Trait‐based ecology at large scales: Assessing functional trait correlations, phylogenetic constraints and spatial variability using open data

The growing use of functional traits in ecological research has brought new insights into biodiversity responses to global environmental change. However, further progress depends on overcoming three major challenges involving (a) statistical correlations between traits, (b) phylogenetic constraints on the combination of traits possessed by any single species, and (c) spatial effects on trait structure and trait–environment relationships. Here, we introduce a new framework for quantifying trait correlations, phylogenetic constraints and spatial variability at large scales by combining openly available species’ trait, occurrence and phylogenetic data with gridded, high‐resolution environmental layers and computational modelling. Our approach is suitable for use among a wide range of taxonomic groups inhabiting terrestrial, marine and freshwater habitats. We demonstrate its application using freshwater macroinvertebrate data from 35 countries in Europe. We identified a subset of available macroinvertebrate traits, corresponding to a life‐history model with axes of resistance, resilience and resource use, as relatively unaffected by correlations and phylogenetic constraints. Trait structure responded more consistently to environmental variation than taxonomic structure, regardless of location. A re‐analysis of existing data on macroinvertebrate communities of European alpine streams supported this conclusion, and demonstrated that occurrence‐based functional diversity indices are highly sensitive to the traits included in their calculation. Overall, our findings suggest that the search for quantitative trait–environment relationships using single traits or simple combinations of multiple traits is unlikely to be productive. Instead, there is a need to embrace the value of conceptual frameworks linking community responses to environmental change via traits which correspond to the axes of life‐history models. Through a novel integration of tools and databases, our flexible framework can address this need.


| INTRODUC TI ON
Trait-based ecology uses the phenotypic characteristics of organisms to study biodiversity responses to environmental change. We define 'functional traits' as characteristics strictly inherent to the organism, requiring no reference to external conditions (Violle et al., 2007), which contribute to fitness (Cadotte & Tucker, 2017). Via their expression within species assemblages, functional traits indirectly control ecosystem functioning (Moore & Olden, 2017;Wilkes et al., 2019).
Two potential advantages of trait-based ecology over traditional taxonomy-based approaches are improved mechanistic understanding of species-environment relationships and greater generality of those relationships over large (e.g. continental-global) extents (Verberk, Van Noordwijk, & Hildrew, 2013). There is a long history of ecological research on the role of traits in predicting shifts in community composition along environmental gradients (Grime, 1977;Poff, 1997;Southwood, 1977;Townsend & Hildrew, 1994;Winemiller, Fitzgerald, Bower, & Pianka, 2015). Several initiatives have collated trait information for different groups of organisms, and multiple traits are routinely measured on individual organisms or referenced from databases after taxonomic identification (BirdLife International, 2019;FishBase Consortium, 2018;Kattge et al., 2011;Schmidt-Kloiber & Hering, 2015;Vieira et al., 2006).

Applications of trait-based ecology have driven important new
insights into biodiversity responses to global environmental change (Brown et al., 2018;Díaz et al., 2016;Newbold et al., 2012;Stuart-Smith et al., 2013). However, to make further progress, there are at least three major challenges that need to be overcome when working at the largest scales. These challenges involve (a) statistical correlations between trait, (b) phylogenetic constraints on the combination of traits possessed by any single species, and (c) spatial effects on trait structure (occurrence probability-or abundance-weighted means of traits in a community) and trait-environment relationships (statistical links between trait structure and environmental variables).

| Statistical correlations and phylogenetic constraints
The vast majority of trait-based studies have searched for single trait-environment relationships, with few results supporting a priori predictions (Hamilton et al., 2019;Peres-Neto, Dray, & Braak, 2017). There is now ample evidence to suggest that such negative results are due to two related issues affecting trait independence. First, statistical correlations between traits (challenge 1) arise because a combination of traits is more adaptive in a given environment (e.g. multivoltinism and small body size) and, therefore, species possessing one trait are more likely to also possess the other (Resh et al., 1988;Usseglio-Polatera, Bournaud, Richoux, & Tachet, 2000). Second, trait non-independence arises due to phylogenetic constraints (challenge 2) whereby closely related species possess a similar combination of traits because of shared ancestry (de Bello et al., 2017;Verberk et al., 2013). Traits do not evolve independently in response to a single selection pressure but as specific combinations of traits linked through evolutionary processes such as trade-offs (investment in one trait reduces the resources available for another) and spin-offs (investment in one trait increases fitness advantages conferred by another trait; Verberk et al., 2013).
Consequently, a causal mechanism by which a trait appears to influence a species' persistence may be related to another, correlated trait. Spurious causal relationships may seriously confound trait-based applications (Moor et al., 2017;Pilière et al., 2016;Poff et al., 2006;Webb, Hoeting, Ames, Pyne, & Poff, 2010;Weemstra et al., 2016). This problem of disentangling cause and correlation can be circumvented by delineating relatively independent sets of traits that respond to key environmental gradients (Verberk et al., 2013). In generalized life-history models, these axes relate to resource availability and resistance and resilience to stress and disturbance (Van Looy et al., 2019;Winemiller et al., 2015).

| Spatial effects on trait structure and traitenvironment relationships
A major purported benefit of using a trait-based approach in largescale ecology is that it offers additional information over taxonomybased analyses because trait structure is less confounded by biogeographical processes limiting species' distributions (Dolédec, Statzner, & Bournaud, 1999;Menezes, Baird, & Soares, 2010). Ideally, trait structure and trait-environment relationships would be consistent across whole continents, or even globally, enabling the development of generalized predictive frameworks (challenge 3). Implicit in the definition of this ideal scenario is a general use of the term 'spatial', which includes both the spatialized environment and 'pure' spatial effects (Clappe, Dray, & Peres-Neto, 2018). Hence, under the current paradigm, if communities are represented using traits, responses to a given environmental change would be similar in any location, regardless of the spatial structures underlying species' distributions.
To take an example from the freshwater realm, previous global (Brown et al., 2018), continental (Blanck & Lamouroux, 2006;Statzner, Bis, Dolédec, & Usseglio-Polatera, 2001;Statzner, Dolédec, & Hugueny, 2004) and river basin scale (Dolédec et al., 1999;Heino, Schmera, & Erős, 2013) meta-analyses have shown that while trait-environment relationships are generally consistent across those scales, systematic spatial effects on trait structure are often clearly evident. For instance, working in alpine streams, Brown et al. (2018) reported a significant contraction of functional richness and a shift in trait structure towards taxa with smaller body sizes and shorter life cycles, as well as shifts in diets, at higher latitudes. In contrast, Statzner et al. (2001) reported low variability in trait structure and trait response to disturbance among a collection of datasets from various stream types in Europe. As yet, however, there have been no explicit, spatially continuous assessments of the variability of trait structure at continental scales, nor of how this spatial variability can confound the results of trait-based studies. Our ability to deliver such a comprehensive analysis has traditionally been hampered by a lack of data at the relevant scales.

| Integrating open data sources to test assumptions of trait-based ecology
There are now >1 billion species' occurrence records in the Global Biodiversity Information Facility (GBIF), and >1.7 million sequences publicly available in the Barcode of Life Database. The Open Tree of Life now has >2.6 million tips in its synthetic phylogenetic tree.
We stipulated three basic criteria for trait independence: (a) there should be minimal statistical correlations between traits; (b) traits should be minimally constrained by phylogeny; and (c) trait structure, and its response to environmental change, should be minimally variable with respect to spatial coordinates at the scale considered.
These criteria correspond to the three challenges introduced above.
Due to the confounding effects of trait correlations and phylogenetic constraints acting on the full set of trait data, we hypothesized that trait-environment relationships would be more strongly evident when trait categories violating these criteria were excluded from the analysis (Van Looy et al., 2019).
By combining openly available environmental data and species' occurrence, trait and phylogenetic records with computational modelling, we establish a new, generalized analytical framework for quantifying trait correlations, phylogenetic constraints and spatial variability at large scales. We demonstrate its application with a case study on freshwater macroinvertebrates and test the implications of our findings for trait-based applications by re-analysing published macroinvertebrate community data from stream sites spanning the major alpine regions of Europe.
Subsequently, we discuss present capabilities and recommend future directions in trait-based ecology.

| Framework development
Our framework applies three separate analyses corresponding to the criteria stipulated above (Figure 1). The steps involved in our analyses are directly applicable to fuzzy coded traits assigned at mixed taxonomic levels but may be adapted for application to any trait types and taxonomic resolutions. Trait correlations are assessed by resampling the observed species × traits matrix a large number of times, respecting the rules of the fuzzy scoring system used to quantify traits ( Figure 1a). This is necessary because the rules used to assign fuzzy scores can vary between trait categories and taxa, potentially introducing correlations which are artefacts of the scoring system.
For each sample of the species × traits matrix, a correlation matrix is produced, generating null distributions of correlation coefficients (Spearman's ρ) for each pair of traits (r null ). Observed pairwise correlations (r obs ) are then ranked among the null distributions to derive two-tailed p values, Bonferroni adjusted for multiple pairwise comparisons. For applications involving exclusively binary or continuously measured traits, this step may be simplified to a standard correlation test. In cases where binary or continuously measured traits are mixed with fuzzy coded traits then our framework may be applied directly.
To enable phylogenetic constraints to be quantified, a phylogenetic tree is constructed with tips corresponding to species within the taxa included in the trait database at mixed taxonomic levels ( Figure 1b). A trait distance matrix is produced from the observed traits and the constraint quantified as the cumulative correlation between traits and phylogenetic distance, compared to a null hypothesis of no phylogenetic autocorrelation. Separate trait distance matrices may be produced to quantify phylogenetic constraints for different groups of traits. The analysis is repeated by sampling species-level branches in the phylogenetic tree a large number of times (represented by grey phylogenetic trees in the background of This procedure is repeated a large number of times to assess the sensitivity of results to spatial sampling. If sufficient abundance data exist at the requisite scale, abundance-based species distribution models (including joint species distribution models) and abundance-weighted mean traits can be used. Further description of the steps involved in applying the framework is given below in the context of our case study.

| Study area
For the purposes of demonstrating our framework, we defined our study area as the extent of European Union Member States, excluding the Outermost Regions held by France, Spain and Portugal (European Union, 2019). To avoid large gaps within this geopolitical region that may have confounded the analyses, we also included the United Kingdom, Norway, Switzerland, Albania, Bosnia and Herzegovina, Montenegro, Macedonia and Serbia. In total, the study area extended to 5,110,076 km 2 .

| Model organism group
In Europe, trait-based freshwater ecology is set to play an increasingly important role in biomonitoring (Reyjol et al., 2014).
All macroscopic (macro)invertebrates, including insects, crustaceans, molluscs and other major groups are considered within the scope of monitoring. This role for trait-based ecology using macroinvertebrates is supported by well-established trait databases (Schmidt-Kloiber & Hering, 2015). The most comprehensive of these databases, compiled by Tachet, Richoux, Bournaud, and Usseglio-Polatera (2010), contains a set of 63 functional trait modalities in 11 categories, covering 484 taxa classified at mixed taxonomic levels and coded using a fuzzy scoring system (Chevenet, Dolédec, & Chessel, 1994; Table S1). For clarity, we use the term 'trait category' (e.g. food) to refer to a group of 'trait modalities'

F I G U R E 1
Overview of the framework, using a hypothetical example of four taxa (s) and four fuzzy coded trait modalities (t) assigned at mixed taxonomic resolutions. (a) Trait correlations are assessed by resampling the observed s × t matrix a large number of times, respecting the rules of the fuzzy scoring system used to assign traits ( Figure S1), to produce null distributions of correlation coefficients for each pair of traits (r null ). Observed pairwise correlations (r obs ) are then ranked among the null distributions to derive p values. (b) Phylogenetic constraints are quantified by constructing a phylogenetic tree with tips corresponding to species represented in the trait database. A trait distance matrix is produced from the observed traits and the constraint quantified as the cumulative correlation between traits and phylogenetic distance. The analysis is repeated by sampling species-level branches (e.g. s2a, s2b, s2c) in the phylogenetic tree a large number of times to quantify the sensitivity of results to the mixed taxonomic resolution of the trait database. (c) Spatial variability is assessed by first fitting species distribution models for each genus represented in the trait database, then randomly sampling a large number of grid cells (x) within the study area to generate predicted probability occurrences (p occ ) in each sampled cell (x = 4 shown here for demonstration purposes). Community (occurrence probability)-weighted means of each trait (p trait ) are then calculated and gradient analysis used to quantify the turnover of taxonomic and trait structure in space. The analysis may be simplified where binary or continuously measured traits are available at the species level Previous work in North America found that many of the available traits for freshwater insects were highly collinear and tightly linked to phylogeny among the 311 taxa considered . A set of traits describing feeding mode, dispersal mode, body size and voltinism (number of generations per year) were identified as relatively unaffected, or 'labile' (i.e. more readily altered through evolutionary processes, independently of other traits). Others have suggested that strong interrelationships among some macroinvertebrate traits mean that traits should not be analysed as independent variables but rather as linked sets or 'syndromes' (Verberk et al., 2008a(Verberk et al., , 2008b(Verberk et al., , 2013. To date, however, there is no consensus on the extent of non-independence among macroinvertebrate traits, despite a decade of debate (Menezes et al., 2010;Statzner & Bêche, 2010).
Based on the full set of data published in Tachet et al. (2010), traits encompassed categories of body size (maximum body length), lifespan, voltinism, aquatic stages (eggs, larva, pupa, adult), reproduction mode, dispersal mode, resistance forms, respiration mode, locomotion, food and feeding mode (Table S1). After removing taxa with incomplete trait information, a total of 443 taxa in 23 orders and 152 families remained. Of these, most had traits assigned at the genus (52%) or species (39%) level. The remainder (mostly Diptera) were assigned at the tribe, subfamily or family levels (9%). This trait database was used as the basis for three sets of analyses designed to assess trait correlations (Table 1), phylogenetic constraints (Table 2) and spatial variability (Table 3). More details on the steps involved are found in the corresponding tables and the Methods S1. The analyses were conducted in R version 4.0.0 (R Core Team, 2020) unless otherwise specified. Key R scripts for estimating trait correlations, phylogenetic constraints and spatial gradients in taxonomic and trait structure are available in the macroTraits GitHub repository (https:// github.com/wilke sma/macro Traits).

TA B L E 1 Steps involved in quantifying trait correlations
Step Description Source data Method (package) a1 Assess the fuzzy scoring system used to quantify traits Tachet et al. (2010) Quantify minimum and maximum possible fuzzy scores for trait modalities within each trait category (see Figure S1). See function fuzzy_trait_correlations in GitHub repository a a2 Generate an ensemble of possible species × trait matrices Step a1 Resample the species × trait matrix 1 × 10 6 times, respecting the rules of the fuzzy scoring system. See function fuzzy_trait_correlations in GitHub repository a a3 Create null distributions of correlation coefficients Step a2 Calculate Spearman's ρ between each pair of trait modalities in the ensemble of possible species × trait matrices a4 Assess the significance of observed correlations Tachet et al. (2010); Step a1 Rank the observed correlations among the null distributions to obtain two-tailed, Bonferroni adjusted p values a https://github.com/wilke sma/macro Traits.

TA B L E 2 Steps involved in quantifying phylogenetic constraints
Step Description Source data Method (package) b1 Obtain genetic data Barcode of Life Database (BOLD) Search BOLD for cytochrome c oxidase subunit I (COI) sequences using taxa names from the trait database (bold v0.8.6 in R) b2 Clean genetic data Step b1 Retain the longest sequences with least missing base pairs for each species b3 Align sequences Step b2 Perform multiple sequence alignment using Clustal Omega (msa v1.14.0 in R) followed by removal of gaps and poorly aligned sequences using Gblocks (ips v0.0-7 in R)

| Re-analysis of alpine macroinvertebrate community data
To test our hypothesis that trait correlations and phylogenetic constraints confound trait-based analyses, we re-analysed macroinvertebrate community (abundance) data from the European subset of alpine sites studied by Brown et al. (2018). For all analyses, we compared the results obtained using all 11 trait categories to those obtained after excluding trait categories that most strongly violated our criteria for trait independence. We calculated commonly used functional diversity (FD) indices for each site (Villéger, Mason, & Mouillot, 2008) and estimated the importance of deterministic community assembly processes (dispersal-and niche-based) along the gradient of glacier cover using the same approach as the original paper (Brown et al., 2018). We also computed turnover (mean Bray-Curtis dissimilarity) in taxonomic and trait structure within and between regions. Finally, we tested for relationships between linear combinations of traits and glacial ice cover using the fourth corner (individual trait-glacial ice cover relationships) and RLQ (overall trait-glacial ice cover relationship) methods (Dray et al., 2014). It should be noted that, while the rest of the analyses reported here are based upon trait occurrence probability, this analysis uses abundance-weighted trait structure.
This difference is justified due to the lack of abundance data at the continental scale and the fact that we do not make direct comparisons between inferences generated using occurrence probability-and abundance-weighted data.

| Trait correlations
The maximum possible number of unique trait combinations, expressed as the product of unique combinations of resampled scores within each trait category ( Figure S4), was >2 × 10 36 . Thus, if all traits were truly independent, we would expect all 443 taxa to exhibit their own unique trait profile, yet only 408 unique trait combinations existed in the trait database. This finding indicates trait non-independence, as previously found for stream insects in North America . Statistically significant correlations between trait modalities were widespread (Figure 2). Negative

TA B L E 3 Steps involved in quantifying spatial variability
Step Fill gaps in GBIF data coverage using national or regional datasets (e.g. RCS monitoring data from France), retaining only genus and species-level occurrence records c4 Fit species distribution models Steps c1-c3 Given the environmental layers, use MaxEnt to predict the occurrence probability in every 2.5 km 2 grid cell for each genus represented in the augmented GBIF data (enmtools.maxent function, ENMTools v0.2 in R) c5

Sample spatial gradients
Step c4 Generate an ensemble of 1,000 random samples of 1,000 grid cells (without replacement) from within the study area. Retrieve the geographical coordinates, elevation and predicted occurrence probabilities for every genus in each sampled grid cell. See function spatial_gradient in GitHub repository a c6 Calculate communityweighted means Step c5 For each sampled grid cell in the ensemble, multiply the predicted occurrence probability by the corresponding trait score for each trait modality and genus. See function spatial_gradient in GitHub repository a c7 Quantify spatial variability (turnover) Steps c4-c6 For each of the 1,000 spatial gradient samples, fit gradient forest models to occurrence probabilities and community-weighted mean traits (gradientForest v0.1-17 in R) to quantify the turnover of taxonomic and trait structure along spatial gradients (northing, easting, elevation). See function spatial_gradient in GitHub repository a correlations were only found between trait modalities within the same trait categories, whereas correlations between different trait categories were always positive. This is because the fuzzy scoring system typically assigns each taxon nonzero values in a limited number of trait modalities within each trait category, damping any signal from negative trait correlations.
Trait modalities for food (describing diet) and feeding mode (how an organism obtains food) were highly intercorrelated, as expected (Cummins & Klug, 1979). The mean number of significant correlations (p adj < .05) per trait modality within each trait category was highest for lifespan (37 correlations), dispersal mode (34) and voltinism (32; Figure S5). Many of these highly intercorrelated traits were linked in intuitive combinations whereby, for example, multivoltine taxa also have ephemeral lives (≤1 year), are present in the aquatic environment at juvenile life stages and disperse passively. These same taxa are also more likely to have smaller bodies. However, body size, along with resistance forms, was the least correlated with other traits (24 correlations each), and also minimally correlated with one another.

| Phylogenetic constraints
All trait categories were significantly constrained by phylogeny but to varying degrees (Figure 3). independence, these trait categories could offer most additional information over taxonomy-based analyses. Food was marginally less constrained by phylogeny than feeding mode, supporting the notion that macroinvertebrates obtain a more generalist diet than expected based on mouthpart morphology and feeding behaviour alone (Dangles, 2002;Tomanova, Goitia, & Helešic, 2006). F I G U R E 2 Statistically significant (Bonferroni adjusted p < .05) correlations between trait modalities, with colour representing the sign of the correlation. See Table S1 for full names of trait modalities and trait categories

| Spatial variability
Taxonomic and trait structure were approximately equal in terms of their total variability (maximum cumulative importance) along spatial gradients (Figure 4c,e,g). However, trait turnover along the latitudinal gradient ('northing' in the reprojection) was much steeper than taxonomic turnover, with a clearer delineation of Mediterranean regions. Across the continent, the environmental correlates of elevation were associated with a more consistent effect on trait structure regardless of location than was the case with taxonomic structure. This can be seen by comparing maps of taxonomic ( Figure 4a) and trait (Figure 4b) turnover. With trait turnover the major mountain regions consistently occupy a similar part of ordination space (green colours in Figure 4b

| Implications for trait-based ecology
While our findings support the notion that trait non-independence (statistical correlations and phylogenetic constraints) is widespread, not all traits were equally affected. Trait categories describing body size, resistance forms and, to a lesser extent, food were the most labile, a similar result to previous work on stream insects in North America . These three trait categories correspond to the typical axes of life-history models, namely resilience (smaller bodies are associated with r-selected species), resistance to disturbance (resistance forms) and resource utilization (food), respectively (e.g. Townsend & Hildrew, 1994;Van Looy et al., 2019). Although these same traits were clearly associated with systematic spatial variation, the gradient forest analysis suggested that patterns in trait structure are more consistently related to strong environmental gradients, such as those associated with elevation, when compared to patterns in taxonomic structure. Low spatial variability in trait structure was evident for major mountain regions. Thus, largescale trait-based analyses across distant mountain regions are less F I G U R E 3 Phylogenetic constraints on traits. Cumulative correlation of trait modalities within 11 trait categories (a-k) with phylogenetic distance (sum of branch lengths) sampled from the species-level phylogenetic tree ( Figure S2)  Table S1 for names of trait modalities  confounded by spatial structure than taxonomy-based analyses, but they may still be affected by trait non-independence. We therefore re-examined the alpine dataset from Brown et al. (2018) to test the hypothesis that trait non-independence confounds trait-based applications at large scales. We compared trait-based indicators calculated using the full set of traits to those calculated using only body size, resistance forms and food.
Functional richness (related to the number of traits represented in the community) and functional evenness (related to both the number of traits and the abundance distribution) were highly sensitive to the set of traits used (Figure 5a,b). In contrast, there was a close relation- Bray-Curtis dissimilarities for trait structure compared to taxonomic structure ( Figure 5g). There was little difference in trait turnover between the scenarios including all traits and labile traits only, although labile trait turnover was less in all cases (Figure 5g). Using the labile subset of traits did not result in a stronger deterministic community assembly response to increasing glacier cover (Figure 5h,i), a gradient of habitat harshness associated with decreasing channel stability, water temperature and organic matter (Jacobsen & Dangles, 2012). This is because the community assembly model is based on reduced dimensionality trait space (Brown et al., 2018), meaning that inclusion of additional, highly intercorrelated traits had a negligible impact on the result, and may have led to overfitting (Figure 5h). Including region and/or stream catchment as random effects failed to improve the fit of generalized additive models of deterministic community assembly processes relative to the global model (Table S2). This indicates the presence of a consistent trait-environment relationship across the major European alpine regions (Table S3). However, the fourth corner (Table S4) and RLQ (

| CON CLUS IONS
We have shown how three major challenges (statistical independence, phylogenetic independence and spatial variability) in largescale trait-based ecology can be better understood using openly available ecological, phylogenetic and environmental data. In the case of freshwater macroinvertebrates, traits were strongly intercorrelated and constrained by phylogeny, although certain traits were less affected (body size, resistance forms, food). Our findings support the applicability of a life-history model for this diverse group comprising axes (and corresponding traits) of resilience (body size), resistance (resistance forms) and resource utilization ( We have harnessed advances in data availability and computational analyses to establish a novel, generalized framework for large-scale trait-based ecology. The framework can be applied to any taxonomic group and habitat type, although the specific approach to assessing statistical correlations, phylogenetic constraints and spatial variability should be adjusted to reflect the type of trait data available (continuous, binary or fuzzy; taxonomic resolution), as well as the availability and coverage of phylogenetic and occurrence or abundance data (Figure 1). While the framework provides quantitative information on the performance of each trait category, background knowledge specific to the taxonomic group and habitats of interest must be used judiciously in future applications. For example, in addition to results gained by applying the framework, reference to existing life-history models for the group of interest will be useful to guide the selection of traits.
Progress towards a universal framework for making robust predictions of ecological responses to environmental change across major habitat types and taxonomic groups depends on identifying traits directly related to niche dimensions (Winemiller et al., 2015). The identification of labile traits, such as that we have demonstrated here, may lead to better indicators of community resilience and resistance to disturbances (Van Looy et al., 2019). Such important work will improve our understanding of the role of traits in controlling ecological stability following disturbance (Donohue et al., 2013;Pennekamp et al., 2018;Radchuk et al., 2019), across time (Yang, Fowler, Jackson, & Donohue, 2019) and space (Zelnik, Arnoldi, & Loreau, 2019).
Improved knowledge of these controls will be central to our ability to forecast future ecosystem dysfunction, and therefore inform efforts to prevent, mitigate and adapt to global environmental change.

DATA AVA I L A B I L I T Y S TAT E M E N T
The R scripts implemented in this study are archived at https://github. com/wilke sma/macro Traits. The existing datasets that support the findings of the case study can be obtained from the original sources cited in the manuscript.