway in the study area showed that the majority of Logistic regression analyses were used to identify which of UVCs occurred on the TCH in the province of Alberta (13.6 the significant parameters best predicted the likelihood of UVC kills/mi), followed by the TCH in Banff National Park (2.6 occurrence within the spatially accurate and mile-marker kills/mi), the TCH in Yoho National Park (2.1 kills/mi), datasets.123 Stepwise (backward) regression procedures were Highway 40 in Kananaskis (2.1 kills/mi), and Highway 93 used to remove variables from the equation until each result- South in Kootenay National Park (1.8 kills/mi). These UVC ing model was not significantly more informative than the rates followed traffic volume trends, which were highest on previous one. The log-likelihood ratio test123 was used to de- the TCH east of Banff National Park in the province of Al- termine the ability of each model to discriminate between berta, followed by the TCH in Banff National Park, TCH in high- and low-kill zones based on location attributes. Signifi- Yoho National Park, Highway 40 in Kananaskis Country, and cance of explanatory variable coefficients was based on the 2 Highway 93 South in Kootenay National Park. of the Wald statistic.123 Standardized estimate coefficients were calculated by multiplying logistic regression coefficients (B) by Spatial Distribution of Roadkills the standard deviation of the respective variables. With this, the research team assessed the relative importance of the explana- The accuracy of the location where site-related variables were tory variables within the model. Odds ratios were examined to measured for the spatially accurate locations was approximately assess the contribution that a unit increase in the predictor less than or equal to 10 m. The UVC distributions from the spa- variable made to the probability of a UVC occurring.229 tially accurate dataset differed significantly from random distri- Hosmer-Lemeshow goodness-of-fit test statistics were in- butions along all five highways in the study area (Kolmogorov- cluded to see how well the model predicted the dependent vari- Smirnov one-sample test: TCHBow River Valley, d = 0.715; able. To validate the high-resolution and mile-marker models, Highway 93 South in Kootenay, d = 0.940; TCHYoho, d = the research team generated cross-validation classification ac- 0.892; Highway 40 in Kananaskis, d = 0.874; all P < 0.01). The curacies for each model. Both the high- and low-resolution distribution of UVCs on Highway 93 South in Kootenay models were validated with a random subset of 20% of the data showed a significant aggregated distribution where the highway not included in their development. traversed the low valley bottom with 60% of the kills occurring Prior to performing the regression analysis, the research along a 24 km (23%) stretch of road (2 = 63.9, P < 0.0001). The team tested potential explanatory variables for multicollinear- TCH in Alberta had the majority of mile markers associated ity.167 Where variables correlated (r > 0.7), the research team with a roadkill (89%), followed by the TCH in Banff National removed one of the two variables from the analysis. Final Park (86%), followed by Highway 40 (84%), followed by High- models and variable coefficients with a P-value less than or way 93 South in Kootenay National Park (61%), and the TCH equal to 0.1 were considered significant. The research team in Yoho National Park (57%). Because of the non-random pat- used the SPSS statistical package version 11.0 for all statistical tern and aggregation of UVCs, the research team explored

OCR for page 57
58 which landscape and road-related factors may be contributing Within the GIS-derived variables, areas of open water to the distribution of collisions in the study area. showed a significant negative correlation to the dependent variable in the spatially accurate dataset, while only the meas- ure of barrier length gave a significant negative correlation in Models both datasets. Less open water and shorter lengths of barriers Univariate tests. Table 24 shows the results of the uni- were associated with high-kill zones. variate tests comparing high- and low-kill locations for each To reduce intercorrelation between the variables,252 the re- environmental variable contributing to the probability of search team omitted the percentage of forest cover from fur- UVCs in each dataset. Each dataset had variables in each ther analyses because it was highly correlated (r > 0.70) with group that were significant in detecting differences between percentage of cleared ground. UVC high- and low-kill zones, however only three of ten vari- ables were significant in the mile-marker dataset. Logistic regression analysis. Both models ranked differ- Within the spatially accurate dataset, Table 24 shows that ently in their ability to predict the observed likelihood for six of the field-based variables were significant: habitat class, UVCs (Table 25). The variables used in each model could col- topography, forest cover, openness, adjacent land slope, and lectively be used to predict where a UVC would occur for the road width. Only two of the field variables (road width and spatially accurate model (P < 0.0001) but not for the mile- topography) were significant from the mile-marker dataset. marker model (P = 0.584) as determined from the log likeli- In both datasets, more UVCs occurred when the topography hood ratio test. For the spatially accurate model, the Hosmer- was flat and the roads were wide. In the spatially accurate Lemeshow statistic was higher than the mile-marker model. dataset, more UVCs occurred than expected in open forest The predictive capabilities of the GPS model correctly classi- habitat and fewer UVCs occurred than expected in conifer- fied 81.8%, while the mile-marker model correctly classified ous forest and rocky areas. only 64.4% of the selected UVC data. Model validation accu- Within the landscape feature variables, distance to drainage racies were 76.9% for the GPS model and 63.3% for the mile- and barrier-guardrail were significant (negatively correlated) marker model. Type of habitat was the most important vari- in the spatially accurate dataset. More UVCs occurred than ex- able in explaining UVCs in the GPS dataset. Ungulatevehicle pected closer to drainages perpendicular to the roadway and collisions were less likely to occur near open water, deciduous closer to barriers-guardrails (including Jersey barriers). No forest, closed coniferous forest, and open forest mix relative to distance-to-landscape features were significantly correlated to open habitat. Kills were 2.7 times less likely to occur in water- the high- and low-kill zones in the mile-marker dataset. dominated habitats (lakes, wetlands) relative to open habitat Table 24. Univariate comparison of factors contributing to UVCs. Spatially Accurate Mile-Marker Variable High Low P-value High Low P-value Field variables Habitat <0.0001 Rock 2 11 Coniferous forest 144 177 Open forest mix 112 54 Topography <0.0001 0.0035 Flat 241 172 24 12 Buried-raised 32 71 Forest cover 46.7 53.3 0.0256 Openness 47.3 41.6 0.0496 Adjacent land slope 11.4 15.9 0.0059 Road width 34.1 24.8 0.0001 19.51 15.2 0.0300 Distance-to-landscape features Drainages 2389.9 3068.9 0.0003 Barrier-guardrail 627.0 1052.2 0.0003 GIS-generated buffer variables Barrier length 272.7 353.2 0.0182 336.51 548.4 0.0036 Open water 49.2 109.8 0.0001 This table shows a comparison using a spatially accurate dataset (n = 499; 391 high- and 108 low-density points) and mile-marker dataset (n = 120; 63 high- and 57 low-density points). Mean values are shown for quantitative variables, and frequencies for each differentiated type are shown for categorical variables, along with their associated P-values. Only those values that were significant at P < 0.05 are displayed.