Revisiting the role of mean annual precipitation in shaping functional trait distributions at a continental scale

Mean annual precipitation (MAP) varies substantially across the globe, impacting the spatial distribution and structure of vegetation (Schimper, 1898). However, evidence for consistent relationships between MAP and the functional traits of the organisms in these ecosystems is equivocal. Indeed, while some early global-scale analyses reported MAP as a key predictor of plant traits (Wright et al., 2005; Moles et al., 2009; Ordoñez et al., 2009), more recent analyses have found relatively weak relationships between traits and MAP (among other broad-scale climatic variables; Moles et al., 2014; Maire et al., 2015; Bruelheide et al., 2018), seemingly at odds with the important role that water availability is predicted to play in determining plant success from first principles (McDowell et al., 2008, 2022). The independent contribution of MAP in shaping the spatial distribution of species traits, at least at the global scale, has therefore remained unclear. An emerging observation is that trait–environment patterns become more pronounced when focussing on regional scales, or within habitat types (Chelli et al., 2019; Guerin et al., 2022; Kambach et al., 2023). Here, focussing on the continent of Australia, we empirically test our theoretical understanding of the relationship between traits and MAP using an unprecedented database, AusTraits – the largest harmonised continent-specific collection of georeferenced trait values globally. Australia represents the ideal laboratory to test trait–MAP relationships for several reasons. First, MAP and mean annual temperature (MAT) are orthogonal in Australia (r = 0.02; Fig. 1a). As such, associations between water availability and traits can be isolated from the effect of MAT, by design. Australia also spans an extraordinary precipitation gradient, encompassing the 22nd (79 mm) and 99th (7625 mm) quantiles of the global distribution of MAP, thereby representing all but the very driest regions of the globe (Fig. 1b). Finally, although Australia is a major land-based carbon sink (accounting for c. 60% of the global terrestrial carbon sink in some years; Poulter et al., 2014), there is also significant uncertainty regarding the effect of precipitation on carbon uptake in this region, which is proposed to emerge from a number of factors including poor representation of drought-adaptation within the highly endemic and structurally distinct vegetation of Australia and significant variation within model ensembles in the simulated or prescribed fraction of woody and herbaceous cover (Teckentrup et al., 2021). Altogether, a re-examination of the relationship between plant traits and MAP in Australia would not only improve our fundamental understanding of the evolution of trait distributions but also yield a timely assessment of the embedded processes in dynamic vegetation models (DVMs) used to simulate ecosystem processes (Teckentrup et al., 2021). We selected eight key functional traits widely considered to capture important physiological processes in vascular plants, and for which sufficient data were available, and generated hypotheses for how each would respond to spatial variation in MAP. Hypotheses were derived from published eco-evolutionary theories explicitly relating traits to MAP or soil moisture and, if these were not available, we inferred predictions from theories based on other moisture-related environmental drivers including vapour pressure deficit (VPD) and site productivity. The selected theories invoke a range of processes including, for example, tissue damage due to leaf overheating, optimisation of plant construction to maximise net carbon uptake and stand-based competition (Table 1). We used bivariate linear regressions to test each of these hypotheses (according to the sign of the relationship) with our overarching hypothesis being that traits with direct theoretical links to MAP would have the strongest correlations. To establish a clearer picture of macroclimatic control on traits in Australia, we also quantify the extent to which trait–MAP relationships are mediated by MAT via interaction (Wright et al., 2017). To account for potential variation in trait responses to the environment due to woodiness, we tested whether observed patterns differed when species were classified as woody or nonwoody. We expected relationships would be stronger in woody taxa because long-lived individuals must function in challenging environmental conditions whereas nonwoody species often avoid dry conditions by surviving as seed. VPD1 Soil moisture2 Wang et al. (2017)1 Paillassa et al. (2020)2 Time to reproductive maturity ↑ Access to light ↑ Productivity1 Moisture stress2 Competition for light greater in mesic environments1 Moisture stress places upper limit on plant height2 Falster et al. (2017)1 Jensen & Zwieniecki (2013)2 Construction cost ↑ Water transport rate ↑ Construction cost ↑ Leaf turnover rate ↓ Photosynthetic rate ↑ Respiration and construction cost ↑ Number of offspring ↓ Competitive ability of offspring ↑ Construction cost ↑ Growth-dependent mortality ↓ For woody taxa, MAP was an excellent predictor (r2 ≥ 30%) of the variation in leaf Δ13C (Δ13C; C3 plants only), leaf mass per area (LMA), maximum plant height (MH) and leaf nitrogen per area (Narea), with LMA and Narea decreasing and MH and Δ13C increasing with MAP (Table 2; Fig. 2). In addition, MAP was a moderate predictor (r2 ≥ 20%) of leaf area (LA) and wood density (WD), being positively and negatively correlated with MAP, respectively. However, MAP was a weaker predictor (r2 ≥ 10%) of the Huber value (SA : LA) and seed mass (SM), being negatively and positively correlated, respectively. Regardless of correlation strength, in all cases the direction of the fitted correlation was consistent with our predictions (Table 1). Combining woody and nonwoody observations tended to cause the amount of variance explained by MAP to decline for the nonwood-related traits (see the Materials and Methods section), with the exception of LA, which increased slightly. Most notably, we observed a reduction in the variance explained for Narea from r woody 2 = 31 % $$ {r}_{\mathrm{woody}}^2=31\% $$ to r overall 2 = 26 % $$ {r}_{\mathrm{overall}}^2=26\% $$ , and a much larger decrease in the variation of LMA explained ( r woody 2 = 38 % $$ {r}_{\mathrm{woody}}^2=38\% $$ vs r overall 2 = 2 % $$ {r}_{\mathrm{overall}}^2=2\% $$ ), to the extent that MAP and LMA were now virtually uncorrelated. These outcomes emerged because, for the most part, trait–MAP relationships were weaker in nonwoody taxa, and in the case of LMA, there was a bimodal distribution in the observations at low MAP due to a greater representation of herbaceous annuals (Fig. 2). In general, including MAT as an interacting predictor with MAP had minimal impact (Δ < 5 percentage points) on the variation explained in each trait for woody or nonwoody taxa, as well as when combined (Table S1). However, there was an c. 10 percentage point increase in r2 in the interaction model compared with the MAP-only model for LA, SM (but not nonwoody) and MH (overall only) which, for LA and MH, mostly reflected a steepening of the MAP slope in warmer climates, and for SM, a mean increase towards warmer climates driven by the main effect of MAT (Figs S1–S3; Table S2). Alone, for woody taxa, MAT was an excellent predictor of MH, a moderate predictor of SM, a weak predictor of LA and SA : LA, and was uncorrelated (r2 ≤ 10%) with the remaining traits (Table S2). In contrast to recent global analyses (Moles et al., 2014; Maire et al., 2015; Bruelheide et al., 2018), we found MAP was a strong predictor of several key functional traits, in a manner consistent with predictions based on the theoretical literature (Table 1). Specifically, as MAP increased, we observed a systematic shift from resource-conservative to resource-acquisitive values for most traits. This response was most prominent in woody taxa, in partial support of our hypothesis, but we also found important exceptions where the explanatory power of MAP was equivalent between growth forms. These findings support an emerging phenomenon that trait responses to macroclimatic gradients become stronger when focussing on distinct regions or within habitats (Buzzard et al., 2019; Chelli et al., 2019; Kambach et al., 2023), as has been shown in Australia for plot-based community-weighted analyses of LA, SM and MH in response to climate (Guerin et al., 2022). Importantly, the rapid expansion of the AusTraits trait database since Guerin et al. (2022) has permitted the significant advancement upon our understanding of Australian trait ecology that we present here by incorporating observations from naturally occurring plant communities across all but the most extreme ranges of the Australian rainfall gradient (including tropical rainforest) for a much greater range of traits, including plant-construction traits which are highly relevant for simulating ecosystem processes in DVMs such as SA : LA, WD and LMA (Sakschewski et al., 2015). The strongest trait–MAP relationship we observed was the positive correlation between plant height and MAP for woody taxa, which was followed closely by when all taxa were analysed simultaneously. This outcome is qualitatively similar to a previous global-scale study of plant height, which found that precipitation in the wettest quarter, and then MAP, were the strongest predictors of plant height (Moles et al., 2009), although MAP explained a substantially greater amount of the total variation in the data in the present study (50% compared to 21%). MH may increase with MAP for several reasons. Maximum achievable height may be biophysically constrained by water availability such that, all else being equal, taller plants can survive in wetter sites (Jensen & Zwieniecki, 2013). An alternative explanation is that on average, natural selection favours taller plants in wetter sites because the benefit of having a higher position in the canopy outweighs the drawback of increased stem construction costs and delayed reproduction as competition for light intensifies (Falster et al., 2017). Similar to MH, we also found that leaf Δ13C and MAP were strongly and positively correlated across all C3 taxa, consistent with a recent, global-scale empirical analysis (Cornwell et al., 2018). Δ13C is a measure of the long-term average of the ratio of the partial pressure of CO2 in the intercellular spaces (Ci) and the atmosphere (Ca). According to least-cost theory, all else being equal, the cost of procuring water for transpiration increases in drier environments (Prentice et al., 2014), leading to reduced stomatal conductance, lower Ci, and causing plants to invest more in photosynthetic capacity relative to water transport capacity (Dong et al., 2017). Indeed, in line with this theory, we also observed that Narea, which is linked to photosynthetic capacity (Dong et al., 2022), was greater in sites with lower MAP regardless of growth form. Remarkably, in contrast to Cornwell et al. (2018), the strength of the correlation between Δ13C and MAP was strong for both woody and nonwoody taxa, suggesting that the processes governing stomatal behaviour are similarly sensitive to MAP in both growth forms. Separately analysing the trait data for woody and nonwoody species also revealed important differences in response to MAP. For example, in line with our expectations based on theory, LMA was found to be strongly and negatively correlated with MAP (i.e. r2 ~ 40%) for woody taxa. LMA is theorised to decline as site productivity increases because the more rapid height growth rates conferred by cheaply constructed leaves, and therefore greater access to light, compensates for high leaf turnover costs associated with the lower leaf lifespan of low LMA leaves (Falster et al., 2017). In addition, higher LMA is hypothesised to be advantageous in more arid sites to maintain leaf function at more negative leaf water potentials (Wang et al., 2023). Qualitatively, our results are consistent with Wright et al. (2004), who found a modest negative relationship (i.e. r2 ~ 10%) between LMA and MAP at the global scale, but only after statistically controlling for MAT. The fact that MAP explained such a large proportion of the variation in LMA in the current study relative to the global scale is intriguing. In Australia, woody trees and shrubs are mostly evergreen, meaning that the well-recognised trade-off between LMA and leaf lifespan is not obscured by a shift to being cold-deciduous. However, even after analysing the data separately for evergreen vs deciduous taxa, Wright et al. (2004) found MAP only explained at most 22% of the variation in LMA. The observed relationship points to the important role of MAP in driving selection on leaf economic strategies for woody taxa across strong rainfall gradients in Australia, and perhaps beyond. By contrast, LMA and MAP were virtually uncorrelated when analysing just the nonwoody taxa and this reflected the general tendency for trait–environment relationships to be weaker within this functional group. Trait interdependencies, or trait trade-offs, which underpin the relationships we observed in woody taxa, may not be present in the nonwoody taxa analysed. For example, an analysis of trait connectivity using data from the TRY database revealed that the direct relationship between leaf lifespan and LMA, which is seen across all taxa and woody taxa, is not present for nonwoody taxa (Flores-Moreno et al., 2019). Alternatively, weaker trait–environment relationships may emerge in nonwoody species because this functional group employs a wider variety of strategies to tolerate harsher or more variable environments, such as occupying relatively benign positions in the understorey or employing dormancy as a bet-hedging strategy in the seed bank, as is the case for many of the annual herbs analysed here (Dwyer & Erickson, 2016). Increasing reliability in DVMs has been accompanied by the development of trait-based approaches, which allow trait values to be predicted directly based on presiding climatic conditions (De Kauwe et al., 2015). Our analysis using a harmonised, data-rich trait database provides clear, albeit correlation-based, evidence for a role of MAP in shaping broad-scale trait gradients in Australia and, consequently, impetus for these relationships to be incorporated into the next generation of DVM simulations of this continent. However, this integration depends on our ability to mechanistically describe the processes that cause these relationships to emerge. To this end, we note that traits with strong correlations with MAP (i.e. LMA, MH, Narea, Δ13C) tended to lack theoretical hypotheses directly linking them to precipitation or soil moisture, being driven instead by site productivity or VPD (Table 1). An open question, however, is the extent to which the strong trait–MAP correlations that we observed are generalisable to other regions, or are instead the outcome of distinct biogeographical processes occurring only within Australia (Gallagher et al., 2023). While it is difficult to address this definitively, we postulate that although the underlying mechanisms driving selection along moisture gradients are present in all systems, other processes may weaken emergent patterns. For example, while Australia is a geologically old continent with minimal recent glaciation, Europe and North America, the focus of much global trait research, experienced substantial glaciation in the Last Glacial Maximum (c. 21 kya). As such, disequilibrium between species distributions and contemporary climate may be greater in these regions, potentially weakening trait–climate associations (Squires, 1988; Svenning & Skov, 2007; Seliger et al., 2021). Trait–MAP relationships may also be more apparent in Australia because, in comparison with the northern hemisphere, a relatively small fraction of the continent experiences freezing winter temperatures (Orians & Milewski, 2007). Consequently, plant growth is not typically limited by seasonal cold periods, emphasising the selective effect of water availability on plant function. We do not rule out temperature as a mediator of the effect of MAP on traits in Australia, however. Indeed, we have demonstrated that for size-based traits like LA and MH, trait–MAP patterns may become more acute in warmer temperatures, as has been shown for LA at the global scale (Wright et al., 2017). The simple analysis that we conducted here also invites further investigation of trait–environment correlations using this unprecedented dataset for Australian taxa. First, while our treatment of species-by-site combinations as independent data points represents the totality of trait variation across sites emerging from both species turnover and across-site intra-specific variation, it does not partition the relative contribution of these components to the observed patterns (Ackerly & Cornwell, 2007). Identifying the sources of this variation would have important implications for our understanding of how trait–climate patterns have emerged historically, and how resilient communities will be to future change (Dong et al., 2020). Second, while our analysis evaluated trait responses to MAP independently of one another, selection operates on multiple traits simultaneously, such that adaption to drought can be achieved through different combinations of strategies (Flores-Moreno et al., 2019). Further analyses using this dataset to investigate how covariance amongst traits varies across environmental gradients (Brown et al., 2022) would yield insight into how these strategies interact to confer plant fitness under different conditions (Fig. S4). In a similar manner, the substantial variation that we observed for some traits across species within sites points to a significant contribution of local-scale species coexistence to trait diversity, the magnitude of which may itself be dependent on climate (Andrew et al., 2021). For example, although we observed a mean negative response of WD to MAP, there was also a strong triangular signal in the data where taxa in mesic sites had a much greater range of WD from very light to very dense wood (Fig. 2). Substantial coexistence of strategies within a given climatic band could also go some way to explain the very weak response of SM to precipitation that we and others (Moles et al., 2005) have observed. The challenge for theoretical frameworks, therefore, is not just to predict the central tendency of trait distributions (Dong et al., 2017; Wang et al., 2017, 2023; Xu et al., 2021), but also the likely diversity of strategies, which may occur under a given climatic regime. Using simple linear correlations, and guided by quantitative predictions from the literature, we have demonstrated that for long-lived, woody taxa, MAP is strongly coordinated with key functional traits. By focussing the scope of this study on Australia, we not only reveal emergent trait–environment patterns for a highly endemic and species-diverse continent (Gallagher et al., 2023), but also demonstrate the potential for continental-scale analyses to act as natural experiments by controlling for distinct phylogenetic history and/or covariation with other environmental factors (e.g. the orthogonality between MAP and MAT in Australia). To this end, we advocate for re-focussing the geographical scope of analyses to regions where observations span strong gradients in the variable of interest independently of other environmental covariates (Kambach et al., 2023). AusTraits (v.4.1.0; Falster et al., 2021) is a harmonised trait database containing over 1000 000 individual records of functional traits for over 30 000 taxa occurring in Australia (including offshore islands and territories). It is therefore the largest collation of trait data for plants growing in Australia and for any whole continent. We queried the AusTraits database for eight functional traits which represent multiple dimensions of plant ecophysiological strategy, are widely distributed across Australia (Supporting Information Figs S5, S6), and for which predictions about the effect of MAP could be derived from the literature (Table 1). These traits were the leaf carbon isotope ratio ( δ l 13 C $$ {\updelta}_{\mathrm{l}}^{13}\mathrm{C} $$ ), maximum height (MH), Huber value (SA : LA), leaf area (LA), leaf mass per area (LMA), leaf nitrogen per area (Narea), dry seed mass (SM) and wood density (WD). Sampling bias is unavoidable in trait data compilations such as AusTraits which assemble datasets from different contributors with their own research interests and sampling strategies (Keller et al., 2023) and, in this case, our analysis does include datasets which intentionally sampled across known climatic gradients in Australia (e.g. Schulze et al., 1998; Fig. S5). Nevertheless, for all traits considered here, the span of the rainfall gradient is not dominated by any one dataset, and thus, we are confident that the benefits of assessing trait–climate relationships at a continental scale across datasets outweighs the costs of potential bias in the data. Because we were interested in how traits respond to variation in climate, we filtered the data to include only observations which had associated geospatial information. In order to sample 'natural' trait observations from AusTraits, we also limited the scope of our study to trait observations made on unmanipulated individuals occurring in situ. In other words, this excluded most data originating from field, laboratory and glasshouse experiments, herbarium records and expert opinion. Finally, we filtered the trait data to remove observations which were recorded as metapopulation (i.e. across sites) or species-level values to ensure that variation within species across sites, if present, was adequately represented. Removing these observations had minimal impact on the total number of observations for each trait (i.e. < 5% of the field-observed traits), with the exception of MH, for which c. 40% of observations were recorded at either the metapopulation or species level. This reflected the fact that some studies in AusTraits assigned the maximum height observed across populations or from published flora descriptions to all observed individuals within the respective dataset. Qualitatively, however, the outcome of the analysis for MH when these observations were removed (Fig. S7) was consistent with when they were retained, although the variation explained by MAP was greater in the former case. Measurement protocols for LA on compound species differed between studies which considered LA at the leaflet vs the whole-leaf scale. The theoretical framework for predicting the response of LA to the environment considered in the present study is based on the rate of conductance through the leaf–atmosphere boundary layer which, for compound species, is likely determined at the leaflet scale (Wright et al., 2017). Thus, to minimise variation in the LA data associated with the varying scale of measurement, we first identified datasets which measured LA on compound taxa and for which LA was measured at the leaf scale, or for which the measurement protocol was unknown. Then, within these datasets only, we used the 'leaf_compoundness' data in AusTraits, which classifies taxa as being 'simple', 'compound' or 'simple compound', to assign taxa as having either simple or compound leaves, assuming 'simple compound' taxa to be compound. LA observations for taxa with compound leaves, or for which compoundness was unknown within these datasets were then omitted. In sum, c. 2000 of the 17 500 independent LA records were removed after this process. We used a number of approaches to check for errors in the data. First, to assess for dataset-level errors, we compared the data distribution for a given dataset against the distribution of all other datasets combined for each trait and visually inspected this comparison for obvious discrepancies. Then, for taxa with multiple observations, especially those with observations from different contributors, we inspected various metrics reflecting the range of likely values that a given taxa could occupy to identify potentially erroneous observations. This process led to the removal of a small number of data points (i.e. n = 15), typically being those which had unusually high or low values relative to other observations of that taxa but also to the distribution of the trait overall. Finally, we inspected the data for duplication of trait values for a given taxon between contributors, which may occur in cases where the data was obtained from a shared data source although no data were removed for this reason. To address our hypothesis regarding the influence of plant growth form on the relationship between MAP and functional traits, we also queried the AusTraits database for the 'woodiness_detailed' trait, using a dataset compiled by Wenk et al. (2023). This trait describes taxa based on both the presence and vertical extent of secondary xylem (i.e. 'true wood') and thus separates taxa into multiple categories. In the case of this study, we did not view plant growth form as a response variable per se. Instead, we pooled taxa with growth form information into two possible groups being either 'woody' or 'nonwoody'. For the 'woody' group, we selected taxa with an erect habit and a lignified stem based on the 'woody' category in the 'woodiness_detailed' dataset. This category was the single largest group in the 'woodiness_detailed' dataset, capturing just over half of the taxa for which information was available. By comparison, the nonwoody group captured a more diverse group of growth forms, including herbs (being the most dominant nonwoody group) but also tussocks, graminoids and palmoid taxa among other classifications. Some taxonomic groups such as monocots and ferns can occasionally produce lignified wood-like material despite lacking secondary xylem. To test whether classification of these taxa into a given group influenced the outcome of our analysis, we conducted our analysis twice, including taxa with the definition 'woody_like_stem' first into the 'nonwoody' and second into the 'woody' group. There was a minor increase in model performance for both the woody and nonwoody group in the latter case, so we primarily report the results of this set of analyses, although there was no qualitative difference in the interpretation of our results in either case (Tables S2, S3). Although our specific aim was to investigate the relationship between traits and MAP, we also selected a range of other climatic variables which are considered to describe soil and atmospheric moisture availability. In addition, we also selected MAT to investigate how temperature mediates responses to MAP. Climate data were obtained from a variety of sources, as no single source provided all variables. MAP, mean precipitation in the wettest quarter, mean precipitation in the driest quarter, mean precipitation in the warmest quarter, mean precipitation in the coldest quarter, mean precipitation seasonality and MAT were obtained from Worldclim at a resolution of 30 arc seconds (c. 1 km at the equator), mean annual potential evapotranspiration from Envirem at a resolution of 30 arc seconds and mean monthly VPD data from 1981 to 2010 from CHELSA at a resolution of 30 arc seconds. For the mean monthly VPD, we averaged the values across months to find a mean annual VPD. Moisture index was calculated by dividing MAP by annual potential evapotranspiration. We then extracted climate data corresponding to the geospatial information associated with trait observations. In peninsular coastal regions where trait observations were recorded, climate data were occasionally not available. In these cases, we used the nearestLand function from the seegsdm package (Golding & Shearer, 2023) to assign these observations with climate data from the nearest non-NA cell within 10 km. If this correction still did not yield climate data, these observations were not included in statistical analyses. Overall, observed relationships tended to be quantitatively similar or weaker when replacing MAP with other variables (Table S2). The most notable exceptions were that SA : LA and MH responded more strongly to precipitation in the warmest quarter. This was especially the case for MH when all taxa were combined, which is explained by the fact that this variable further distinguishes short-statured Mediterranean annual forbs experiencing winter rainfall from productive, summer monsoonal rainforests (Fig. S8). Moreover, there was some evidence to suggest that Narea responded more negatively to MI and LA more positively to precipitation in the wettest quarter and seasonality compared with MAP in both woody and nonwoody taxa when separated as well as when combined. Thus, although we report these results in the Supporting Information, we did not focus on these findings. Our analyses considered that trait variation may occur at two spatial scales being within-site variation and across-site variation. Here, a site is the scale across which broad-scale climatic gradients influence trait variation. Variation within sites is instead assumed to be due to other factors including within-site environmental heterogeneity, as well as local coexistence of multiple functional strategies. Here, we assigned trait observations to sites by overlaying a 30 arc second grid on the continent, corresponding to the resolution of the climate data, where each grid cell represents a site. Before statistical analysis, we found the site-level mean (or maximum, in the case of MH) for each taxon for each trait, yielding 400–5000 unique species-by-site combinations depending on the trait and growth form. In other words, for a given trait, each site has a number of observations equivalent to the number of taxa recorded in it. This means that species-by-site combinations sometimes summarised across observations from different data contributors and through time. Importantly, this approach accounted for the effect of uneven sampling intensity on analyses by summarising taxon-level information to one observation per grid cell. Before statistical analysis, we transformed trait and climate data to meet the assumption of normality for linear regression. Specifically, we log-transformed all traits, log-transformed VPD, MAP, precipitation in the wettest and warmest quarter, moisture index and precipitation seasonality and sqrt-transformed precipitation in the driest quarter. To first test trait responses to MAP across all available taxa, we regressed the species-by-site values of each functional trait against MAP using ordinary least-squares regression. We then assessed the Pearson correlation coefficient as a measure of the strength and direction of correlations as well as the r2 value to understand the percentage of variation explained by MAP. Given the large sample sizes involved in this study, P-values for even weak correlations are likely to be significant, so we do not report them. To assess the explanatory power of MAP compared with other moisture-related variables, we regressed each trait as a bivariate linear function of each of the remaining climatic variables. To test whether the relationship between traits and MAP is mediated by temperature, we also fitted each trait as an interaction between MAP and MAT and assessed the increase in variation explained by the more complex model as a comparative measure of model fit. To test our hypothesis that woody taxa will exhibit stronger responses to the climate than nonwoody taxa, we then repeated the process as described above for the woody and nonwoody groups separately. In all cases, taxa with unknown growth forms were omitted from analyses to permit comparison between patterns when taxa were combined and when separated by growth form. To explore the possibility of nonlinear curvature in the relationship between traits and climate, we fit all of the above models again with an additional quadratic parameter and assessed the r2 value to identify whether including the quadratic term significantly improved model fit. For the most part, model fits were not improved substantially (Tables S2–S4). However, there was some evidence that Narea had a concave-up relationship with MAP whereby there was a steep decline in Narea from arid to mesic regions until c. 1200 mm rainfall where it began increasing again slightly (Fig. 2). Discussions with Will Cornwell, Mark Westoby, David Coleman and John Dwyer enriched the analysis and interpretation of the results. This research was funded by a grant awarded by Eucalypt Australia (Eucalypt futures: using functional traits to predict species distributions and responses to environmental change) to DSF and PAV and grants awarded by the Australian Research Council to DSF and PAV (DP200100555, FT160100113). We also thank diverse contributors to AusTraits. The AusTraits project received investment (doi: 10.47486/TD044, 10.47486/DP720) from the Australian Research Data Commons (ARDC). The ARDC is funded by the National Collaborative Research Infrastructure Strategy (NCRIS). Open access publishing facilitated by University of New South Wales, as part of the Wiley - University of New South Wales agreement via the Council of Australian University Librarians. None declared. IRT and DSF conceived the original idea with further development of the concept by PAV, EHW, RVG, SMW and IJW. IRT conducted the review of trait theory. IRT conducted statistical analysis with assistance from SMW. IRT led the writing of the manuscript. DSF, PAV, EHW, RVG, SMW and IJW assisted with drafting the manuscript. All authors approved of the final draft of the manuscript. Data and code are available on GitHub (https://github.com/traitecoevo/aus_trait_gradients). AusTraits v.4.1.0 is available at https://zenodo.org/doi/10.5281/zenodo.3568417. Fig. S1 Interaction of the trait-mean annual precipitation relationship with mean annual temperature for all taxa. Fig. S2 Interaction of the trait-mean annual precipitation relationship with mean annual temperature for woody taxa. Fig. S3 Interaction of the trait-mean annual precipitation relationship with mean annual temperature for nonwoody taxa. Fig. S4 Pair-wise trait correlation matrix for woody taxa. Fig. S5 Geographical distribution of functional trait data based on dataset. Fig. S6 Geographical distribution trait observations with respect to the Australian climate and Whittaker's biomes. Fig. S7 Comparison of the relationship between maximum height and mean annual precipitation when metapopulation and species-level observations are removed vs retained in the analysis. Fig. S8 Comparison of the relationship between maximum height and mean annual precipitation vs precipitation in the warmest quarter. Table S1 Comparison of model fit in mean annual precipitation (MAP)-only models vs MAP-mean annual temperature interaction models. Table S2 Model fit for trait–environment linear regressions when woody-like taxa are grouped with woody taxa. Table S3 Model fit for trait–environment linear regressions when woody-like taxa are grouped with nonwoody taxa. Table S4 Model fit for trait–environment quadratic regressions. Please note: Wiley is not responsible for the content or functionality of any Supporting Information supplied by the authors. Any queries (other than missing material) should be directed to the New Phytologist Central Office. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.