Raptor Nest Survey.

We identified occupied/maintained Golden Eagle nests within the study area using a helicopter from 18 April through 26 April 2011, when Golden Eagles were expected to be incubating eggs or brooding nestlings. We recorded the locations of all occupied and unoccupied Golden Eagle nests identified during the survey and tentatively assigned territory status. We conducted a follow-up survey on 18 May 2011 to confirm status of each identified territory and associated nest(s). A Golden Eagle nesting territory was defined as “an area that contained, or historically contained, one or more nests (or scrapes) within the home range of a mated pair… a confined area where nests are found, usually in successive years, and where no more than one pair is known to have bred at one time” (Steenhof and Newton 2007). A territory was designated as occupied if the area encompassed a nest or nests or potential nest sites and was defended by a mated pair of Golden Eagles. Each nest within a territory was designated as: (1) occupied if we observed an adult eagle in an incubating position, eggs or young, a pair of eagles at or near the nest, newly constructed or refurbished stick nest, freshly molted feathers, or current year's whitewash (Pagel et al. 2010, Steenhof and Newton 2007) or (2) unoccupied if we did not observe any of these.

Capturing Golden Eagles.

We deployed GPS PTT units on four resident Golden Eagles within the study area in late winter/early spring 2011. We targeted Golden Eagles in four territories where individuals had laid eggs in previous years because captures occurred prior to establishment of nests and prior to the 2011 aerial survey. Two nests were along the primary China Mountain ridge near the center of the study area. One nest was in a small canyon on the western slope of China Mountain, and the fourth nest was in the southeastern portion of the study area along Cottonwood Creek. Distances between each targeted nest ranged from 9.6 km to 23.0 km. We targeted one adult resident breeding Golden Eagle from each territory.

Capture efforts were timed to take place before egg-laying to minimize the influence on reproductive success. A USFWS-permitted biologist with a capture permit issued by the U.S. Geological Survey (USGS) Bird Banding Lab, as well as relevant state scientific collection permits, captured all Golden Eagles. We used a padded leg-hold trapping technique for all captures (Bloom 1987), with fresh, natural prey items, including mule deer (Odocoileus hemionus) carcasses, as bait. We placed trap sets close to the four target nests in locations that we could monitor from long distances via binoculars. Upon capture, each Golden Eagle was hooded, placed into an aba to prevent struggling, and banded with USGS size 9 rivet leg band. We recorded mass, wing chord length, toe-pad length, and hallux length of each captured Golden Eagle. Nontargeted individuals (e.g., nonbreeders) were banded, measured, and released at the point of capture.

Captured adult Golden Eagles were fitted with GPS PTT-100 transmitters weighing 86 grams (Microwave Telemetry, Columbia, Maryland, U.S.A.). Transmitters were equipped with a solar recharger, auxiliary battery, and UHF transmitter. We attached transmitters using a backpack-style harness constructed of 0.84 cm Teflon ribbon (Bally Ribbon Mills, Bally, Pennsylvania, U.S.A.). The estimated precision of the location data collected via these units was 18 m for horizontal GPS positions (Microwave Telemetry pers. comm.). We programmed the GPS units to obtain one location per hour (the maximum fix rate for the devices) from at least one hour before sunrise to at least one hour after sunset. Data from each backpack were automatically downloaded every two days via the Argos satellite data collection relay system (CLS America, Largo, Maryland, U.S.A.).

Model Covariates.

We used a geographical information system (GIS) and 30-m × 30-m pixel level resolution digital elevation model to obtain measures of covariates associated with recorded Golden Eagle locations and a sample of random locations including: elevation, rugged, and slope. Terrain ruggedness (rugged) measured the variation in slope and aspect across the landscape on a scale of 0 to 1 (1 being most rugged) and was calculated using the “Vector Ruggedness Measure” tool in ArcMap 10.1 (Hobson 1972, Sappington et al. 2007, ESRI 2011). Solar radiation accounted for atmospheric effects, elevation, slope, aspect, sun angle, and shadows cast by surrounding topography as measured by the Area Solar Radiation tool (ESRI 2011). We also considered distance from a location to the China Mountain ridge line (rim edge) as a potential RSF covariate because this was a prominent ridge line within the study area and was used by Golden Eagles for perching, nesting, and exploitation of thermals/updrafts (C. LeBeau unpubl. data).

Vegetation covariates at Golden Eagle locations were based on Landsat Thematic Mapper satellite images (USGS 2012). We used images that represented the vegetation present during each season (spring–fall; e.g., 27 May 2009, 20 July 2011, and 22 September 2011, respectively) and had the highest quality (e.g., lowest percentage of cloud cover). To evaluate aboveground biomass using ArcMap 10.1, we used the Normalized Difference Vegetation Index, NDVI:

(Deering 1978, Vogelmann et al. 2001, Chander et al. 2009) and the Soil Adjusted Vegetation index, SAVI:

where L  =  0.5 and is a soil brightness correction factor (Huete 1988, Vogelmann et al. 2001, Chander et al. 2009). Measures of NDVI, represented the amount of visible light absorbed by plants as well as the amount of near-infrared light plants reflected (Deering 1978). SAVI is similar to NDVI, but attempts to accommodate for some of the limitations of NDVI by using a soil adjustment factor, which reconciles the influence of soil in spectral features by the percentage of green cover (Huete 1988). NDVI and SAVI values ranged from −1 to 1 with larger values measuring areas of greater vegetation vigor or biomass. We also applied Tasseled Cap Transformations to the Landsat images to calculate greenness, wetness, and brightness vegetation indices to indicate vegetation canopy, moisture, and bare soil, respectively (Kauth and Thomas 1976, Crist 1985). Values for greenness, wetness, and brightness ranged from −1 to 1 where higher values related to increased presence of vegetation and a greater magnitude of moisture and soil reflectance, respectively.

We obtained occupied Greater Sage-Grouse (Centrocercus urophasianus) lek locations from Idaho Fish and Game and Nevada Department of Wildlife to characterize potential Golden Eagle prey habitat. We calculated the distance to nearest lek (lek) from Golden Eagle and random locations to identify areas of sage-grouse concentrations during spring lekking. In addition, we used proportion of shrub cover within a 1-km² area (shrub) surrounding each eagle and random location generated from LANDFIRE (Landscape Fire and Resource Management Planning Tools) landcover dataset, to describe the potential black-tailed jackrabbit (Lepus californicus) habitat (Steenhof et al. 1997, LANDFIRE 2012). Lastly, distance from a Golden Eagle and random location to the associated individual's nest (nest) was considered for the final RSF because it is an important feature in occupied Golden Eagle habitats (McGrady et al. 2002, McLeod et al. 2002), and can account for some of the spatial correlation in habitat selection (i.e., locations farther away have a lower probability of use).

Model Development.

Due to low GPS fix success (71%) attributable to limitations of the backpacks' solar charging system as a result of topography, we used only locations obtained from 2 hr after sunrise to 2 hr before sunset, when Golden Eagles were most active and when GPS fix success was highest (98%; Soutullo et al. 2006a). We defined GPS fix success as the ratio of regular GPS fix attempts and the actual GPS fixes. We defined available habitat for each individual as a minimum convex polygon (MCP) around all observed GPS locations for that individual within the boundary of the study area because landscape-level habitat availability should be based on the distribution of radio-collared animals (McClean et al. 1998). We excluded 8% of all locations that were located outside of the study area from the analysis because we did not have habitat covariate data associated with these locations. We generated sets of available locations by taking simple random samples of locations within each individual's MCP. The number of available locations selected from each home range was two times the number of used locations for each Golden Eagle.

Golden Eagles are a landscape-level species (Kochert et al. 2002, Katzner et al. 2012) and it is important to consider multiple scales when estimating Golden Eagle habitat selection. Thus, we calculated the average value of each terrain covariate and vegetation indices within 1 km of each used and available location (e.g., slope_mean; Table 1). Covariate values for elevation, slope, and rugged were obtained from the 30-m × 30-m grid cell that encompassed each used and available location (Table 1). Many of the habitat features within the study area consisted of small rocky outcrops or steep canyons that may have been suitable perch locations. Use of these features by Golden Eagles would not be detectable on a larger scale (e.g., 1 km), so we included the value at the smallest extent possible, believing any errors in the underlying grid and the Golden Eagle GPS locations would be random and negligible. Along with linear terms for slope, rugged, brightness, and wetness, we considered quadratic terms for these covariates based on hypotheses that Golden Eagles may have preferred locations with moderate, rather than extreme, covariate values.

Table 1.

Explanatory covariates measured at each used and available location for modeling Golden Eagle seasonal habitat selection in southern Idaho, U.S.A., 2011. All data layers were represented at a 30-m resolution and included quadratic terms where indicated.

We incorporated the covariates into a binary logistic regression equation in a use–availability framework with a maximized likelihood to estimate an exponential RSF (Manly et al. 2002, Johnson et al. 2006, McDonald 2013) that predicted the relative probability of Golden Eagle habitat selection. We focused on the time of year when Golden Eagles breed and raise their young and assumed habitat selection patterns could differ across this time period. Therefore, we developed separate models during three different biologically meaningful seasons in 2011: spring: late February to early March–30 June); summer (1 July–15 September); and fall (16 September–1 November). In addition, the covariates related to Golden Eagle habitat selection used in the analysis had different distributions within the three seasons. For example, brightness values ranged from 0.314 to 0.375 during spring, 0.204 to 0.447 during summer, and 0.234 to 0.460 during fall.

Our evaluation of Golden Eagle habitat selection within the study area consisted of a multistep process, where we: (1) developed RSFs for each individual within the individual's home range during each season, (2) reestimated individual RSFs using a combined dataset (used and available) with locations from all monitored Golden Eagles, (3) predicted a model-averaged (Burnham and Anderson 2002) relative probability of Golden Eagle habitat selection within the study area during each season, and finally (4) validated the predictive ability of each RSF estimated in Step (3) using a leave-one-out technique (Johnson et al. 2006). For example, we developed a unique RSF (i.e., different combinations of covariates; Step 1) for each Golden Eagle resulting in four unique RSFs for each season. We considered each RSF to be equally plausible at predicting habitat selection for the average individual Golden Eagle within the study area, and so reestimated each individual's RSF using the used and available locations from all four Golden Eagles resulting in 12 total RSFs, three for each season (Step 2). In Step 3, we averaged the predictions across seasonal RSFs estimated in Step 2. Finally, we validated the resulting RSFs by testing their predictive ability in a leave-one-out approach in Step 4.

The lack of published, landscape-level Golden Eagle habitat selection studies precluded identification of a list of a priori models for model estimation and comparison. Thus, we first developed a habitat selection model for each individual using forward, stepwise selection (Neter et al. 1996) and the Bayesian information criterion (BIC; Burnham and Anderson 2002). Models with smaller BIC values had more support in the data and were considered parsimonious (Burnham and Anderson 2002). We built models using forward variable selection via improvements in BIC values using R language for statistical computing (R Development Core Team 2012). For example, the covariate selected first resulted in the lowest BIC score among other univariate models. We added remaining covariates to the first selected covariate and reevaluated the model to see if the BIC score could be lowered. If the model BIC was further reduced, then the model-building process continued looking forward (adding covariates) until the BIC value could not be further reduced or until the model reached a maximum of five covariates resulting in a RSF for each Golden Eagle. We limited each RSF to five covariates to maintain simplicity and consistency among the four individuals, although we acknowledge that we may have missed some relationships.

Correlations among covariates may result in erroneous inferences (Neter et al. 1996). Prior to model building, we conducted a pairwise correlation analysis to identify potential colinearities between covariates. Based on results of the correlation analysis, we did not allow any mutually correlated variables in any one model. For example, if two covariates were correlated based on Pearson's correlation coefficient (|r|) being >0.6, we dropped that covariate that was chosen later in the model selection process. In addition, we calculated variance-inflation factors (Fox and Monette 1992) to test for multicollinearity between covariates in the final individual models. Based on convention, we only considered quadratic versions of covariates in conjunction with their linear counterparts.

During development of individual RSFs (Step 1), we sought covariates that were related to habitat selection within an individual's home range. Step 2 of our modeling process consisted of using those identified covariates within individual models to predict the relative probability of selection by the average individual Golden Eagle within the study area. This was done to account for the variability in habitat selection patterns among the four individuals. For example, the habitat selection analysis for one Golden Eagle might indicate that a particular covariate (e.g., slope) was an important predictor of habitat selection by that Golden Eagle; however, this covariate may have had little relationship to the selection patterns for other Golden Eagles, which suggests some level of variability among the individuals monitored. We accounted for this variability in habitat selection by treating the individual as the experimental, or primary, sampling unit (Thomas and Taylor 2006), and applying each individual model to a combined dataset that included all Golden Eagles' used and available data.

We estimated final RSFs from the reapplication of each individual model to the combined dataset (Step 2). We estimated confidence intervals for each coefficient in each of the final 12 RSFs using bootstrapping, where the individual Golden Eagles were randomly sampled with replacement and the final model was refit to the new sample of used locations and available locations (Manly 2007). We used 200 bootstrap iterations to calculate standard errors of model coefficients (estimated as SDs of the bootstrap distribution of estimates). We calculated lower and upper confidence limits for each estimate from the SEs using the estimate +/− 1.645(SE). Coefficient estimates with 90% confidence intervals (CIs) not containing 0 were considered statistically significant at the alpha  =  0.10 level. We calculated odds ratios [(exp()−1)*100] from coefficients in the final RSF models and used them to interpret the relationship between each covariate and Golden Eagle habitat selection (McDonald et al. 2006). Odds ratios describe the estimated percent change in relative probability of selection for a 1-unit change in a predictor variable. We did not calculate odds ratios for covariates with both linear and quadratic effects. We created marginal effects plots using the estimated parameters and their associated CIs from the top model in each season and study area to show the marginal effect of selected variables (McDonald et al. 2006).

We used the resulting 12 final RSFs (Step 2) to estimate the relative probability of selection by the average Golden Eagle within the study area (Step 3). We used a model averaging process to account for the variability of selection among individuals. Model averaging is commonly used to address model uncertainty among a set of models estimating habitat selection (e.g., Arnold 2010). This approach minimizes the effect of uninformative parameters among individual models, particularly if covariates are included in one model and not in another (Burnham and Anderson 2002). This is particularly important when using multiple RSFs to make an overall prediction of the relative probability of use by the average Golden Eagle within the study area. To make an overall prediction of use within the study area, we placed a 100-m × 100-m grid over the study area and extracted covariate values for each cell. Using these values we predicted the relative probability of selection using each individual final RSF. When we made predictions across the study area, we used the distance to the nearest known occupied nest observed during the aerial nesting inventory survey (i.e., it was not limited to the four monitored Golden Eagles). The average predictions from the four Golden Eagle RSFs were classified into five equal-area bins (low, medium-low, medium, medium-high, and high use) using percentiles.

Following model estimation and creation of a predictive map for the study area, the model estimation and prediction processes were evaluated (Step 4), which is an important part of determining the quality of a RSF (Johnson et al. 2006, McDonald et al. 2006). We sequentially left one individual out of the model estimation and prediction mapping process (Steps 2–3) and investigated the ability of the data from the other three Golden Eagles to predict use for the individual that was held out. Predictions for the used and available locations for the individual not included in developing predictions were binned into 20 equal area classes – i.e., the same number of sampled locations (available and/or used) were assigned to each class based on percentiles of the predicted values. We then calculated Spearman's rank correlation coefficient to estimate the strength of the relationship between prediction class (1–20, with 20 being “best”), and the number of used locations within each bin. Higher correlations indicate the data and process used to create the final predictive map were robust for among-animal variability in habitat use.

Nesting Inventory.

We identified 32 occupied Golden Eagle territories (0.10 occupied territories per km²) during the nest surveys conducted from 18 April to 18 May 2011. Three of the 32 occupied territories had nests classified as unoccupied because no adults were observed near the nests during aerial nesting surveys, but GPS/Argos data showed adult use around these nest locations indicating an occupied territory. Eagle 1 was the only monitored eagle that had an occupied nest; however, this nest was unsuccessful. Some nesting attempts may have been undetected due to the timing of the aerial survey. Overall, occupied territories appeared evenly distributed throughout the study area with an average inter-nest distance of 4.71 km during 2011 (Fig. 1).

Captures.

Capture efforts occurred over 11 d within four different territories between 17 February and 16 March 2011. We captured and banded five Golden Eagles. Two adult (defined by plumage characteristics; Bloom 2001, Kochert et al. 2002) males and two adult females were fitted with backpack satellite telemetry packages via a harness system. One of the captured Golden Eagles was determined to be 2 yr old based on plumage (Bloom 2001). Although this subadult was banded and measured, it was released without a GPS backpack. Most Golden Eagle captures occurred where mule deer carcasses were bait.

Habitat Selection.

We estimated daytime habitat selection between 0800–1900 H during spring and summer and 0900–1700 H during fall. The total number of locations recorded during each season varied for each individual because we excluded locations recorded outside of the study area from the analysis (Table 2). The size of each individual's MCP varied among the three seasons, but the smallest MCPs were estimated during spring, ranging from 6771 ha (Golden Eagle 2) to 32 818 ha (Golden Eagle 3; Table 2).

Table 2.

Size of minimum convex polygon (MCP) in ha for each individual Golden Eagle during each season in southern Idaho, U.S.A., 2011.

The final seasonal RSFs for each individual contained numerous covariates; however, nest, slope, rugged, and brightness were identified repeatedly as important predictors of Golden Eagle habitat selection. The variable nest was in all final RSFs, except for Golden Eagle 1, where nest was only included in the spring model. Rugged was included in many of the seasonal RSFs, and both the linear and quadratic terms occurred in all spring models and in the fall model for Golden Eagle 1 (Tables 3–Table 4.5), indicating a moderate level of rugged (0.025–0.030) was generally selected. Brightness and slope also were common in the Golden Eagle models (Tables 3–Table 4.5), and marginal plots indicated that habitats with moderate to higher levels of brightness and slope were commonly selected. Selection for other covariates was less consistent (Tables 3–Table 4.5). Unique covariates selected for the final spring, summer, and fall RSFs ranged from 8 (summer and fall) to 10 (spring; Tables 3–Table 4.5). Slope measured at the pixel level occurred in each summer RSF except for Golden Eagle 1, where the average slope_mean measured within 1 km of each used and available location was selected first during model building. Wetness, radiation, and greenness were included in multiple final RSFs (Tables 3–Table 4.5), but their relationships with habitat selection were variable across seasons and individuals.

Table 3.

Final resource selection functions for resident Golden Eagle daytime habitat selection during the spring of 2011 in southern Idaho, U.S.A.

Table 4.

Final resource selection functions for resident Golden Eagle daytime habitat selection during the summer of 2011 in southern Idaho, U.S.A.

Table 5.

Final resource selection functions for resident Golden Eagle daytime habitat selection during the fall of 2011 in southern Idaho, U.S.A.

Estimated coefficients for covariates found in >1 individual RSF within a season were generally consistent among individuals, indicating similar habitat preferences (Tables 3–Table 4.5). During spring, the odds ratio for nest indicated that the relative probability of a Golden Eagle selecting habitat within the study area decreased by an average of 35.0% (range  =  31.7–38.6%) for every 1 km increase in nest (Tables 3–Table 4.5). The final models for two of the individuals (Golden Eagles 1 and 3) contained linear and quadratic effects for brightness in the spring, and estimates predicted that on average, relative probability of selection increased up to ≈0.327 brightness units, then declined (range  =  0.321–0.332; Table 3). The linear and quadratic terms for rugged were included in each Golden Eagle's final RSF and the relative probability of selection increased on average up to 0.280 rugged units, then declined (n  =  4; range  =  0.269–0.311; Table 3).

During summer, the relative probability of selection decreased on average by 45.9% for every 1 km increase in nest (n  =  3; range  =  45.83–45.94%). Selection increased on average by 31.6% for every one unit increase in brightness (n  =  2; range  =  28.7–34.7%; Table 4). The final models for two of the individuals (Golden Eagles 2 and 4) contained linear and quadratic effects for slope, and estimates predicted that on average, relative probability of selection increased up to 31.3° of slope, then declined (n  =  2; range  =  26.5–36.0; Table 4). Although the linear term for slope was in the final model for Golden Eagle 3, the quadratic term was not selected during model building. Based on the final model for Golden Eagle 3, the relative probability of selection increased by 12.9% for every 1° increase in slope (Table 4).

During fall, the relative probability of selection decreased on average by 30.9% for every 1 km increase in nest (n  =  3; range  =  30.6–31.1%; Table 5). As in summer, selection increased by 12.360% for every 1° increase in slope for Golden Eagle 3 (Table 5). The linear and quadratic term for slope was included in the final RSF for Golden Eagle 4 and selection increased up to 36.0°, then declined (Table 5).

Validation of the four seasonal models from each Golden Eagle indicated the final RSFs had strong predictive abilities when estimated by pooling data from three of the four individuals. Correlations between observed use and the 20 prediction classes ranged from 0.879 to 0.989 during spring, 0.822 to 0.971 during summer, and 0.787 to 0.946 during fall (Table 6).

Table 6.

Spearman's rank correlation coefficients estimated during validation of final resource selection functions. For example, 0.929 is the correlation between observed use for Golden Eagle 1 in the Spring of 2011 and use predicted by averaging predictions from the final spring RSFs for Golden Eagles 2, 3, and 4.

The relative probability of selection was highest around all identified nests within the survey area. Selection also appeared to be highest around the steep canyons and cliff faces during all three seasons (Fig. 1).