The study was done in the Biosphere Reserve of Ñacuñán (34°03′S, 67°54.5′W), in the central Monte desert, Province of Mendoza, Argentina. The climate is dry, with wide variations in annual precipitation between years (mean: 348.9 mm, range: 192.6–585.4 mm, 1972–2002). It is also highly seasonal, with warm and rainy summers (> 20 °C; 269 mm) and cold and dry winters (< 10 °C; 80 mm). A complete description of the area can be found in [81]).
The main habitat of the Reserve is the algarrobal, an open woodland of algarrobo (Prosopis flexuosa) trees 3–6 m high scattered in a matrix of perennial Larrea divaricata tall shrubs (1–3 m high, horizontal cover > 35%). Other woody species are Geoffroea decorticans trees, tall shrubs such as Capparis atamisquea, Condalia microphylla and Atriplex lampa (usually > 1 m high), and low shrubs (~ 20% cover, usually < 1 m high) such as Lycium spp., Verbena aspera and Acantholippia seriphioides. There is also an important cover (> 25%) of perennial grasses (Pappophorum spp., Trichloris crinita, Digitaria californica, Aristida spp., Setaria spp., Sporobolus cryptandrus). Most of the reserve has been closed to cattle ranching and other significant human activities since 1971. About a third of the surface of the algarrobal lacks perennial vegetation in the form of open patches of variable size (from centimetres to metres). Forb cover is highly variable among seasons and years, usually an order of magnitude lower than grass cover. Forbs were not considered in the description of the vegetation structure, following criteria in local studies of the seed bank [21, 22] and bird foraging [30].
The guild of small seed-eating birds that forage mainly from the ground is formed by 4–5 species: Zonotrichia capensis, Saltatricula multicolor, Diuca diuca, Phrygilus carbonarius, and Poospiza ornata only in spring–summer [30, 111, 112]. The guild changes in abundance and composition along the year: the abundant southern subspecies Zonotrichia capensis australis is only present during autumn and winter (other subspecies are resident) [113] and Poospiza ornata arrives for the breeding season [114]. We expect these species to remove experimental seeds from the ground according to their mean proportional abundances in this habitat [95]: autumn–winter: 60.4% Z. capensis, 23.9% P. carbonarius, 8.1% D. diuca, 7.6% S. multicolor; spring–summer: 49.4% P. ornata, 18.7% S. multicolor, 15.5% D. diuca, 11.3% P. carbonarius, and 5.2% Z. capensis. Seeds of herbaceous plants are the staple diet of these granivorous birds (75–99% of their granivorous diet is grasses and one forb species [56]). All postdispersal granivorous birds prefer medium–large grass seeds [53, 54] and reduce their seed intake in spring and summer to include insects and fruit in their breeding diets [30, 112]. Other birds able to consume seeds do not forage on the ground, are rare or only visit this habitat occasionally (e.g., Poospiza torquata, Catamenia analis, Zenaida auriculata, Columba maculosa, Columbina picui, Carduelis magellanica, Molothrus bonariensis, Passer domesticus, Eudromia elegans, Nothura darwinii; [95]).
Primary seed dispersal starts in late spring and finishes by winter, with maximum seed availability in the soil during autumn–winter and a minimum when summer begins [22, 57]. Seed abundance in the soil is very heterogeneous at small scales, with close patches of extreme abundances. Seeds are consistently more abundant under trees and shrubs and in depressions of the soil where litter accumulates, mostly because of the persistent seed bank of forb seeds [22, 24]. The abundance of grass seeds is less heterogeneous though still higher under woody cover, with some intra- and inter-annual variability. As a consequence, our hypotheses and their predictions should apply for any of the main granivorous bird species and for all of them analysed as a guild.
Single seeds were offered on top of each of 300 devices made of upside-down feet of plastic flute glasses with their stems buried so the top (originally the base of the glass) remained 2–3 cm over the ground. This configuration prevented access by granivorous ants and other arthropods that cannot walk upside-down on the smooth plastic surface ([49, 115], pers. obs.). Devices were arranged 5-m apart on three 10 × 10 grids (“J”, “F” and “V”; area ≈ 2000 m2 each) located 80–400 m apart within the algarrobal. A single Setaria italica seed (a commercial species bigger than the otherwise similar local Setaria leucopila, both readily consumed by birds [53]) was offered on each device for two consecutive days from sunrise to sunset (standardised following data and civil twilight criteria by U.S. Naval Observatory [116]), once per season (ranging from ~ 23.3 h in Autumn to ~ 29.5 h in Summer). The top surface of the device, 6 cm in diameter, was covered with local soil for a similar visual appearance to the surrounding ground. According to estimations from simultaneous soil samples (L. Marone, unpublished), our experimental seed offer was not particularly attractive against background seed offer in this habitat: from 2.4 (bare soil in spring) to 19.3 (beneath shrubs in winter) grass seeds are expected in a similar sized area of soil, with a similar biomass density to the expected average for grass seeds during winter (~ 0.1 mg/cm2) and 36% of the biomass of all consumable seeds (those in the diet of granivorous birds [56]).
We were not able to identify which animal species removed the seed from each device. However, rodents that may have remove seeds in the area are mainly nocturnal [49]. Two additional trials were done under a modified protocol to test the assumption that birds were the (only) diurnal organisms removing seeds from these devices. Clayish local soil was smoothed around 50 devices in grid F before offering seeds during an extra day after the main summer and winter trials. Footprints of birds, mammals and other taxa (insects and lizards) were easily recognised, though we were not able to confidently distinguish among the focal granivorous bird species. In most cases where the seed has been removed only bird footprints were detected, in both winter (32/35 = 91%) and summer (7/9 = 78%) trials. The rest had mixed footprints of birds and other taxa or not recognisable footprints, but no device without seed had footprints of other taxa exclusively. On the other side, no bird footprints were found around devices where the seed remained, suggesting that it was not rejected when closely approached by walking birds. Moreover, removal of single seeds did not differ from removal of groups of ten seeds during a pilot study, suggesting that birds completely remove small groups of accessible experimental seeds when detected.
Since seed removal from a device was not independent between the two consecutive days in any of the four seasons (Fisher exact tests for 2 × 2 contingency tables: χ2 > 14.15, P < 0.001, n = 300, with more observations of “double removal” and “never removed” than expected by chance), a device was defined as “used” in each season if the seed was removed at least once during the 2 days it was offered. Independence of seed removal among seasons was tested for each grid by comparing the distribution of observed frequencies of the number of seasonal trials in which each device was used (0–4) against the expected frequencies calculated as the product of four Bernoulli trials with n = 100 (each seasonal experiment) with variable probabilities of success, estimated as the proportion of used devices in that grid for each season. Goodness of fit was evaluated with a χ2 statistic.
Temporal and spatial heterogeneity in intensity of seed removal (proportion of used devices per grid) was tested with binomial Generalised Linear Models (logit link) with G and S as independent categorical variables. Significance of predictors was assessed comparing the change in deviance of nested models obtained through stepwise backwards elimination, asymptotically distributed as χ2. The ratio of deviance to degrees of freedom in the minimum adequate model was 1.74, but corrections for overdispersion did not change interpretation of results (F– vs. χ2-tests).
Studies on use of space through short-term observations rely on asymmetric evidence: while patch visit can be inferred from seed removal, non-removal does not imply the patch is not to be explored eventually. This lack of a proper “no use” group should raise concerns on simple statistical analyses based on a priori classification into exclusive groups (e.g., discriminant analyses), particularly if they assume similar variability in both (homoscedasticity), or based on assigning zero probability of use (e.g., classical logistic regression). Identification of explanatory variables and predictive value of the statistical models can both suffer from the unrecognized probability of false-negatives [117,118,119]. Though more complex modelling strategies can incorporate or simulate incomplete evidence on absences in a spatially explicit context (e.g., species distribution and Bayesian models; see [119,120,121]), we opted for an indirect strategy of analysis that best matches how we developed our hypotheses and predictions. First, we detected and simplified the main structural and floristic characteristics defining habitat heterogeneity at the microhabitat scale. Then, we evaluated if those characteristics and the spatial positions of used microhabitats (i.e., those where the seed had been removed) were a random (no selection) or a skewed (selective) sample of available microhabitats. This is a similar approach to that previously reported in [30] at the smaller spatial scale of analysis.
Vegetation and soil cover around each device were measured with a vertical 1-m long pole (2 cm diameter) positioned every 10 cm along four 50-cm transects (= 20 points per microhabitat) from the device to each cardinal direction. At each point, perennial plants touching the pole were identified to genus level. The presence of vegetation > 1 m and the presence of dense litter (when it prevented from seeing the mineral soil below) or its absence (bare soil) were also recorded. Percentage cover per plant group (grasses, standing dry grasses, low shrubs, tall shrubs, and trees) and of bare soil and deep litter were calculated after those measurements.
A Principal Components Analysis (PCA) with Varimax rotation of the selected axes was done to reduce the number of dimensions of the ten variables measured at the microhabitat scale. Some variables were previously transformed (arcsin, square root or logarithm) to improve the symmetry of their distributions and then standardised into the PCA correlation matrix. Alternative analyses at the level of plant genera gave similar but noisier results on the main axes. Three components (PC1–PC3) were retained following the Kaiser criterion (eigenvalue > 1), the broken-stick model, and the scree-plot [122]. Before multidimensional analyses, scores on each axis were multiplied by its eigenvalue to weight them according to the variability they represent (see [123]); though applied for correctness, these variable transformations and weights did not change results significantly. Separate PCA analyses for each grid resulted in similar principal components and scores correlated with those of the grouped analysis (Pearson correlation, all cases n = 100, r > 0.8, P < 0.001, except for PC3 in grid V: r = 0.21, P = 0.03) confirming that heterogeneity of main characteristics at this scale were similar for the three sites. Therefore, only PCA scores based on all microhabitats were used for subsequent analyses. Scores based on microhabitat characteristics were compared among grids with Kruskal–Wallis tests.
Microhabitats around seed devices were also categorized at field a priori following the same criteria used on previous studies [21, 22]: beneath trees, beneath tall shrubs, beneath low shrubs, beneath grasses (under no woody cover), and bare soil (no perennial cover). Microhabitats around 79 grid points (26%) did not fit neatly into any of those categories and were assigned to an “intermediate” category (e.g., shrub borders).
Big trees (trees > 3 m and algarrobos > 4 m high) in and around the grids were mapped, measuring the distance between each device and the nearest tree canopy. Other minimum distances to vegetation in several height strata and to closest canopy of each plant group were measured but resulted highly correlated with measurements of horizontal cover (since most distances were smaller than a microhabitat radius, e.g., 96% of distances to nearest grass or 78% to vegetation 1–2 m high, see [60]). Therefore, only distances to tall trees were informative in addition to the measured characteristics at the microhabitat scale. Distances were compared among grids with Kruskal–Wallis tests.
Selection at the microhabitat scale was evaluated (1) multidimensionally, with a spatial technique applied to results of the PCA ordination and (2) unidimensionally, with a randomization test for each of the three retained PCA components. The first test is analogue to the representation of used and available microhabitats in multidimensional scatterplots and the evaluation of the spatial segregation between two classes of points. The test was a 3-D extension of a 2-D point pattern spatial analysis that classifies each point by its type and that of its nearest neighbour and compares the proportion of each kind of pair with that expected by chance (i.e., a join-counts analysis of a binary label according to a nearest-neighbour matrix, testing for differences against a random labelling model [124, 125]). A number of pairs of equally labelled points greater than expected by chance indicates that used or available points were aggregated in the PCA space. Global spatial segregation between classes of points was evaluated with a 2 × 2 contingency table, with expected frequencies and statistic (C, asymptotically distributed as Chi-square with 2 d.f.) as proposed by Dixon (NN test [126]). When significant evidence of global segregation was found, each type of pair was tested with an asymptotically normal Z statistic [126]. Statistical analysis were done in R [127], modifying the functions provided by Dixon [128] to a multidimensional case to obtain the matrix of Euclidean distances between points and identify nearest neighbours in a 3-D case.
To evaluate selection on each principal component, the null hypothesis that microhabitat characteristics of used devices are a random sample of those of the available ones was tested with randomization tests [129,130,131]. Statistics of central tendency (mean, median) and dispersion (variance) were calculated from 4999 or 1999 samples, respectively, of the same size as observed (used), taken without replacement from the available values, to evaluate selection consisting in a skewed use of lower or higher values of the environmental variable (resulting in lower or higher mean or median) and selection consisting in avoiding extreme or central values (lower or higher dispersion, respectively; see [120, 132, 133]). Results based on the median of the distributions (not reported) were very similar to those based on the mean. A pseudo-P value associated to the hypothesis that the observed statistic was obtained by chance was calculated as double the number of equal or more extreme values than the observed in the distribution, divided by the number of samples taken including the observed (i.e., a two-tailed test). PCA scores were spatially independent on most axes (see “Results”), so no correction was applied to statistically evaluate selection hypotheses at the microhabitat scale.
Two sets of confidence intervals and probabilities were calculated, based on different null models. First, randomly chosen values were obtained in each iteration from the 300 values of all available microhabitats to evaluate selection assuming no selective use of space at bigger scales (up to the extent of the study). Second, a stratified null model was done to control for a possible habitat selection at the grid scale (i.e., assuming a hierarchical use of space based on the observation that grids differed systematically in the proportion of used devices, see “Results”). Under this model, random samples of the same size as observed were taken from each grid, so the expected mean value was an average of the mean values of the three grids weighted for the number of used devices in each of them.
Seed removal in each 10 × 10 grid was examined for spatial autocorrelation using spatial analyses for non-continuous data, assuming an isotropic process. Spatial autocorrelation of seasonal seed removal and of microhabitat characteristics were evaluated comparing a measure of similarity between pairs of points given by their position with another determined by the focal variable (agreement between two matrices of similarity [124]). Discrete Euclidean distances between devices (from regular grids) were aggregated in four distance classes: < 8.5 m (nearest 8 neighbours of a focal central grid point), 8.5–12.5 m (12 neighbours), 12.5–17 m (16), and 17–22 m (24). Relationship between points given by distance (known as matrix of weights, W) were considered binary, with 1 indicating that two points were separated by a distance in the focal distance class, and 0 otherwise. Two similar spatial analyses were used according to the type of variable [124, 134, 135]: join-counts for binary variables (seasonal seed removal in each distance class) [136], and Moran’s I statistic for continuous variables (values of the main PCA axes, distance to trees). Correlograms from both analyses can be interpreted in a similar way: values higher than expected indicate positive spatial autocorrelation at the focal distance class (i.e., aggregated pattern), and values significantly low indicate negative spatial autocorrelation (i.e., overdispersed or regular pattern).
The number of pairs of devices that were both used (1–1 joins) was compared against expectations from two different models: (1) Complete Spatial Randomness (CSR), in which the probability of use of a device follows a homogeneous Poisson process depending only on the observed number of devices used in a grid, with no spatial interaction (i.e., all devices have the same chance of being used independently of its neighbours), and (2) Distance-To-Tree Model (DTT), a very simple heterogeneous Poisson producing aggregation of seed removal from first order or induced autocorrelation associated with distance to tall trees. This analysis tests the observed configuration keeping the observed composition, edge effects and potential habitat selection at bigger scales (i.e. different use of grids), assuming no second order autocorrelation (i.e., the use of a device is independent of that of its neighbours except for the modelled first order autocorrelation). The expected distribution of the statistic under each model was estimated with 1999 random samples of grid points of the same size as observed (with no replacement). For CSR, all points shared the same probability to be selected. For DTT, probability of use varied inversely with distance to the nearest tall tree: a device had a relative probability of use of 0.6 if at < 5 m or 0.3 if at 5–10 m of an algarrobo > 4 m high and a probability of 0.1 if at < 10 m to the nearest tree > 3 m high; points at > 10 m of any tall tree (none in grid J, 4 in F and 26 in V) had no chance of being selected (probability = 0). Expected values (median) and limits of confidence intervals (percentiles 2.5% and 97.5%) were obtained from each of the generated distributions, estimating the probability P that the observed value belongs to the distribution under the expected model as two times the proportion of equal or more extreme values than the one observed, including it (i.e., a two-tailed test with n = 2000). Simple algorithms for resampling and join-counts estimation were programmed in R [127]. Moran’s I statistics for continuous variables were analysed in a similar way, comparing them against a CSR as implemented in the software [136]. Complementary spatial analyses with alternate techniques (SADIE [137]) provided similar evidence (see Additional file 1, and [60] for details).
