Aim: Environmental and spatial factors are broadly recognized as important predictors of beta diversity patterns. However, the scale at which beta diversity patterns are evaluated will affect the outcoming results. For example, studies at larger scales will usually find spatial processes as the main predictor of beta diversity patterns. In this study we evaluate how beta diversity patterns change when analyses are conducted at different scales by reducing the scale of analysis in a hierarchical manner.

Taxon: Chiroptera.

Location: Atlantic Forest biome.

Methods: Information on the occurrence of 59 bat species were obtained from the Atlantic Bats and Species Link database. We partitioned beta diversity into its two components (nestedness and turnover), and calculated these indexes hierarchically: the biome in its entirety (all ecoregions); between larger regions (north, central and south); and between ecoregions within each region. We performed a Generalized Dissimilarity Model (GDM) to identify and predict the turnover of bat species in the Atlantic Forest based on geo-climatic predictors. We obtained 19 geo-climatic data from AMBDATA, an environmental dataset based on different data sources commonly used in species distribution modeling.

Results: We found that turnover was the main component influencing a latitudinal gradient when the biome was analysed in its entirety. However, when the scale of the analysis was reduced, we found that species loss (nestedness component) had a large effect in determining beta diversity dissimilarity. We also found that nestedness was the main pattern explaining beta diversity dissimilarity along a longitudinal gradient.

Main conclusions: Beta diversity patterns changed with the scale of analysis, which indicates that bat species composition does not follow the same pattern throughout the Atlantic Forest. This corroborates the importance of analysing beta diversity patterns at different scales in order to understand how environmental dissimilarity across geographic space can influence species distribution patterns.

Occurrence and geo-climatic data

            Our study was based on 2,626 bat occurrence data points for 525 sites (coordinates) within the Atlantic Forest. Our data came from two sources. We extracted 1,795 occurrence data points for 160 sites from Muylaert et al. (2017). This dataset was compiled by bat specialists who also reviewed the taxonomy of the species and the coordinates of sampling sites. We then used the species list provided in Muylaert et al. (2017) to search for other occurrence records in speciesLink (data downloaded from http://splink.cria.org.br/). We obtained 830 occurrence records of bat species for 364 sites. We reviewed the dataset obtained from speciesLink according to reliability of information regarding: i) coordinates and site correspondence (we used google maps to check if the coordinates were referring to the places indicated), ii) correct taxonomy (we excluded species with “sp”, “ssp”, “cf” and “aff”), and iii) voucher specimens (we only considered records with specimens that were deposited in a museum). We also included a single occurrence record of Natalus macrourus (Trajano, 1984) from Parque Estadual Turístico do Alto Ribeira (PETAR), which was not considered by Muylaert et al. (2017) or SpeciesLink. Occurrence records belonging to the bat families Molossidae, Vespertilionidae and Embalonuridae were not included in our study because they are seldomly captured in mist-nets (Nogueira, Pol &, Peracchi, 1999; Nogueira, Pol, Monteiro &, Peracchi, 2008), which was the predominant method used for sampling bat species represented in our data sources.

            We obtained geo-climatic data from AMBDATA (available at http://www.dpi.inpe.br/Ambdata/index.php). The AMBDATA is an environmental dataset systematized from different data sources and commonly used in species distribution modelling.  It consists of 19 bioclimatic variables at 30 arc-sec resolution (approx. 1 km). These are: 1) annual mean temperature (ºC); 2) mean diurnal range (ºC); 3) isothermality (mean diurnal range divided by annual temperature range, and multiplied by 100); 4) temperature seasonality (standard deviation *100); 5) maximum temperature of warmest month (ºC); 6) minimum temperature of coldest month (ºC); 7) temperature annual range (ºC); 8) mean temperature of wettest quarter (ºC); 9) mean temperature of driest quarter (ºC); 10) mean temperature of warmest quarter (ºC); 11) mean temperature of coldest quarter (ºC); 12) annual precipitation (mm); 13) precipitation of wettest month (mm); 14) precipitation of driest month (mm); 15) precipitation seasonality (coefficient of variation); 16) precipitation of wettest quarter (mm); 17) precipitation of driest quarter (mm); 18) precipitation of warmest quarter (mm); and 19) precipitation of coldest quarter (mm). We also included three non-climatic environmental variables from AMBDATA: 1) tree cover at a 500 m resolution (percentage); 2) elevation (m) at 3 arc-sec horizontal resolution (about 90 m) and a vertical resolution of 1 m and lastly, 3) declivity (degrees) generated from the elevation grid.

Beta diversity and the Generalized Dissimilarity Model (GDM)

            There are various dissimilarity indices to measure changes in species composition between assemblages. We used the Sorensen index (βsØr) as implemented in ‘Betapart package’ – ‘R-project’ (Baselga & Orme, 2012). The input data table consists of the presence and absence of bat species for each study site (latitude and longitude). The package computes the total dissimilarity across all sites, and calculates turnover (Simpson’s index, βsim) and nestedness (the difference between the Sorensen and Simpson index, βsne) components. ‘Betapart’ returns cluster and dissimilarity matrices (between pairwise sites, and pairwise matrices of shared and non-shared species between sites) of turnover and nestedness.

First, we computed total beta diversity and its two components, nestedness and turnover, among the ten Atlantic Forest ecoregions proposed by Olson et al. (2001). Then we split the Atlantic Forest into three larger regions (southern, central and northern). Lastly, nestedness and turnover were calculated among the ecoregions making up each of the three regions. Each region was treated separately.

            We used a species presence data frame with the coordinates of the occurrence sites to perform Generalized Dissimilarity Modeling (GDM), which analyses spatial patterns of pairwise dissimilarity in species composition between sites, using a nonlinear regression matrix. GDM quantifies dissimilarity using the Soresen Index (total beta diversity), then associates the turnover component (βsim) with biological distance (predictor variables) between sites (Fitzpatrick et al., 2013). The GDM procedure was used to predict bat species turnover across the Atlantic Forest based on environmental data. We used the ‘R packageGDM’ (Fitzpatrick & Lisk, 2016) to fit a GDM with the 22 environmental variables and the geographical distance (decimal degrees) between occurrence sites. The latter was calculated using the option ‘geo=T’ in the function ‘gdm’ of the GDM package. We used the parameter ‘weightType= richness’ to weight sites relative to the number of species to minimize sampling bias. We chose not to exclude sites with few species (i.e. less than five) because in our database over 360 occurrence sites had five or fewer species, and 90 sites had 10 species or fewer. Less than 11 sites recorded 50% or more of the total number of species (59), so that excluding sites with few species would lead to a significant loss of data. In addition, the average number of species per site was equal to five, and sites with few species are evenly distributed throughout the biome. Therefore, maintaining all sites while correcting for species richness, even with a low number of species, does not weaken our model (see Fig. S1.2). Patterns of species turnover can be visualized on a raster with RGB colour standards; areas with similar colours contain similar assemblages. The GDM matrix regression used was I-spline with three basic functions, meaning that we used three points (the minimum) to form the I-spline curve (Fitzpatrick & Lisk, 2016). I-Splines can be visualized in a graph showing the relationship of predicted biological distance versus observed biological distance, providing an indication of how species composition changes along each environmental gradient (Fitzpatrick & Lisk, 2016). The selection of the best subset of predictors for our model followed Williams, Belbin, Austin, Stein, & Ferrier (2012): the initial model included all predictors; variables that contributed less than 2% to model explanation were iteratively removed. Variable removal was done on a stepwise basis beginning with the elimination of the variable that contributed the least to model explanation. Variables were reassessed regarding their importance and significance during each step of model reduction (i.e., backward elimination). Our model started with 23 predictor variables and ended with 11.