We collected data in Big Bend National Park (BBNP), a 3205-km2 area on the Texas–Mexico border (29°19′42″N, 103°12′21″W), during 1995–97. Lying within the subtropical Chihuahuan Desert, BBNP is covered by eight basic land-cover types: surface water (0·1%), developed areas (0·2%), bare ground (1·4%), montane woodland (2·6%), riparian vegetation (4·4%), limestone grassland (25·4%), igneous grassland (30·6%) and shrubland (35·3%). Details of the topography, altitudes, soils, rainfall, temperatures, definitions of land-cover types and characteristic plant species are available in Gutzwiller & Barrow (2001). BBNP is natural or semi-natural, with the latter conditions the result primarily of legacies of commercial ranching practised until 1944, when the Park was established. Human development in BBNP is sparse (Gutzwiller & Barrow 2003).
Each year, between late February and late May, a 12-week period that coincided with the breeding season in BBNP (Wauer 1996), observers conducted 20-min unlimited-distance point counts (Ralph, Sauer & Droege 1995), once each week, at the centres of 70 sampling sites. Data were collected under standard conditions of wind speed, air temperature and precipitation (Gutzwiller & Barrow 2003). Details of the techniques used to avoid potential time-of-day biases, enhance detection of flushed birds, avoid double counting of birds and preclude potential observer effects associated with clothing colour, bird-identification ability and the range of habitat types sampled can be found in Gutzwiller & Barrow (2003).
bird species’ traits
Factors that may affect broad-scale habitat selection influenced our consideration of body size, dispersal ability and cowbird-host status as potential predictors of species’ responses to broad-scale boundary length. As birds fly over landscapes, they may be able to sense broad-scale environmental features, including general boundary-length conditions. If broad-scale boundary length influences initial decisions in the habitat-selection process (Svärdson 1949; Hildén 1965; Hutto 1985), it has the potential to affect subsequent landscape occupancy by species and hence avian community structure. Through barrier or filter effects, boundaries may contribute to a landscape's resistance to animal movement (Forman 1995), which may constrain feeding, breeding and dispersal.
If smaller species cross boundaries less easily than larger species (Wiens, Crawford & Gosz 1985), smaller species may be more likely to avoid landscapes with high boundary length. If wide-ranging species perceive less heterogeneity and hence fewer boundaries than less-mobile species (Wiens 1992), stronger dispersers such as neotropical migrants may be less likely to avoid landscapes with high boundary length (cf. Bélisle & St Clair 2001). The brown-headed cowbird Molothrus ater is a brood parasite that specifically searches habitat edges for host nests, deposits its eggs in nests at or near boundaries, and causes host species to produce fewer young (Norman & Robertson 1975; Paton 1994; Askins 1995). Through impacts on host nest success, host species may be more likely to avoid landscapes with high boundary length than non-host species.
We used body mass (Dunning 1984) as an index of body size. Neotropical migrant status was used as an indicator of dispersal ability; a species was considered a neotropical migrant if any of its populations breed north of and winter south of the Tropic of Cancer (23°27′N) (DeGraaf & Rappole 1995). We used information in DeGeus & Best (1991) and in Ortega (1998) to determine whether a species was a host for the brown-headed cowbird. The bird species involved in the present analysis were native and typical of the habitat in BBNP.
boundary-length and landscape variables
For each sampling site, we used a global positioning system (GPS) with 0–1-m accuracy to obtain latitude and longitude coordinates. Land-cover data layers were derived from satellite images with 30-m resolution, and we used 1:24 000 scale TIGER digital line graph files for 1995 to develop data layers for roads; birds associate significantly with land-cover and road features measured at these resolutions in our study area (Gutzwiller & Barrow 2002, 2003). ArcView geographical information system software (ESRI 1998) was used to obtain landscape measurements.
Within a 2-km radius circle centred on each sampling site, we measured six boundary-length variables (Table 1). The 2-km radius circle is within the range of spatial extents for which significant bird–landscape relations have been detected in our study area (Gutzwiller & Barrow 2001) and in other regions (Van Dorp & Opdam 1987). All boundary lengths were measured to the nearest 0·1 km. Because road surfaces and verges were physically and biologically quite different from adjoining land-cover types, and the road surface separated the verges on either side of a road, we treated the two sides of a road as separate roadside boundaries. Reed, Johnson-Barnard & Baker (1996) argued that edge habitat exists on both sides of a road and applied this logic to estimate amounts of road-related edge habitat. Thus, lengths of roadside boundaries included both roadsides. When riparian vegetation bordered a river, as it did at several sites along the Rio Grande, lengths of riparian vegetation boundaries included riparian vegetation on both sides of the river.
Table 1. Boundary-length and landscape variables measured within 2 km of sampling sites
|TBL||Total boundary length (km), including boundaries between eight land-cover types (see Study area) plus roadside boundaries|
|LCB||Length (km) of all land-cover boundaries between eight land-cover types (see Study area)|
|RDB||Roadside boundary length (km)|
|SGB||Shrubland–grassland boundary length (km), igneous and limestone grassland combined|
|SRB||Shrubland–riparian vegetation boundary length (km)|
|RGB||Riparian vegetation–grassland boundary length (km), igneous and limestone grassland combined|
|NLT||Number of types of land cover|
|DSP||Number of patches of shrubland|
|IGP||Number of patches of igneous grassland|
|LGP||Number of patches of limestone grassland|
|RVP||Number of patches of riparian vegetation|
|DSC||Shrubland coverage (%)|
|IGC||Igneous-grassland coverage (%)|
|LGC||Limestone-grassland coverage (%)|
|ELV||Altitude at the site centre (m)|
Boundary length may be influenced by other landscape variables, such as the number of land-cover types, the number of patches of different types and the percentage of the landscape occupied by different types. Because these conditions may also affect bird numbers (Freemark et al. 1995; Flather & Sauer 1996), control for such effects was necessary to determine whether boundary-length variables themselves were important in structuring bird assemblages. Road length, and therefore the length of roadside boundaries, may be associated with certain landscape features (e.g. altitude, number of land-cover types and number of patches of different types). Thus, to assess effects of boundary-length variables per se, landscape conditions associated with road length had to be accounted for as well.
To control for these various conditions, we included in our analyses landscape variables (Table 1) that co-varied with the six boundary-length variables (rS = 0·62–0·20, P≤ 0·0001–0·100, n= 70, a priori α= 0·10) or had the potential to affect bird assemblage structure in BBNP (Gutzwiller & Barrow 2001, 2002). We measured landscape coverage to the nearest 0·1%. We used a GPS to measure the altitude at the centre of the sampling site to the nearest 5 m. Through its association with temperature, rainfall and soil type, and thus with vegetation structure and floristic composition (Wauer 1971), altitude served as an integrated measure of a variety of environmental conditions. From maps (Barnes 1979) we derived digital geological data useful for identifying limestone and igneous grasslands.
Mean abundance for a species was calculated as the mean number of individuals observed per count at a particular site during a given year. To study the probability of occurrence of a species, we analysed presence–absence data, which indicated whether or not a species was observed at least once during all of the counts for a particular site during a given year.
Biases in bird detection have the potential to influence assessments of bird–landscape relations (Gutzwiller & Barrow 2001). We controlled analytically for three conditions (hereafter extraneous variables) that were outside the focus of our study but had the potential to affect bird detection: observer identity (OBx, where x was an identifying number for an observer); whether or not an observer at the site centre could see farther than 100 m in all directions (BLK); and the number of weekly bird counts (NCS). We controlled statistically for variation in dependent variables associated with extraneous variables by including the latter in regression models when they were significantly correlated with dependent variables.
We used OBx to control analytically for effects that may arise from differences in observer experience or ability to detect birds. BLK was used to control for local conditions, such as tall or dense vegetation or abrupt topography near the site centre, that could block aural or visual detection of birds. At a few sites, observers were not able to sample during every week of the 12-week study season because of inclement weather or vehicle difficulties (Gutzwiller & Barrow 2003). For analyses involving probability of occurrence, we used NCS to control for variation in the detected presence of species associated with number of weekly counts. We did not use NCS as a variable in analyses of mean abundance because this variable's computation already included the number of weekly counts at a site.
The assemblage-level focus of our analysis compelled us to study as many bird species as possible. To detect some uncommon species, it was necessary to sample 12 times at each site using 20-min counts. Our need to include uncommon species, and our interest in drawing inferences for the entire breeding season, were not compatible with the assumptions and data requirements of other techniques used to control for species’ detection probabilities. Although these approaches are appropriate in many circumstances, the primary problems they presented for our situation were: untenable assumptions about detection of individuals and dynamics of site occupancy (capture–recapture, occupancy and removal-model methods); estimators that would have been impossible to apply or would have provided poor precision because of insufficient detections, particularly for uncommon species (distance and double-observer methods); and personnel demands that would have required us to use fewer sites, fewer sampling visits to sites or shorter count periods (double-sampling, double-observer and removal-model methods), all of which would have made it much more difficult to draw reliable inferences about uncommon species.
The details of these problems for our circumstances are presented in Gutzwiller & Barrow (2003). In short, we adjusted our dependent variables for species’ detection probabilities using an approach (statistical control for extraneous variables via regression) for which the assumptions were clearly met, and we were able to draw biologically meaningful inferences because our adjustment approach enabled us to include data from the entire breeding period for both common and uncommon species.
Spatial and temporal autocorrelations have the potential to inflate the statistical significance of regression coefficients. We computed spatial-trend variables, which were third-order polynomial terms based on the easting (E) and northing (N) for sites (Buckland & Elston 1993), and included these variables in regression analyses. These variables enabled us to control for broad-scale spatial trends, which is necessary before testing for site-to-site spatial autocorrelation (Kaluzny et al. 1998; Lichstein et al. 2002). We tested and adjusted for spatial autocorrelation using mixed-model procedures (Littell et al. 1996). Statistical problems caused by within-year autocorrelations were precluded by using a site (not a weekly bird count) as the unit of analysis, and problems caused by among-year autocorrelations were precluded by analysing data for each year separately. This latter step was also required to determine the number of years that species were associated with boundary-length variables (see below).
We used forward stepwise least-squares regression (SAS 1999) to determine whether a species’ mean abundance was related to explanatory variables (boundary-length variables, extraneous variables, landscape variables and spatial-trend variables). Species involved in analyses of mean abundance were those that were detected during a given year at ≥ 90% of the 70 sites. Uncommon species had so many zero abundances that the normality assumption for least-squares regression (Neter, Wasserman & Kutner 1989) could not be met. For these species we assessed relations between a species’ presence and absence (probability of occurrence) and explanatory variables with forward stepwise logistic regression (SAS 1999; Hosmer & Lemeshow 2000). For some logistic regression models, complete or quasi-complete separation occurred when species occupied < 20% or ≥ 90% of the 70 sites; this problem can generate invalid results (SAS 1999; Hosmer & Lemeshow 2000). Hence, logistic regression was used for a species only when it was present on ≥ 20% but < 90% of the 70 sites during a given year (Gutzwiller & Barrow 2001). We applied standard diagnostic and remedial methods (Neter, Wasserman & Kutner 1989; Hosmer & Lemeshow 2000) to preclude numerical problems (e.g. multicollinearity, separation, zero cell counts and imprecise regression estimates) and ensure that all assumptions of least-squares and logistic regression were met. We reported the relative importance of each explanatory variable in a model as the percentage of variation in the dependent variable that was associated with that variable; within a model, these percentages summed to the model R2.
Based on the regression results, we tallied the number of years during which a species’ mean abundance or probability of occurrence was positively or negatively associated with each boundary-length variable. We used one-tailed binomial tests (SAS 1999) to determine whether boundary-length variables were correlated with the abundances or occurrences of a significant majority (proportion > 0·50) of species in the assemblage, whether a significant majority of species was correlated positively, and whether a significant majority was correlated negatively (Hofor each test, proportion = 0·50; Zar 1999). Two-tailed binomial tests (Zar 1999) were used to assess whether each boundary-length variable's association with the assemblage was disproportionate to its length at our study sites. We used two-tailed binomial tests for paired-sample data (Zar 1999) to determine whether different boundary-length variables were related to significantly different proportions of species; paired-sample tests were necessary because the associations being compared in a given test involved the same species. The unit of analysis for binomial tests was a species, and all assumptions of these methods (Zar 1999) were met.
Kendall tau correlation coefficients (SAS 1999) were used to test whether the number of years during which species were associated with boundary-length variables was correlated with body size, dispersal ability and cowbird-host status. The positive or negative influence of a trait may vary depending on other traits and interactions among traits (Henle et al. 2004; Ewers & Didham 2006). In response to this possibility, and as a follow-up to our Kendall correlation analyses, we used Poisson and negative-binomial regression to explore whether actual relations were masked by co-variation or interactions among traits. We assessed relations between species’ associations with boundary-length variables and the main and interaction effects of the three traits; the number of years of positive relations and the number of years of negative relations with each boundary-length variable were analysed as separate dependent variables. The unit of analysis for these correlation and regression analyses was a species, and all assumptions of these methods (Allison 1999; Zar 1999) were met.
An a priori α= 0·10 was employed in all analyses to avoid type II errors (Gutzwiller & Barrow 2001). For a given least-squares or logistic regression model, we assessed the significance of each explanatory variable after using the sequential Bonferroni method (Rice 1989) to adjust a family-wide α= 0·10. This approach reduced the chance of inflated type I error rates (and overfitted regression models) that can result from multiple related tests (Miller 1981; Gutzwiller & Barrow 2001). This same adjustment was applied separately to determine the significance of the binomial tests for disproportionate effects, significance of the paired-sample binomial tests, significance of the Kendall correlation coefficients for a given biological trait, and significance of the explanatory variables in each Poisson and negative-binomial regression model.
For the least-squares and logistic regression analyses, we could not legitimately use confirmatory methods that presuppose a relatively small set of a priori candidate models (e.g. information-theoretic modelling; Burnham & Anderson 2002). Meaningful formulation of such models for each species and our suite of explanatory variables was not feasible because combined theoretical or empirical effects of the explanatory variables were not evident from the literature or other information, and influences of the individual variables on desert birds have not been studied sufficiently. Identification of candidate models from among the large number of potential models would therefore have been largely arbitrary. When relatively little is known about a system, the stepwise procedures we used are appropriate for initial analyses (Neter, Wasserman & Kutner 1989; Hosmer & Lemeshow 2000). Adequate theoretical and empirical information was also not available for variables involved in the binomial, Kendall correlation and Poisson and negative-binomial regression analyses. Thus, consistent with the first-step nature of this study, our analyses were exploratory.