1. Introduction

Understanding how biodiversity is shaped across spatial scales is still hotly debated in ecology and biogeography (Lamanna et al., 2014; Violle et al., 2014; Xu et al., 2023). This topic has mainly been addressed at local and regional scales (Wang et al., 2012) but has recently been expanded to continental (Andrew et al., 2021) and global scales (Xu et al., 2023). Biodiversity patterns depend highly on spatial scales, such as local sites or regions (Karadimou et al., 2016; Zhang et al., 2018; Keil and Chase, 2019; Suárez-Castro et al., 2022). Regions refer to large land units delimited by the distribution of distinct natural communities (Olson et al., 2001). Studies also suggest that data types (i.e., species occurrence and abundance) affect biodiversity patterns, with consequences for our interpretation of how diversity relates to geoclimatic filters (Anderson et al., 2011; Zhao et al., 2021; Liu et al., 2022).

Occurrence-based diversity indices place substantial weight on rare species, whereas abundance-weighted indices emphasize the role of numerically dominant species (Anderson et al., 2011; Legendre and Legendre, 2012). Although predicting how spatial scale influences diversity is relatively straightforward (Karadimou et al., 2016; Zhang et al., 2018; Oliveira et al., 2023; Xu et al., 2023), predicting the influence of data types is somewhat more challenging, and often overlooked by studies (Baselga, 2016; Liu et al., 2022). Nevertheless, comparing results from different data types can help elucidate the complex nature of biodiversity. Still, few studies have explicitly addressed the importance of data type for diversity indices (Anderson et al., 2011; Baselga, 2016; Zhao et al., 2021; Liu et al., 2022).

Biodiversity has multiple complementary components, with functional diversity (FD hereafter) at the center of most recent studies (Mori et al., 2018; Graco-Roza et al., 2022). FD reflects the range and variation of functional traits within or between local communities (Petchey and Gaston, 2006). Functional traits are characteristics of individuals' size, form, shape, and weight that impact individual fitness (Violle et al., 2007). The emergence of FD has become central to spatial comparisons because it helps to scale up processes from species to multiple ecosystem functions better than species identity diversity (Graco-Roza et al., 2022), particularly changes that occur at biogeographical scales (Violle et al., 2014). FD encompasses the complementary components of alpha and beta variation. The former has a longer history of research. Still, the number of studies centered on β-FD has grown in recent years as it is critical to elucidate large-scale biodiversity patterns (Mori et al., 2018). α-FD reflects the richness, dispersion, and evenness of trait values from local communities without reference to other sites, whereas β-FD reflects functional differences among a set of sites that are created by trait replacement and nestedness (Whittaker, 1960; Anderson et al., 2011; Mammola and Cardoso, 2020). Trait replacement appears when geoclimatic gradients are sufficiently long, and species adaptations are favored in distinct environmental positions (Bevilacqua and Terlizzi, 2020). Furthermore, local communities can be a smaller subset of the biota of richer communities due to gains/losses of species and traits, and consequently, a nested pattern emerges (Ulrich et al., 2009; Baselga, 2010).

Numerous underlying processes can govern α- and β-FD gradients, analogous to those operating on species identities. FD can mirror gradients dictated by biotic interactions, present-day climate, topography, and soil properties as species adapt to particular conditions (Cadotte and Tucker, 2017). In most cases, environmental heterogeneity can increase FD. For instance, α richness is expected to increase in more benign and stable environments, such as hot and humid tropical regions, due to a greater number of viable niches, whereas it may decrease in stressful climates and more variable environments due to physiological constraints and narrower niche spaces (Lamanna et al., 2014; Echeverria-Londono et al., 2018; de la Riva et al., 2018; Andrew et al., 2021). Also, β replacement is expected to increase in regions with greater environmental stability, whereas nestedness would increase in regions with high environmental variability (Pinto-Ledezma et al., 2018). Nestedness can also result from dispersal-related processes and nested patterns of resource availability (Ulrich et al., 2009; Baselga, 2010). Besides present-day climates, biodiversity patterns are expected to track historical events such as long-term climatic stability (Xu et al., 2023). Regions with relatively stable climates over glacial-interglacial cycles would present high levels of species replacement due to more opportunities for species speciation, whereas regions with climatic oscillations would present high levels of nestedness due to high species extinctions (Pinto-Ledezma et al., 2018; Moulatlet et al., 2023). Furthermore, there would be a correspondence between biodiversity patterns and human pressures due to losses of small-ranged species in combination with the expansion of large-ranged species, leading to a reduction in β replacement, for example (Xu et al., 2019).

FD patterns are very dependent on spatial scale effects (Karadimou et al., 2016; Zhang et al., 2018; Suárez-Castro et al., 2022), making it ideal for comparisons of site- and regional-scale patterns (Keil and Chase, 2019). For instance, it has been shown that increasing the geographical area of the study implies the inclusion of a greater number of rare species and the dominance of common species, which will ultimately increase α richness but decrease α evenness (Karadimou et al., 2016). Also, FD gradients can be governed by geoclimatic filters at larger spatial scales due to longer environmental and historical gradients, and a greater diversity of trait values common at this scale (Harrison et al., 2020; Suárez-Castro et al., 2022).

Here, we investigated how much FD gradients and the detection of underlying geoclimatic filters depend on data type (occurrence-based vs. abundance-weighted) and spatial scales used (sites vs. regions). The Atlantic Forest woody flora was examined to address the following research questions and hypotheses: (a) How much do data types and spatial scales change FD estimates? We hypothesized that data types would change FD estimates regardless of the spatial scale as species abundances would describe FD differently than species occurrence, with occurrence data often increasing FD indices since rare and common species have the same weight in the analyses (Baselga, 2016; Zhao et al., 2021; Liu et al., 2022); and (b) How much does the effect of geoclimatic filters on FD gradients depend on data type and spatial scale? We hypothesized that geoclimatic filters would be related to FD gradients regardless of data type but be scale-dependent, manifesting stronger relationships at a larger scale (Harrison et al., 2020). Sufficiently long geoclimatic gradients, as well as a greater diversity of trait values common at regional scales, make the detection of ecological filters stronger as species adaptations are favored in distinct environmental positions (Bevilacqua and Terlizzi, 2020; Harrison et al., 2020; Suárez-Castro et al., 2022).

2. Material and methods 2.1. The vegetation domain studied

The Atlantic Forest of South America is a global biodiversity hotspot (Mittermeier et al., 2011), home to over 15, 000 angiosperm species, half endemic (Zizka et al., 2017). It is the Brazilian vegetation domain with the highest historical deforestation rate, with only 12%–28% of the original remaining (Rezende et al., 2018; Ribeiro et al., 2009), ca. 10% of which is protected (Lima et al., 2020). The Atlantic Forest originally covered an area of 1, 300, 000 Km² from the coast to 700 km to the west (Ribeiro et al., 2009). In addition, the Atlantic Forest is one of the three biodiversity hotspots most vulnerable to global change (Bellard et al., 2014). The latitudinal gradient (5° N to 29° S) along which the vegetation domain spans is marked by seven distinct Köppen climatic zones along tropical and subtropical areas (Alvares et al., 2013); a topography from sea level to > 3000 m; and soils from deep and nutrient-rich clay soils to rocky or waterlogged shallow soils, as well as nutrient-poor sandy soils (Scarano, 2009; Neves et al., 2017; Cantidio and Souza, 2019). Abiotic heterogeneities may impose limiting conditions on the vegetation's function, from semi-deciduous and deciduous forests that face water-shortage periods to dense rainforests with high water availability (Neves et al., 2017; Cantidio and Souza, 2019; Vitória et al., 2019; Silva et al., 2021a). Vegetation types surrounding the dense rainforest are often considered environmentally marginal forests (Scarano, 2009; Neves et al., 2017).

2.2. Species and trait datasets

Species occurrence (presence-absence) and abundance data (number of individuals) were derived from the Caaporãn database, which contains 320 literature sources and herbarium records as described by Cantidio and Souza (2019). The literature sources are stored in Silva et al. (2024). Further details regarding the species datasets, description of the selected literature, studied sites, and sampling-inclusion criteria are in Species datasets (Appendix A). Due to the broad geographical scope of our study and the inclusion of different vegetation types, a mixture of sampling procedures was used in floristic inventories that compose the datasets, all with varying sampling protocols (Silva et al., 2024). The most common sampling-inclusion criteria reported by studies was the breast-height diameter (27% of the studies; the majority used BHD = 5 cm), followed by breast-height circumference (18% of the studies; the majority used BHC ≥15 cm). In 41% of the studies, the sampling-inclusion criteria did not apply or were not informative, and the remaining 14% used a variety of other criteria (e.g., plant height, reproductive age, soil-height diameter, or circumference or perimeter). It is worth noting that differences in sampling-inclusion criteria are a common issue faced by current biogeographical studies. Still, using relative abundances instead of the raw data, among other standardized procedures, helps diminish these differences (See Species datasets in Appendix A). We conducted two independent cleaning processes for these datasets, considering several criteria. After these processes, the occurrence matrix ended up with 4001 species and 695 sites, and the abundance matrix ended up with 2714 species and 346 sites. These were not the final matrices used in the main analyses because we had to exclude many species lacking trait information, as detailed below.

An extensive trait dataset was built by compiling over 40, 000 trait records from 188 sources, including published research articles and freely available databases. Literature sources are stored in Silva et al. (2024). Further details regarding the trait dataset, description of the selected literature, and data filtering process are in the Trait datasets (Appendix A). Seven functional traits were studied: maximum height (Hmax), woody density (WD), specific leaf area (SLA), leaf N content (N), leaf dry matter (LDMC), leaf thickness (Lthick), and seed dry mass (Seed) (Table S1). These traits capture key aspects of how plants function and deal with competition for light (Hmax), maintain photosynthesis (SLA and N), regulate water flow and safety (WD, LDMC, and Lthick), produce offspring and disperse (Seed), which is why they are present in most ecological strategy schemes such as the C-S-R triangle (e.g., SLA and LDMC; Grime, 1977), the L-H-S scheme (e.g., maximum height and seed mass; Westoby, 1998), and multidimensional spaces (e.g., WD and N; Wright et al., 2004; Díaz et al., 2016). Also, these traits are adequately represented in the Atlantic Forest woody flora literature to be used in a large-scale study like ours (Silva et al., 2021b).

It is worth noting that a lack of knowledge about biodiversity often hampers the goals of using extensive trait datasets (Hortal et al., 2015), and a high lack of trait information, particularly inflated by rare species, may compromise FD estimates. Therefore, we proceeded with additional cleaning steps of compositional matrices, taking trait matrices as guideless, aiming to exclude several species lacking trait records. After assigning trait values to each compositional matrix, we started by removing species with fewer than three traits, one species at a time, except if (1) seed dry mass was present or if (2) the species removal implied a decrease in the species richness of sites to below nine (minimum richness found in 1-ha sites) or if (3) the species removal decreased the richness of any site to less than 70% of its initial richness (Silva et al., 2024). Species removed were the rarest in both datasets, as suggested by low values of occurrence frequency and relative abundance (Table S2). After these cleaning processes, we were left with an occurrence matrix composed of 1690 species in 690 sites and an abundance matrix consisting of 1198 species in 343 sites, which were further used in the main analyses (Fig. S1). Trees represented 80% of all species, while others were shrubs.

2.3. Trait imputation and estimation of FD

To determine whether the observed trait values were sufficient for use in the datasets (Fig. S2), for each trait we separately calculated their coverage in the entire dataset, as well as the local trait coverage (i.e., the proportion of species with trait values compared to the total number of species per site) (See Missing data and trait coverage in Appendix A). To fill gaps in the trait matrices, we performed one of the most resolved phylogenetic interpolations using the Rphylopars package (Goolsby et al., 2017) for occurrence and abundance data separately in R (R Core Team, 2022). Phylogenetic trees used for trait imputations (Figs. S3 and S4) were reconstructed using non-molecular phylogenetic information with the phylo.maker function in the v.PhyloMaker package (Jin and Qian, 2022), which are explained in detail in Phylogenetic tree reconstruction (Appendix A). The quality of imputations was assessed by checking if general patterns of pairwise trait correlations for observed data differed from those for mixed data (observed and imputed values together). Correlation coefficients were very similar, suggesting that imputations had reliable data (Table S3).

Seven indices of α- and β-FD were computed to cover the multifaceted nature of functional diversity and complementarity between indices (Mammola et al., 2021). α-FD indices were estimated by the framework based on the kernel density of n-dimensional hypervolumes through richness (α_rich - the total amount of trait space available), dispersion (α_disp - how spread and dense the trait space is), and evenness (α_even - how regular the trait space is) (Mammola and Cardoso, 2020). β-FD was estimated by the framework based on probabilistic hypervolumes, which decomposes the overall differentiation among kernel hypervolumes (β_total - total functional dissimilarity between sites) into the replacement of space between hypervolumes (β_replac - differences due to replacement of functional space), and gains/losses of space enclosed by each hypervolume (β_nest - differences in the amount of functional space enclosed) (Mammola and Cardoso 2020). Transformed and z-scored trait values (mean = 0, SD = 1) were used to control for differences in measurement units. α- and β-FD were estimated using functions of the BAT package (Cardoso et al., 2023). We checked and filtered out sites where critical values could underestimate community diversity and cause biases in FD estimates (See Accuracy and biases in FD indices in Appendix A). These indices did not change regardless of the level of species richness and local trait coverages (Figs. S5-S7), with results supporting our decision to use the FD information of the entire set of sites to run the main analyses.

2.4. Determination of the largest spatial scale

We structured the study area in discrete landscape units to create the regional scale. Regions of vegetation physiognomy, species composition, and trait dominance have been proposed for the Atlantic Forest woody flora (Cantidio and Souza, 2019; Silva et al., 2022). Still, these regions do not reflect the dimension of trait diversity that interests us here. For this aim, we applied a regionalization procedure on FD indices exclusively to produce the largest spatial scale of the study rather than to propose a new regionalization for the Atlantic Forest. We used the information on α-FD estimated from occurrence data because it covered the largest number of sites (Fig. S8a).

Regionalization followed three steps. First, we produced three spatially contiguous surfaces, one for each α-FD index, using the geostatistical interpolation approach of Ordinary Kriging (Fig. S8b). Ordinary Kriging interpolations were run by fitting three alternative semivariogram models (Gaussian, exponential, and spherical), with optimal fitting chosen by the lowest Root Mean-Square Standardized, and the highest correlations between observed and predicted values, as well as predicted values and residuals (Wackernagel, 2003). Semivariograms were run using the fit.variogram function of the gstat package (Gräler et al., 2016). Contiguous surfaces (band raster datasets) were created by considering the Exponential semivariogram using the krige function since the best spatial fits were achieved by exponential models in all cases. A cell resolution of 2.5 arc-min (ca. 5 km²) was used. Second, we produced a single-band raster dataset from combining the three contiguous surfaces (Fig. S8c) in ArcGis v.10.1 (ESRI, 2012). Third, a regionalization analysis based on the composite Kriging map was performed using the unsupervised ISODATA classification, following procedures that Silva et al. (2022) applied. Regionalization was run in ArcGIS using the "Iso cluster unsupervised classification" tool with most parameters set as default (minimum class size = 20 and sample interval = 10). However, the ISODATA algorithm requires the suggestion of how many clusters the input band raster would be classified (Memarsadeghi et al., 2007). To simplify this process, we compared regionalizations with a few clusters, starting with three and finishing when key elements of α diversity appeared on the map based on the contiguous surfaces (Fig. S8d). The best regionalization for FD was achieved with six clusters because it was when a region near the southeastern coast appeared (Fig. S8d–iv). This part of the Atlantic Forest is known to have a differentiated compositional flora, phylogenetic pool, and community function (Oliveira-Filho and Fontes, 2000; Rezende et al., 2020; Silva et al., 2021b, 2022), aspects we considered essential to also predict the formation of a FD region.

2.5. Explanatory data

Each site was characterized by 12 variables related to climate, soil properties, elevation, forest stability, and human pressures (Table S4). Climate and elevation data were obtained from the WorldClim project v.2.0 at a spatial resolution of ~1 km² (Fick and Hijmans, 2017). Soil variables were obtained from the Soil Grids database v.1.0 and v.2.0 at a resolution of 250 m² (Hengl et al., 2017). We used averaged soil values from seven standard depths in this database. Forest stabilities over the glacial cycle (120 kyr before the present) and at the glacial maximum (21 kyr) were obtained from Carnaval et al. (2014) and refer to the frequency each studied site changed between forest and non-forest states at 1000-yr intervals regardless of changes in species composition (Carnaval et al., 2014). The human pressures variable was represented by the human footprint index, which summarizes present-day pressures from crop and pasture lands, population density, and roadways, among other factors (Venter et al., 2016).

2.6. Statistical analyses

To verify how much data types and spatial scales change α- and β-FD indices (question 1), we performed two independent analyses. First, for the site scale, we compared occurrence-based and abundance-weighted FD through scatterplots, then quantified how many sites deviated from the total equivalence between indices, excluding 5% of the values around zero. Equivalence reaches along the diagonal with intercept = 0 and slope = 1. These assessments were limited in sampling size by the abundance data. Second, for the regional scale, indices of α- and β-FD were compared for each region through boxplots and one-way ANOVA followed by Tukey's multiple comparisons of means, using the aov and TukeyHSD functions of the stats package (R Core Team, 2022).

To test how much the effect of geoclimatic filters on FD gradients depended on data types and spatial scales (question 2), we performed two analyses. First, for the site scale, we tested the relationships between α-FD and explanatory variables through mixed-effect models using FD regions as the random factor. These models control for spatial autocorrelation among local sites and, therefore, avoid the type I Error common to traditional statistics (Zuur et al., 2009). For β-FD, the relationships were tested through multiple regressions on distance matrices (MRM), which use the Euclidean dissimilarity between pairs of sites for all explanatory variables. Mixed models were not used for β-FD because the comparisons of sites occur not only within regions but also between regions, making the random structure not applicable. Therefore, we included the dissimilarity of spatial distances, computed from geographical coordinates, as a controlled factor in MRMs to account for distance-decay effects (Baselga and Gómez-Rodríguez, 2021). To avoid multicollinearity and inflation effects in mixed and MRM models, we used a small subset of explanatory variables, which were selected based on the variance inflation factor – VIF (Table S5) in combination with a Pearson correlation matrix (Table S6), and an automated model selection using α- or β-FD indices as response variables and geoclimatic variables as predictors (Table S7). VIF was run with the usdm function (Naimi et al., 2014), and the automated model selection with the dredge function (Bartoń 2023). Mean annual temperature, daily temperature range, soil sand content, and elevation were removed due to VIFs ≥ 3 and r ≥ 0.70 (Zuur et al., 2009). Furthermore, we ran single models for the best predictor, with the greatest R² in each case. Variables were box-cox transformed whenever necessary before performing models. Mixed models were performed by the lmer function of the lme4 package (Bates et al., 2015), and MRM models by the MRM function of the ecodist package with 999 permutations (Goslee and Urban, 2007). The Akaike Information Criteria was used to compare non-spatial and mixed models (Zuur et al., 2009). The percentage of the total explained variance of mixed models was obtained by extracting the total effect (conditional R²), fixed effect (marginal R²), and the difference between them (random R²) (Nakagawa and Schielzeth, 2013).

For the regional scale, we performed pairwise correlations between the average value of FD indices and explanatory variables, which were extracted for each region in each data type. Therefore, the sampling size in each correlation was equal to the number of regions. For the significant correlations only, we compared how the patterns of FD variation mirrored the patterns of explanatory variables through boxplots.

3. Results 3.1. Site scale:  effects of data types and geoclimatic filters

Data types influenced the patterns of both α- and β-FD at the site scale (Fig. 1). The occurrence-based α-FD was greater than the abundance-weighted counterparts in 88% of the sites for α_rich, and 81% for α_disp (Fig. 1a and b). The number of sites with such characteristics dropped to 61% for α_even, because abundance data increased rather than decreased α_even for a substantial number of sites (34%) (Fig. 1c). In addition, abundance-weighted indices were greater than the occurrence-based counterparts in 69% of the sites for β_total and 61% for β_replac (Fig. 1d and e). β_nest was equally influenced by data types (Fig. 1f).

Precipitation seasonality and soil depth were the variables more often retained by automated model selections, as well as the best predictors of FD indices (Tables S8 and S9). However, they had weak to no relationships regardless of data type used, as the maximum explanatory power of fixed effects was R² = 0.11 (Tables S8 and S9). The fixed effect remained reduced regardless of whether multivariate or single models were used. However, the random effect explained up to R² = 0.31 (Table S8). The increase in precipitation seasonality slightly increased α_rich and α_disp while decreasing α_even for both occurrence and abundance data in single models (Fig. 2a–f and Table S8). Also, increasing the dissimilarity in soil depth slightly increased β_total and β_replac, while decreasing β_nest (Fig. 2g-l and Table S9).

3.2. Regional scale: effects of data types and geoclimatic filters

The combination of the three α-FD indices through the regionalization procedure revealed six regions of functional diversity in the Atlantic Forest (Figs. 3 and S8a). Up to ca. latitude 25°S, FD regions covered large areas, and no abrupt substitution between regions in space was observed (Fig. 3). Above this latitude, a marked gradient of regions appeared, especially in the southeastern part, where regions replaced each other on relatively short spatial scales from east to west (Fig. 3).

Data types influenced the patterns of both α- and β-FD at the regional scale (Fig. 4). Occurrence data often increased all three α-FD indices (Fig. 4a–c). For α_rich and α_disp, index values were higher in Region 1 (northeastern and southeastern coastal areas) and progressively decreased until reaching lower values in Region 6 (the southernmost part) (Fig. 4a and b). No progressive change among regions was observed for α_even or any β-FD indices (Fig. 4c–f). Abundance data increased β_total in Regions 1, 2, and 3 (up to ca. latitude 25°S) but decreased it in Regions 4, 5, and 6 (below 25°S) (Fig. 4d). Also, abundance data increased β_replac in four cases, but no difference was observed for Regions 5 and 6 (Fig. 4e). Contrary to other indices, abundance data decreased β_nest in all regions (Fig. 4f).

Contrary to site-scale analyses, explanatory variables had strong relationships with α-FD indices at the regional scale, regardless of the data type (Fig. 5 and Table S10). α_rich was influenced by the greatest number of explanatory variables, including soil depth, precipitation seasonality, and forest stability over 120 kyr for both occurrence and abundance data (Fig. 5a–e). α_disp was influenced by soil depth while α_even by cation exchange capacity for either occurrence and abundance data (Fig. 5a–d). These α-FD indices strongly tracked the pattern of variation from Region 1 to 6 in the explanatory variables, as revealed by comparing Fig. 4a–c and Fig. 6a–d. No influence of explanatory variables was observed for β-FD at the regional scale (Table S10).

4. Discussion

Biogeographical studies do not entirely elucidate whether and how much functional diversity is sensitive to different data types (occurrence vs. abundance) and spatial scales (sites vs. regions) (Baselga, 2016; Anderson et al., 2011; Harrison et al., 2020; Liu et al., 2022). The present paper helps in this discussion by showing that the choices behind the use of data types and spatial scales profoundly affect ecological conclusions drawn from functional diversity indices. Furthermore, to our knowledge, this study is the first to investigate α- and β-FD components simultaneously to answer questions related to data types and spatial scales.

4.1. Changes in FD patterns

A key message repeated over time by biogeographical studies is that biodiversity should be ideally studied and mapped across multiple components and spatial scales (Keil and Chase, 2019; Mammola et al., 2021; Suárez-Castro et al., 2022). We would like to add to this message the need to consider different data types, as pointed out by a few other studies (Anderson et al., 2011; Zhao et al., 2021; Liu et al., 2022). Data type was expected to change FD estimates regardless of spatial scale, as species abundances would describe FD differently than species occurrence, with occurrence data often increasing FD indices because rare and common species have the same weight in the analyses (Baselga, 2016; Zhao et al., 2021; Liu et al., 2022). This hypothesis was supported by our findings as occurrence data often increased α-FD while abundance data often increased β-FD at both spatial scales.

The framework used to estimate FD, based on the kernel density of n-dimensional hypervolumes, allows for the weight of species occurrence or abundance (Mammola and Cardoso, 2020). A divergence between occurrence-based and abundance-weighted indices may reveal how sensitive functional diversity is to differences in species abundances. We found that all FD indices were sensitive and dependent on abundance data, particularly α evenness, β total, and β replacement, suggesting that occurrence should not replace abundance for the estimation of FD as they are complementary rather than redundant. However, data type portrayed similar regional-scale patterns in most indices, which, in applied terms, can help studies overcome cases where abundance data are still missing, but only if the aim is to detect regional-scale patterns (Shen et al., 2013). An exception in the regional scale was found for total β-FD, as the effects of data types were context-dependent. In the three regions distributed up to ca. latitude 25°S in the Atlantic Forest, the occurrence-based total β was lower than the abundance-weighted counterparts, but the opposite pattern emerged from the other three regions above this latitude. Overall, these results suggest that plant communities are not subject to the same processes along the latitudinal gradient of the Atlantic Forest (5° N to 29° S).

The comparison of data types also helps us make inferences about the contribution of rare and dominant species to FD despite it being beyond the scope of our study. Occurrence-based FD lays substantial weight on rare species, whereas abundance-weighted FD emphasizes the role of numerically dominant species (Anderson et al., 2011; Legendre and Legendre, 2012). Our results suggest that rare species may play key and consistent roles for α-FD along the entire north-to-south gradient in the Atlantic Forest. However, as abundance-weighted β-FD indices are less influenced by the turnover of rare species (Barwell et al., 2015), we can assume from the total β-FD pattern that community differentiation is more sensitive to the influence of rare species in the subtropical part of the vegetation domain but less toward the north.

4.2. Effects of geoclimatic filters

Our understanding α- and β-FD in the Atlantic Forest woody flora and how FD links to geoclimatic filters is still lacking, as evidence for taxonomic and phylogenetic diversities may not resemble the functional dimension as observed for palms (Freitas et al., 2021), frogs (Paz et al., 2022), and mammals (de la Sancha et al., 2020). It has been shown that the relationship between FD and geoclimatic filters may differ between site- and regional-scale comparisons (Harrison et al., 2020). Therefore, we expected geoclimatic filters to relate to FD gradients regardless of data types but be scale-dependent, manifesting stronger effects at the larger scale. This hypothesis was supported as occurrence-based and abundance-weighted FD indices were both related to explanatory variables, but the strongest relationships appeared for the regional scale. Various finer filters, such as microenvironmental gradients and biotic interactions, can operate at the site scale, among many other unaccounted effects, which together can create different levels of FD, even if regional geoclimatic filters do not differ (Cadotte and Tucker, 2017; Bruelheide et al., 2018). On the other hand, sufficiently long geoclimatic gradients, as well as a greater diversity of trait values common at regional scales, make the detection of ecological filters stronger as species adaptations are favored in distinct environmental positions (Bevilacqua and Terlizzi, 2020; Harrison et al., 2020; Suárez-Castro et al., 2022).

Gradients of functional trait dominance and trait integration (i.e., correlation of multiple traits to each other) have been shown to track geoclimatic filters at the large spatial scale of the Atlantic Forest (Silva et al., 2021a, b). We showed that this line of reasoning also applies to the dimension of functional diversity. Overall, FD was maximized in deeper soils with higher cation exchange capacity, where the climate is marked by greater precipitation seasonality, and where forests have been more stable over the last 120, 000 years. These findings reinforce other evidence of larger trait hypervolumes (equivalent to greater FD) in more productive, hot/humid climates compared to cold/dry climates, likely because in more benign and stable environments, a greater number of viable niches are available (Lamanna et al., 2014; Echeverria-Londono et al., 2018; de la Riva et al., 2018; Andrew et al., 2021).

It is worth noting that explanatory variables associated with α-FD differed from those associated with β-FD. Interestingly, soil factors such as soil depth and cation exchange capacity had, if not greater, at least equal effects than climate on FD patterns. This result is not consistent with the main filters related to temperature and precipitation that govern taxonomic and phylogenetic diversity in the Atlantic Forest (Cantidio and Souza, 2019; Brown et al., 2020; Rezende et al., 2020; Klipel et al., 2022). Across the Americas, β diversity of tree communities tends to decrease toward higher latitudes following climatic gradients, as shown by functional, taxonomic, and phylogenetic data (Lamanna et al., 2014; McFadden et al., 2019). In South America, the climate is a strong force via precipitation regimes in structuring the diversity of tree lineages across the large scale of vegetation domains (Neves et al., 2020). Specifically, in the Atlantic Forest, the climate has affected taxonomic and phylogenetic replacement alone or via its interaction with topographic complexity due to elevations from sea level to > 3000 m (Cantidio and Souza, 2019; Brown et al., 2020; Rezende et al., 2020; Klipel et al., 2022). The fact that edaphic gradients are central to the promotion of FD expands the idea that environmental filters govern the biodiversity patterns in the Atlantic Forest in South America and reinforces other evidence of biogeographical-scale drivers beyond climate (Neves et al., 2017; Cupertino-Eisenlohr et al., 2021).

4.3. Cautionary notes

We would like to discuss results that were not central to answering the research questions but still deserve consideration. The local communities studied differed in trait coverages (i.e., the relative number of trait records compared to the lack of information per site). Still, it did not influence α- and β-FD estimates. These results do not imply that trait coverage effects can be further neglected, but on the contrary, they draw attention to the fact that an in-depth investigation is recommended. It is also worth noting that local trait coverages would likely be a problem for the FD estimation if the non-peer-reviewed grey literature did not exist, as most of the trait data for the Atlantic Forest woody flora (39%) is not yet available in databases or published papers, but only exists, for example, in master dissertations and doctoral theses.

Another point refers to the many rare species missing trait records that were removed from our datasets before FD estimation. The lack of influence of trait coverages on FD estimates also suggests that removing rare species was not problematic for representing the functional diversity of local communities. As our results may not apply to other systems (Leitão et al., 2016), we reinforce the need to consider the contribution of rare species whenever possible or at least be aware of the effects of species removal on FD estimates.

The type of trait used to calculate functional diversity can have strong effects on understanding the relationships between FD and geoclimatic filters. In this study, we used seven traits from different plant organs simultaneously to describe FD, from leaves to wood and seed. We understand that it would be possible to explore traits from different organs separately, and then relate specific cases of FD to specific explanatory variables. For example, excluding seed mass from FD when relating it to the historical forest stability could be of interest, since the causal-effect relationship between them is still unclear. We emphasize that this was not the aim of our study, mainly to avoid complexity in the results, but this type of approach could be adopted by further studies. There seems to be a knowledge gap regarding this approach in research on biogeography and community ecology. Even using species basal area or biomass to represent local community compositions, rather than species abundances, is another avenue of investigation for further studies.

5. Conclusions

Data types change α- and β-FD estimates no matter the spatial scale (occurrence data often increases α-FD while abundance data often increases β-FD). As all FD indices investigated were sensitive and dependent on abundance data, occurrence should not replace abundance for the estimation of FD. However, if the study aims to detect regional-scale patterns, the data type is a minor issue, except for the total β-FD, as the effects of data types can strongly diverge along the latitudinal gradient. Furthermore, the effect of geoclimatic filters on FD gradients depends on data types and is scale-dependent, manifesting stronger effects at the regional scale. The strong role of edaphic gradients expands the idea of biogeographic drivers beyond climate. These findings caution functional biogeographic studies to ideally consider the effect of data types and spatial scales before designing and reaching conclusions about the complex nature of FD. Few large-scale studies have explicitly addressed the importance of data type for α- and β-FD gradients like ours. Therefore, further research is needed to generalize the effect of data types across various other taxa and ecosystems. In the Atlantic Forest context, a change in FD patterns from north to south (a tropical-to-subtropical gradient) implies that conservation efforts should not be generalized.

Acknowledgments

This work was supported by FAPERJ – Fundação Carlos Chagas Filho de Amparo à Pesquisa do Estado do Rio de Janeiro through a post-doctoral fellowship and scientific grant for José Luiz Alves Silva [E-26/204.257/2021], and by CNPq – Conselho Nacional de Desenvolvimento Científico e Tecnológico through a grant for Angela Pierre Vitória [nº 302325/2022–0]. We are also grateful to Professor Thiago Motta Venancio and Dr. Augusto César da Silva for their help in data analyses and processing.

Data statement

All the references used to compile functional traits and species composition data (occurrence and abundance) were stored in Silva et al. (2024).

CRediT authorship contribution statement

José Luiz Alves Silva: Writing – original draft, Visualization, Validation, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Alexandre Souza: Writing – original draft, Validation, Data curation. Angela Pierre Vitória: Writing – original draft, Supervision, Funding acquisition.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.pld.2024.06.004.