DISTRIBUTION AND HABITAT SELECTION PATTERNS OF MOUNTAIN SHEEP IN THE LARAMIE RANGE

DISTRIBUTION AND HABITAT SELECTION PATTERNS OF MOUNTAIN SHEEP IN THE LARAMIE RANGE Hall Sawyer 1, Ryan Nielson 1, and Martin Hicks 2 1 Western Ecosystems Technology, Inc. 2003 Central Ave. Cheyenne, WY 82001 2 Wyoming Game and Fish Department 5400 Bishop Blvd. Cheyenne, WY 82006 June 7, 2009 Suggested Citation: Sawyer, H., R. Nielson, and M. Hicks. 2009. Distribution and habitat selection patterns of mountain sheep in the Laramie Range. Western Ecosystems Technology, Inc. Cheyenne, Wyoming.

INTRODUCTION Considerable effort and resources have been dedicated to restoring mountain sheep (Ovis canadensis) to portions of their historic range. Restoration attempts generally involve translocation of mountain sheep from a source population, coupled with a combination of disease prevention, domestic sheep management, and habitat improvement. The success or failure of translocation efforts may be influenced by a number of factors, including habitat quality (Risenhoover et al. 1988, Cook et al. 1990, McKinney et al. 2003), contact with domestic sheep (Coggins and Matthews 1992, Martin et al. 1996, Singer et al. 2000a, Zeigenfuss et al. 2000), disease transmission (Risenhoover et al. 1988, Gross et al. 2000), the source stock (Douglas and Leslie 1999, Singer et al. 2000a), and cougar predation (Wehausen 1996, Hayes et al. 2000, Holl et al. 2004, McKinney et al. 2006). Singer et al. (2000a) estimated that only 41% of mountain sheep translocation efforts were considered successful. Given the difficulties with restoring mountain sheep populations, the persistence of their populations and associated management strategies have generated much debate (Berger 1999, Wehausen 1999, Ostermann-Kelm 2005, Turner et al. 2005). Nonetheless, most mountain sheep in desert or low-elevation environments occur in small (<100) populations in isolated habitats (Krausman and Leopold 1986, Etchberger et al. 1989, Bleich et al. 1990), where population persistence requires long-term management obligations that involve multiple translocations and/or efforts designed to create movement corridors and encourage dispersal. Typically, the monitoring of translocated animals is aimed at evaluating demographic characteristics (e.g., survival, lamb recruitment) and identifying habitat selection patterns. Ultimately, the success or failure of the translocation will be determined by the long-term population performance and demography. However, quantifying habitat preferences may assist managers with identifying management prescriptions or habitat manipulations (e.g., water development, prescribed burn) that are most likely to benefit the translocated population. Habitat selection is commonly evaluated using resource selection functions (RSF; Manly et al. 2002). In general, an RSF estimates the relative probability that a sample unit will be used by an animal, as a function of several habitat variables (e.g., elevation, slope, distance to water). RSFs are used to infer preference or avoidance of habitat variables and provide a basis for visualizing selection patterns through predictive maps. It is generally assumed that RSF coefficients are unbiased and accurately reflect the selection patterns of the sampled population. However, if RSFs are estimated from global positioning system (GPS) data with missing observations, it is likely that the coefficients will be biased if the missing observations are habitat-induced or non-random relative to the habitat characteristics (Frair et al. 2004, Nielson et al. 2009; e.g., Figure 1). In this situation, the successful GPS fixes do not accurately represent the full range of animal locations and are a biased set of sample units, which in turn produce biased coefficients and may ultimately lead to incorrect inference of preference or avoidance. Nielson et al. (2009) recently showed that data loss of only 10% (i.e., 90% GPS success) may lead to erroneous results if GPS fix success is strongly related to the habitat characteristics that influence selection. It is widely recognized that habitat features such as canopy cover and rugged terrain can reduce the fix success of GPS collars (Rempel et al. 1995, Moen et al. 1996, Hansen and Riggs 2006, Heard et al. 2008), yet the potential for missing observations to produce biased results is rarely considered (but see Rettie and McLoughlin 1999, D Eon 2003, Frair et al. 2004). Given that mountain sheep habitat is 1

characterized by rugged terrain (Krausman and Shackleton 2000), GPS studies of mountain sheep may be particularly vulnerable to habitat-induced data loss. Here, we examine habitat selection patterns from sample (n=23) of mountain sheep translocated from Montana into southeast Wyoming, and we implement a new modeling procedure developed by Nielson et al. (2009) to estimate habitat selection when GPS success is less than 100%. Figure 1. Illustration of habitat-induced bias where the fix success (locations 1, 2, 3...8) of a GPS collar is reduced when it is in Habitat B. Selection coefficients estimated from these data would be biased high for Habitats A and C, and biased low for Habitat B. Thus, the relative importance of Habitat B would be underestimated and the value of Habitats A and C would be overestimated. STUDY AREA AND BACKGROUND Our study area is located in the Laramie Peak mountain sheep data analysis unit (DAU) and Hunt Area 19 which is located in the central portion of the Laramie Range. Mountain sheep were historically abundant in Laramie Range, but were extirpated during European settlement. Translocation efforts aimed at restoring mountain sheep in the region began in 1964 and continued at a rate of 1 to 3 per decade through 1989 (Hurley 1996; Table 1). These efforts eventually resulted in several small populations in the Duck Creek drainage, the Laramie Peak Wildlife Habitat Management Area, Sybille Canyon, and the Johnson Creek WHMA. Several smaller, ram-dominated sub-populations were also established near: 1) Seller s Mountain, 2) along the North Laramie River and Bearhead, Sheep and Johnson Mountains, 3) Labonte Canyon, 4) upper Horseshoe Creek, 5) Laramie Peak, 6) Albany Peak and upper Cottonwood Creek, 7) and the area surrounding Marshall. Table 1. Transplant history of mountain sheep into the Laramie Range, 1964 2007. Year Number Release Location 1964 40 North Laramie River Canyon 1965 36 Labonte Canyon 1966 21 Labonte Canyon 1973 42 Duck Creek Canyon 1982 27 Marshall 1989 20 Marshall 2007 42 Hay and Duck Creek Canyon Total 228 2

After nearly two decades of stagnant growth and limited interchange among populations, a large translocation effort was implemented in 2007. The purpose of the translocation was to improve population persistence and interchange among sub-populations. The 2007 translocation followed a series of large natural fires that occurred in 2001 and 2002 that burned approximately 39,000 acres, including 4,000 acres in and around the Duck Creek Canyon population. These fires presumably improved mountain sheep habitat and migration corridors by opening areas that had previously been encroached by conifers. In January of 2007, 42 mountain sheep (31 ewes, 6 lambs, and 5 yearling rams) from the Paradise-Perma Herd in Montana, were transplanted into the Duck Creek Canyon area (Figure 2). Interestingly, this was the only transplant effort not to use sheep from the Whiskey Mountain Herd south of Dubois. Thirty (Table 2) of the 42 sheep were fitted with GPS radio collars to monitor subsequent distribution and habitat selection patterns. Figure 2. Location of Duck Creek Study Area in the Laramie Range, southeast Wyoming. 3

METHODS GPS Data Forty two mountain sheep were transplanted into the study area on January 30, 2007 and 30 were equipped with GPS radio-collars programmed to collect locations every 9 hours for approximately 20 months. The GPS collars were equipped with remote-release mechanisms programmed to release collars on August 24, 2008. Unfortunately, only 11 of the release mechanisms functioned properly and several collars had battery problems. Helicopter net-gunning was used to capture and recover collars from 13 animals, but 6 collars were never recovered. Of the 30 collars, 24 were eventually recovered and 23 (19 ewes, 4 rams) had recoverable data that resulted in 30,006 GPS locations collected between February 1, 2007 and October 13, 2008. Overall GPS fix success was 86%. Table 2. Animal ID, sex, age, collar status, survival, and number of GPS locations collected for each collared mountain sheep. Animal ID Sex Age Collar Retrieved? Survive study period? GPS locations 150.016 Ewe 1.5 Yes Yes 1,543 150.035 Ewe 4+ Yes Yes 1,584 150.056 Ewe 1.5 Yes Yes 1,527 150.076 Ewe 4+ No NA NA 150.096 Ram 1.5 Yes Yes 1,012 150.116 Ram 1.5 Yes Yes 1,532 150.135 Ewe 2.5 No NA NA 150.156* Ewe 3.5 Yes Yes 1,463 150.176 Ewe 4+ Yes Yes 1,380 150.196 Ewe 4+ Yes Yes 1,525 150.216 Ewe 3.5 Yes Died December 2007 607 150.236 Ewe 4+ No NA NA 150.256 Ram 1.5 Yes Yes 1,423 150.275 Ram 1.5 Yes Yes 1,487 150.296 Ewe 4+ Yes Yes 1,580 150.316 Ewe 1.5 Yes Yes 1,536 150.336 Ewe 2.5 No NA NA 150.356 Ewe 2.5 Yes Yes 1,380 150.376 Ewe 4.5 Yes Yes 1,404 150.396 Ewe 3.5 Yes Yes 1,445 150.416 Ewe 3.5 No NA NA 150.435 Ewe 3.5 Yes Yes none 150.456 Ewe 4+ Yes Yes 1,340 150.476 Ewe 4+ Yes Yes 893 150.496 Ewe 1.5 Yes Died February 2008 403? 150.517 Ewe 4+ Yes Yes 1,595 150.537 Ewe 4+ Yes Yes 1,420 150.556 Ewe 4+ Yes Yes 1,528 150.576 Ewe 4+ No NA NA 150.596 Ewe 4+ Yes Yes 1,550 * Indicates migratory mountain sheep 4

Distribution Patterns Similar to other polygynous ungulates, mountain sheep are sexually segregated during the non-breeding months (Geist 1971, Bleich et al.1997, Krausman and Shackleton 2000). Accordingly, we examined distribution and habitat selection patterns of ewes and rams separately. We evaluated the distribution patterns of mountain sheep using three general approaches. First, we mapped GPS locations of rams (n=5,454) and ewes (n=24,582). Second, we delineated core-use areas for rams (n=4) and ewes (n=19) by calculating a 95% kernel home range for all 30,006 GPS locations collected between February 1, 2007 and October 13, 2008. Lastly, we delineated seasonal ranges for rams and ewes and by calculating 95% kernel home ranges for GPS locations that occurred in winter (November 1 April 30), summer (June 1 September 30), and lambing (May 1 30). Lambing range was calculated only for ewes. Habitat Selection Patterns We estimated habitat selection for rams and ewes during three biologically meaningful seasons (DeCesare and Pletscher 2006), including winter (November 1 April 30), summer (June 1 September 30), and lambing (May 1 31). We excluded GPS data collected during the first 60 days of the transplant to allow animals a period of acclimation. Study areas for each sex and season were delineated by calculating a minimum convex polygon (MCP) from all GPS locations occurring in each designated period. Based on the literature, we identified five variables that likely influence the resource selection patterns of mountain sheep, including distance to escape terrain, distance to recently burned habitat, distance to timber, distance to perennial water, and elevation (Risenhoover and Bailey 1985, Krausman and Leopold 1986, Wakelyn 1987, Krausman et al. 1989, Etchberger and Krausman 1999, Dicus 2002, McKinney et al. 2003, Turner et al. 2004, DeCesare and Pletscher 2006). Elevation was calculated from a 30 x 30-m resolution digital elevation model (DEM; USGS 1999). We used the DEM to calculate distance to escape terrain by identifying 30-m cells with slopes > 27 (DeCesare and Pletscher 2006). We calculated distance to water from a perennial stream map developed by Wyoming Geographic Information Science Center (ftp://piney.wygisc.uwyo.edu/data/water/hydro100k). Recently burned areas were identified by from aerial photos (WGFD, unpublished data) and an updated version of the NW ReGAP project (Lennartz et al. 2007). We defined recently burned as 6-7 years old. We did not include vegetation type as a predictor variable because the best available data (i.e., NW ReGAP; Lennartz et al. 2007) did not accurately reflect conditions on the ground for all vegetation types (e.g., exposed rock was classified as grassland). However, as a group, conifer classification in the NW ReGAP appeared to match the ground conditions, so we used those data to calculate distance to timber in each cell. We used two types of modeling procedures to estimate RSFs, including: 1) the standard discrete choice (DC) model (McDonald et al. 2006), which ignores missing GPS locations, and 2) a modified discrete choice (MDC) model that attempts to account for missing GPS locations (Nielson et al. 2009). We estimated RSFs using both DC and MDC models so that we could verify whether or not missing GPS locations were associated with habitat features that mountain sheep were selecting for and biasing selection coefficients. For both the DC and MDC models we first estimated the relative probability of use for each 5

animal as a function of predictor variables. We then averaged the coefficients from models of individuals to develop a population-level model. Lastly, we mapped predictions of the population-level model. This approach treats the marked animal as the experimental unit, thereby eliminating 2 of the most common problems with resource selection analyses: pooling data across individuals and ignoring spatial or temporal correlation in animal locations (Thomas and Taylor 2006). An additional benefit of treating each animal as the experimental unit is that inter-animal variation can be examined (Thomas and Taylor 2006), while still providing population-level inference via averaging coefficients (Marzluff et al. 2004, Millspaugh et al. 2006, Sawyer et al. 2006). This approach which fits one model with all the variables to each animal, recognizing that the estimated coefficients may be treated as random variables and represent independent, replicated measures of resource use. This approach quantifies the resource selection of individuals and provides a valid method of assessing population-level use by averaging coefficients (Marzluff et al. 2004). Specifically, we estimated coefficients for the population-level model by averaging the coefficients of s individual animal RSFs, using n ˆ 1 β ˆ t = βts, (1) n s= 1 where ˆts β was the estimate of coefficient t (t =1, 2,, p) for individual s (s = 1, 2,, n). We estimated the variance of each population-level model coefficient using the variation between individual animals and the equation ( ˆ n 1 ) ( ˆ ˆ ) 2 var βt = βts βt (2) n 1 s= 1 We then calculated P-values for the coefficients in the population-level model to assess the consistency of selection at the population-level. Our sample size was the number of marked mountain sheep, not sampling units or GPS locations. This approach to calculating P-values or t-tests is considered conservative because the inter-animal variation is included in the calculation of the variance. Each seasonal MCP (e.g., summer ewes) was divided into 189 x 189 m grids in which values for each predictor variable were averaged. The average values within each 189 x 189 m grid were used as covariates in modeling. We standardized predictor variables by subtracting the mean and dividing by the standard deviation prior to analysis. Standardization can improve convergence of maximum likelihood estimation routines and allows the relative importance of variables to be assessed. For example, a statistically significant variable with a larger estimated coefficient explains more variation in the data and thus is a more important predictor for habitat selection by bighorn sheep. For the summer and lambing season models we used the following predictor variables for habitat selection: elevation (m), distance to escape terrain (m), distance to fire (m), distance to water (m), and distance to timber (m). As per Nielson et al. (2009) we included distance from the animal s previous location (km) as a variable in each model to help refine the available habitat units at each time step. We omitted distance to water from the winter models because mountain sheep can meet most of their water requirements from snow or ice during the winter (Krausman and Shackleton 2000). The primary difference between the DC and MDC models is that the MDC estimates both resource selection and detection probability, while the DC only estimates resource selection. In order to estimate detection probability, the MDC can include variables that potentially influence probability of detection (i.e., GPS fix 6

success). We considered elevation, distance to escape terrain, distance to fire, distance to timber, and terrain ruggedness as potential variables influencing probability of detection. Terrain ruggedness was calculated using the vector ruggedness measure (VRM) described by Sappington et al. (2007). Because we considered five different variables (one at a time) in the probability of detection component of the MDC, we had to fit 5 models to each of the 5 seasons: winter rams, winter ewes, summer rams, summer ewes, and lambing. We then used Akaike s information criterion (AIC; Burnham and Anderson 2002) to evaluate the support for each model. We summed the AIC values for each model across the individual animals to determine which had the lowest sum (AIC). We used the model with the lowest sum (AIC) to create the population-level model and predictive map for each season. We mapped predictions of population-level models for each winter on a 27 27 m grid that covered the seasonal study areas. We achieved this by centering a 189 x 189 m grid on each 27 x 27 m cell and calculating the average for each predictor variable within the larger grid. Importantly, we standardized predictor variables to ensure they mapped on the same scale as used for model estimation. The model prediction for each grid cell was then assigned a value of 1 to 5 based on the 20 th, 40 th, 60 th, and 80 th quantiles of the distribution of predictions for each map, and we classified areas as high use, medium-high use, medium use, medium-low use, or low use. All models and associated predictive maps were based on GPS data collected in 2008. To evaluate model performance we used GPS locations collected in 2007 to calculate a Spearman-rank correlation (r s ) that characterized the number of GPS locations that occurred in 10 equal-sized prediction bins based on each of the population-level models (Boyce et al. 2002, Sawyer et al. 2009). To illustrate the estimated effect of each habitat variable we created figures showing how relative probability of selection changed as a function each variable, while holding all other variables constant. We refer readers to Nielson et al. (2009) for further details of the MDC model for GPS bias. All statistical analyses were performed in the R language and environment for statistical computing (R Development Core Team 2008). 7

RESULTS Distribution Patterns Rams used a core-use area that covered approximately 17-mi 2 and contained the smaller 6-mi 2 core-use area used by ewes (Figures 3 and 4). Rams utilized habitat further north than ewes, including Split Rock, Pine Mountain, and Collins Peak. Within their core-use area, ram distribution shifted seasonally as they utilized the entire area during the winter, but occupied a smaller region in the north during the summer (Figure 5). In contrast, the seasonal ranges of ewes contained considerable overlap, with the exception of summer utilization of Pine Mountain, a small area of winter use along Cherry Creek and upper Duck Creek, and several small lambing areas located from Reese Mountain southeast to Moonshine Peak (Figure 6). Only one of the 23 collared sheep migrated outside of the study area. Ewe #156 migrated north out of the study area on March 21, 2007 and moved approximately 13 miles north to the North Laramie, near Indian Head and Chimney Rock (Figure 7). The migration route appeared to go through a recently burned area on Green Mountain (Figure 8). Figure 3. Distribution of 30,006 GPS locations collected from 4 rams and 18 ewes, February 2007 October 2008. 8

Figure 4. Core-use areas delineated for ewes and rams, February 2007 October 2008. Figure 5. Seasonal ranges (winter and summer) delineated for rams. 9

Figure 6. Seasonal ranges (winter, summer, and lambing) delineated for ewes. 10

Close-up of summer range (left) Figure7. Seasonal distribution and migration route of mountain sheep #156. 11

Figure 8. Seasonal distribution and migration route of mountain #156 relative recently burned areas. 12

Habitat Selection Patterns Summer Ewes Average GPS fix-success for ewes during summer was 84% (n=17, SE=0.02). Ewes selected for summer habitats close to burned areas, high elevation, close to escape terrain, close to water, and close to timber. The most important variables were distance to burned area, elevation, and distance to escape terrain. Selection coefficients estimated from the MDC model suggest that GPS fix success, or probability of detection, was influenced by distance to timber. Comparing coefficients between the DC (uncorrected model) and MDC (corrected model) suggest selection coefficients may be over or under-estimated if probability of detection is not accounted for (Table 3). Specifically, the importance of distance to escape terrain (-0.893 vs. -0.967) and distance to water (-1.037 vs. -0.927) were under and overestimated, respectively, when probability of detection was not considered. Table 3. Coefficients, standard errors (SE), and associated P-values estimated for discrete choice (DC) and modified discrete choice (MDC) resource selection models for ewes during the summer (June 1 September 30). RSF Variable Discrete Choice Modified Discrete Choice Coefficient SE P-value Coefficient* SE P-value Distance to burn -1.115 0.061 0.000-1.119 0.063 0.000 Elevation 1.066 0.040 0.000 0.979 0.048 0.000 Distance to escape -0.893 0.119 0.000-0.967 0.115 0.000 Distance to water -1.037 0.056 0.000-0.927 0.042 0.000 Distance to timber -0.077 0.027 0.013-0.120 0.023 0.000 P[Detection] Variable Intercept NA NA NA 3.202 0.697 0.000 Distance to timber NA NA NA 2.089 1.003 0.056 *Coefficients for MDC model are standardized and are listed in their order of importance. We found a strong correlation (rs= 0.99) between the predictions of the populationlevel model and ewe locations from 2007 (Figure 9), indicating that model may effectively predict mountain sheep distribution in years other than that for which the model was developed. The predictive map suggests that the majority of high quality summer ewe habitat occurred in the northern portion of the ewe study area (Figure 9). 13

Figure 9. Population-level model predictions and associated levels of mountain sheep use of ewes during the summer of 2008 and distribution of ewe locations (n=7,605) from the summer of 2007 that were used to evaluate model predictions. Winter Ewes Average GPS fix-success for ewes during winter was 81% (n=17, SE=0.01). Ewes selected for winter habitats close to escape terrain, high elevation, close to timber, and close to burned areas. The most important variables were distance to escape terrain and elevation. Selection coefficients estimated from the MDC model suggest that GPS fix success, or probability of detection, was influenced by distance to burned area. Comparing coefficients between the DC (uncorrected model) and MDC (corrected model) suggest selection coefficients may be over or under-estimated if probability of detection is not accounted for (Table 4). Specifically, the importance of distance to escape terrain (-1.936 vs. -1.984) and distance to burn (-0.309 vs. -0.355) were underestimated when probability of detection was not considered. 14

Table 4. Coefficients, standard errors (SE), and associated P-values estimated for discrete choice (DC) and modified discrete choice (MDC) resource selection models for ewes during the winter (November 1 April 30). RSF Variable Discrete Choice Modified Discrete Choice Coefficient SE P-value Coefficient SE P-value Distance to escape -1.936 0.143 0.000-1.984 0.166 0.000 Elevation 0.914 0.160 0.000 0.797 0.141 0.000 Distance to burn -0.309 0.065 0.000-0.355 0.042 0.000 Distance to timber -0.172 0.044 0.001-0.197 0.048 0.001 P[Detection] Variable Intercept NA NA NA 0.223 0.030 0.000 Distance to burn NA NA NA 0.353 0.064 0.000 *Coefficients for MDC model are standardized and are listed in their order of importance. We found a strong correlation (rs= 0.87) between the predictions of the populationlevel model and ewe locations from 2007 (Figure 10), indicating that model may effectively predict mountain sheep distribution in years other than that for which the model was developed. The predictive map suggests that the majority of high quality winter ewe habitat occurred in the northeast portion of the ewe study area (Figure 10). Figure 10. Population-level model predictions and associated levels of mountain sheep use of ewes during the winter of 2008 and distribution of ewe locations (n=2,334) from the winter of 2007 that were used to evaluate model predictions. 15

Lambing Ewes Average GPS fix-success for ewes during lambing was 86% (n=17, SE=0.01). Ewes selected for lambing habitats that were close to escape terrain, high elevation, close to burned areas, close to water, and close to timber. The most important variables were distance to escape terrain, elevation, and distance to burned area. Selection coefficients estimated from the MDC model suggest that GPS fix success, or probability of detection, was influenced by distance to burn. Comparing coefficients between the DC (uncorrected model) and MDC (corrected model) suggest selection coefficients may be over or under-estimated if probability of detection is not accounted for (Table 5). Specifically, the importance of distance to escape terrain (-1.583 vs. -1.666) and elevation (0.924 vs. 0.856) were under and over-estimated, respectively, when probability of detection was not considered. Table 5. Coefficients, standard errors (SE), and associated P-values estimated for discrete choice (DC) and modified discrete choice (MDC) resource selection models for ewes during the lambing period (May 1 31). RSF Variable Discrete Choice Modified Discrete Choice Coefficient SE P-value Coefficient SE P-value Distance to escape -1.583 0.299 0.000-1.666 0.312 0.000 Elevation 0.924 0.184 0.000 0.856 0.208 0.001 Distance to burn -0.665 0.216 0.008-0.661 0.160 0.001 Distance to water -0.446 0.216 0.059-0.427 0.212 0.063 Distance to timber -0.228 0.109 0.056-0.186 0.101 0.087 P[Detection] Variable Intercept NA NA NA 5.081 1.883 0.017 Distance to burn NA NA NA 3.679 1.988 0.086 *Coefficients for MDC model are standardized and are listed in their order of importance. We found a strong correlation (rs= 0.94) between the predictions of the populationlevel model and lambing locations from 2007 (Figure 11), indicating that model may effectively predict mountain sheep distribution in years other than that for which the model was developed. The predictive map suggests that the majority of high quality lambing habitat occurred in the northern portion of the lambing study area (Figure 11). 16

Figure 11. Population-level model predictions and associated levels of mountain sheep use of ewes during the lambing period of 2008 and distribution of ewe locations (n=2,334) from the lambing period in 2007 that were used to evaluate model predictions. Summer Rams Average GPS fix-success for rams during summer was 87% (n=3, SE=0.01). Rams selected for summer habitats with higher elevation, close to escape terrain, close to burned areas, and away from water. The most important variables were elevation, distance to escape terrain, and distance to burn. Selection coefficients estimated from the MDC model suggest that GPS fix success, or probability of detection, was moderately influenced by distance to burned area. Comparing coefficients between the DC (uncorrected model) and MDC (corrected model) suggest selection coefficients may be over or under-estimated if probability of detection is not accounted for (Table 6). Specifically, the importance of elevation (0.511 vs. 0.602) and distance to burn (-0.238 vs. -0.271) were underestimated when probability of detection was not considered. Interestingly, accounting for probability of detection reduced the SEs of elevation and distance to water such that estimates became statistically significant (P<0.10), yet the coefficient for probability of detection was not significant. This lack of statistical significance likely reflects the high variability among individual rams that could not be accounted for with the small sample size. 17

Table 6. Coefficients, standard errors (SE), and associated P-values estimated for discrete choice (DC) and modified discrete choice (MDC) resource selection models for rams during the summer (June 1 September 30). RSF Variable Discrete Choice Modified Discrete Choice Coefficient SE P-value Coefficient SE P-value Elevation 0.511 0.234 0.080 0.602 0.026 0.001 Distance to escape -0.301 0.085 0.035-0.308 0.080 0.031 Distance to burn -0.238 0.352 0.284-0.271 0.152 0.109 Distance to water 0.015 0.111 0.453 0.073 0.022 0.040 Distance to timber 0.157 0.306 0.329 0.065 0.075 0.239 P[Detection] Variable Intercept NA NA NA 12.487 9.023 0.150 Distance to burn NA NA NA 17.773 14.619 0.174 *Coefficients for MDC model are standardized and are listed in their order of importance. We found a strong correlation (rs= 0.87) between the predictions of the populationlevel model and ram locations from 2007 (Figure 12), indicating that model may effectively predict mountain sheep distribution in years other than that for which the model was developed. The predictive map suggests that the high quality summer ram habitat occurred throughout the ram study area (Figure 12). Figure 12. Population-level model predictions and associated levels of mountain sheep use of rams during the summer of 2008 and distribution of ram locations (n=1,792) from the summer of 2007 that were used to evaluate model predictions. 18

Winter Rams Average GPS fix-success for rams during winter was 86% (n=4, SE=0.03). Rams selected for winter habitats close to escape terrain, close to burned areas, high elevation, and away from timber. The most important variables were distance to escape terrain and distance to burn. Selection coefficients estimated from the modified discrete choice model suggest that GPS fix success, or probability of detection, was moderately influenced by distance to burned area. Comparing coefficients between the discrete choice (uncorrected model) and modified discrete choice (corrected model) suggest selection coefficients may be over or underestimated if probability of detection is not accounted for (Table 7). Specifically, the importance of distance to escape terrain (-1.243 vs. -1.316) and distance to burn (-0.648 vs. - 0.674) were underestimated when probability of detection was not considered. Interestingly, accounting for probability of detection reduced the SEs of distance to burn, elevation, and distance to timber such that estimates became statistically significant (P<0.10), yet the coefficient for probability of detection was not significant. This lack of statistical significance likely reflects the high variability among individual rams that could not be accounted for with the small sample size. Table 7. Coefficients, standard errors (SE), and associated P-values estimated for discrete choice (DC) and modified discrete choice (MDC) resource selection models for rams during the winter (November 1 April 30). RSF Variable Discrete Choice Modified Discrete Choice Coefficient SE P-value Coefficient SE P-value Distance to escape -1.243 0.188 0.004-1.316 0.184 0.003 Distance to burn -0.648 0.472 0.132-0.674 0.375 0.085 Elevation 0.288 0.157 0.083 0.299 0.102 0.030 Distance to timber 0.150 0.097 0.110 0.196 0.049 0.014 P[Detection] Variable Intercept NA NA NA 2.101 0.331 0.004 Distance to burn NA NA NA -0.061 0.311 0.428 *Coefficients for MDC model are standardized and are listed in their order of importance. We found a strong correlation (rs= 0.88) between the predictions of the populationlevel model and ram locations from 2007 (Figure 13), indicating that model may effectively predict mountain sheep distribution in years other than that for which the model was developed. The predictive map suggests that most high quality winter ram habitat occurred throughout the northern and central portions of the ram study area (Figure 13). 19

Figure 13. Population-level model predictions and associated levels of mountain sheep use of rams during the winter of 2008 and distribution of ram locations (n=749) from the winter of 2007 that were used to evaluate model predictions. Relative Probability of Selection Plots Plots illustrating how habitat variables influence the relative probability of selection for rams and ewes during the summer, winter, and lambing periods indicate differences in habitat selection between the sexes and across seasons (Figures 14-18). Figure 14 indicates that both rams and ewes rarely occupy habitats >200 m from escape terrain, with the exception of rams during the summer. Escape terrain appears most important during the lambing period. Figure 15 indicates relative probability of use steeply declines as distance from burned areas increases. With the exception of wintering ewes, sheep rarely used areas > 1.5 km from a burned area. Rams prefer to be closer to burned areas in the summer compared to the winter. Ewes demonstrate a similar preference for burned areas during summer and lambing periods, but may utilize areas further from burned areas during the winter. Figure 16 indicates a consistent decline in the relative probability of use as elevation decreases, except for rams during the summer. Sheep rarely used areas < 2000 m (6,560 ft.) in elevation. Figure 17 indicates that distance to water influences the relative probability of selection differently for ewes and rams; ewes typically occupy areas within 1 km of water, while rams prefer habitat further away from perennial water sources. Figure 18 indicates that distance to timber influences the relative probability of selection differently for ewes and rams; ewes typically 20

occupy areas within 500 m of timber, while rams prefer habitat > 500 m away from timber, particularly in the winter. Figure 14. Influence of distance to escape terrain on the relative probability of selection for ewes and rams during summer, winter, and lambing periods. Figure 15. Influence of distance to burned area on the relative probability of selection for ewes and rams during summer, winter, and lambing periods. 21

Figure 16. Influence of elevation on the relative probability of selection for ewes and rams during summer, winter, and lambing periods. Figure 17. Influence of distance to water on the relative probability of selection for ewes and rams during summer, winter, and lambing periods. 22

Figure 18. Influence of distance to timber on the relative probability of selection for ewes and rams during summer, winter, and lambing periods. DISCUSSION Mountain sheep require access to escape terrain and high quality forage in areas with unobstructed visibility (Risenhoover and Bailey 1985). Evaluating the habitat selection patterns of translocated animals may improve management efforts aimed at restoring mountain sheep to historic ranges by providing detailed information as to how animals utilize local habitat features and identifying potential ways to improve existing habitat conditions (e.g., prescribed burns, water development). The most common contemporary approach to evaluating habitat selection is to equip animals with GPS radio-collars, monitor them for a number of years, and then use resource selection functions (RSF) to assess habitat preference and avoidance. The RSF approach provides a rigorous and standard methodology by which habitat selection patterns can be evaluated (Manly et al. 2002). However, the RSF assumes that animals are detected in all the habitat types in which they occur. When the fix success of GPS collars is reduced because of canopy cover, dense vegetation, or rugged terrain, this assumption may be violated. This type of data loss is referred to as habitat-induced GPS bias (Hebblewhite et al. 2007, Nielson et al. 2009) because it may produce biased estimates of habitat use. Current GPS studies in Wyoming suggest that mountain sheep may be particularly vulnerable to habitat-induced GPS bias because of their propensity for rugged terrain. For example, the seasonal GPS fix success of mountain sheep in our study area ranged from 81 to 87%. This success rate was consistent with GPS collars tested from both Devil s Canyon (86%; A. Rutledge, Wyoming COOP Unit, pers. commun.) and Grand Teton National 23

Park (70-80%; A. Courtemanch, Wyoming COOP Unit, pers. commun.) populations. The level of potential bias in a resource selection model estimated from GPS data depends on both the proportion of missing locations (e.g., 10%, 20%, etc.) and the animal s selection of underrepresented habitats (Frair et al. 2004). If missing GPS locations occur randomly and are not influenced by habitat choices of the marked animal, then estimated RSF models should not be biased and correcting for missing locations unnecessary. In cases where the degree of GPS bias is unknown, the MDC modeling procedure of Nielson et al. (2009) allows researchers to assess whether GPS bias is a concern, and if so, provides unbiased estimates of selection coefficients. For our purposes, the MDC method was useful for identifying habitat variables that influenced GPS success and correcting selection coefficients accordingly. Given the small sample sizes associated with the ram models, we focus our discussion of detection probability on the more robust ewe models. RSFs estimated for ewes suggested that distance to timber and burned areas affected probability of detection during the summer and winter, respectively. In general, detection probability decreased as mountain sheep moved closer to timber and away from burned areas. Intuitively, GPS success is expected to decrease in timbered habitat or areas with canopy cover that limit satellite reception (Remple et al. 1995, Moen et al. 1996, D Eon et al. 2002, Heard et al. 2008). We assume that distance from burned areas affects probability of detection because mountain sheep are likely moving out of open areas and into habitats with more vertical structure (e.g., shrubs) that may obscure satellite reception. Had we not used the MDC model and corrected for GPS bias, our selection coefficients would have underestimated the importance of escape terrain and burned areas, and overestimated the importance of water and elevation. The most important habitat features for mountain sheep appeared to be escape terrain, elevation, and burned areas, although the relative importance of each varied across seasons and sexes. While there are not many management options available to improve escape terrain or elevation gradients, there is ample opportunity to modify the size, shape, and spatial arrangement of burned areas. A combination of prescribed and natural fires in the summers of 2001 and 2002 burned approximately 4,000 acres in and adjacent to the Duck Creek study area. These fires presumably improved mountain sheep habitat by clearing areas that were previously dominated by lodgepole pine (Pinus contorta), douglas fir (Pseudotsuga menziesii), and mixed-mountain shrub. Prescribed fire has often been used as a management tool to benefit mountain sheep and is known to improve diet quality and foraging efficiency (McWhirter et al. 1992). Hobbs and Spowart (1984) found that prescribed burning significantly improved the winter diets of mountain sheep and that the nutritional benefits of fire are more pronounced for mountain-shrub habitats compared to grasslands. Interestingly, our results suggest selection for burned areas was stronger during the summer and lambing periods than the winter. In addition to the nutritional benefits of fire (Hobbs and Spowart 1984, Seip and Bunnell 1985), the open habitat created by fires may establish movement corridors that encourage dispersal or migration to other available habitat (Risenhoover et al. 1988). Similar to other species that occupy fragmented landscapes and occur in disjunct populations, the ability of mountain sheep to disperse and colonize is key to their long-term persistence (Risenhoover et al. 1988, Bleich et al. 1990, Goodson et al. 1996, Singer et al. 2000b,c, DeCesare and Pletscher 2006). Anecodotal evidence in the Laramie Range suggests seasonal migrations of mountain sheep have traditionally been limited to small ram groups, but since the 2001-02 fires, observations of migrating ewes have increased (M. Hicks, WGFD, pers. commun.). Our data suggest that at least one GPS-collared ewe established a 24

seasonal migration route from Duck Creek to the North Laramie within the first year of translocation. This migration route appeared to be associated with burned areas and field observations suggest this ewe traveled with several other animals (M. Hicks, WGFD, pers. commun.). Wakelyn (1987) implicated fire suppression and subsequent conifer encroachment as a major factor in the decline of mountain sheep populations and the reduction of their occupied range. Although the benefits of fire to mountain sheep habitat have been well-documented (Hobbs and Spowart 1984, Seip and Bunnell 1985, Risenhoover et al. 1988, Etchberger et al. 1989, McWhirter et al. 1992, Smith et al. 1999, DeCesare and Pletscher 2006, Bleich et al. 2008), our understanding of the frequency of fires needed to maintain open visibility, quality forage, and movement corridors is poorly understood. Bleich et al. (2008) recently studied mountain sheep response to fire history in a chaparral ecosystem in California and found sheep preferred areas that had been burned within 15 years. Bleich et al. (2008) suggested that successful restoration of mountain sheep to their chaparral ecosystem may depend on more frequent fire intervals to ensure sheep have a continuous availability to recently burned areas. Our results suggest that burned areas were an important habitat component for translocated sheep in the Laramie Range, however preferential use of those areas may deteriorate over time, as shrub and conifer communities mature. Further monitoring of this herd may allow managers to establish a recommended fire interval that is appropriate for the persistence of mountain sheep and the mixed-mountain shrub habitats that they rely upon. MANAGEMENT IMPLICATIONS Despite recent advances with GPS telemetry and resource selection analyses, estimating resource selection when GPS fix success is less 100% remains problematic (Frair et al. 2004, Nielson et al. 2009). Resource selection models estimated for mountain sheep from GPS data should consider the proportion of missing locations and how they may affect habitat selection inferences. The modified discrete choice (MDC) model developed by Nielson et al. (2009) provides a method to verify how, or to what degree, missing GPS locations affect selection coefficients. GPS fix success for mountain sheep in the Laramie Range averaged 86% and estimates of habitat selection that did not account for missing locations tended to underestimate the importance of escape terrain and burned areas, while overestimating the importance of elevation and water. Probability of detection (GPS success) appeared to increase as sheep moved away from timber or closer to burned areas. Across the western U.S., fire suppression has reduced the amount of high-visibility habitat and quality of forage needed by mountain sheep (Wakelyn 1987, Etchberger et al. 1989, Bleich et al. 2008). Our results suggest that mountain sheep selected for areas that were burned in 2001 and 2002, particularly those at higher elevations (>2,000 m) and in close proximity ( 200 m) to escape terrain. Limited evidence also suggests that the fires may have created a movement corridor between Duck Creek and the North Laramie, and that several ewes have already established seasonal migrations between the two areas. Given that migration between disjunct subpopulations is critically important for the long-term persistence of translocated mountain sheep (Bleich et al. 1990), current management would benefit from additional monitoring to determine what proportion of the mountain sheep in the Duck Creek area establish seasonal migrations in the wake of the recent fires. 25

Our results were consistent with other studies that indicate mountain sheep are sexually segregated during the non-breeding months (Geist 1971, Bleich et al. 1997, Krausman and Shackleton 2000). Accordingly, management prescriptions aimed at improving mountain sheep habitat should consider the range and preferences of both sexes. ACKNOWLEDGEMENTS Governor s Big Game License Coalition WY FNAWS, Eastern FNAWS, Minnesota/Wisconsin FNAWS, National FNAWS, Iowa FNAWS Kennedy Family Turtleback Ranch Wynonna Johnson Bill Jones True Ranches Platte County Resource District Laird Flying Service Kevin Hurley Dooley Oil Montana Fish, Wildlife and Parks LITERATURE CITED Berger, J. 1999. Intervention and persistence in small populations of mountain sheep. Conservation Biology 13:432 435. Bleich, V. C., J. D. Wehausen, and S. A. Holl. 1990. Desert-dwelling mountain sheep: conservation implications of a naturally fragemented distribution. Conservation Biology 4:383 390. Bleich, V. C., R. T. Bowyer, and J. D. Wehausen. 1997. Sexual segregation in mountain sheep: resources or predation? Wildlife Monographs 134:1 50. Bleich, V. C., H. E. Johnson, S. A. Holl, L. Konde, S. G. Torres, and P. R. Krausman. 2008. Fire history in a chaparral ecosystem: implications for conservation of a native ungulate. Rangeland Ecology Management 61:571 579. Boyce, M. S., P. R. Vernier, S. E. Nielsen, F. K. A. Schmiegelow. 2002. Evaluating resource selection functions. Ecological Modelling 157:281 300. Burnham, K. P., and D. R. Anderson. 2002. Model selection and multimodel inference. Second edition. Springer Verlag, City, New York, USA. Coggins, V. L., and P. E. Matthews. 1992. Lamb survival and herd status of the Lostine Bighorn Herd following a Pasteurella die-off. Proceedings from Biennial Symposium of the Northern Wild Sheep and Goat Council 8:147 154. Cook, J. G., E. B. Arnett, L. L. Irwin, and F. G. Lindzey. 1990. Population dynamics of two transplanted mountain sheep herds in southcentral Wyoming. Proceedings from Biennial Symposium of the Northern Wild Sheep and Goat Council 7:19 30. DeCesare, N. J., and D. H. Pletscher. 2006. Movements, connectivity, and resource selection of Rocky Mountain mountain sheep. Journal of Mammalogy 87:531 538. 26

D Eon, R. G., R. Serrouya, G. Smith, and C. O. Kochanny. 2002. GPS radiotelemetry error and bias in mountainous terrain. Wildlife Society Bulletin 30:430 439. D Eon, R. G. 2003. Effects of stationary GPS fix-rate bias on habitat selection analyses. Journal of Wildlife Management 67:858 863. Dicus, G. 2002. GIS-based habitat models for mountain sheep winter range in Glacier National Park, Montana. Proceedings from Biennial Symposium of the Northern Wild Sheep and Goat Council 13:110 127. Douglas, C. L. and D. M. Leslie, Jr. 1999. Management of mountain sheep. Pages 238 262 in R.Valdez and P. R. Krausman, editors. Mountain sheep of North America. The University of Arizona Press, Tucson, Arizona, USA. Etchberger, R. C., P. R. Krausman, and R. Mazaika. 1989. Mountain sheep habitat characteristics in the Pusch Ridge Wilderness, Arizona. Journal of Wildlife Management 53:902 907. Etchberger, R. C., and P. R. Krausman. 1999. Frequency of birth and lambing sites of a small population of mountain sheep. Southwestern Naturalist 44:354 360. Frair, J. L., S. E. Nielsen, E. H. Merrill, S. R. Lele, M. S. Boyce, R. H. M. Munro, G. B. Stenhouse, and H. L. Beyer. 2004. Removing GPS collar bias in habitat selection studies. Journal of Applied Ecology 41:201 211. Geist, V. 1971. Mountain sheep: a study in behavior and evolution. University of Chicago Press, Chicago, IL, USA. Goodson, N. J., D. R. Stevens, and S. King. 1996. Establishment of altitudinal migration in a reintroduced mountain sheep population. Proceedings from Biennial Symposium of the Northern Wild Sheep and Goat Council 10:45 56. Gross, J. E., F. J. Singer, and M. E. Moses. 2000. Effects of disease, dispersal, and area on bighorn sheep restoration. Restoration Ecology 8:25 37. Hansen, M. C., and R. A. Riggs. 2006. Accuracy, precision, and observation rates of global positioning system telemetry collars. Journal of Wildlife Management 72:518 526. Hayes, C.L., E.S. Rubin, M.C. Jorgensen, and W.M. Boyce. 2000. Mountain lion predation of mountain sheep in the Peninsular Ranges, California. Journal of Wildlife Management 64:954-959. Heard, D. C., L. M. Ciarniello, and D. R. Seip. 2008. Grizzly bear behavior and global 4 positioning system collar fix rates. Journal of Wildlife Management 72:596 602. Hebblewhite, M., M. Percy, and E. H. Merrill. 2007. Are all global positioning system collars created equal? Correcting habitat-induced bias using three brands in the central Canadian Rockies. Journal of Wildlife Management 71:2026 2033. Hobbs, N. T., and R. A. Spowart. 1984. Effects of prescribed fire on nutrition of mountain sheep and mule deer during winter and spring. Journal of Wildlife Management 48:551 560. Holl, S. A., V. C. Bleich, and S. G. Torres. 2004. Population dynamics of bighorn sheep in the San Gabriel Mountains, California, 1967-2002. Wildlife Society Bulletin 32:412 426. Hurley, K. P. 1996. History of transplanting mountain goats and mountain sheep Wyoming. Proc. Biennial Symposium of the Northern Wild Sheep and Goat Council 10:205 209. Krausman, P. R., and B. D. Leopold. 1986. Habitat components for desert mountain sheep in Harquahala Mountains, Arizona. Journal of Wildlife Management 50:504 508. Krausman, P.R., G. Long, R.F. Seegmiller, and S.G. Torres. 1989. Relationships between desert mountain sheep and habitat in western Arizona. Wildlife Monograph 102:1 66. 27