Black-tailed Prairie Dog Habitat Suitability Modeling for the Southern Great Plains: Cross-scale Analysis of Soils, Topography and Climate David J. Augustine, Research Ecologist, USDA-Agricultural Research Service, 1701 Centre Ave, Fort Collins, CO 80526; David.Augustine@ars.usda.gov Willam E. Armstrong, GIS Specialist, USDA-Agricultural Research Service, 1701 Centre Ave, Fort Collins, CO 80526; Billy.Armstrong@ars.usda.gov Jack F. Cully, Assoc. Professor of Biology and Assistant Wildlife Unit Leader, Kansas Coop. Fish and Wildlife Research Unit, 204 Leasure Hall, KSU, Manhattan, KS 66506, bcully@ksu.edu Michael F. Antolin, Professor, Department of Biology, Colorado State University, Fort Collins, CO 80523-1878; Michael.antolin@colostate.edu 1 ABSTRACT We developed multi-scale habitat suitability models for black-tailed prairie dogs (BTPD) in the southwestern Great Plains, corresponding to the western region of the Great Plains LCC. We used long-term (10-yr), high-resolution datasets on BTPD colony boundary locations collected at 7 study areas distributed across the region to develop resource selection functions based on colony locations and expansion patterns. Models are based on (1) soil maps and associated Ecological Sites (NRCS SSURGO database), (2) a topographic wetness index based upon water runoff and solar insolation patterns (TWIsi) that tests a priori hypotheses for topographic controls on BTPD, and (3) broad climatic gradients in temperature and mean annual precipitation. We show that BTPD habitat suitability is positively associated with soil organic matter, pH, clay content and depth to a restricted layer as well as TWIsi. BTPD habitat suitability is negatively associated with slope and soil sand content. The negative influence of slope is stronger on soils with high organic matter content. The positive influence of TWIsi is greater for soils with low sand content. Habitat suitability is positively associated with soil clay for areas with mean annual precipitation of 400 – 500 mm, but where mean annual precipitation declines to 350 mm, habitat suitability becomes negatively associated with soil clay content. Resulting models and map products provide a basis for land managers to compare and prioritize areas of conservation importance for BTPD and evaluate habitat for a suite of associated species of concern at scales from pastures to broad landscapes. We also provide the first assessment BTPD habitat suitability relative to Ecological Site Descriptions, which is essential for incorporating BTPD into associated state and transition models being developed and used by NRCS, USFS and BLM. We present the relative value of different Ecological Sites for BTPD in each of 3 regions based on Major Land and Resource Areas (MLRAs): MLRAs 67B/69 (eastern CO), MLRAs 72/77A (southwestern KS), and MLRAs 77A/B (northeast NM; OK and TX panhandles). Models and maps have immediate utility for land managers in the GPLCC and provide a tool for evaluation of plague mitigation strategies and future BTPD and plague management in response to climate change. 2 INTRODUCTION Because black-tailed prairie dogs (BTPD) function as ecosystem engineers and keystone species in Great Plains grasslands, their conservation and management lies at the core of many conservation efforts in the region. BTPD management is challenging and controversial because they may compete with livestock (Derner et al. 2006) and are severely affected by epizootic plague outbreaks caused by the bacterium Yersinia pestis (Cully et al. 2010). Furthermore, large BTPD colony complexes are needed to achieve conservation goals for multiple associated species including black-footed ferrets (Mustela nigripes; Roelle et al. 2005), mountain plovers (Charadrius montanus; Dinsmore et al. 2010, Augustine 2011) and burrowing owls (Tipton et al. 2009). Management such as dusting with insecticides to control plague transmission, poisoning to control prairie dog populations, and translocations to establish new populations are expensive (Andelt 2006), emphasizing the need to ensure they are applied in a spatially optimized manner to provide multiple ecosystem goods and services. Black-tailed prairie dogs are broadly distributed in central North America, and hence adapted to range of temperature and precipitation regimes and plant communities. Although many social and economic factors influence where BTPD complexes can be conserved or expanded, a suite of critical abiotic and biotic factors also controls BTPD habitat suitability. In particular, climate, soils, topography and vegetation structure vary widely across the GPLCC and directly influence BTPD persistence and expansion. At the eastern edge of their range, BTPD can be limited by tall vegetation and increased predation risk, while forage and water limitations may be constraining in the western portion of their range (Koford 1958, Hoogland 1995). The influence of precipitation regimes (long-term mean precipitation; seasonal and interannual variability) on BTPD colony expansion rates has direct relevance to contemporary management and long-term conservation planning in the face of climate change, but has never been systematically assessed. The ability to evaluate and map BTPD habitat within the GPLCC planning area would provide a valuable tool for optimizing use of scarce BTPD conservation funds. Research on landscape-scale patterns and controls of plague in BTPD complexes over the past 15 year (Cully et al. 2006, Antolin et al. 2006, Cully et al. 2010) has highlighted the need for an empiricallybased, landscape-scale habitat suitability model to assist in evaluating plague mitigation strategies and understanding BTPD and plague responses to climate change. Such an effort would also improve our understanding of local versus large-scale constraints on BTPD distribution and abundance. Past efforts to model BTPD habitat suitability focused on the northern Great Plains, and lacked high-resolution data on BTPD colony locations and expansion rates (Proctor et al. 2006). Belak (2001) examined high-resolution BTPD colony maps for two sites in South Dakota, but did not assess the influence of climate. Over the past 15 years, research on BTPD ecology (Stapp et al. 2004, Antolin et al. 2006, Augustine et al. 2008, Cully et al. 2010) and USFS monitoring have generated high-resolution, long-term datasets on BTPD colony boundaries for National Grasslands encompassing more than 1 million acres of the GPLCC. These data provided a unique opportunity to develop and test a quantitative habitat suitability model for BTPD because (1) measurements were repeated annually, thus quantifying colony expansion pattterns during plague-free periods, (2) sites are distributed across a broad north-south temperature gradient and east-west precipitation gradient (Fig. 1), and (3) analysis of a subset of these data demonstrated that 10-12 years of measurements provides a substantially different perspective on BTPD distribution than short-term (1-3 year) surveys (Augustine et al. 2008). 3 We developed and tested quantitative BTPD habitat suitability models for the southwestern Great Plains that examined the influence of climate, soils, and topography. We evaluated soils in two ways. First, we examined resource selection functions (RSFs) based upon quantitative measures of soil texture, organic matter content, pH and depth to a restricted layer. Second, we examined RSFs that aggregated these soil attributes at the level of Ecological Sites recently developed by the Natural Resource Conservation Service (NRCS). We controlled for the influence of land use by focusing on National Grasslands consistently managed with moderate cattle stocking rates, and thereby independently evaluated the influence of climate, soil and topography on BTPD habitat. Thus, our models do not incorporate the influence of land use, but will be essential for future incorporation of land use effects into modeling and conservation planning efforts. METHODS Study Area We studied BTPD habitat suitability in the western portion of the Great Plains LCC (Figure 1). Our analyses focused on BTPD colonies occurring with 7 study sites consisting of National Grasslands or geographically distinct sub-units of National Grasslands in Colorado, Kansas, Oklahoma, New Mexico, and Texas. The Kiowa and Rita Blanca NGs encompass a precipitation gradient, with colonies in New Mexico (Kiowa) exhibiting different plague epizootic patterns than colonies in Texas and Oklahoma (Rita Blanca, Cully et al. 2010); these Grasslands were therefore treated as separate sites. Similarly, the eastern and western units of the Pawnee National Grassland encompass a precipitation gradient and colonies in the two units exhibit different plague epizootic patterns (Stapp et al. 2004), hence were treated as separate sites. Within the administrative boundaries of each study site (Figure 1, blue boundaries), land ownership consists of a mosaic of private, state and federal lands. Analyses of prairie dog colony locations and surrounding locations lacking prairie dogs were based on data from ~900,000 ha (2.1 million acres) of federal land occurring within the administrative boundaries of the 7 study areas. Figure 1. Locations of study sites in the southern Great Plains with long-term, high-resolution datasets on blacktailed prairie dog colonies. 4 We extrapolated habitat suitability models derived from the 7 study sites to the western portion of the Great Plains LCC consisting of those counties within the shortgrass steppe ecoregion (see Lauenroth et al. 1999) that encompassed the precipitation and temperature gradients represented by our study sites (Figures 1-3, Table 1). These models are based upon topographic and soil attributes and do not address land use change or the influence of grazing management, and hence are intended to represent variation in habitat in the absence of anthropogenic effects on soils and vegetation. Climate data Given the goal of examining the influence of precipitation and temperature gradients on habitat suitability, we examined two sources of spatially interpolated annual climate data for the southern Great Plains. First we focused on precipitation and maximum/minimum daily temperature data compiled using the TOPS (Terrestrial Observation and Prediction System; http://ecocast.arc.nasa.gov/topwp/) model as data are available online for any specified region of the country (www.coasterdata.net), and were developed in coordination with efforts of the Great Plains LCC (www.greatplainslcc.org/resources/). However, for climate data compiled at the scale of our 7 National Grassland study areas, our preliminary analyses revealed unusually low predicted annual precipitation at the easternmost study area (Cimarron National Grassland; TOPS predicted mean annual precipitation for 1980-2009 = 368 mm), which was similar to predictions for our westernmost site (Timpas Unit, Comanche National Grassland; predicted mean annual TOPS precipitation for 1980-2009 = 361). Data from the long-term meterological station at the Cimarron National Grassland (Elkhart, KS www.ncdc.gov) showed that TOPS consistently underpredicted actual precipitation for a region of southwestern KS and southeastern CO, for unknown reasons. We therefore considered a second source of spatially interpolated long-term weather datasets generated by the PRISM model (Parameter-elevation Regressions on Independent Slopes Model; http://www.prism.oregonstate.edu). We found these data to be in much stronger agreement with point data from meterological stations located on or near the National Grasslands. All subsequent analyses of precipitation and temperature gradients were therefore based on the PRISM database (Figures 2 and 3; Table 1). We used databases of longterm (1980-2010) mean annual precipitation (mm) and maximum daily temperature (degrees C) at a resolution of 8 km x 8 km. BTPD Colony Locations BTPD colonies have been mapped annually with global positioning systems (GPS) units on National Grasslands (NGs) encompassing ~900,000 ha of federally managed grassland in the western portion of the Great Plains LCC. GPS mapping began as early as 1993 on Pawnee NG, and occurred nearly annually at all study areas during 2001-2010 (Stapp 2004, Augustine et al. 2008, Cully et al 2010; Table 1). Annual datasets for each of the 7 study sites were screened for mapping errors including datum accuracy and consistency in the resolution of boundary 5 Table 1. Mean annual precipitation and temperature during 1980-2010 (PRISM database), hectares of National Grassland, and years of black-tailed prairie dog colony mapping for each of 7 study sites in the western portion of the Great Plains LCC. Study Site Carrizo Cimarron Kiowa Pawnee East Pawnee West Rita Blanca Timpas Hectares 102692 43978 23777 37901 46841 37952 69738 Mean Annual Precipitation (mm) 20.1 21.0 20.1 17.6 16.9 20.8 20.6 Mean Daily Maximum Temperature (˚C) 418 440 415 380 348 417 352 6 Years of BTPD Colony Mapping 2001-2006, 2008-2010 2001-2010 2001-2006,2009-2010 2001-2010 2001-2010 2001-2006,2009-2010 2001-2010 Figure 2. Map of the distribution of study sites in within the western portion of the Great Plains LCC relative to variation in mean annual precipitation. 7 Figure 3. Map of the distribution of study sites in within the western portion of the Great Plains LCC relative to variation in mean annual precipitation. 8 mapping. Kiowa and Rita Blanca National Grasslands were not mapped in 2007 or 2008. Data for the Carrizo Unit of the Comanche National Grassland for 2007 were excluded due to mapping errors. Modeling Approach Traditional habitat suitability models relied on simple functions to relate the distribution of an organism to limiting factors such as food and cover, based on knowledge of the organism’s ecology. An index of habitat suitability was derived as an integrated function of these limiting factors. For black-tailed prairie dogs, Clippinger (1989) and Proctor et al. (2006) focused on limitations imposed by soil texture, slope, and composition of plant species on a site. Qualitative relationships between these factors and BTPD distribution are evident throughout their range (Koford 1955, Clippinger 1989), but quantitative relationships have only been tested for a few specific locations in the northern Great Plains (Reading and Matchett 1997, Belak 2001). When evaluating wildlife habitat, a key consideration is the scale of habitat selection being evaluated. Habitat selection can be categorized into four hierarchical scales of analysis (Johnson 1980): First-order habitat selection = selection of the geographical range of a species. Second-order habitat selection = selection of a home range by an individual or social group within the available area defined by the geographical range. Third-order habitat selection = selection of habitat components within the immediate vicinity of an individual or social group’s home range Fourth-order habitat selection = selection or procurement of resource items (e.g. food items) from those available at a given location We evaluated habitat selection using two different metrics: colony presence and colony expansion pattern. These two metrics correspond to analyses of second-order and third-order habitat selection respectively. The second-order habitat selection analysis based on colony presence defined the area of available habitat at a broad spatial scale (allotments in which BTPD colonies have been mapped; Fig 2.) and examined the influence of soils, topography, and climate on colony presence. The third-order analysis of habitat selection based on colony expansion pattern analysis defined available habitat at a finer spatial scale based on the direction and extent of colony expansion over a plague-free interval of 3 or more years. For the broad-scale analysis of BTPD colony occupancy, we quantified the maximum cumulative extent of all colony locations mapped during 2001 – 2010. A screening process following Augustine et al. (2008) was applied to exclude allotments potentially affected by incomplete colony mapping. For each allotment, we generated a set of random locations to quantify availability of habitat attributes, where the number of randomly selected pixels was equal to the number of pixels encompassed by the colony boundaries (used pixels). Available pixels were selected at two spatial scales: within a 2 km buffer of colony boundaries, and within a 0.5 km buffer of colony boundaries. Nearby colonies could potentially overlap in the area from which available pixels were selected, thereby inducing non-independence among colonies within the dataset. To address this issue, we implemented an ArcGIS script that identified all colonies whose boundaries were within the buffer distance (2 km or 0.5 km depending on scale of available habitat) of one another. Pairs of colonies located less than the buffer distance to one another were then grouped together into a single colony cluster, and the process repeated until all colony clusters within the dataset were separated by more than the buffer distance. Colony clusters, rather than individual colonies, were then used as independent subjects in the logistic 9 regression. This method dramatically reduced but did not eliminate the possibility that available habitat associated with two different colony clusters could overlap. To prevent an available pixel from being included in the set of available pixels for two different colonies, we selected available pixels randomly and without replacement. Prior to analysis, we excluded all colony clusters that were < 10 ha to reduce influence of small colonies that may not yet have expanded sufficiently to express selection relative to topographic position or soil characteristics. Colony cluster polygons were converted to 30-m resolution rasters, where each used pixel (value = 1) was contained completely within a colony boundary. Clusters where the amount of surrounding available habitat (i.e. within the buffer distance) on NSF lands was less than the area of the colony cluster were also removed from analysis. This was done because most colony clusters meeting this criteria had expanded to the point where they occupied nearly all of the NSF land in that area, leaving little or no available habitat for comparison. Table 2. Number of colony clusters or colonies used in analyses of BTPD habitat suitability at 3 different spatial scales. Scale of Analysis 2-km Buffer 0.5-km Buffer Local Expansion Pattern Study Site Area (Ha) Colony Clusters Pixels Colony Clusters Pixels Colonies Carrizo 102,692 28 142,275 71 139,097 66 120,153 Pawnee West 46,841 15 60,122 38 58,097 22 45,110 Cimarron 43,978 12 59,267 26 33,876 21 32,703 Rita Blanca 37,952 19 49,953 28 49,052 6 9,857 Timpas 69,738 19 14,738 23 14,088 13 10,725 Pawnee East 37,901 10 24,238 16 12,957 15 14,877 Kiowa 23,777 10 24,243 14 23,243 9 15,047 113 374,836 216 330,410 152 248,490 Study Site Total Pixels We also analyzed patterns of colony expansion relative to soils, topography, and local climatic conditions. We calculated annual changes in boundaries of 164 colonies across the 7 study sites during 2001 – 2005, when colonies in all or a majority of each site did not experience plague epizootics (Cully et al. 2010, Cully and Antolin, unpublished). For each colony in each year, we determined whether the colony was expanding, stable, shifting, or declining based on the following definitions: Expanding: colony area increased by more than 20% between years 1 and 2, and area occupied in year 1 makes up at least 80% of area occupied in year 2. Stable: colony area changed by less than 20%, and colony area in year 1 makes up at least 80% of area in year 2. Shifting or declining: colony area increased by less than 20%, and colony area in year 1 makes up less than 80% of area in year 2 10 Decreasing: colony area declined by more than 20%. We identified those colonies expanding and/or stable for a sequence of at least 3 consecutive years, and used these colonies to evaluate expansion patterns relative to topoedaphic and climate variables. Colonies less than 10 ha in size were excluded from analysis. For each colony in the expansion dataset, we identified the centroid of the colony in the first year of the sequence and the distance from the centroid to the maximum extent of the colony at the end of the expansion period. This distance plus 90 m was used to establish a buffered area around the centroid that defined the area of available habitat (e.g. see Augustine et al. 2007). We added 90 m to the distance between colony centroid and maximum colony boundary extent in order to be able to sample available habitat surrounding those colonies with minimal expansion (i.e. consistently stable colonies) which may not have expanded because they were surrounded by low-quality habitat. We identified all 30-m resolution pixels that were within the area into which the colony expanded (used pixels = 1) and randomly selected the same number of pixels from the buffered area into which the colony did not expand (available pixels = 0). Numbers of colony clusters or colonies used in habitat suitability model fitting at each of the 3 spatial scales at which we defined available habitat (2-km buffer, 0.5 km buffer, local expansion pattern) are summarized in Table 2. Model Predictors Vegetation Most assessments of wildlife habitat suitability are based on vegetation characteristics. However, this approach is problematic for species that substantially modify vegetation in areas they inhabit. BTPD are well-known to modify their habitat by burrowing, grazing and clipping tall vegetation. As a result, variables such as vegetation cover (e.g. Whicker and Detling 1988, Hartley et al. 2009) and remotely-sensed greenness indicies (e.g. the Normalized Difference Vegetation Index [NDVI]) differ substantially between grassland on versus off BTPD colonies for reasons unrelated to habitat selection or suitability. Furthermore, maps of vegetation characteristics other than remotely-sensed cover and NDVI are typically unavailable for broad landscapes, or if available have low resolution and accuracy. To derive predictions that are of greatest utility to land managers and conservation planning, our RSF models did not consider vegetative predictors. Rather, they are based on climatic, topographic and edaphic variables that are available across the entire Great Plains LCC. The parameters we used are correlated with regional variation in grass species (C4 shortgrasses vs. C3 mid-height grasses; Epstein et al. 1997) and local variation in site potential for different plant communities (including variation in shrub presence and density) and hence vegetation structure (Dodd et al. 2002; USDA-NRCS Ecological Site Descriptions), but do not explicitly include vegetation structure or species composition. At a local scale, we note that vegetation structure can potentially have a strong influence on colony expansion patterns, but such influences are not incorporated into our habitat models. Topography Traditional habitat suitability models often use slope and aspect, but these parameters provide an incomplete measure of topography. For example, ridges and swales can have the same slope and aspect, but differ in value as BTPD habitat. We used 10-m resolution digital elevation models (DEMs) for each study area to derive a a topographic wetness index (TWI) for 11 each site. TWI was calculated in ArcGIS using the Landscape Connectivity and Pattern (LCaP) tool (Theobald 2007). We computed TWI in two ways: (1) excluding any effect of aspect on the index (TWIn), and (2) incorporating aspect using a weighting from 0 (xeric) to 1.0 (mesic) based on relative solar insolation (TWIsi). TWIsi quantifies differences between ridges, slopes and swales and north- and south-facing slopes independent of soil texture effects. Initial model fitting for datasets with varying definitions of available habitat showed that TWIsi consistently outperformed TWIn, so all subsequent model fitting and selection analyses only considered TWIsi. We also used the 10-m DEMs to calculate slope for each pixel across each study site (Table 2). The TWI and Slope rasters were then resampled to a 30-m resolution aligned with the soil and BTPD rasters and used for subsequent model fitting. Soils We used the Soil Survey Geographic (SSURGO) database created by the USDA’s Natural Resources Conservation Service to quantify a suite of soil attributes. We used the USDA’s Soil Data Viewer tool to derive quantitative maps of soil properties rather than categorical maps of soil types or ecological sites, including percent sand to 1 m depth (SAND), percent clay to 1 m depth, average soil depth to bedrock or a restrictive layer, soil organic matter content, and soil pH, all at a 30-m pixel resolution (Table 3). Use of these quantitative soil properties allowed us to model habitat suitability across all 8 study sites, even though specific soil series may only be found at one or two sites. Each map unit within the SSURGO database (i.e. each polygon) is typically composed of one or more “components”, where the components represent the major soil types within a map unit. Differences in soil properties can exist over short distances between map unit components, but these are not represented spatially in the SSURGO database. For each map unit, an estimate of the percent composition of each component is provided in the database. To obtain a single value for each quantitative soil attribute for each map unit, we used a “dominant condition” aggregation method where we first grouped together components with like attribute values in a map unit. For each group, percent composition was set to the sum of the percent composition of all components participating in that group. Soil horizon attributes were aggregated to 1 m depth at component level, before components were aggregate to the map unit level. The attribute value for the group with the highest cumulative percent composition was then assigned to each map unit. As a result, our analyses are contingent upon the accuracy of the soil mapping process, and do not reflect the potential influence of fine-scale spatial variation in soil components within map units. Our second approach used NRCS Ecological Site Descriptions to assess variation in BTPD habitat suitability. Ecological Site Descriptions (ESD’s) are becoming a key tool guiding rangeland management in the Great Plains because they are based on the SSURGO database, and NRCS has developed detailed descriptions of plant communities, potential site productivity, models linking livestock management to plant community states and transitions for each ESD. In 2010, the NRCS, US Forest Service, and Bureau of Land Management signed a MOU establishing that all three agencies would collectively use the ESD framework to guide rangeland management. At present, however, most first-round ESD’s do not incorporate prairie dogs. For models evaluating BTPD habitat selection for Ecological Sites, we did not include quantitative soil attributes, because these attributes are used to define the ESD boundaries. We used the dominant ESD within each SSURGO database poloygon in our models, but note that ESD’s do not necessarily map 1:1 to soil components, as discussed above for quantitative soil attributes. 12 Table 3. Summary of topographic, soil and climate attributes used in modeling black-tailed prairie dog habitat suitability. Parameter Units Slope Derived from 10-m digital elevation model, degrees TWIsi Index ranging from ~1-30; derived from 10-m digital elevation model following Theobald ( 2007) Sand % by weight to 1 m depth Clay % by weight to 1 m depth Organic Matter % by weight to 1 m depth pH Result of 1:1 soil:water method Depth Depth to impermeable layer, cm Precipitation Mean annual amount, 1980-2010, mm Temperature Mean monthly maximum, 1980-2010, ˚C Ecological site definitions vary among Major Land and Resource Areas (MLRAs) within the Great Plains, and hence types of ecological sites varied among some study sites. We therefore analyzed ESD’s in three clusters of study sites, based on consistency in ESD definitions: (1) MLRA 067B/69: Pawnee East, Pawnee West, Timpas, and Carrizo, (2) MLRA 77A/B: Rita Blanca and Kiowa, and (3) MLRA 72: Cimarron. Model Fitting and Selection: We used general linear mixed models fit with the Laplace approximation method (Bolker et al. 2009) to assess relative BTPD habitat suitability. With this modeling approach, we generated population-level resource selection functions (RSFs) across two orders of selection and 7 BTPD populations based upon the used-available designs of 2nd and 3rd order habitat selection (Johnson 1980), where the probabilities generated by the RSFs are proportional to the probability of use by BTPD (Manly et al. 2002). We used logistic regression with a binary response variable with values of 1 for used pixels and 0 for available pixels. All models included a random intercept term that treated each colony cluster (clusters defined as a group of colonies within a 0.5 or 2 km neighborhood of one another) as a subject to account for the nesting of used and available pixels within colony clusters, and to account for variation in sample sizes among colony clusters (Gillies et al. 2006). All models were fit using the GLIMMIX procedure in SAS v9.3. Models based on quantitative soil and topographic variables considered 8 possible predictors: slope (SL), topographic wetness index incorporating the effect of solar insolation on evaporation (TWIsi), mean soil sand content to 1 m depth (SAND), mean soil clay content to 1 m depth (CLAY), soil organic matter content (OM), soil pH (pH), and soil depth to a restricted layer (DTR). We compared the suite of potential models based on two criteria: minimization of AIC (Burnham and Anderson 2002), and maximization of the area under the Receiver Operating Characteristics curve (Area under ROC curve; Hanley and McNeil 1982; Gonen 2006). Our use of general linear mixed models requires that each candidate model be fit individually without the aid of automated model comparison procedures available for general linear models in some statistical packages. We therefore used a 3-stage approach for considering and selecting within 13 sets of candidate models with and without interaction terms. First, we evaluated the set of candidate models that only included the 7 possible topoedaphic predictors (no interaction terms) using backward selection and minimization of AIC. Second, we evaluated a set of candidate models that included all predictors in the best model identified in the first step, but that also considered interactions between topographic variables (TWIsi and Slope) and those soil characteristics that could influence soil moisture and hence site productivity (SAND, CLAY, OM). In this second step, we identified the best models with single interaction terms for TWIsi and Slope, and then also considered a model with both the TWIsi interaction term that minimized AIC and the Slope interaction term that minimized AIC. Third, we evaluated a set of candidate models that included all predictors in the model identified in the second step, but that also considered interactions between 4 topoedaphic variables (TWIsi, SAND, CLAY, OM) and the climatic variables that vary across the study region (PRECIP = mean annual precipitation, and TEMP = mean maximum monthly temperature). We hypothesized that large-scale variation in temperature and precipitation could influence BTPD habitat selection via their influence on moisture availability and hence forage productivity in this water-limited ecosystem. The four topoedaphic variables above were selected for tests of interactions with climate because they all influence moisture availability at the local level. TWIsi is a direct measure of topographic effects on moisture, with highest values in swales and drainages. SAND, CLAY and OM influence moisture availability through water infiltration and soil water holding capacity. We evaluated all possible TEMP x topoedaphic interactions (4 models), all possible PRECIP x topoedaphic interactions (3 models; PRECIP x SAND not considered due to high covariance), and models that included an interaction term for both TEMP and PRECIP. Mixed models generate coefficients for prediction at both the colony-specific level (conditional model) and for prediction at the level of the population of colonies within the study region (marginal or population model). Because our goal was prediction at the population level, we examined model fit using a method that included assessing the model’s prediction accuracy at the population level. When assessing the prediction accuracy of a model, true positive and false negative rates are two widely used indicies (Wang et al. 2011). For a binary test, a threshold cutoff can be defined where values above the threshold are assigned a positive outcome, and values below the threshold are assigned a negative outcome. The receiver operator characteristic (ROC) curve is the entire collection of true positive and false negatives for varying thresholds from 0 to 1. A summary index of model performance (i.e. predictive abilility) can then be defined as the area under the ROC curve (AuROC), which is equivalent to the probability that model predictions for a randomly selected pair of used and available pixels are correctly ordered. Wang et al. (2011) note that on the basis of results from Pepe (2005), and Pepe et al. (2006), “when using a combined linear test as a decision rule, the ROC-based approach may outperform the likelihood-based approach in terms of prediction performance. On the other hand, it is possible that when prediction is of interest, allowing some variables with weaker association to stay in a model may improve prediction accuracy (Pinsky 2005).” For these reasons, when comparing models with versus without climate variables (and hence comparing models with different random coefficients), we considered both AuROC (following Gonen 2006) and AIC in model selection. Specifically, we only considered models including interactions with precipitation and temperature when they increased AuROC relative to the model lacking interactions with climate, and then used AIC to compare and select among the set of models that increased AuROC relative to the best model without climate interactions. 14 Figure 4. Example of colony clusters defined by the 2-km linkage rule and the associated distribution of pixels representing available habitat for a portion of the Pawnee West study site. The green background shows the distribution of the National Grassland property. Each colony cluster is represented by points of a different color. In this example, there are 7 colony clusters. Within each color, dense concentrations of points show used pixels located on colonies, and sparsely distributed points are available pixels. 15 Figure 5. Example of colony clusters defined by the 0.5-km linkage rule and the associated distribution of pixels representing available habitat for each cluster. Area shown is a portion of the Pawnee West study site. The green background shows the distribution of the National Grassland property (National Forest System lands). Each colony cluster and its associated available habitat are represented by points of a different color. In this example, there are 9 colony clusters. The distribution of used pixels is the same as in Figure 5, except that a group of small colonies in the northern portion of the area of Figure 5 were not included with the 0.5-km rule because they each became a separate cluster < 10 ha in size, and hence fell below the colony size cutoff. 16 This model fitting approach was applied to 3 different datasets where available habitat surrounding colonies was defined at different scales. The first two datasets correspond to an analysis of second-order habitat selection: (1) used and available pixels defined based on a 2 km buffer around each colony cluster, (2) used and available pixels defined based on 0.5 km buffer around each colony cluster. The third dataset corresponds the third-order habitat selection, where used and available pixels were defined based on the local pattern of colony expansion over >3 consecutive years. For the first dataset, the model fitting procedure was applied to (a) the full dataset combining colony clusters from all 7 National Grasslands (referred to as global models hereafter), and (b) each National Grassland modeled separately (referred to as local models hereafter). Finally, to evaluate BTPD habitat selection relative to Ecological Sites, we analyzed the 2-km buffer and the expansion pattern databases for the 3 groups of study sites defined based on MLRAs. We also included SLOPE and TWIsi in the ecological site models. In these analyses, we did not consider interactions with climate variables due to limited variation in temperature and precipitation within the different MLRAs. Model Mapping: We used the selected models to generate maps of relative BTPD habitat suitability at the scale of the 7 National Grassland study sites, and at the scale of the broader shortgrass steppe study region encompassing 74 counties in Colorado, New Mexico, Oklahoma, Kansas, and Texas. For ease of reference, raster files are organized by study site and county (Appendix A). Following Manly et al. (2002), we calculated a relative value for each pixel based on the selected model’s coefficients and intercept, exponentiated these values, and then used a linear stretch of exponentiated values to obtain rescaled RSF predicted values between 0 and 1 (see also Johnson et al. 2006, DeCesare et al. 2012). We refer to these as the RSF probability maps. Specifying how different probability values correspond to classes of suitable versus unsuitable habitat depends upon the level and types of error that one is willing to accept. Given the design of our sampling, where locations of BTPD colonies represent used habitat and locations lacking BTPD colonies represent available habitat, the “available” habitat is likely to include both areas of high quality (or potentially suitable) habitat that has not yet been colonized, and areas of low quality (or unsuitable) habitat that is being avoided by colonizing prairie dogs. In this view, false negative model predictions (i.e. where pixels occurring within known BTPD colony locations are predicted to not have BTPD present) are a more egregious error than false positive model predictions (i.e. where pixels within “available” habitat are predicted to have BTPD present). We therefore mapped RSF probability categories based on cutoff values corresponding to low and fixed false negative error rates of 5, 10 and 15%, and then present the false positive error rates corresponding to each of these cutoff values. In all of the category maps we present, we use the following categories of probabilities: Category 1: RSF probability values below the cutoff for a 5% false negative rate Category 2: RSF probability values below the cutoff for a 10% false negative rate but not included in category 1, Category 3: RSF probability values below the cutoff for a 15% false negative rate but not included in category 1 or 2 Category 4: RSF probability values above the cutoff for a 15% false negative rate. 17 Thus, category 1 depicts areas consistently predicted to represent low quality habitat even under a stringent false negative error rate and category 4 represents areas consistently predicted to be high quality habitat, even with considerable relaxation of the false negative error rate (and correspondingly lower false positive rate). Categories 2 and 3 represent areas of intermediate habitat value. We compared the best local models (fit to a specific study site using data only from that study site) with selected global models (fit using data from all study sites combined) in terms of the proportion of the landscape predicted to be in each of the 4 categories above (relative value comparison) and in terms of the proportion of the landscape predicted to be in category 1 by one model but in category 4 by the other model. We used spatial differences in model predictions as our primary means of comparing the models, as interpretation of differences in coefficients is difficult when models contain multiple and differing interaction terms. Results Second-order habitat selection Our assessment of second-order habitat selection measured available habitat at two scales: a 2 km buffer surrounding the maximum extent of each colony cluster, and a 0.5 km buffer surrounding the maximum extent of each colony cluster. The 2 km buffer distance was originally selected as an appropriate compromise between larger distances, which would cause increasing overlap among nearby colony buffers, and shorter distances, which would sample a less extensive area of the landscape. However, we also conducted the same analyses using the 0.5 km buffer to assess whether our selection of buffer distance notably affected the habitat suitability model, in particular the direction of the effect of different topoedaphic parameters. We first present detailed findings for the 2 km buffer modeling effort, as these findings form the basis for our final, large-scale mapping of habitat suitability, and then present the comparable models based on the 0.5 km buffer distance. For the second-order habitat selection analysis, we first examined global models based on the full dataset (all colonies from all 7 study sites), and then also fit local models for each of the 7 study sites for comparison. Global second-order models For the set of models that did not include interaction terms, the most parsimonious model included all 7 topoedaphic predictors (TWIsi, Slope, % Sand, % Clay, pH, % Organic matter, and Depth to a restricted layer; Table 4), which was a substantial improvement of all competing models with 6 or fewer predictors (Δ AIC > 558). Of the potential models including interactions between slope and soil parameters, the most parsimonious included a Slope x Organic Matter interaction (Table 4; Δ AIC relative to no interaction model = 251.3). Of the potential models including interactions between TWIsi and soil parameters, the most parsimonious model included a TWIsi x Sand interaction (Table 4; Δ AIC relative to no interaction model = 748.0). 18 Table 4. Summary of second-order RSF model set including direct effects of up to 7 topoedaphic predictors, and potential interactions between topographic predictors (TWIsi and/or Slope) and soil predictors that influence soil moisture holding capacity (Clay, Sand, Organic matter). Included in the model set are the best model excluding interaction terms (a), the best model with a single Slope interaction term (b) the best model with a single TWIsi interaction term (c) and the selected model with both Slope and TWIsi interaction terms (d). All models included a random intercept. No. of Parameters AIC Δ AIC pH 1 519179.1 31400.5 TWIsi 1 514935.7 27157.1 Slope 1 514527.5 26748.9 Sand 1 497848.3 10069.7 Clay 1 500988.7 13210.1 OM 1 513868.4 26089.8 Restr 1 518314.7 30536.1 TWIsi Slope Clay OM pH DTR 6 492838.6 5060.0 TWIsi Sand Clay OM pH DTR 6 490398.9 2620.3 TWIsi Slope Sand Clay pH DTR 6 490141.3 2362.7 Slope Sand Clay OM pH DTR 6 490133.1 2354.5 TWIsi Slope Sand OM Clay DTR 6 489298.6 1520.0 TWIsi Slope Sand OM pH DTR 6 489287.5 1508.9 TWIsi Slope Sand Clay OM pH 6 489276.1 1497.5 TWIsi Slope Sand Clay OM pH DTR (a) 7 488717.4 938.8 TWIsi Slope Sand Clay OM pH DTR Slope*Clay 8 488693.3 914.7 TWIsi Slope Sand Clay OM pH DTR Slope*Sand 8 488683.5 904.9 TWIsi Slope Sand Clay OM pH DTR Slope*OM (b) 8 488466.1 687.5 TWIsi Slope Sand Clay OM pH DTR TWIsi*Clay 8 488637.7 859.1 TWIsi Slope Sand Clay OM pH DTR TWIsi*OM 8 488583.6 805.0 TWIsi Slope Sand Clay OM pH DTR TWIsi*Sand (c) TWIsi Slope Sand Clay OM pH DTR TWIsi*Sand Slope*OM (d) 8 487969.4 190.8 9 487778.6 0.0 Model Parameters 19 Table 5. Summary of models that include interactions with mean annual precipitation and/or mean monthly maximum temperature. Letters in parentheses show the best model including an interaction with precipitation (a), the best model including an interaction with temperature (b), and the best model with both temperature and precipitation (c). The final selected global model for second-order habitat selection is shown in bold. Best model without climate interactions: AuROC AIC # of Random Coefficients TWIsi Slope Sand Clay OM pH DTR TWIsi*Sand Slope*OM 0.6343 487778.6 0 0.6377 487213.8 1 2631.5 0.6392 485862.0 1 1279.7 0.6288 484624.4 1 42.1 0.6418 484582.3 1 0.0 0.6390 486253.5 1 1671.2 0.6359 485629.0 1 1046.7 0.6419 487771.3 1 3189.0 0.6067 482098.0 2 Interactions with Precipitation: TWIsi Slope Sand Clay OM pH DTR TWIsi*Sand SL*OM Precip -TWIsi*Precip TWIsi Slope Sand Clay OM pH DTR TWIsi*Sand SL*OM Precip Precip OM*Precip TWIsi Slope Sand Clay OM pH DTR TWIsi*Sand SL*OM Precip Precip Sand*Precip TWIsi Slope Sand Clay OM pH DTR TWI*Sand SL*OM Precip Precip Clay*Precip (a) Interactions with Temperature: TWIsi Slope Sand Clay OM pH DTR TWIsixSand SL*OM Temp Precip OM*Temp TWIsi Slope Sand Clay OM pH DTR TWIsixSand SL*OM Temp Precip Clay*Temp TWIsi Slope Sand Clay OM pH DTR TWIsixSand SL*OM Temp Precip TWIsi*Temp (b) Interaction with Precipitation and Temperature: TWIsi Slope Sand Clay OM pH DTR TWIsixSand SL*OM Precip Precip Clay*Precip Temp OM*Temp (c) 20 Δ AIC Table 6 Coefficients and associated standard errors for the best topoedaphic model (see model selection statistics in Table 4) and the best model including topoedaphic predictors plus interactions with mean annual precipitation (model selection statistics in Table 5). For a summary of habitat suitability maps based on the Topoedaphic + Precipitation model, see Table 7. Topoedaphic + Precipitation Model Topedaphic Model Coefficient Std Error Coefficient Std Error Intercept -2.9322 0.09884 5.6411 0.4356 TWIsi 0.08353 0.002025 0.06968 0.002067 Slope -0.05225 0.008034 -0.08081 0.007802 Sand -0.01129 0.000487 -0.01085 0.000495 Clay 0.01582 0.000664 -0.2531 0.005908 OM 0.3941 0.01101 0.4315 0.01087 pH 0.2504 0.0105 0.3203 0.01075 0.00198 0.000095 0.000555 0.000099 TWIsi x Sand -0.00119 0.000046 -0.00083 0.000047 Slope x OM -0.08895 0.006393 -0.06223 0.006028 Precipitation -0.02261 0.001121 Precipitation x Clay 0.000675 -- Restr 21 Including both interaction terms further reduced AIC by 190.8 relative to the best model with a single interaction term (Table 4). The final selected model based on topoedaphic predictors had an area under the ROC curve of 0.6343, with coefficients presented in Table 6. Consideration of an expanded model set that allowed for interactions between the precipitation gradient and topoedaphic predictors showed the most parsimonious model to include an interaction between precipitation and soil clay content (Table 5). This model both increased model predictive ability (AuROC = 0.6418) and was substantially more parsimonious relative to the best topoedaphic-only model (Δ AIC = 3196.3). Including interactions between the temperature gradient and topoedaphic predictors model increased model predictive ability to a similar degree (AuROC = 0.6419) but with substantially less parsimony AIC (Δ AIC = 7.3). The validity of using AIC to compare models with different numbers of random coefficients (e.g. model with no climate interactions vs. model with temperature interaction) is unclear based upon the current statistical literature, due to varying approaches in calculating the degrees of freedom for models with different numbers of random coefficients. However, both the temperature and precipitation models include a random intercept and one random coefficient (either temperature or precipitation respectively, analyzed at the study site scale), and thus the same degrees of freedom regardless of the method of calculation. The precipitation and temperature models had similar predictive ability, but the model including precipitation was more parsimonious than the model including temperature. Models including interactions with both precipitation and temperature yielded lower AIC (reflecting the inclusion of an additional random coefficient), but had substantially reduced predictive ability and thus were rejected from consideration. Our final Figure 6. Predicted relative BTPD habitat suitability as a function of slope for varying levels of soil organic matter content based on the final selected Topoedaphic + Precipitation model (Table 6). Figure 7. Predicted relative BTPD habitat suitability as a function of Topographic Wetness Index with aspect correction (TWIsi) for varying levels of soil sand content based on the final selected Topoedaphic + Precipitation model (Table 6). 22 selected second-order RSF for prairie dog habitat therefore included 7 topoedaphic predictors, precipitation, and TWIsi x Sand, Slope x Organic matter, and Precipitation x Clay interactions (Table 5 and 6). Coefficients of the selected model including precipitation (Table 6) show that BTPD habitat suitability increases with increasing soil pH and depth to a restricted layer across all levels of the other predictors. BTPD habitat suitability declines with increasing slope, but does so more rapidly on soils with high organic matter content than on soils with low organic mattercontent (Figure 6). The TWIsi x Sand interaction shows that BTPD habitat suitability is positively associated with the topographic wetness index (i.e. greater suitability for swales and draws), but this positive association is greater for soils with low sand content than for soils with high sand content (Figure 7). Thus, high-quality Figure 8. Predicted relative BTPD habitat suitability as a habitat is associated with lowlands with function of soil clay content for varying levels of mean high silt+clay content, whereas sandy annual precipitation (see model coefficients in Table ?). lowlands have lower relative habitat value. The Precipitation x Clay interaction shows that BTPD habitat suitability is positively associated with soil clay content for regions with 400 – 500 mm precipitation, but the strength of this association increases with increasing mean annual precipitation (Figure 8). At the lowest end of the precipitation gradient (as precipitation declines from 400 to 350 mm) the association with soil clay content switches from positive to negative, i.e. declining habitat quality with increasing clay content at 350 mm mean annual precipitation (Figure 3). Habitat suitability maps were generated for each of the 7 study areas where suitability is measured as a probability (varying from 0 to 1) derived from the best global RSF including topoedaphic predictors and precipitation. Maps are referenced in Table 7. At some study sites, in particular the Cimarron National Grassland, a striped pattern is evident in the predictions for BTPD habitat suitability in areas of relatively low or zero slopes. This striping pattern is an artifact of the algorithm used to model water flow patterns when calculating the Topographic Wetness Index. The artifact was most notable at the Cimarron site due to the lower quality of the DEM for this site, presumably resulting from differences in the method used to create the DEM for this county. As resolution of DEMs improves and more accurate methods are used to 23 Table 7. Index of maps of BTPD habitat suitability generated based on the final selected global model including topoedaphic predictors, precipitation and an interaction between precipitation and soil clay content (see Table 6 for coefficients). Study Site Map # Output Type Title Carrizo 1 RSF Probability Map of Global Topoedaphic + Precipitation RSF Model: Carrizo Study Area Cimarron 2 RSF Probability Map of Global Topoedaphic + Precipitation RSF Model: Cimarron Study Area Kiowa 3 RSF Probability Map of Global Topoedaphic + Precipitation RSF Model: Kiowa Study Area Pawnee East 4 RSF Probability Map of Global Topoedaphic + Precipitation RSF Model: Pawnee East Study Area Pawnee West 5 RSF Probability Map of Global Topoedaphic + Precipitation RSF Model: Pawnee West Study Area Rita Blanca 6 RSF Probability Map of Global Topoedaphic + Precipitation RSF Model: Rita Blanca Study Area Timpas 7 RSF Probability Map of Global Topoedaphic + Precipitation RSF Model: Timpas Study Area Carrizo 8 RSF Category Map of Global Topoedaphic + Precipitation Model Categories: Carrizo Study Area Cimarron 9 RSF Category Map of Global Topoedaphic + Precipitation Model Categories: Cimarron Study Area Kiowa 10 RSF Category Map of Global Topoedaphic + Precipitation Model Categories: Kiowa Study Area Pawnee East 11 RSF Category Map of Global Topoedaphic + Precipitation Model Categories: Pawnee East Study Area Pawnee West 12 RSF Category Map of Global Topoedaphic + Precipitation Model Categories: Pawnee West Study Area Rita Blanca 13 RSF Category Map of Global Topoedaphic + Precipitation Model Categories: Rita Blanca Study Area Timpas 14 RSF Category Map of Global Topoedaphic + Precipitation Model Categories: Timpas Study Area Table 8. RSF probability cutoff values that correspond to false negative error rates of 5, 10 and 15%. These probability cutoffs were used to generate the RSF category maps (Maps 8-14 in Table 7). False Negative Error Rate 5.0 10.0 15.0 0.06010 0.07695 0.09585 False Positive Rate 36.7 30.4 25.6 Sensitivity 90.0 80.0 70.0 1-Specificity 73.5 61.0 51.4 RSF Probability Cutoff 24 generate them (e.g. high-resolution LiDAR), such artifacts can be removed from habitat models based upon topographic indicies. We classified the RSF probabilities into 4 categories based on cutoff probabilities that correspond to different false negative error rates (Table 8). Category 1 (low habitat suitability) corresponds to locations where BTPD are predicted to be absent based on a relatively stringent false negative rate of 5%. Category 4 (high habitat suitability) corresponds to locations where BTPD are predicted to be present based on a less stringent false negative rate of 15%, which is associated with a lower false positive rate (Table 8), and hence a lower rate of incorrectly predicting BTPD presence. Maps depicting the distribution of the 4 probability categories for each study site are referenced in Table 7. Local second-order models Our analysis of local models first evaluated the set of candidate models that included up to 7 topoedaphic predictors, and then examined an expanded model set that included potential interactions between topography (TWIsi, Slope) and soil characteristics that influence waterholding capacity (Sand, Clay, or OM), following the same process as the global model analysis. Because our analysis of interactions with precipitation and temperature in the global models was based on among-site variation in climate, precipitation and temperature were not considered in local models. Selected local models included all 7 topoedaphic predictors at 4 sites, 6 predictors at Kiowa and Pawnee West, and 4 predictors at Rita Blanca. All selected local models included an interaction between slope and one soil parameter (either clay or organic matter), and 5 of 7 local models included an interaction between TWIsi and one soil parameter (either clay or organic matter). The magnitude and sign of the best local models were largely consistent with the best global model, with the exception that the global model included a TWIsi x Sand interaction rather than with clay or organic matter (Table 16). When the main effect and interaction term coefficients are considered together, all local models and the global model predict that habitat suitability increases with increasing TWIsi, soil organic matter content, and soil clay content (except under low mean annual precipitation in the global model; Fig. 3). All local and the global models predict that habitat suitability decreases with increasing slope. Most (7 of 8) models predict that habitat suitability increases with increasing soil depth to a restricted layer, and with increasing soil pH (Table 16). Predictions of the global topoedaphic + precipitation model showed a high degree of consistency with the best models fit to each local dataset (Table 19). Disagreement between the global versus local models was less than 5% of the landscape for 5 of 7 study sites: Carrizo, Cimarron, Kiowa, Pawnee West, and Timpas (Table 19). The greatest disagreement occurred at the Pawnee East study site, where the local model was based on a small sample size (10 colony 25 Table 9. Summary of model set for BTPD colonies on the Carrizo Unit of the Comanche National Grassland including direct effects of up to 7 topoedaphic predictors, and potential interactions between topographic predictors (TWIsi and/or Slope) and soil predictors that influence soil moisture holding capacity (Clay, Sand, Organic matter). Included in the model set are the best model excluding interaction terms (a), the best model with a single TWIsi interaction term (b) the best model with a single Slope interaction term (c) and a model with both Slope and TWIsi interaction terms (d). The selected model (c) is highlighted in bold. All models included a random intercept. Carrizo Parameters Δ AIC OM 1 11889.6 pH 1 11700.2 TWIsi 1 11634.4 Restr 1 11241.0 Slope 1 9471.2 Sand 1 5615.0 Clay 1 2668.0 TWIsi Slope Sand Clay OM Restr 6 1655.7 TWIsi Sand Clay OM pH Restr 6 1470.1 TWIsi Slope Sand Clay pH Restr 6 712.8 TWIsi Slope Clay OM pH Restr 6 655.1 Slope Sand Clay OM pH Restr 6 575.4 TWIsi Slope Sand Clay OM pH 6 546.6 TWIsi Slope Sand Clay OM pH Restr (a) 7 545.9 Model (a) + TWIsi x OM 8 533.8 Model (a) + TWIsi x Sand 8 486.7 Model (a) + TWIsi x Clay (b) 8 469.6 Model (a) + Slope x OM 8 475.7 Model (a) + Slope x Sand 8 46.5 Model (a) + Slope x Clay (c) 8 0.0 Model (a) + Slope x Clay + TWIsi x Clay 9 69.0 26 Table 10. Summary of model set for BTPD colonies on the Cimarron National Grassland including direct effects of up to 7 topoedaphic predictors, and potential interactions between topographic predictors (TWIsi and/or Slope) and soil predictors that influence soil moisture holding capacity (Clay, Sand, Organic Matter). Included in the model set are the best model excluding interaction terms (a), the best model with a single TWIsi interaction term (b) the best model with a single Slope interaction term (c) and a model with both Slope and TWIsi interaction terms (d). The selected model (c) is highlighted in bold. All models included a random intercept. Cimarron Number Δ AIC Restricted 1 26898.3 Slope 1 26577.5 TWIsi 1 23501.5 pH 1 19329.0 Clay 1 8664.3 OM 1 7123.2 Sand 1 2257.8 TWIsi Slope Sand Clay pH Restr 6 837.4 Slope Sand Clay OM pH Restr 6 831.3 TWIsi Slope Sand OM pH Restr 6 618.3 TWIsi Sand Clay OM pH Restr 6 340.6 TWIsi Slope Clay OM pH Restr 6 308.0 TWIsi Slope Sand Clay OM Restr 6 276.4 TWIsi Slope Sand Clay OM pH 6 274.7 TWIsi Slope Sand Clay OM pH Restr (a) 7 264.8 Model (a) + TWIsi x Clay 8 244.0 Model (a) + TWIsi x Sand 8 214.4 Model (a) + TWIsi x OM (b) 8 1.4 Model (a) + Slope x OM 8 254.6 Model (a) + Slope x Sand 8 249.5 Model (a) + Slope x Clay (c) 8 247.2 Model (a) + Slope x Clay + TWIsi x OM (d) 9 0.0 27 Table 11. Summary of model set for BTPD colonies on the Kiowa National Grassland including direct effects of up to 7 topoedaphic predictors, and potential interactions between topographic predictors (TWIsi and/or Slope) and soil predictors that influence soil moisture holding capacity (Clay, Sand, Organic Matter). Included in the model set are the best model excluding interaction terms (a), the best model with a single TWIsi interaction term (b) the best model with a single Slope interaction term (c) and a model with both Slope and TWIsi interaction terms (d). The selected model (c) is highlighted in bold. All models included a random intercept. Kiowa Parameters Δ AIC pH 1 1863.0 Restricted 1 1776.0 Sand 1 1433.0 Slope 1 1363.9 TWIsi 1 1351.7 OM 1 1161.1 Clay 1 1045.7 TWIsi Slope Clay pH Restr 5 510.3 Slope Clay OM pH Restr 5 281.5 TWIsi Clay OM pH Restr 5 256.0 TWIsi Slope OM pH Restr 5 235.9 TWIsi Slope Clay OM pH 5 138.9 TWIsi Slope Clay OM Restr 5 89.2 TWIsi Slope Sand Clay pH Restr 6 359.6 Slope Sand Clay OM pH Restr 6 282.3 TWIsi Sand Clay OM pH Restr 6 256.8 TWIsi Slope Sand OM pH Restr 6 205.0 TWIsi Slope Sand Clay OM pH 6 127.8 TWIsi Slope Sand OM Clay Restr 6 91.2 TWIsi Slope Clay OM pH Restr (a) 6 88.4 TWIsi Slope Sand Clay OM pH Restr 7 88.6 Model (a) + TWIsi x Clay 7 89.1 Model (a) + TWIsi x OM (b) 7 67.2 Model (a) + Slope x Clay 7 3.3 Model (a) + Slope x OM (c) 7 0.0 Model (a) + Slope x OM + TWIsi x OM (d) 8 1.3 28 Table 12. Summary of model set for BTPD colonies on the Eastern Unit of the Pawnee National Grassland including direct effects of up to 7 topoedaphic predictors, and potential interactions between topographic predictors (TWIsi and/or Slope) and soil predictors that influence soil moisture holding capacity (Clay, Sand, Organic Matter). Included in the model set are the best model excluding interaction terms (a), the best model with a single TWIsi interaction term (b) the best model with a single Slope interaction term (c) and a model with both Slope and TWIsi interaction terms (d). The selected model (d) is highlighted in bold. All models included a random intercept. Pawnee East Parameters Δ AIC pH 1 923.4 TWIsi 1 792.6 Clay 1 903.1 OM 1 841.6 Sand 1 925.7 Restricted 1 579.6 Slope 1 281.7 Slope Sand OM Restr 1 176.2 TWIsi Sand Clay OM pH Restr 6 387.1 TWIsi Slope Clay OM pH Restr 6 173.6 TWIsi Slope Sand OM pH Restr 6 169.9 TWIsi Slope Sand Clay OM pH 6 112.1 TWIsi Slope Sand OM Clay Restr 6 66.3 Slope Sand Clay OM pH Restr 6 64.9 TWIsi Slope Sand Clay pH Restr 6 63.7 TWIsi Slope Sand Clay OM pH Restr (a) 7 61.2 Model (a) + TWIsi x Clay 8 56.2 Model (a) + TWIsi x Sand 8 56.2 Model (a) + TWIsi x OM (b) 8 12.8 Model (a) + Slope x OM 8 58.3 Model (a) + Slope x Sand 8 40.8 Model (a) + Slope x Clay (c) 8 39.0 Model (a) + Slope x Clay + TWIsi x OM (d) 9 0.0 29 Table 13. Summary of model set for BTPD colonies on the Western Unit of the Pawnee National Grassland including direct effects of up to 7 topoedaphic predictors, and potential interactions between topographic predictors (TWIsi and/or Slope) and soil predictors that influence soil moisture holding capacity (Clay, Sand, Organic Matter). Included in the model set are the best model excluding interaction terms (a), the best model with a single TWIsi interaction term (b) the best model with a single Slope interaction term (c) and a model with both Slope and TWIsi interaction terms (d). The selected model (d) is highlighted in bold. All models included a random intercept. Pawnee West Parameters Δ AIC Restricted 1 1306.4 pH 1 1267.5 TWIsi 1 979.3 Clay 1 913.9 Sand 1 843.7 Slope 1 883.4 OM 1 674.2 TWIsi Sand Clay OM Restr 5 349.0 TWIsi Slope Sand Clay Restr 5 260.7 Slope Sand Clay OM Restr 5 245.4 TWIsi Slope Clay OM Restr 5 211.9 TWIsi Slope Sand Clay OM 5 189.4 TWIsi Slope Sand OM Restr 5 189.1 TWIsi Sand Clay OM pH Restr 6 350.0 Slope Sand Clay OM pH Restr 6 247.1 TWIsi Slope Sand Clay pH Restr 6 243.9 TWIsi Slope Clay OM pH Restr 6 213.7 TWIsi Slope Sand Clay OM pH 6 191.4 TWIsi Slope Sand OM pH Restr 6 190.2 TWIsi Slope Sand Clay OM Restr (a) 6 186.3 TWIsi Slope Sand Clay OM pH Restr 7 186.8 Model (a) + TWIsi x OM 7 187.0 Model (a) + TWIsi x Sand 7 175.0 Model (a) + TWIsi x Clay (b) 7 171.4 Model (a) + Slope x Sand 7 183.4 Model (a) + Slope x Clay 7 158.4 Model (a) + Slope x OM (c) 7 44.3 Model (a) + Slope x OM + TWIsi x Clay (d) 8 0.0 30 Table 14. Summary of model set for BTPD colonies on the Rita Blanca National Grassland including direct effects of up to 7 topoedaphic predictors, and potential interactions between topographic predictors (TWIsi and/or Slope) and soil predictors that influence soil moisture holding capacity (Clay, Sand, Organic Matter). Included in the model set are the best model excluding interaction terms (a), the best model with a single TWIsi interaction term (b) the best model with a single Slope interaction term (c) and a model with both Slope and TWIsi interaction terms (d). The selected model (d) is highlighted in bold. All models included a random intercept. Rita Blanca Parameters Δ AIC Sand 1 1571.48 OM 1 1564.61 pH 1 1229.80 TWIsi 1 1171.54 Restricted 1 1127.85 Clay 1 871.62 Slope 1 558.88 TWIsi Sand Clay 3 506.48 TWIsi Slope Sand 3 449.70 Slope Sand Clay 3 116.28 TWIsi Slope Clay 3 152.02 Slope Sand Clay OM 4 118.02 Slope Sand Clay pH 4 117.90 TWIsi Slope Clay Restr 4 105.88 Slope Sand Clay Restr 4 118.19 TWIsi Slope Sand Clay (a) 4 56.67 TWIsi Slope Sand Clay Restr 5 58.67 TWIsi Slope Sand Clay OM 5 58.61 TWIsi Slope Sand Clay pH 5 58.59 TWIsi Sand Clay OM pH Restr 6 507.03 TWIsi Slope Sand OM pH Restr 6 214.18 Slope Sand Clay OM pH Restr 6 121.24 TWIsi Slope Clay OM pH Restr 6 97.16 TWIsi Slope Sand Clay OM Restr 6 60.60 TWIsi Slope Sand Clay OM pH 6 60.48 TWIsi Slope Sand Clay pH Restr 6 60.43 TWIsi Slope Sand Clay OM pH Restr 7 62.34 Model (a) + TWIsi x Sand 5 52.4 Model (a) + TWIsi x Clay (b) 5 13.1 Model (a) + Slope x Sand 5 50.9 Model (a) + Slope x Clay (c) 5 22.8 Model (a) + Slope x Clay + TWIsi x Clay (d) 6 0.0 31 Table 15. Summary of model set for BTPD colonies on the Timpas Unit of the Comanche National Grassland including direct effects of up to 7 topoedaphic predictors, and potential interactions between topographic predictors (TWIsi and/or Slope) and soil predictors that influence soil moisture holding capacity (Clay, Sand, Organic Matter). Included in the model set are the best model excluding interaction terms (a), the best model with a single TWIsi interaction term (b) the best model with a single Slope interaction term (c) and a model with both Slope and TWIsi interaction terms (d). The selected model (d) is highlighted in bold. All models included a random intercept. Parameters Number Δ AIC Clay 1 2535.7 OM 1 2454.3 pH 1 2352.7 TWIsi 1 2336.6 Sand 1 1812.8 Restricted 1 1674.7 Slope 1 984.1 TWIsi Sand Clay OM pH Restr 6 827.9 TWIsi Slope Sand Clay OM pH 6 343.6 TWIsi Slope Sand OM pH Restr 6 325.0 TWIsi Slope Clay OM pH Restr 6 180.8 TWIsi Slope Sand Clay pH Restr 6 148.9 TWIsi Slope Sand Clay OM Restr 6 120.8 Slope Sand Clay OM pH Restr 6 55.5 TWIsi Slope Sand Clay OM pH Restr (a) 7 53.6 Model (a) + TWIsi x Sand 8 8 8 8 8 8 55.6 53.1 7.5 44.8 43.3 43.3 9 0.0 Model (a) + TWIsi x OM Model (a) + TWIsi x Clay (b) Model (a) + Slope x Sand Model (a) + Slope x Clay Model (a) + Slope x OM (c) Model (a) + Slope x OM + TWIsi x Clay (d) 32 Table 16. Summary of BTPD habitat suitability models selected for each of 7 study sites on the basis of BTPD colonies locations monitored at the site during 2001-2010. We also present coefficients of the global model (i.e. fit to data from all 7 sites combined = topoedaphic model in Table 6) for comparison. Study Site AuROC Carrizo 0.6316 Intercept -8.9027 TWIsi -0.00945 Slope -0.6292 Sand 0.007488 Clay 0.05429 OM 0.223 pH 0.9234 DTR -0.00034 Slope*Clay 0.01672 Slope*OM TWIsi*Clay TWIsi*OM TWIsi*Sand Cimarron 0.8384 Kiowa 0.6087 -8.6946 0.1233 -0.2287 -0.01474 0.1092 2.26 -0.3456 0.03213 0.006239 -4.6525 0.0744 0.1981 0.0309 1.1364 0.2156 0.002723 Local Models Pawnee East Pawnee West 0.5829 0.5705 -4.0264 0.06736 -0.2961 0.02758 0.03551 0.3721 0.1498 0.002009 0.005261 -0.4113 -0.06504 0.08898 0.0735 0.1085 -0.01575 0.004683 0.4591 Rita Blanca 0.5889 Timpas 0.6694 Global Model 0.6343 -1.504 0.08225 -0.8128 0.008761 0.04179 -8.8533 0.3212 -0.2114 -0.0340 -0.0326 -0.1724 1.1344 0.0126 -2.9322 0.08353 -0.0523 -0.0113 0.01582 0.3941 0.2504 0.00198 -0.7616 -0.0118 -0.089 0.000475 0.01435 -0.18 -0.00219 -0.00192 -0.03816 -0.0012 33 Table 17. RSF probability cutoff values that correspond to false negative error rates of 5, 10 and 15% for each of the local models. These probability cutoffs were used to generate the RSF category maps (Maps 22-28 in Table 18). False Negative Error Rate Carrizo 5% 10% 15% 0.19620 0.32000 0.40790 35.93 30.55 26.07 0.02858 0.03377 0.04955 22.17 12.21 7.47 0.02368 0.03035 0.03549 42.00 34.66 27.76 0.21770 0.28570 0.33380 False Positive Rate 39.04 33.74 28.69 Probability Cutoff 42.63 37.67 32.40 False Positive Rate 42.63 37.67 32.40 0.21300 0.26594 0.29285 40.18 33.68 28.64 0.10050 0.17333 0.22265 32.81 26.29 22.08 Probability Cutoff: False Positive Rate: Cimarron Probability Cutoff False Positive Rate Kiowa Probability Cutoff False Positive Rate Pawnee East Pawnee West Rita Blanca Probability Cutoff Probability Cutoff False Positive Rate Timpas Probability Cutoff False Positive Rate 34 Table 18. Index of maps of BTPD habitat suitability generated based on the final selected local model fit to data from each study site separately (see Table 15 for coefficients). Study Site Map # Model Output Type Map Title Carrizo 15 Local (Table 15) RSF Probability Map of Local Topoedaphic RSF Model: Carrizo Study Area Cimarron 16 Local (Table 15) RSF Probability Map of Local Topoedaphic RSF Model: Cimarron Study Area Kiowa 17 Local (Table 15) RSF Probability Map of Local Topoedaphic RSF Model: Kiowa Study Area Pawnee East 18 Local (Table 15) RSF Probability Map of Local Topoedaphic RSF Model: Pawnee East Study Area Pawnee West 19 Local (Table 15) RSF Probability Map of Local Topoedaphic RSF Model: Pawnee West Study Area Rita Blanca 20 Local (Table 15) RSF Probability Map of Local Topoedaphic RSF Model: Rita Blanca Study Area Timpas 21 Local (Table 15) RSF Probability Map of Local Topoedaphic RSF Model: Timpas Study Area Carrizo 22 Local (Table 15) RSF Category Map of Local Topoedaphic Model Categories: Carrizo Study Area Cimarron 23 Local (Table 15) RSF Category Map of Local Topoedaphic Model Categories: Cimarron Study Area Kiowa 24 Local (Table 15) RSF Category Map of Local Topoedaphic Model Categories: Kiowa Study Area Pawnee East 25 Local (Table 15) RSF Category Map of Local Topoedaphic Model Categories: Pawnee East Study Area Pawnee West 26 Local (Table 15) RSF Category Map of Local Topoedaphic Model Categories: Pawnee West Study Area Rita Blanca 27 Local (Table 15) RSF Category Map of Local Topoedaphic Model Categories: Rita Blanca Study Area Timpas 28 Local (Table 15) RSF Category Map of Local Topoedaphic Model Categories: Timpas Study Area 35 Table 19. Summary of the percent of the landscape in each of 4 habitat suitability categories based on the best global model (fit to all colonies at all study sites) versus the best local model (fit only to colonies within each study site). The magnitude of spatial inconsistency between local versus global models is shown as the percent of land area predicted to be in category 1 by one model but in category 4 by the other model. Global Topoedaphic + Precipitation Model Local Topoedaphic Model Spatial Inconsistency between Global vs. Local Models % of Landscape in RSF Category % of Landscape in RSF Category % of Landscape Total Hectares 102692 1 2 3 4 1 2 3 4 1 Global, 4 Local 1 Local, 4 Global Sum 47 4 4 45 45 11 8 36 0.04 0.16 0.2 Cimarron 43978 61 5 8 26 79 1 3 17 2.30 0.00 2.3 Kiowa 23777 28 11 18 43 51 19 5 25 0.16 4.00 4.2 Pawnee East Pawnee West 37901 31 17 15 37 39 9 8 44 3.60 7.80 11.4 11 17 25 47 14 10 10 66 1.20 0.60 1.8 Rita Blanca 46841 37952 23 14 4 59 19 12 10 59 3.00 5.10 8.1 Timpas 69738 9 5 8 78 20 5 3 72 0.25 4.20 4.5 Study Site Carrizo 36 clusters). This disagreement was due to the local model including a large positive coefficient for Sand, which was in contrast to the negative or small positive coefficients for Sand in all other local and global models (Table 16). The second-highest rate of local versus global model disagreement was for the Rita Blanca National Grassland (8.1% Table 19). This error is most likely related to the fact that the Rita Blanca site spans two different counties and states, the effects of which are addressed in greater detail in the Discussion. Overall, the generally strong match between global and local models provides strong support for the application of the global topoedaphic + precipitation model in predicting habitat suitability across the broader project area. Ecological Site-based second-order models An ecological site is defined as “a distinctive kind of land with specific physical characteristics that differs from other kinds of land in its ability to produce a distinctive kind and amount of vegetation” (USDA 1997). The quantitative soil parameters considered in the previous BTPD habitat suitability models (e.g. Sand, Clay, Organic matter, pH and Depth to restricted layer) are among the soil characteristics that have been used to classify map units in the SSURGO database into different ecological sites. Types of ecological sites and their definitions vary across the Great Plains, and have typically been standardized by NRCS across counties at the level of Major Land and Resource Areas (MLRAs). As noted by NRCS (http://esis.sc.egov.usda.gov/), MLRA’s are used by the Natural Resources Conservation Service “in the planning, design, implementation, and evaluation of natural resource management activities. MLRA boundaries reflect nearly homogenous areas of landuse, elevation, topography, climate, water resources, potential vegetation, and soils.” In some cases, a portion of the ecological sites in a given MLRA also occur in adjacent MLRAs, but it is also possible for adjacent MLRAs to have largely different sets of ecological sites. We analyzed BTPD habitat suitability relative to ecological sites by first comparing the types of ecological sites present (i.e. as defined by NRCS) at each of the 7 study sites (Table 18). We note that in a few cases, we combined two rare ecological sites with strong similarities into a single category for these analyses (e.g. “Draw” and “Swale” combined into Draw/Swale and Sandstone and Sandstone Breaks combined in the Sandstone/Sandstone Breaks). Based on this analysis, study sites were placed into 3 groups corresponding to (1) 4 sites in MLRA 67B and 69 that included a total of 14 ecological sites, of which 7 were present across 3 or all of the 4 study sites, (2) 2 sites in MLRAs 77A/77B, which included 18 ecological sites, of which 7 occurred in both study sites, and (3) 1 site occurring at the boundary of MLRAs 72/77A, which had 6 ecological sites, of which 4 were unique to the study site. For each group we considered a model that only included Ecological Site as a categorical predictor of BTPD habitat quality, and a second model that included the topographic predictors (TWIsi and Slope) in addition to Ecological Site. Coefficients for each ecological site provide a measure of that ecological site’s value as habitat relative to a “reference” ecological site. For each analysis, we identified the ecological site for which the ratio of used:available pixels was closest to 1.0, and used it as the 37 reference site. It is possible for ecological sites with a negative coefficient to be widely used by BTPD, as the coefficient’s value represents a ranking relative to the reference group. For Group 1 (MLRAs 67B/69), the model based on ecological site as the sole predictor was a substantial improvement over the null model (ΔAIC = 8,617.0; AuROC = 0.5711). Five ecological sites had neutral (not different from zero) or positive coefficients: Salt Flat, Alkaline Plains, Overflow, Clayey Plains, and Loamy Plains. Large negative coefficients were observed for the Shallow Siltstone, Limestone Breaks, Shaly Plains, Sandy Bottomland, Deep Sand, and Sandstone/Sandstone Breaks. Including TWIsi and Slope in addition to Ecological Site further improved the model (ΔAIC relative to the Ecological Site-only model = 2263.8; AuROC = 0.5969), but had minimal influence on relative rankings of the ecological sites other than Salt Flat having a small but significant negative coefficient relative to Alkaline Plains (Table 20). For Group 2 (MLRAs 77A/B), the model with ecological site as the sole predictor was a substantial improvement over the null model (ΔAIC = 3,581; AuROC = 0.5866; Table 21). Seven ecological sites had neutral (not different from zero) or positive coefficients: Sandy Loam, Loamy Upland, Deep Hardland, Salt Flat, Loamy Bottomland, Gravelly Loam, and Draw/Swale. Large negative coefficients were found for the Shallow Siltstone, Sandy Plains, Sand Hills, Hardland Slopes, Sandy Bottomland, and Gravelly ecological sites. Smaller but significantly negative coefficients were observed for the Very Shallow, Playa, High Lime, and Limy Upland ecological sites. The Malpais Upland ecological site was too rare in the dataset to be analyzed effectively. Including TWIsi and Slope in addition to Ecological Site further improved the model (ΔAIC relative to the Ecological Site-only model = 1412, AuROC = 0.6054), but had minimal influence on relative rankings of the ecological sites (Table 22). For Group 3 (Cimarron National Grassland; MLRAs 72/77A), the model based on ecological site as the sole predictor was a substantial improvement over the null model (ΔAIC = 18,221; AuROC = 0.7574; Table 23). Two ecological sites had neutral or positive coefficients: Limy Upland and Loamy Upland. Negative coefficients were observed for all ecological sites with soils of high sand content: Sandy, Sands, Sandy Lowland, and Choppy Sands. Including TWIsi and Slope in addition to Ecological Site further improved the model (ΔAIC relative to the Ecological Site-only model = 3518, AuROC = 0.8253, but did not influence relative rankings of the ecological sites (Table 24). 38 Table 20. Number of BTPD colony clusters in which different ecological sites occurred (either in used and/or available pixels) for each study site. Based on variation in MLRAs and the types of ecological sites present at each study site, the sites were grouped into 3 separate datasets for analysis: (1) Pawnee E, Pawnee W, Carrizo and Timpas (ecological sites in light grey shading), (2) Kiowa and Rita Blanca (ecological sites in bold type), and (3) Cimarron (ecological sites in dark grey shading). Pawnee E Pawnee W Carrizo Timpas Kiowa Rita Blanca Cimarron MLRA(s): 67B 67B 67B 69 77A/B 77A/B 72/77A Loamy Plains 10 15 28 19 Gravel Breaks 4 9 18 1 Sandstone/Sandstone Breaks 3 5 9 5 Shaly Plains 4 11 Clayey Plains 3 7 Shallow Siltstone 2 1 Overflow 1 8 Site: Deep Sand 2 Limestone Breaks 8 10 1 1 17 Saline Overflow 17 Alkaline Plains 15 Sandy Bottomland 1 2 Salt Flat 3 4 Sandy Plains 6 14 5 18 6 4 9 8 Deep Hardland 18 4 Sandy Loam 16 9 Very Shallow 14 5 High Lime 5 10 Draw/Swale 3 3 Sand Hills 3 1 Playa 10 Hardland Slopes 5 Loamy Bottomland 3 26 1 Gravelly Loam 8 Shallow Sandstone 3 Gravelly 2 Malpais Upland 1 Limy Upland 17 Loamy Upland 7 7 9 Sandy 9 Sands 8 Choppy Sands 7 Sandy Lowland 4 39 Table 21. Coefficients for a resource selection function based upon ecological sites in eastern Colorado, fit to BTPD colonies at 4 study sites (Group 1 in Table 20), where available habitat was defined using a 2 km buffer around colony clusters. Coefficient estimates for each ecological site reflect that sites value as habitat relative to the Alkaline Plains site. One site (Salt Flat) did not differ significantly in value from Alkaline Plains. Alkaline Plains was selected as the reference group because this ecological site had a ratio of used:available pixels of 1.04, which was closest to 1 of all ecological sites. Ecological Site Estimate Standard Error t Value Pr > |t| Shallow Siltstone -6.6143 Limestone Breaks -2.9473 2.3315 -2.84 0.0046 0.1522 -19.36 <.0001 Shaly Plains -2.0058 0.1355 -14.8 <.0001 Sandy Bottomland -1.6221 0.0863 -18.8 <.0001 Deep Sand -1.4945 0.113 -13.22 <.0001 Sandstone -1.3706 0.07744 -17.7 <.0001 Gravel Breaks -0.8519 0.06799 -12.53 <.0001 Sandy Plains -0.481 0.06203 -7.76 <.0001 Saline Overflow -0.3689 0.08973 -4.11 <.0001 Salt Flat -0.02793 0.07536 -0.37 0.711 Alkaline Plains Loamy Plains 0 0.2462 Reference Group 0.06091 4.04 <.0001 Overflow 0.6303 0.07562 8.33 <.0001 Clayey Plains 0.7216 0.1002 7.2 <.0001 40 Table 22. Coefficients for a resource selection function based on ecological sites in eastern Colorado (MLRAs 67B/69) plus two topographic parameters (TWIsi and Slope), fit to BTPD colonies at 4 sites (Group 1 in Table 18), where available habitat was defined using a 2 km buffer around colony clusters. Coefficient estimates for each ecological site reflect value as habitat relative to the Alkaline Plains ecological site. Alkaline Plains was selected as the reference group because this ecological site had a ratio of used:available pixels of 1.04, which was closest to 1 of all ecological sites. Predictor Estimate Standard Error t Value Pr > |t| TWIsi 0.01231 0.001748 7.04 <.0001 Slope -0.1867 0.004618 -40.42 <.0001 Shallow Siltstone -6.6152 2.2194 -2.98 0.0029 Limestone Breaks -2.6315 0.1481 -17.77 <.0001 Shaly Plains -1.9425 0.1362 -14.26 <.0001 Sandy Bottomland -1.6157 0.08684 -18.6 <.0001 Deep Sand -1.4349 0.1132 -12.68 <.0001 Sandstone -0.9952 0.0786 -12.66 <.0001 Gravel Breaks -0.6851 0.06873 -9.97 <.0001 Sandy Plains -0.4711 0.06267 -7.52 <.0001 Saline Overflow -0.4151 0.09068 -4.58 <.0001 Salt Flat -0.1509 0.07605 -1.98 0.0472 Alkaline Plains Loamy Plains 0 0.2206 Reference Group 0.0615 3.59 0.0003 Overflow 0.4991 0.07633 6.54 <.0001 Clayey Plains 0.5752 0.1008 5.71 <.0001 41 Table 23. Coefficients for a resource selection function based upon ecological sites occurring in counties of northeast New Mexico, Oklahoma panhandle, and Texas Panhandle (MLRAs 77A/B; Group 2 in Table 18), where available habitat was defined using a 2 km buffer around colony clusters. Coefficient estimates for each ecological site reflect that sites value as habitat relative to the Sandy Loam site. Sandy Loam was used as the reference group because this ecological site had a ratio of used:available pixels of 1.11, which was closest to 1 of all ecological sites. Ecological Site Model Estimate Standard Error t Value Pr > |t| Shallow Sandstone -6.7203 1.7911 -3.75 0.0002 Malpais Upland -6.6749 5.9933 -1.11 0.2654 Sandy Plains -3.3249 0.1353 -24.58 <.0001 Sand Hills -2.841 0.7733 -3.67 0.0002 Hardland Slopes -1.628 0.1497 -10.87 <.0001 Sandy Bottomland -0.7117 0.2467 -2.89 0.0039 Gravelly -0.6677 0.06216 -10.74 <.0001 Very Shallow -0.4296 0.04977 -8.63 <.0001 Playa -0.3956 0.1101 -3.59 0.0003 High Lime -0.2988 0.04113 -7.27 <.0001 Limy Upland -0.19 0.02821 -6.73 <.0001 Sandy Loam 0 . . . Loamy Upland 0.05808 0.08136 0.71 0.4753 Deep Hardland 0.351 0.0308 11.4 <.0001 Salt Flat 0.5965 0.06467 9.22 <.0001 Loamy Bottomland 0.6459 0.1529 4.22 <.0001 Gravelly Loam 0.665 0.0408 16.3 <.0001 Draw and Swale 1.941 0.09697 20.02 <.0001 42 Table 24. Coefficients for a resource selection function based upon ecological sites occurring in counties of northeast New Mexico, Oklahoma panhandle, and Texas Panhandle (MLRAs 77A/B; Group 2 in Table 18) plus topographic parameters (TWIsi, Slope), where available habitat was defined using a 2 km buffer around colon clusters. Coefficient estimates for each ecological site reflect that sites value as habitat relative to the Sandy Loam site. Sandy Loam was used as the reference group because this ecological site had a ratio of used:available pixels of 1.11, which was closest to 1 of all ecological sites. Ecological Site + Topography Model Estimate Standard Error t Value Pr > |t| TWIsi 0.04308 0.002666 16.16 <.0001 Slope -0.4234 0.01654 -25.59 <.0001 Shallow Sandstone -6.1457 1.4424 -4.26 <.0001 Malpais Upland -5.8808 5.8672 -1 0.3162 Sandy Plains -3.3641 0.1362 -24.69 <.0001 Sand Hills -2.7484 0.7573 -3.63 0.0003 Hardland Slopes -1.1604 0.1533 -7.57 <.0001 Playa -1.0369 0.1128 -9.2 <.0001 Sandy Bottomland -0.856 0.2622 -3.26 0.0011 Gravelly -0.4357 0.06389 -6.82 <.0001 High Lime -0.3222 0.04159 -7.75 <.0001 Limy Upland -0.2623 0.02854 -9.19 <.0001 Very Shallow -0.2216 0.05063 -4.38 <.0001 Sandy Loam 0 Reference Group Deep Hardland 0.2051 0.03122 6.57 <.0001 Loamy Upland 0.2123 0.08394 2.53 0.0115 Loamy Bottomland 0.514 0.154 3.34 0.0008 Gravelly Loam 0.6631 0.041 16.17 <.0001 Salt Flat 0.6715 0.06559 10.24 <.0001 Draw and Swale 1.6951 0.09755 17.38 <.0001 43 Table 25. Coefficients for a resource selection function based upon ecological sites in southwestern Kansas (Cimarron National Grassland), where available habitat was defined using a 2 km buffer around colony clusters. Coefficient estimates for each ecological site reflect that sites value as habitat relative to the Limy Upland site. Limy Upland was used as the reference group because this ecological site had a ratio of used:available pixels of 1.02, which was closest to 1 of all ecological sites. Ecological Site Model Choppy Sand Sandy Lowland Sands Sandy Limy Upland Loamy Upland Estimate -4.9985 -3.9526 -3.0221 -1.5382 0 1.85 Standard Error 0.2108 0.1032 0.08972 0.07884 Reference Group 0.02997 t Value -23.71 -38.29 -33.68 -19.51 Pr > |t| <.0001 <.0001 <.0001 <.0001 61.73 <.0001 Table 26. Coefficients for a resource selection function based upon ecological sites in southwestern Kansas (Cimarron National Grassland), where available habitat was defined using a 2 km buffer around colony clusters. Coefficient estimates for each ecological site reflect that sites value as habitat relative to the Limy Upland site. Limy Upland was used as the reference group because this ecological site had a ratio of used:available pixels of 1.02, which was closest to 1 of all ecological sites. Ecological Site + TWIsi + Slope Model TWIsi Slope Choppy Sands Sandy Lowland Sands Sandy Limy Upland Loamy Upland Estimate 0.1101 -0.08377 -5.2822 -4.6736 -3.2556 -1.7566 0 1.489 44 Standard Error 0.002011 0.008445 0.2119 0.1055 0.09135 0.08186 Reference Group 0.03106 t Value 54.77 -9.92 -24.92 -44.31 -35.64 -21.46 Pr > |t| <.0001 <.0001 <.0001 <.0001 <.0001 <.0001 47.93 <.0001 Comparison of spatial scales for evaluating second-order models All previous results are based on models where colony clusters were defined by a 2-km linkage rule, and then available habitat surrounding the cluster was defined by a 2-km buffer distance. We evaluated how reducing this distance influenced model results by comparing the 2km model results to the same model evaluation process based on a 0.5-km linkage rule for colony clusters and an associated 0.5-km buffer distance for available habitat. By reducing this distance to 0.5 km, the number of different colony clusters in the dataset increases substantially because colonies separated by 0.5 – 2.0 km are now considered separate (independent) relative to one another. This can potentially increase the power of our tests of model likelihood and fit. At the same time, we decrease the area of the landscape from which available pixels are selected (using 0.5 km buffer distance rather than 2.0 km), thereby potentially reducing our ability to discriminate features within the landscape which characterize the most suitable BTPD habitat. Using the 0.5-km linkage and buffer distance, we identified a total of 216 colony clusters for analysis (Table 2) which was double the number of colony clusters using the 2-km linkage rule. Model selection based on AIC identified a model with all 7 topoedaphic predictors plus interactions between TWIsi x Clay and Slope x OM as the most parsimonious model within the set of models that did not include climate interactions (Table 26). Incorporating climate variables yielded a model that included interaction terms for Precipitation x Clay (as in the 2-km model) plus Temperature x OM (not included in the 2-km model; Table 27 and 28). The models differ in terms of the sign of the Clay coefficient because the 0.5-km model includes an interaction term for TWIsi x Clay interaction term, while the 2-km model included a TWIsi x Sand interaction. The models differ in terms of the sign of the OM coefficient because the 0.5km model included a Temperature x OM interaction while the 2-km model does not. Overall, however, both models show high similarity in that both include an interaction between TWIsi and soil texture, an interaction between slope and soil organic matter content, and an interaction between soil clay content and the precipitation gradient (Table 28). Although the Temperature x OM interaction term was retained in the 0.5-km model based on minimization of AIC, the magnitude of the effect of this term on habitat suitability was small. Predictions of the two models showed a high degree of consistency for 6 of the 7 study sites (Table 29b), and indicating that our models are robust across a range of linkage and buffer distances. The one notable exception was on the Cimarron National Grassland, where 18% of the landscape that was mapped as high-quality habitat 0.5-km model was mapped as low-quality habitat by the 2-km model. Inspection of the map outputs shows that these two models produced similar predictions for the upland region north of the Cimarron River, but differed in some areas of sandy soils south of the Cimarron River. Colonies on soils south of the river are more restricted in extent and show lower expansion rates than colonies on soils north of the river, which is more in accord with the 2-km model’s prediction that the region north of the river was largely suitable habitat, while the region south of the river was a more complex mosaic of habitat in categories 1, 2 and 3. We used the 2-km model as our final selected second-order model 45 because relative to the 0.5-km model, it was based upon available pixels sampled from a larger proportion of the landscape, had the greater predictive ability (greater AuROC), and included fewer parameters. Table 27. Summary of the set of topoedaphic RSF models considered for the dataset based on a 0.5 km linkage and buffer distance. The selected model is shown in bold. Model Parameters AIC Δ AIC pH 1 457721.1 13040.0 DTR 1 456956.6 12275.5 OM 1 456849.5 12168.4 TWIsi 1 456310.1 11629.0 Slope 1 452920.3 8239.2 Sand 1 452745 8063.9 Clay 1 449364.2 4683.1 TWIsi Sand Clay OM pH DTR 6 447405.5 2724.4 TWIsi Slope Sand OM pH DTR 6 447276 2594.9 TWIsi Slope Sand Clay OM DTR 6 445437.3 756.2 TWIsi Slope Sand Clay pH DTR 6 445401.5 720.4 Slope Sand Clay OM pH DTR 6 445260.6 579.5 TWIsi Slope Clay OM pH DTR 6 445206.9 525.8 TWIsi Slope Sand Clay OM pH 6 445171.4 490.3 TWIsi Slope Sand Clay OM pH DTR (a) 7 445142.9 461.8 (a) + TWIsi x OM 8 445128.6 447.5 (a) + TWIsi x Sand 8 445102.5 421.4 (a) + TWIsi x Clay 8 444799.5 118.4 (a) + Slope x Sand 8 445140.8 459.7 (a) + Slope x Clay 8 445102.6 421.5 (a) + Slope x OM 8 445055.2 374.1 (a) + TWIsi x Clay + Slope x OM 9 444681.1 0.0 46 Table 28. Summary of models that include interactions with mean annual precipitation and/or mean monthly maximum temperature for colony clusters and available habitat defined based on the 0.5-km linkage and buffer distance. Letters in parentheses show the best model including an interaction with precipitation (a) , the best model including an interaction with temperature (b), and the best model with both temperature and precipitation (c). The final selected global model for the 0.5-km linkage/buffer distance is shown in bold. AuROC 0.5841 AIC 444681.1 # of Random Coefficients 0 Δ AIC TWIsi Slope Sand Clay OM pH Restr TWIsi*Clay SL*OM Precip Precip Precip*TWIsi 0.5851 444521.8 1 148.2 TWIsi Slope Sand Clay OM pH Restr TWIsi*Clay SL*OM Precip Precip Precip*OM 0.5847 444411.8 1 38.2 TWIsi Slope Sand Clay OM pH Restr TWI*Clay SL*OM Precip Precip Precip*Clay 0.5863 444377.9 1 4.3 TWIsi Slope Sand Clay OM pH Restr TWIsi*Clay SL*OM Precip Precip Precip*Sand (a) 0.5858 444373.6 1 0.0 TWIsi Slope Sand Clay OM pH Restr TWIsi*Sand SL*OM Temp Precip Temp*Clay 0.5848 444560.5 1 186.9 TWIsi Slope Sand Clay OM pH Restr TWIsi*Sand SL*OM Temp Precip Temp*TWIsi 0.5843 444479.1 1 105.5 TWIsi Slope Sand Clay OM pH Restr TWIsi*Sand SL*OM Temp Precip Temp*OM (b) 0.5845 444392.5 1 18.9 0.5861 443219.4 2 71.3 0.5869 443148.1 2 0.0 Best model without climate interactions: TWIsi Slope Sand Clay OM pH Restr TWIsi*Clay Slope*OM Interactions with Precipitation: Interactions with Temperature: Interactions with Precipitation and Temperature: TWIsi Slope Sand Clay OM pH Restr TWIsi*Sand SL*OM Precip Precip Precip*Sand Temp Temp*OM TWIsi Slope Sand Clay OM pH Restr TWIsixSand SL*OM Precip Precip*Clay Temp Temp*OM (c) 47 Table 29a. Comparison of final selected topoedaphic and topoedaphic + climate models for datasets based on a 2km versus 0.5-km colony linkage rule and buffer distance. Best Topoedaphic Model Best Topoedaphic + Climate Model 2 km 0.5 km 2 km 0.5 km 113 216 113 216 AuROC 0.6343 0.5841 0.6418 0.5869 Intercept -2.9322 -2.7656 5.6411 -4.0769 TWIsi 0.08353 0.08166 0.06968 0.07852 Slope -0.05225 -0.1242 -0.08081 -0.1089 Sand -0.01129 -0.00301 -0.01085 -0.00336 Clay 0.01582 0.05127 -0.2531 0.1327 OM 0.3941 0.2258 0.4315 -0.3983 0.3203 0.217 0.000555 0.000728 # of Colony Clusters pH 0.2504 0.158 Restr 0.00198 0.000738 TWIsi x Sand -0.00119 -0.00083 -0.00241 TWIsi x Clay -0.06223 -0.00231 -0.09991 Precipitation -0.02261 0.007059 Precipitation x Clay Temperature 0.000675 -0.00022 Slope x OM -0.08895 -0.08192 -0.06428 Temperature x OM 0.03672 48 Table 29b. Comparison of mapped distribution of BTPD habitat suitability categories based on the final selected topoedaphic + climate models for datasets derived from a 2-km versus a 0.5-km linkage rule and buffer distance. Study Site Carrizo Cimarron Kiowa Pawnee East Pawnee West Rita Blanca Timpas 2km Global Topoedaphic + Precipitation Model 500 m Global Topoedaphic + Precipitation Model (% of Landscape in RSF Category) (% of Landscape in RSF Category) Spatial Inconsistency between 2 km and 0.5 km models (% of Landscape) 1 for 2km, 1 for 0.5 4 for 0.5 km, 4 for 2 km km 1.3 0.0 18.5 0.0 0.7 0.0 1 47 61 28 2 4 5 11 3 4 8 18 4 45 26 43 1 24 46 12 2 16 4 14 3 12 1 18 4 48 49 56 31 17 15 37 39 23 15 23 0.0 1.0 1.1 11 23 9 17 14 5 25 4 8 47 59 78 35 8 25 28 10 18 13 12 25 24 70 32 0.0 0.0 0.0 6.7 0.0 6.3 6.7 0.0 6.3 49 Sum 1.3 18.5 0.7 Third-order Habitat Selection Our assessment of third-order habitat selection defined the habitat available to a colony locally on the basis of the direction and extent of that colony’s expansion over a series of more than 3 consecutive years. We view this analysis as being similar to the selection of habitat within an animal’s home range, where the home range is defined on the basis of the outermost positions in a set of an animal’s locations over a period of time. Across all 7 study sites, we identified a total of 152 colonies meeting the criteria of having been mapped for a series of 4 or more consecutive years where the colony was stable or expanding. For the set of models that did not include interaction terms, the most parsimonious model included all 7 topoedaphic predictors (TWIsi, Slope, % Sand, % Clay, pH, % Organic matter, and Depth To a Restricted layer; Table 30), which was an improvement of all competing models with 6 or fewer predictors (Δ AIC > 6). Of the potential models including interactions between slope and soil parameters, the most parsimonious included a Slope x Organic Matter interaction (Table 4; Δ AIC relative to no interaction model = 251.3). Of the potential models including interactions between TWIsi and soil parameters, the most parsimonious model included a TWIsi x Sand interaction (Table 4; Δ AIC relative to no interaction model = 191.0). Including both interaction terms further reduced AIC by 49.2 relative to the best model with a single interaction term (Table 4). The selected model based on topoedaphic predictors had an area under the ROC curve of 0.5928, with coefficients presented in Table 32. Consideration of an expanded model set that allowed for interactions between climate (precipitation and temperature) and topoedaphic predictors showed the most parsimonious model to include an interaction between precipitation and soil clay content plus an interaction between temperature and soil organic matter (Table 31). This model both increased model predictive ability (AuROC = 0.5960) and was more parsimonious relative to the best topoedaphic-only model (Δ AIC = 1033.6). Our final selected third-order RSF for prairie dog habitat therefore included 7 topoedaphic predictors, Precipitation, Temperature, and TWIsi x Sand, Slope x Organic matter, Temperature x Organic Matter, and Precipitation x Clay interactions (Tables 31 and 32). 50 Table 30. Summary of third-order RSF model set including direct effects of up to 7 topoedaphic predictors, and potential interactions between topographic predictors (TWIsi and/or Slope) and soil predictors that influence soil moisture holding capacity (Clay, Sand, Organic matter). Included in the model set are the best model excluding interaction terms (a), the best model with a single TWIsi interaction term (b), the best model with a single Slope interaction term (c) and the selected model with both Slope and TWIsi interaction terms (d). All models included a random intercept. Predictors Parameters AIC ΔAIC pH 1 342943.1 11244.1 DTR 1 342621 10922.0 TWIsi 1 342018.5 10319.5 OM 1 341257.6 9558.6 Slope 1 339409.5 7710.5 Sand 1 336926.4 5227.4 Clay 1 334799.9 3100.9 TWIsi Slope Sand OM pH DTR 6 333460.9 1761.9 TWIsi Sand Clay OM pH DTR 6 333441.3 1742.3 TWIsi Slope Sand Clay pH DTR 6 332221 522.0 TWIsi Slope Clay OM pH DTR 6 332052.9 353.9 TWIsi Slope Sand Clay OM DTR 6 332011.3 312.3 Slope Sand Clay OM pH DTR 6 331964.4 265.4 TWIsi Slope Sand Clay OM pH 6 331945.2 246.2 TWIsi Slope Sand Clay OM pH DTR (a) 7 331939.2 240.2 (a) + TWIsi x OM 8 331930.5 231.5 (a) + TWIsi x Sand 8 331918.1 219.1 (a) + TWIsi x Clay (b) 8 331748.2 49.2 (a) + Slope x Clay 8 331933.2 234.2 (a) + Slope x Sand 8 331941.2 242.2 (a) + Slope x OM (c) 8 331901.3 202.3 (a) + TWIsi x Clay + Slope x OM (d) 9 331699 0.0 51 Table 31. Summary of third-order RSF models that include interactions with mean annual precipitation and/or mean monthly maximum temperature. The final selected global model for second-order habitat selection is shown in bold. Parameters # of Random Coefficients AuROC AIC Best model without climate interactions (d) 9 0 0.5928 331699.0 (d) + Temp + Temp x TWIsi 11 1 0.5960 331493.5 674.5 (d) + Temp + Temp x OM 11 1 0.5965 331188.5 369.5 (d) + Temp + Temp x Clay 11 1 0.5949 330990.2 171.2 (d) + Precip + Precip x TWIsi 11 1 0.5943 331545.9 726.9 (d) + Precip + Precip x Sand 11 1 0.5870 331137.6 318.6 (d) + Precip + Precip x OM 11 1 0.5949 331045.9 226.9 (d) + Precip + Precip x Clay 11 1 0.5956 330819.0 0.0 (d) + Temp + Temp x Clay + Precip + Precip x Clay 13 2 0.5965 330806.7 141.3 (d) + Temp + Temp x OM + Precip + Precip x Clay 13 2 0.5960 330665.4 0.0 Predictors ΔAIC Table 32. Coefficients and associated standard errors for the best model including topoedaphic predictors plus interactions with mean annual precipitation and mean maximum monthly temperature (model selection statistics in Tables 30 and 31). For a summary of habitat suitability maps based on this model, see Table 33. Coefficient Intercept Standard Error 2.8934 2.7608 TWIsi 0.07622 0.005554 Slope -0.1316 0.02084 Sand -0.0051 -- Clay -0.1576 0.00893 OM -0.4982 0.1529 pH 0.2499 0.01883 DTR -0.00042 -- Temperature -0.06196 0.09925 Precipitation -0.01275 0.003679 TWIsi x Clay -0.00251 -- Slope x OM -0.05797 -- Temperature x OM 0.04676 0.007904 Precipitation x Clay 0.000528 -- 52 Coefficients of the selected third-order model including climate (Table 32) show that BTPD habitat suitability increases with increasing soil pH and decreases with increasing soil sand content across all levels of the other predictors. BTPD habitat suitability declines with increasing slope, but does so more rapidly on soils with high organic matter content than on soils with low organic matter content (Figure 8). Note that this relationship between slope, organic matter, and habitat suitability is nearly identical to that predicted by the global second-order model (Figure 6). Figure 9. Predicted relative BTPD habitat suitability as a function of slope for varying levels of soil organic matter content based on the final selected third-order Topoedaphic + Climate model (Table 32). Figure 10. Predicted relative BTPD habitat suitability as a function of TWIsi for varying levels of soil organic matter content based on the final selected third-order Topoedaphic + Climate model (Table 32). The TWIsi x Clay interaction (Figure 10) shows that habitat suitability increases with increasing TWIsi for soils with clay content less than 30%. However, for soils with clay content greater than 30%, the relationship between habitat suitability and TWIsi becomes negative (Figure 10), suggesting that on soils with high water holding capacity, the need to select for areas that receive runon is alleviated. The relationship between habitat suitability, soil clay content, and TWIsi depicted in Figure 10 is the inverse of the relationship between habitat suitability, soil sand content, and TWIsi represented in the global second-order model (Figure 7). Predicted habitat suitability increases with increasing soil clay content, but due to the Precipitation x Clay interaction, the strength of BTPD selection for high-clay soils increases with increasing mean annual precipitation. Conversely, the selection for low-clay soils is greatest at the dry end of the precipitation gradient. This relationship is similar to that predicted by the global second-order 53 model (Figure 8), except that in the third-order model, the slope of the relationship between clay and habitat suitability remains positive across the full range of mean annual precipitation found within the study region. Figure 11. Predicted relative BTPD habitat suitability as a function of soil clay content for varying levels of mean annual precipitation based on the final selected thirdorder Topoedaphic + Climate model (Table 32). Figure 12. Predicted relative BTPD habitat suitability as a function of soil organic matter content for varying levels of mean maximum monthly temperature based on the final selected third-order Topoedaphic + Climate model (Table 32). Finally, predicted habitat suitability increases with soil organic matter content, and due to the temperature x Organic Matter interaction, the slope of this relationship increases with increasing mean monthly maximum temperature. However, in contrast to the Precipitation x Clay interaction where a large shift in Clay slope occurs across the precipitation gradient, we found only a minor shift in the Organic Matter slope across the temperature gradient (Figure 11 vs. 12). Thus, although the Temperature x OM interaction was retained in the final selected third-order RSF model, its influence on predicted habitat suitability values is minor compared to the other three interaction terms. 54 Table 33. Summary of the percent of the landscape in each of 4 BTPD habitat suitability categories based on the final selected second-order RSF including topoedaphic and climate predictors (Table 6) for 73 counties in the southwestern Great Plains. State Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas County Adams Arapahoe Baca Bent Boulder Broomfield Cheyenne Crowley Denver Douglas El Paso Elbert Huerfano Jefferson Kiowa Kit Carson Larimer Las Animas Lincoln Logan Morgan Otero Phillips Prowers Pueblo Sedgwick Washington Weld Yuma Cheyenne Finney Gove Grant Greely Grey Haaskell Hamilton Kearny Total Area 2 (km ) 3,062 2,083 6,623 3,989 694 156 4,615 2,070 149 1,461 4,741 4,788 1,979 592 4,627 5,601 1,454 9,519 6,692 4,775 3,361 3,284 1,783 4,259 5,776 1,423 6,534 10,389 6,142 2,644 3,368 2,759 1,489 2,017 2,242 1,496 2,585 2,256 Square km in Habitat Category 1 2 3 4 635 341 424 1,662 787 235 196 866 2,217 368 407 3,631 500 292 324 2,874 226 45 42 381 15 9 12 121 1,144 367 387 2,716 83 223 251 1,512 40 15 18 76 1,112 69 54 226 3,125 590 415 611 2,707 317 296 1,468 970 122 109 778 167 24 29 372 683 185 259 3,499 1,030 403 456 3,712 387 121 169 777 2,885 600 671 5,362 1,362 743 837 3,750 1,869 511 580 1,815 1,234 463 337 1,326 267 91 155 2,772 379 120 60 1,224 822 210 187 3,040 790 368 490 4,127 531 92 55 745 1,832 453 586 3,663 2,391 1,266 1,665 5,066 4,082 436 240 1,383 1,184 206 129 1,126 1,162 140 196 1,871 761 491 512 995 388 52 47 1,002 33 32 78 1,874 724 100 76 1,342 182 21 28 1,265 503 212 534 1,336 804 113 157 1,181 55 % of County in Habitat Category 1 2 3 4 21 11 14 54 38 11 9 42 33 6 6 55 13 7 8 72 33 6 6 55 9 5 8 77 25 8 8 59 4 11 12 73 27 10 12 51 76 5 4 15 66 12 9 13 57 7 6 31 49 6 6 39 28 4 5 63 15 4 6 76 18 7 8 66 27 8 12 53 30 6 7 56 20 11 13 56 39 11 12 38 37 14 10 39 8 3 5 84 21 7 3 69 19 5 4 71 14 6 8 71 37 6 4 52 28 7 9 56 23 12 16 49 66 7 4 23 45 8 5 43 34 4 6 56 28 18 19 36 26 4 3 67 2 2 4 93 32 4 3 60 12 1 2 85 19 8 21 52 36 5 7 52 Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Nebraska Nebraska Nebraska Nebraska Nebraska Nebraska Nebraska Nebraska Nebraska New Mexico New Mexico New Mexico New Mexico Oklahoma Oklahoma Oklahoma Texas Texas Texas Texas Texas Wyoming Total Lane Logan Meade Morton Rawlins Scott Seward Sherman Stanton Stevens Thomas Wallace Witchita Chase Cheyenne Deuel Dundy Hayes Keith Kimball Lincoln Perkins 1,850 2,779 2,526 1,892 2,756 1,858 1,657 2,736 1,763 1,884 2,777 2,368 1,861 2,325 3,085 1,135 2,379 1,835 2,851 2,460 6,638 2,290 360 474 1,129 963 1,305 225 1,104 336 200 1,064 328 396 117 1,626 2,271 627 2,168 1,573 2,336 1,648 5,852 1,355 128 187 222 77 551 63 22 331 66 222 285 266 68 94 248 209 18 164 118 286 154 100 82 274 121 109 420 72 27 243 52 64 729 263 111 60 144 96 36 62 68 270 144 114 1,280 1,843 1,054 743 481 1,497 505 1,826 1,445 534 1,435 1,444 1,564 545 421 203 157 36 329 256 488 721 19 17 45 51 47 12 67 12 11 56 12 17 6 70 74 55 91 86 82 67 88 59 7 7 9 4 20 3 1 12 4 12 10 11 4 4 8 18 1 9 4 12 2 4 4 10 5 6 15 4 2 9 3 3 26 11 6 3 5 8 2 3 2 11 2 5 69 66 42 39 17 81 30 67 82 28 52 61 84 23 14 18 7 2 12 10 7 31 Colfax 4,443 1,014 324 310 2,795 23 7 7 63 Harding 5,477 2,833 395 500 1,749 52 7 9 32 Mora 3,269 989 360 248 1,673 30 11 8 51 Union Beaver Cimarron Texas Dallam Hansford Lipscomb Ochiltree Sherman Laramie 9,794 4,694 4,768 5,302 3,886 2,377 2,398 2,371 2,386 6,580 3,487 2,867 2,633 2,013 917 485 1,731 417 443 3,147 868 357 389 179 956 46 222 57 108 967 1,252 389 233 216 337 55 93 94 44 897 4,186 1,081 1,513 2,894 1,676 1,790 352 1,803 1,791 1,570 36 61 55 38 24 20 72 18 19 48 9 8 8 3 25 2 9 2 5 15 13 8 5 4 9 2 4 4 2 14 43 23 32 55 43 75 15 76 75 24 244,927 90,448 19,553 19,695 115,230 36.9 8.0 8.0 47.0 56 Figure 12. Map of predicted habitat suitability categories for 73 counties in the southwestern Great Plains, based on the final selected second-order habitat suitability model. 57 Table 34. Summary of the percent of the landscape in each of 4 BTPD habitat suitability categories based on the final selected third-order RSF including topoedaphic and climate predictors (Table 32) for 73 counties in the southwestern Great Plains. State Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Colorado Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas County Adams Arapahoe Baca Bent Boulder Broomfield Cheyenne Crowley Denver Douglas El Paso Elbert Huerfano Jefferson Kiowa Kit Carson Larimer Las Animas Lincoln Logan Morgan Otero Phillips Prowers Pueblo Sedgwick Washington Weld Yuma Cheyenne Finney Gove Grant Gray Greeley Hamilton Haskel Kearny Lane Logan Meade Morton Total Area 2 (km ) 3,062 2,083 6,623 3,989 694 149 4,615 2,070 149 1,461 4,741 4,786 1,979 592 4,627 5,601 1,454 9,519 6,692 4,775 3,361 3,285 1,783 4,259 5,776 1,423 6,534 10,389 6,142 2,644 3,368 2,759 1,489 2,242 2,017 2,584 1,496 2,256 1,850 2,779 2,526 1,892 Square km in Habitat Category 1 2 3 4 698 254 258 1,852 624 193 196 1,069 1,931 376 313 4,002 796 377 553 2,263 202 65 49 377 8 3 5 132 777 494 626 2,718 146 338 471 1,115 42 18 15 75 954 86 67 355 3,125 590 415 611 2,395 364 271 1,755 758 169 158 895 148 20 20 403 634 197 494 3,301 773 444 434 3,950 428 160 191 676 2,937 691 660 5,231 1,522 753 787 3,630 1,682 508 417 2,168 1,103 307 319 1,632 443 291 575 1,976 251 53 86 1,392 726 234 288 3,011 1,153 661 952 3,010 411 62 68 882 1,678 548 500 3,807 2,985 1,858 1,831 3,714 3,273 362 277 2,230 956 305 166 1,217 894 82 189 2,203 297 171 214 2,078 327 53 44 1,065 563 30 56 1,592 34 34 72 1,877 517 373 436 1,258 145 9 14 1,328 725 92 66 1,372 151 68 83 1,548 403 169 161 2,047 771 61 85 1,610 902 18 60 912 58 % of County in Habitat Category 1 2 3 4 23 8 8 60 30 9 9 51 29 6 5 60 20 9 14 57 29 9 7 54 5 2 3 89 17 11 14 59 7 16 23 54 28 12 10 50 65 6 5 24 66 12 9 13 50 8 6 37 38 9 8 45 25 3 3 68 14 4 11 71 14 8 8 71 29 11 13 46 31 7 7 55 23 11 12 54 35 11 9 45 33 9 9 49 13 9 18 60 14 3 5 78 17 5 7 71 20 11 16 52 29 4 5 62 26 8 8 58 29 18 18 36 53 6 5 36 36 12 6 46 27 2 6 65 11 6 8 75 22 4 3 72 25 1 3 71 2 2 4 93 20 14 17 49 10 1 1 89 32 4 3 61 8 4 4 84 14 6 6 74 31 2 3 64 48 1 3 48 Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Kansas Nebraska Nebraska Nebraska Nebraska Nebraska Nebraska Nebraska Nebraska Nebraska New Mexico New Mexico New Mexico New Mexico Oklahoma Oklahoma Oklahoma Wyoming Texas Texas Texas Texas Texas Total Rawlins Scott Seward Sherman Stanton Stevens Thomas Wallace Wichita Perkins Lincoln Kimball Keith Hayes Dundy Deuel Cheyenne Chase Union Mora Harding Colfax Texas Cimarron Beaver Laramie Sherman Ochiltree Lipscomb Hansford Dallam 2,756 1,858 1,657 2,736 1,763 1,884 2,777 2,368 1,861 2,291 6,638 2,461 2,851 1,835 2,379 1,135 3,085 2,324 9,794 3,269 5,477 4,443 5,301 4,768 4,693 6,580 2,363 2,371 2,250 2,377 3,886 928 153 787 210 831 362 179 288 82 916 5,534 1,438 2,081 1,413 2,043 443 2,044 1,305 2,926 507 2,554 762 1,089 1,938 2,085 3,126 204 203 1,106 295 373 165 36 199 270 10 144 107 273 51 246 158 290 115 52 83 111 394 166 637 126 405 302 610 655 230 930 182 57 96 70 464 177 46 80 337 46 456 135 263 62 100 121 227 128 70 19 220 205 134 868 169 429 329 302 454 172 934 118 50 72 67 973 1,486 1,622 591 1,918 876 921 2,356 1,544 1,666 1,028 825 506 527 300 233 362 442 720 5,363 2,468 2,088 3,051 3,301 1,721 2,207 1,590 1,859 2,062 975 1,945 2,075 34 8 47 8 47 19 6 12 4 40 83 58 73 77 86 39 66 56 30 15 47 17 21 41 44 48 9 9 49 12 10 6 2 12 10 1 8 4 12 3 11 2 12 4 3 3 10 13 7 7 4 7 7 12 14 5 14 8 2 4 3 12 6 2 5 12 3 24 5 11 3 4 2 9 4 4 1 19 7 6 9 5 8 7 6 10 4 14 5 2 3 3 25 54 87 36 70 50 49 85 65 90 45 12 21 19 16 10 32 14 31 55 75 38 69 62 36 47 24 79 87 43 82 53 244,745 77,492 19,578 20,706 126,969 31.7 8.0 8.5 51.9 59 DISCUSSION Based on our second-order RSF model, BTPD habitat suitability was positively associated with soil organic matter, pH, clay content and depth to a restricted layer as well as with increasing values of the topographic wetness index. BTPD habitat suitability was negatively associated with slope and soil sand content. The negative influence of slope was stronger on soils with high organic matter content. The positive influence of TWIsi was greater for soils with low sand content. Habitat suitability was positively associated with soil clay for areas with mean annual precipitation of 400 – 500 mm, but where mean annual precipitation declined to 350 mm, habitat suitability became negatively associated with soil clay content. The third-order RSF model yielded relationships with topoedaphic and climatic predictors that were strikingly similar to the second-order model. The strong concordance in predictions across both spatial scales of the second-order analysis (0.5 km and 2 km buffer distances) as well as across both second and third order analyses indicates that model predictions are robust. In all of these models, the 7 topoedaphic predictors have a significant influence on habitat suitability, and the strongest modifying effect of climate was via the influence of mean annual precipitation on the selection of habitats with varying levels of soil clay content. Our BTPD habitat suitability models based upon resource selection functions are consistent with, but incorporate substantially more detail than, previous habitat suitability indicies developed for the northern Great Plains (Clippinger 1989; Proctor et al. 2006). Clippinger’s (1989) index was based on slope, soil texture, and vegetation cover and height. Use of the index requires measures of vegetation height, which are typically not available from landscape-scale measurements or from remote sensing. Proctor et al. (2006) simply used a broad classification of “suitable” habitat as having slopes of less than 10% and soil that support potential vegetation consisting of grassland, salt-desert shrub, dry salt flats, or mixed barren sites. In contrast, our models provide an empirically-based, quantitative prediction of relative habitat quality (RSF probability maps) as well as predicted categories of habitat quality (RSF category maps), and clearly show that within areas classified as “suitable” by previous indicies, habitat value varies substantially. Rather than relying on broad categories of potential vegetation and slope, our model shows how 5 soil attributes and 2 topographic variables can collectively predict relative habitat value for BTPD. The habitat categories predicted by our models are also directly linked to measures of model error. Category 1, which we define as “low quality habitat” represents those areas predicted to be unoccupied by (hence unsuitable for) BTPD given a 5% false negative rate. This means that such locations are predicted to be unsuitable when only 5% of the area known to be occupied BTPD during 2001-2010 is incorrectly classified as being unsuitable. While the selection of this false negative error rate is arbitrary, the 5% value is relatively low and hence conservative in predicting low-quality habitat. For our second-order model, the probability cutoff corresponding to a 5% false negative rate is associated with a correspondingly high false 60 positive rate of 36.7% (Table 8), i.e. that 36.7% of “available” pixels in the dataset are classified as suitable habitat when using this probability cutoff. Thus, areas classified into habitat category 1 in our models represent areas classified as unsuitable based on a model where there is little chance (5%) that areas known to support BTPD will be classified as unsuitable, and there is a much higher chance (36.7%) that areas not known to support BTPD in the past decade will be classified as suitable. Conversely, habitat category 4, which we define as “high quality habitat” represents areas classified as suitable based on a model where there is a high chance (15%) that areas known to support BTPD will be classified as unsuitable and there is a correspondingly smaller chance (25.6%) that unoccupied locations will be classified as suitable. Categories 2 and 3 represent intermediate quality habitat associated with intermediate error rates. Our approach of defining categories by fixed false negative error rates is based on the fact that false negatives are a known error, whereas false positives represent the prediction that BTPD could occur in a particular location where they do not currently occur, hence are not necessarily an error. Because our models are based in large part on soil attributes derived from the national Soil Survey Geographic (SSURGO) database, they are limited by the accuracy of the SSURGO database. These soil maps were developed on a county by county basis by staff that varied among counties and states, such that methods for delineation of soil polygon boundaries and the assignment of soil attribute values can vary. In addition, point data collected in the field were used to populate databases representing much broader areas, and areas with spatially variable soil attributes are represented in SSURGO by a single value. An example of the effects of among-county variation in soil mapping is evident in habitat suitability maps for the Rita Blanca National Grassland (Maps 6, 13, 20 and 27). The northern half of this study site is located in Cimarron County, Oklahoma, while the southern half is located in Dallam County, Texas. An abrupt line in habitat suitability probabilities and categories is evident at the county line in Maps 6 and 13, which are based on the global model. This line is not evident in Maps 20 and 27, which are based on the local model. The best local model for Rita Blanca included only four predictors: TWIsi, Slope, Sand and Clay (Table 14), whereas the global model additionally included OM, pH, and DTR. These differences between the local versus global model maps suggest that measurement or mapping of OM, pH and DTR differed across the county/state line, and as a result, they were not valuable in predicting BTPD colony locations in the local model. Inspection of the input maps for these two counties showed that OM was consistently mapped at higher values in Dallam County, TX, compared to Cimarron County, OK, and hence the global model consistently predicted greater habitat suitability in Dallam County. These county-level differences in methods/measurements of soil organic matter are the reason for the discrepancies between local and global models for the Rita Blanca National Grassland (8.1% of the landscape; Table 19). Use of the local model appears to me more appropriate for Rita Blanca National Grassland. We also note that the Carrizo study area encompassed portions of two counties (Baca and Las Animas Counties, Colorado). Unlike Rita Blanca, we did not observe artifacts associated 61 with the county line (i.e. no evident differences in mapping of soil attribute values between the two counties) in the Carrizo study area map. Further inspection of model predictions across county and state boundaries showed the following. First, model predictions were spatially consistent across county boundaries within a given state. Second, model predictions were spatially consistent across state boundaries between Colorado, Wyoming and Nebraska as well as between Colorado, New Mexico and Oklahoma. The only notable spatial inconsistencies occurred between Prowers County, CO vs. Hamilton County, KS, Baca County, CO vs. Morton County, KS and where Texas (in particular, Dallam County) bordered on Oklahoma and New Mexico. As noted previously, differences along the Texas border appear to be the result of methodologies for measuring and mapping soil organic matter. The unusual striping patterns in habitat suitability predictions for the northern half of Morton County, KS were an artifact of the lower-quality DEM available for that county, and did not affect counties in the study region. As DEMs and soil survey maps continue to be updated and improved, including improved coordination across state lines, habitat suitability models can be refit and improved. Our models are unique in accounting for regional variation in climate. We found that precipitation has a large modifying effect on BTPD selection for soils of varying texture, while temperature has only minor effects on habitat selection. The interaction between the precipitation gradient and BTPD selection of soils with varying clay content can be understood on the basis of how both precipitation and clay influence soil moisture availability and hence potential aboveground plant production. In semiarid regions, the influence of soil texture on the retention of soil moisture at a site varies as a function of mean annual precipitation, a relationship known as the ‘inverse texture hypothesis’ (Noy-Meir 1973, Sala et al. 1988). For our study region in particular, Sala et al. (1988) showed that sandy soils are more productive than fine-textured soils when mean annual precipitation is less than 370 mm, whereas the opposite is true when precipitation is greater than 370 mm. Our model predicts that in the more arid, western portion of their range, BTPD select for low-clay sites (second-order model) or at least do not avoid high-sand sites (third-order model). This pattern suggests that in these arid regions, plant production on sites with relatively more sand and less clay may be important for BTPD in terms of sustaining food production, while not reaching such a high level of production that plant height affects visibility. As mean annual precipitation increases, BTPD show increasing avoidance of high-sand sites, where plant production and height may become sufficient to impact visibility and impair predator detection. In particular, for mean annual precipitation in the range of 400 – 500 mm, sandy soils (low clay) are increasingly associated with taller-structured, unpalatable vegetation such as sand sagebrush (Artemisia filifolia), which can be a deterrent to BTPD colonization. For the 7 National Grassland units that we studied, the proportion of the landscape classified as low-quality habitat (category 1) varied from a high of 61% of the Cimarron National Grassland to a low of 9% of the Timpas Unit of the Comanche National Grassland (Table 19). Conversely, the largest expanses of National Grassland classified as highly suitable BTPD 62 habitat (category 4) occur on the Timpas Unit (72% of area; 54,396 ha) and Carrizo Unit (36% of area; 46,211 ha) of the Comanche National Grassland. In total, we modeled black-tailed prairie dog habitat suitability for >2.4 billion pixels corresponding to an area larger than 244,000 km2 in the southern Great Plains. Within this area, the second-order model classified 47.0% as high-quality habitat, 36.9% was classified as lowquality, and 16.1% was in intermediate quality categories (Table 33). Sixteen counties had >70% of the land base classified as high-quality habitat: Prowers, Pueblo, Bent, Crowley, Kiowa, Broomfield, and Otero counties in Colorado; Sherman, Hansford, and Ochiltree counties in Texas; and Scott, Stanton, Witchita, Haskell, Finney, and Greely counties in Kansas. Conversely, 7 counties had >70% of the land base classified as low-quality habitat: Chase, Cheyenne, Keith, Hayes, and Lincoln counties in Nebraska, Lipscomb county in Texas, and Douglas county in Colorado. The most extensive contiguous areas of high-quality BTPD habitat occur in the region extending from east-central and southeastern Colorado into west-central Kansas (Figure 12). Large areas of low-quality BTPD habitat were primarily associated with sandy soils along major drainages (South Platte, Arkansas, Big Sandy, Cimarron and Purgatoire) as well as regions of sandy soil and/or rugged topography in El Paso, Douglas and Yuma counties of Colorado, most of southwest Nebraska, and the region extending along the borders between New Mexico, Colorado, Oklahoma and Kansas (Figure 12). The third-order model classified 31.7% of the region as low-quality habitat, and 51.9% as high-quality habitat (Table 34). Similar to the second-order model, large expanses of low-quality habitat were found in southwestern Nebraska, and in El Paso, Douglas and Yuma counties of Colorado. The largest expanses of high-quality habitat extended from east-central and southeastern Colorado into westcentral Kansas. These maps provide a basis for conservation planning for black-tailed prairie dogs and associated species of conservation concern at both local and regional spatial scales. They also provide a basis for assessing how areas that currently exist as native rangeland and areas that have been converted to cropland or other land uses compare in terms of their value as habitat for black-tailed prairie dogs. Finally, both the maps and underlying RSF models provide a basis for assessing how projected changes in precipitation and mean annual temperature within this region may modify habitat suitability for prairie dogs. 63 LITERATURE CITED Andelt, W. 2006. Methods and economics of managing prairie dogs. Pages 129-138 in J. L. Hoogland, editor. Conservation of the Black-tailed Prairie Dog. Island Press, Washington D.C. Antlolin, M., L. Savage, and R. Eisen. 2006. Landscape features influence genetic structure of black-tailed prairie dogs (Cynomys ludovicianus) Landscape Ecology 21:867-875. Augustine, D. J. 2011. Habitat selection by mountain plovers in shortgrass steppe. Journal of Wildlife Management 75:297-304. Augustine, D. J., J. F. Cully Jr, and T. L. Johnson. 2007. Influence of fire on black-tailed prairie dog colony expansion in shortgrass steppe. Rangeland Ecology and Management 60:538542. Augustine, D. J., M. R. Matchett, T. P. Toombs, J. F. Cully Jr, T. L. Johnson, and J. G. Sidle. 2008. Spatiotemporal dynamics of black-tailed prairie dog colonies affected by plague. Landscape Ecology 23:255-267. Belak, J. R. 2001. Modeling the effects of habitat quality on black-tailed prairie dog habitat occupancy using spatially correlated data. Colorado State University, Fort Collins, CO. Bolker, B., M. Brooks, C. Clark, S. Geange, J. Poulsen, M. Stevens, and J. White. 2009. Generalized linear mixed models: a practical guide for ecology and evolution. Trends in Ecology and Evolution 24:127-134. Clippinger, N. W. 1989. Habitat Suitability Index Models: Black-tailed prairie dog. U.S. Fish and Wildlife, U.S. Department of the Interior Biological Report 82 (10.156). Cully Jr, J. F., D. E. Biggins, and D. B. Seery. 2006. Conservation of prairie dogs in areas with plague. Pages 157-168 in J. L. Hoogland, editor. Conservation of the Black-tailed Prairie Dog. Island Press, Washington D.C. Cully Jr, J. F., T. L. Johnson, S. K. Collinge, and C. Ray. 2010. Disease limits populations: Plague and black-tailed prairie dogs. Vector-Borne and Zoonotic Diseases 10:7-15. DeCesare, N., M. Hebblewhite, F. Schmiegelow, D. Herviex, G. McDermid, L. Neufeld, M. Bradley, J. Whittington, K. Smith, L. Morgantini, M. Wheatley, and M. Musiani. 2012. Transcending scale dependence in identifying habitat with resource selection functions. Ecological Applications 22:1068-1083. Derner, J., J. Detling, and M. Antolin. 2006. Are livestock weight gains affected by black-tailed prairie dogs? Frontiers in Ecology and the Environment 4:459-464. Dinsmore, S. J., M. B. Wunder, V. J. Dreitz, and F. L. Knopf. 2010. An assessment of factors affecting population growth of the mountain plover. Avian Conservation and Ecology 5:5. [online] URL: http://www.ace-eco.org/vol5/iss1/art5/. Gillies, C., M. Hebblewhite, S. Nielsen, M. Krawchuk, C. Aldridge, J. Frair, D. Saher, C. Stevens, and C. Jerde. 2006. Application of random effects to the study of resource selection by animals. Journal of Animal Ecology 75:887-898. Gonen, M. 2006. Receiver Operating Characteristic (ROC) Curves.in Proceedings of the Thirdty First Annual SAS Users Group International Conference. SAS Institute, Inc, Cary, NC. Hanley, J., and B. McNeil. 1982. The meaning and use of the area under a receiver operating characterisic (ROC) curve. Radiology 143:29-36. Johnson, C. J., S. E. Nielsen, E. H. Merrill, T. L. McDonald, and M. S. Boyce. 2006. Resource selection functions based on use-availability data: theoretical motivation and evaluation methods. Journal of Wildife Management 70:347-357. 64 Johnson, D. 1980. The comparison of usage and availability measurements for evaluating resource preference. Ecology 1980:65-71. Koford, C. B. 1958. Prairie dogs, whitefaces, and blue grama. Wildlife Monographs 3:3-78. Lauenroth, W. K., I. C. Burke, and M. P. Gutmann. 1999. The structure and function of ecosystems in the central North American grassland region. Great Plains Research 9:223 - 260. Manly, B., L. McDonald, D. Thomas, T. McDonald, and W. Erickson. 2002. Resource Selection by Animals: Statistical Design and Analysis for Field Studies, Second Edition. Kluwer Academic Publishers, Dordrecht, The Netherlands. Noy-Meir, I. 1973. Desert ecosystems: environment and producers. Annual Review of Ecology and Systematics 4:23-51. Pepe, M. 2005. Evaluating technologies for classification and prediction in medicine. Statistics in Medicine 24:3687-3696. Pepe, M., T. Cai, and G. Longton. 2006. Combining predictors for classification using the area under the receiver operating characteristic curve. Biometrics 62:221-229. Pinsky, P. 2005. Scaling of true and apparent ROC AUC with number of observations and number of variables. Communications in Statistics: Simulation and Computation 34:771781. Proctor, J., B. Haskins, and S. C. Forrest. 2006. Focal areas for conservation of prairie dogs and the grassland ecosystem. Pages 157-168 in J. L. Hoogland, editor. Conservation of the Black-tailed Prairie Dog. Island Press, Washington D.C. Reading, R. P., and R. Matchett. 1997. Attributes of black-tailed prairie dog colonies in northcentral Montana. Journal of Wildlife Management 61:664-673. Roelle, J., B. Miller, J. Godbey, and D. Biggins. 2005. Recovery of the black-footed ferret: Progress and continuing challenges. Page 288 p. in U. G. S. US Department of Interior, editor. US Geological Survey. Sala, O., W. Parton, L. Joyce, and W. Lauenroth. 1988. Primary production of the central grassland region of the United States. Ecology 69:40-45. Soil Survey Staff, Natural Resources Conservation Service, United States Department of Agriculture. Soil Survey Geographic (SSURGO) Databases for Colorado, Kansas, Oklahoma, New Mexico, and Texas. Available online at http://soildatamart.nrcs.usda.gov . Stapp, P., M. F. Antolin, and M. Ball. 2004. Patterns of extinction in prairie dog metapopulations: plague outbreaks follow El Nino events. Frontiers in Ecology Environment 2:235-240. Theobald DM. 2007. LCaP v1.0: Landscape Connectivity and Pattern tools for ArcGIS. Colorado State University, Fort Collins, CO (http://warnercnr.colostate.edu/~davet/LCaP.html). Tipton, H., P. Doherty, and V. Dreitz. 2009. Abundance and density of mountain plover (Charadrius montanus) and burrowing owl (Athene cunicularia) in eastern Colorado. Auk 126:493-499. Truett, J., J. Dullum, M. Matchett, E. Owens, and D. Seery. 2001. Translocating prairie dogs: a review. Wildlife Society Bulletin 29:863-872. Wang, Y., H. Chen, R. Li, N. Duan, and R. Lewis-Fernandez. 2011. Prediction-based structured variable selection through the Receiver Operating Characterisic curves. Biometrics 67:896-905. 65