Functional trait trade-offs define plant population stability across different biomes

Abstract

Ecological theory posits that temporal stability patterns in plant populations are associated with differences in species' ecological strategies. However, empirical evidence is lacking about which traits, or trade-offs, underlie species stability, especially across different biomes. We compiled a worldwide collection of long-term permanent vegetation records (greater than 7000 plots from 78 datasets) from a large range of habitats which we combined with existing trait databases. We tested whether the observed inter-annual variability in species abundance (coefficient of variation) was related to multiple individual traits. We found that populations with greater leaf dry matter content and seed mass were more stable over time. Despite the variability explained by these traits being low, their effect was consistent across different datasets. Other traits played a significant, albeit weaker, role in species stability, and the inclusion of multi-variate axes or phylogeny did not substantially modify nor improve predictions. These results provide empirical evidence and highlight the relevance of specific ecological trade-offs, i.e. in different resource-use and dispersal strategies, for plant populations stability across multiple biomes. Further research is, however, necessary to integrate and evaluate the role of other specific traits, often not available in databases, and intraspecific trait variability in modulating species stability.

1. Introduction

Identifying the drivers of temporal stability in plant populations and communities has consequences for maintenance of multiple ecosystem functions over time, including carbon sequestration, fodder resources for livestock and nutrient cycling [1–3]. One of the main determinants of community stability is the cumulative temporal variability in the abundances of individual species’ populations [4–6]. Lower temporal variability in individual population abundances at a given site, and particularly for dominant species, generally increases overall community stability [7–10]. Accordingly, assessing the drivers of temporal variability in populations is necessary to understand and forecast the potential consequences of increasingly common environmental perturbations [11,12].

While empirical evidence is still scarce and ambiguous, theoretical predictions suggest that the drivers of temporal variability in single plant populations are related to different ecological characteristics of species (e.g. r/K life-history strategies [13]). These differences can be described through functional traits that determine how plants respond to environmental factors, affect other trophic levels and influence ecosystem properties [14–16]. Specifically, differences in functional traits among species result in varied responses to the environment that might lead to different patterns of demography, adaptation and distribution, thus giving rise to different population fluctuations over time (e.g. [6,17–19]).

Assessing differences in functional traits between species, as well as the relationship of these differences to specific ecological patterns, has been a long-standing focus in plant ecology leading to a search for general trait trade-offs across taxa and ecosystems (e.g. [20]). Trait trade-offs are generally understood as a shift in the balance of resource allocation to maximize fitness within the constraints of finite resources (e.g. Grime's C-S-R strategy scheme [21]). Mostly, such trade-offs have been assessed within the context of community assembly theory and eco-evolutionary models for niche differentiation (e.g. [22–24]). Ultimately, traits linked to specific axes of ecological differentiation are key to understanding major trade-offs in plant strategies, such as the trade-off between leaf maximum photosynthetic rate and leaf longevity, also known as the leaf economic spectrum [25].

At the same time, different specific trade-offs can also underlie differences in temporal variations in species' abundances, both within and between community types. For example, species that are able to respond quickly to environmental variability, i.e. acquisitive resource-use strategy, fast-growing species that invest in organs for rapid resource acquisition and/or high dispersal ability, should sustain higher temporal variation in population size, and will be favoured in sites where disturbance and/or environmental instability determine a fluctuation in resources [13,26,27]. By contrast, species adapted to endure environmental variability (i.e. conservative resource-use strategy, slow-growing and long-lived species that invest in structural tissues and permanence) are thought to persist during unfavourable periods due to resources stored from previous, more favourable years [28], and will exhibit less temporal variability [13,29]. These species are expected to be favoured in more stable and predictable environments [30].

It remains unclear though whether the potential relationship between species' traits and species’ stability would be detected across different biomes and through differences in single traits or combined axes of differentiation that incorporate multiple traits [20,26,31]. Several ecological strategy schemes, such as the classic r/K selection [13] and C-S-R [21] theories, as well as the leaf-height-seed (LHS) scheme [26], can theoretically help predict how functional trade-offs determine species' temporal strategies and their fitness across different types of environments. The LHS scheme for instance is based on three independent plant traits which should provide key proxies for independent trade-offs in plants (stress adaptation, competition and response to disturbance respectively; [26]). Interestingly, only a few empirical studies have linked differences in temporal strategies to functional traits within plant communities [6,17,18,32,33]. For example, Májeková et al. [6] empirically confirmed that herbaceous species with a more conservative resource-use strategy (i.e. those with higher leaf dry matter content—LDMC) have more stable populations over time. A similar relationship was found at the community level, where communities including a greater abundance of species with high LDMC were more stable [34,35]. A recent global meta-analysis of sown grasslands, although based on short-term experiments, suggested that an increase in the abundance of rapidly growing species can destabilize community biomass over time [33]. This is supported by empirical demonstrations that, in natural vegetation, community stability is predicted by the functional traits of the dominant species rather than by species diversity per se [7]. Further, only Májeková et al. [6] tested whether trait-based predictions of population temporal variability were consistent across different management regimes, i.e. fertilization and competitor-removal treatments, generally finding minor differences and consistent predictions for LDMC. Ultimately, global empirical evidence of a general link between quantitative functional traits and the temporal variability of populations, and whether this link is maintained despite differences in community types and environmental conditions, is still missing [27].

Here, using an extended compilation of long-term, recurrently monitored vegetation plots, encompassing different habitat types around the world (https://lotvs.csic.es) [36], we determine which plant traits better predict the temporal stability of plant populations. We expect that populations of species with more acquisitive and higher dispersal-ability traits will tend to be more variable over time, while those of species with more conservative trait values and lower dispersal ability will tend to be more stable over time. We also expect to find empirical evidence of the generality of these relationships.

2. Materials and methods

(a) . Plots and population stability

We used 78 datasets contained in the LOTVS collection of temporal vegetation data. These consist of a total of 7396 permanent plots of natural and semi-natural vegetation that have been consistently sampled for periods of between six and 99 years, depending on the dataset (electronic supplementary material, table S2) [36,37]. These datasets were collected from study sites in different biomes that span the globe, in 18 different countries: Australia, China, Czech Republic, Estonia, France, Germany, Hungary, Kenya, Mongolia, Netherlands, New Zealand, Norway, Russia, South Africa, Spain, Switzerland, United Kingdom and USA. They differ in sampling method (e.g. abundance measured as above-ground biomass, visual species cover estimates and species individual frequencies), plot size and study duration. The studies that generated the datasets sampled different types of vegetation (predominantly grasslands but also shrublands and forests) and covered a wide array of biomes, with mean annual precipitation spanning from 140 mm to 2211 mm, highest temperature of the warmest month spanning from 11.3°C to 35.7°C and lowest temperature of the coldest month spanning from −35.3°C to 7.7°C (electronic supplementary material, table S2).

First, for each plot, we quantified the inter-annual variability in the size of each species's population using the coefficient of variation (CV) of abundance over time, i.e. the standard deviation (s.d.) of species abundance over mean species abundance [6,27]. Since a fundamental differentiation between growing strategies corresponds to whether a species is woody or non-woody [20,28], we focused the main analyses on non-woody species only. This meant we excluded any species belonging to forest overstoreys (i.e. trees and shrubs), woody species’ seedlings and any other species defined as woody when present in the plots. Moreover, based on the collected data available, in many plots, we could not distinguish adult woody individuals from seedlings, with seedlings being the most likely cause of high variability in woody species' CV values (figure 2a). Nevertheless, we tested differences in CV values between woody and non-woody species in our data, and we considered a possible influence of the presence of woody overstorey on the CV values (see data analysis).

Figure 2.

Effect of categorical traits on species variability (CV). Here we show results of the models fitted using single-categorical traits as predictors for the mean species CV at dataset level (i.e. analogous models as the reduced model in the main text): (a) woodiness; (b) lifespan; (c) life form (Ch: chamaephyte, Cr: cryptophyte, H: hemicryptophyte, P: phanerophyte, T: therophyte); (d) growth form. Estimates and respective confidence intervals (95% by the thin line and 68% by the thick line) are shown in red, which correspond to the summary statistics of each category. Intercept was excluded from the model to better understand the differences across trait categories. The subpanels represent, on the left side, the closeup of the estimates, on the right side, the violin plot for the data used in each model.

To avoid using biased CV values for very sporadic species (increased CV), we also excluded those species that occurred in fewer than 30% of the sampling events across the time series for a given plot [6]. Further, to account for variability in CV values between and within the datasets, mostly due to differences in abiotic, biotic and management conditions, we calculated the average CV value for each species in each dataset, standardizing and scaling these averages within each dataset (z-scores). This resulted in a total of 3397 species per dataset CV values. To account for potential effects of temporal directional trends in vegetation affecting CV [38], we also computed a detrended version of CV (CVt3) which gave very similar results to the basic CV calculations (see electronic supplementary material, figure S1).

(b) . Functional traits

For all the species in our dataset, we collected trait information from the TRY global database [39]. We considered different functional traits representing different components of major plants’ growing strategies [26]. Regarding categorical traits, we considered lifespan (annual and non-annual), life form, woodiness (woody and non-woody) and growth form. For continuous traits, we analysed plant height, seed mass, specific stem density, LDMC, specific leaf area (SLA), leaf nitrogen content per unit mass and leaf phosphorus content per unit mass (see [40] for trait name nomenclature and definitions). Besides considering the effects of these traits separately, we also evaluated the effect of both categorical traits and quantitative traits together (see electronic supplementary material, figure S4) and the effect of quantitative traits beyond categorical traits. Furthermore, considering phylogeny as a proxy of conserved functional traits, we considered the effect of potentially unmeasured traits (see electronic supplementary material, figure S5b).

For each species, we averaged trait values across all standard measurements obtained from TRY, excluding those performed under explicit treatments or on juveniles, and outliers. The traits that were log-transformed (using natural logarithm) to achieve a normal distribution. For details on the traits used, their summary statistics, their correlations and their coverage in each dataset, see electronic supplementary material, table S3. To take into account multi-variate trade-offs between species, we also considered axes of functional variation derived from multi-variate analyses (principal coordinates analysis, PCoA). The traits considered were weakly inter-correlated, with the two major axes of trait differentiation from PCoA, linked mainly to LDMC and seed mass (see electronic supplementary material, table S1 for details). The taxonomic names follow the nomenclature of ‘The Plant List’ (www.theplantlist.org). Nomenclature was standardized using the R package ‘Taxonstand’ [41].

(c) . Data analyses

To quantify how the considered traits were linked to species CV, we used linear mixed effect models (‘lmer’ function in R package ‘lme4’, [42]). As a response variable, we used the mean CV for each species in each dataset, standardized as mentioned above. To analyse the effect of the continuous traits, we fitted a single model. As predictors, we included all the continuous traits listed above, scaled and centred. To account for the taxonomic and spatial structure of the data, we included both species identity and dataset identifier as random intercept factors in all of the models. We visually checked the compliance of all of the models residuals with normality and homoscedasticity. To assess the goodness-of-fit of the full model, fixed (i.e. marginal) and total (i.e. conditional) R² were calculated [43,44]. To define which among the continuous traits were more relevant for species stability, we compared the fixed R² of different models, each differing in the subset of predictors that were included. These different models were fitted to different datasets because of the presence of missing values in the trait data. We used R² as a unifying measure of goodness of fit (i.e. as a measure of how well the different models explain the variability in the different datasets). Using this approach, we selected the model that had the highest fixed R². In the present work, we focused on significant terms in the reduced model. For completeness, we also compared AIC of full and reduced models by fitting them to the same subset of the data (i.e. we fitted the reduced model to the dataset of the full model). We found that the AIC was indeed lower when using a subset of the trait variables (AIC of the full model was 1939.2, AIC of the reduced model using the same data frame was 1934.6). Separate models were fitted to clarify the influence of categorical traits on the stability of species, each using either woodiness, lifespan, life form or growth form as predictors. In these models, we excluded the intercept, to better see the differences between the trait categories. In addition, analogous models were run also on the two components determining species' CV separately, i.e. mean abundance and s.d. of abundance in time, also standardizing these variables within each dataset (electronic supplementary material, figure S1).

A series of analogous models were fitted using a different set of predictors, all shown in the electronic supplementary material. To examine the influence of differentiation axes based on multiple traits, instead of using single separate traits, models were run using two multi-variate PCoA axes that resulted from the combination of traits. We also fitted separate models using each single trait of those emerging as significant in the reduced multi-variate model (See electronic supplementary material, table S1). To explore the consistency of the stability-trait relationships across datasets, we also fitted models using each single trait and adding a random slope effect for the datasets (electronic supplementary material, figure S3). We also tested the interaction between the most influential categorical trait, namely lifespan, and the other continuous traits (see electronic supplementary material, figure S4). Finally, a set of models was fitted to assess the possible effect of phylogenetic relatedness on the results found. Specifically, we tested to what extent considering phylogeny modified the effect of the considered traits and whether phylogeny, considered as a proxy of unmeasured traits, improved the main models emerging from the analyses of quantitative traits (see electronic supplementary material, figure S5 for all the details regarding these models).

3. Results

By focusing initially on continuous traits, we were able to detect two sets of key functional traits playing a consistent role in species’ population temporal stability: one linked exclusively to seed mass, and the other linked to the leaf economic spectrum, i.e. LDMC, SLA and Leaf N content. Based on the reduced linear mixed effect model, these two sets of traits had the most influence on species CV among the continuous traits considered (table 1; figure 1).

Table 1.

Effects of continuous traits on species variability (CV), models comparison. Model's summary for both the full model and the reduced model, which test the influence of continuous traits on the species variability (coefficient of variance in time, CV). The full model contains all the predictors while the reduced model contains only a subset of the initial predictors. Estimates and relative s.e. (in brackets) are shown. R² (fixed): variation explained by fixed factors; R² (total): variation explained by both fixed and random factors. p-values calculated using Satterthwaite approximation for degrees of freedom. ***p ≤ 0.001; **p ≤ 0.01; *p ≤ 0.05.

	full model	reduced model
(intercept)	−0.10 (0.06)	−0.03 (0.04)
plant height	−0.01 (0.09)
leaf N content	0.03 (0.08)	0.06 (0.04)
leaf P content	0.04 (0.07)
seed mass	−0.12 (0.08)	−0.08 * (0.04)
SLA	0.02 (0.09)	0.09 * (0.04)
LDMC	−0.23 ** (0.07)	−0.21 *** (0.04)
SSD	0.06 (0.06)
N	676	1630
species	93	395
datasets	67	77
R2 (fixed)	0.05	0.07
R2 (total)	0.13	0.18

Figure 1.

Effects of continuous traits on species variability (CV). Regression plots of the reduced model showing the effects of LDMC (a), SLA (b), seed mass (c) and leaf N (d) content on the CV of species.

We found significant negative coefficients with species CV for LDMC and for seed mass (table 1; figure 1). These coefficients indicate that species with greater LDMC and greater seed mass were more stable (i.e. lower CV values; figure 1a). By contrast, we found positive coefficients for SLA and Leaf N content, although the effect was statistically significant only for SLA. For these traits, the larger the trait value, the higher the species CV and therefore the less stable the species populations (figure 1b,d). The effect of these traits was consistent across datasets (low deviation of the datasets' random slope effect compared to the main effect slope for both the models using LDMC and seed mass; electronic supplementary material, figure S3). Since the variability explained by individual traits was low (R² = 0.07 for fixed effects in the reduced model using the quantitative traits, table 1), we assessed the role of combining quantitative traits into multi-variate axes, categorical traits, or by considering phylogeny.

Similar results to individual traits were found using either of the two first PCoA axes based on multiple traits (electronic supplementary material, table S1), although with a lower predictive power (R² fixed was 0.05 compared to 0.07 in the reduced model that used individual traits). We also fitted models using the single PCoA axis and the single traits. In this case, single-trait models again explained more variability compared to the models with the single PCoA axis (PCoA Axis 1 model's R² fixed was 0.040 versus 0.050 when using LDMC; PCoA Axis 2 model's R² fixed was 0.003 versus 0.005 when using seed mass; electronic supplementary material, table S1). Although we realize that these models are fitted to subsets of the database having different species numbers and datasets, R², as a generic measure of goodness of fit, gives us an indication that the models using functional traits perform better than the ones using aggregated axes of functional differentiation. Moreover, using R² to compare models with PCoA axes and the single traits is not problematic because the models have the same number of degrees of freedom. Finally, when the two components determining species' CV were analysed separately, i.e. species’ mean abundance and s.d. of abundance over time, the model predicting mean abundance was stronger than the model using s.d. of abundance over time (with significant results and a higher R² fixed; see electronic supplementary material, figure S2) although LDMC predicted significantly both mean abundance and its s.d.

Categorical traits provided some improved predictions compared of using continuous traits, both influencing CV alone (table 2) and in combination with quantitative traits (electronic supplementary material, figure S4). Herbaceous species with longer lifespan (i.e. perennial and biennial) tended to have a lower CV (fixed R² = 0.04; table 2). Adding lifespan to the models with quantitative traits, however, did improve predictions only to 0.10 (fixed R²). Most importantly, the interaction between lifespan and the quantitative traits considered was not significant, indicating that, for example, LDMC was a good predictor of stability for both non-annual and annual species. Woody species, trees and shrubs also had low CV scores (although with very low fixed R²= 7.04 × 10⁻⁰⁷). Finally, after accounting for phylogeny (i.e. adding phylogenetic eigenvectors to ‘correct’ CV values), there was no evidence for an overall improvement in model explanatory power (fixed R² was 0.01) nor did this substantially modify the results (see electronic supplementary material, figure S5). At the same time, the phylogenetic signal not accounted for by the considered traits (decoupled phylogenetic information; [45]; electronic supplementary material, figure S5), used here as a proxy of unmeasured traits, did not change the original explained variability (fixed R² stayed at 0.07).

Table 2.

Effects of categorical traits on species variability (CV): comparison of the models testing the influence of categorical traits on the species variability (coefficient of variance in time, CV). Estimates and relative s.e. (in brackets) are shown. R² (fixed): variation explained by fixed factors; R² (total): variation explained by both fixed and random factors. p-values calculated using Satterthwaite approximation for degrees of freedom. ***p ≤ 0.001; **p ≤ 0.01; *p ≤ 0.05. Ch: chamaephyte, Cr: cryptophyte, H: hemicryptophyte, P: phanerophyte, T: therophyte.

	woodyness	lifespan	life form	growth form
non-woody	0.03
	(0.02)
woody	0.03
	(0.05)
annual		0.49 ***
		(0.05)
not-annual		−0.06 *
		(0.02)
Ch			−0.03
			(0.08)
Cr			−0.09
			(0.09)
H			−0.06
			(0.04)
P			0.18
			(0.10)
T			0.55 ***
			(0.05)
fern				−0.27
				(0.16)
graminoid				−0.13 ***
				(0.04)
herb				0.12 ***
				(0.03)
herb/shrub				−0.21
				(0.11)
shrub				−0.01
				(0.06)
shrub/tree				−0.03
				(0.13)
tree				0.30 *
				(0.13)
N	3869	3869	2492	3849
species	1794	1794	990	1779
datasets	78	78	73	78
R2 (fixed)	7.04 × 10−07	0.04	0.06	0.02
R2 (total)	0.23	0.23	0.14	0.22

4. Discussion

By analysing a large worldwide compilation of permanent vegetation plot records, we confirmed the generality and consistency of theoretical predictions relating key functional traits to plant population stability over time. We specifically found that the species with greater LDMC and a larger seed mass were the most stable over time. Ultimately, these results suggest that common functional trade-offs related to resource use and dispersal consistently influence herbaceous plant population stability across different biomes worldwide. While the results demonstrate that simple plant traits can help, consistently, in predicting the stability of individual species, and ultimately of plant communities, the variability explained by these traits was low, despite accounting for other key traits like lifespan or using phylogeny as a proxy of unmeasured traits. Further research is therefore necessary to integrate and evaluate the role of intraspecific trait variability and other potentially relevant traits, generally not available in trait databases, in modulating species stability.

We identified two functional trade-offs that influence species stability. Specifically, differences associated with the leaf economic spectrum (in our case linked to LDMC, SLA and N content values) define trade-offs in terms of slow-fast resource acquisition [20,25]. Differences in seed mass values represent the competition-colonization (seedling establishment) trade-off [46] related to the species' dispersal and establishment strategy. Moreover, when analysing multi-variate functional differentiation in herbaceous species, these sets of traits were the ones most strongly associated with the two first principal axes (electronic supplementary material, table S1), further confirming the importance of these two functional differentiation axes. These findings are broadly consistent with Díaz et al. [20], who found that the main differentiation between species was related to size-related (whole plant and seed) and leaf traits.

Ultimately, the individual functional traits related to the populations’ temporal patterns are intrinsically linked to how the species adapt to patterns of resource availability and disturbance, both if we analyse the effect of single traits or multi-trait effects (PCoA axes). At the same time, it is interesting to notice that, in our case, combined trait information in the form of plant spectra (i.e. via the PCoA axes) lost some ecological explanatory power compared to specific trait effects. If, on one hand, such multi-trait trade-offs are essential to distinguish the major axis of differentiation among organisms [20], on the other, the independent effect of individual traits might be even more relevant ecologically. This suggests that, for predicting species stability, using single functional traits could be more effective than using axes of functional variation based on multiple traits. By using axes of functional variation, the traits' individual effects could be blurred or could be missed because both additive and non-additive effects of single traits [47] are ecologically more relevant than combined multi-trait effects.

Leaf traits relate to species adaptations to resource availability. Higher LDMC values, as well as smaller SLA and N content values, correspond to a slow return on investments in nutrients, lower potential relative growth rate and longer leaf and whole-plant lifespan [16,25]. This implies higher potential of buffered population growth. In fact, slow-growing and long-lived species, for example with higher values of LDMC, could have an advantage in unfavourable years due to resources stored from previous, more favourable years, thus maintaining buffered population growth and consequently more stable populations [6,28]. Different leaf traits, although broadly linked, capture different aspects of leaf function [16]. It follows that they would be differently linked to species growth strategies and their temporal dynamics. Our results show that, although SLA and Leaf N do have an influence, it seems to be secondary (i.e. they have a weaker effect; table 1; figure 1) when compared to LDMC, which is consistently and strongly related to species temporal variability. One explanation is that LDMC is better related to growth rate, compared to the other leaf traits (e.g. [48]). Another explanation could be that LDMC is probably a trait whose measurement is less likely to be influenced by measurement precision/protocols and therefore it might show less intraspecific variability due to data measurements. At the same time, LDMC was also the trait selected, over SLA and Leaf N, in Májeková et al. [6], where leaf trait measurements from a single location and single working group were more comparable. Possibly LDMC reflects, to a greater extent, a stronger trade-off in growth and defence, and ultimately plant productivity (which is likely linked to the denominator of CV), while SLA and Leaf N are possibly linked to trade-offs more tightly linked to photosynthetic strategies [49]. Alternatively, LDMC can be also interpreted as a better indicator of response to water stress, which might be an underlying cause of inter-annual variability (see [6]). More locally based research is certainly required to define the relative effects of different traits associated with the leaf economic spectrum on population temporal dynamics.

Similarly, seed mass consistently appears to have an influence on species temporal variability (table 1; figure 1). This trait relates to the species’ adaptations to disturbance patterns and colonization. Larger seed mass means greater resources stored to help the young seedling establish and survive in the face of stress with the cost of short-distance dispersal, while smaller seeds (also in combination with seed shape) are typically related to greater longevity in seed banks and dispersal over longer distances [46,50,51]. Therefore, species germinating from seeds with a larger mass are more likely to survive during adverse years and so their populations are more stable in a given site compared to species with smaller seeds, which will tend to maintain their populations through permanence in seed banks, which enables proper germination timing [18,52]. In addition, species with greater seed mass might be favoured in communities where gaps are scarce, which are usually dominated by perennial species (with higher LDMC values) and are more stable. Large seeds will tend to remain closer to the mother plant than small seeds, thus increasing the stabilizing effects on populations. Small-seeded species still maintain buffered population growth [53], yet their above-ground abundance will be more variable over time, because they usually germinate only in favourable years. This explanation is particularly supported, for example, for short-lived plants (annuals and biennial species together, electronic supplementary material, table S3), which tend to be less stable over time (figure 2b) and are generally associated with the small-seed strategy at a global scale [26].

It is important to consider that the same traits that predicted species variability, using CV, also predicted the components of CV, i.e. species means and s.d. Clearly the s.d. in species fluctuation is inherently increasing with species means, following the so-called Taylor's power law [54]. This leads to the use of CV in the study of stability, as a more ‘scaled’ measure of species variability. At the same time, when the CV is negatively correlated to species mean abundance, as in our case (r = −0.46, which corresponds to the case of a slope in the Taylor's power law being lower than 2), it implies that more dominant species tend to fluctuate comparatively less than subordinate species. This is an important observation because this scenario implies that the same type of species that are dominant and likely with greater abundance, e.g. with high LDMC [49], is also the more stable ones. Since dominant species were key drivers of the stability of the communities considered in our study [37], the results of the present study indicate that the same traits that determine species dominance also determine species stability, which is a key message for any attempt to predict both community structure and its potential to buffer environmental fluctuations [27].

Despite low R² values, our models found consistent evidence of the relationship between continuous traits related to leaf and seed economics and species temporal stability across different biomes (electronic supplementary material, figure S4). While we did consider other traits that affected the stability of species, these did not substantially improve the predictive power of models. In particular, adding lifespan in interaction with the continuous traits analysed in our models improves their performance only to a small extent (see electronic supplementary material, figure S4). Beyond the obvious effect of lifespan on species temporal stability, the results in electronic supplementary material, figure S4 indicate that although the seed mass effect seems to be obscured or encompassed by the lifespan trait, our original results linked to traits on the leaf economic spectrum were still relevant for species stability. Further, adding ‘unmeasured’ traits (using phylogeny as a proxy of unmeasured conserved traits, see electronic supplementary material, figure S5b) did not substantially change the original explained variability. Results showed that some effect from additional traits could be detected, supporting the need for research to identify other important traits that could be related to species stability, for example, those linked to vegetative propagation and reproduction, like those specifically related seed dispersal and seed dormancy traits. Importantly, the results where phylogeny was considered were otherwise completely consistent with the original results. This is a first indication that additional (not considered here) quantitative traits might not increase the explanatory power of the models in a qualitatively important way. As such, further tests using other potentially relevant trait, or traits measured directly in the biomes and locations under study, are surely needed to expand the findings of the present study. Very often traits available in database represent only a small portion of traits actually determining species fitness and the values obtained for those available (generally an average value) might not represent the phenotypic expression in the specific study site under observation. Indeed, one missing factor that could explain the observed variability in species CV could be intraspecific variability in both trait values and species CV, as indicated also by the higher R² values when considering the random effects species and dataset. Because of these effects, the present study was not necessarily focused on maximizing the explained variability but in detecting the most consistent patterns across different biomes, which were detected in the effects of LDMC and seed mass, and opening a new field of research focused on the search of the best traits, and their combinations, in predicting species stability.

An important point to acknowledge is that the compilation of datasets used here is biased towards more temperate biomes, with a predominance of grasslands and open shrublands. This is an artefact of historic sampling bias and dictates available ecological datasets to study inter-annual ecological stability. Such sampling bias is typically a widespread problem for analyses integrating diverse datasets, where available information can be affected by regional research preferences and funding opportunities for research. These issues are particularly pronounced in long-term experiments, where the presence of vegetation with woody species, and particularly tree species, can cause confounding factors in the analyses of temporal dynamics. We dealt with this by using datasets as a random factor and focusing on herbaceous vegetation only, which resulted in patterns apparently consistent across different vegetation types, i.e. also in vegetation with woody species (electronic supplementary material, figure S4).

Finally, our results show worldwide evidence that species with more conservative leaf economics and greater seed mass are generally more stable, i.e. less variable over time, and therefore confirm theoretical assumptions and are consistent with previous localized empirical evidence on the interdependence between these traits, their relative trade-offs and population temporal stability (e.g. [6,13]). In addition, our results show the global validity of these trade-offs, found across a variety of abiotic and biotic conditions. Overall, our findings contribute to a better understanding of the drivers of plant population temporal stability, which has important implications for the conservation of ecosystem functions over time across the world.

Data accessibility

The metrics used in the analyses are available at https://doi.org/10.5281/zenodo.7978254 [55] under CCBY licence. For access to the LOTVS datasets in full please refer to https://lotvs.csic.es/

The data are provided in the electronic supplementary material [56].

Authors' contributions

L.C.: conceptualization, data curation, formal analysis, investigation, methodology, validation, visualization, writing—original draft and writing—review and editing; E.V.: conceptualization, data curation, methodology and writing—review and editing; T.G.: conceptualization, data curation, methodology and writing—review and editing; L.G.: conceptualization, data curation, formal analysis, funding acquisition, methodology and writing—review and editing; J.L.: conceptualization, funding acquisition, methodology and writing—review and editing; A.E.: data curation, methodology and writing—review and editing; C.P.C.: data curation, methodology and writing—review and editing; M.M.: methodology and writing—review and editing; J.Da.: data curation and writing—review and editing; J.De.: data curation and writing—review and editing; D.J.E.: data curation and writing—review and editing; M.E.: data curation and writing—review and editing; R.G.-G.: data curation and writing—review and editing; E.G.: data curation and writing—review and editing; D.G.: data curation and writing—review and editing; V.H.: data curation and writing—review and editing; S.P.H.: data curation and writing—review and editing; T.H.: data curation and writing—review and editing; R.I.: data curation and writing—review and editing; A.J.: data curation and writing—review and editing; N.J.: data curation and writing—review and editing; M.K.: data curation and writing—review and editing; K.K.: data curation and writing—review and editing; F.K.: data curation and writing—review and editing; F.L.: data curation and writing—review and editing; R.H.M.: data curation and writing—review and editing; G.Ó.: data curation and writing—review and editing; R.J.P.: data curation and writing—review and editing; M.P.: data curation and writing—review and editing; B.P.: data curation and writing—review and editing; J.P.: data curation and writing—review and editing; M.R.: data curation and writing—review and editing; W.S.: data curation and writing—review and editing; U.S.: data curation and writing—review and editing; M.Sc.: data curation and writing—review and editing; H.S.: data curation and writing—review and editing; P.Š.: data curation and writing—review and editing; M.Š.: data curation and writing—review and editing; C.S.: data curation and writing—review and editing; M.So.: data curation and writing—review and editing; M.St.: data curation and writing—review and editing; J.V.: data curation and writing—review and editing; V.V.: data curation and writing—review and editing; D.W.: data curation and writing—review and editing; K.W.: data curation and writing—review and editing; S.K.W.: data curation and writing—review and editing; B.A.W.: data curation and writing—review and editing; T.P.Y.: data curation and writing—review and editing; F.-H.Y.: data curation and writing—review and editing; M.Z.: data curation and writing—original draft; F.d.B.: conceptualization, data curation, formal analysis, funding acquisition, methodology, project administration, resources, supervision, writing—original draft and writing—review and editing.

All authors gave final approval for publication and agreed to be held accountable for the work performed therein.

Conflict of interest declaration

We declare we have no competing interests.

Funding

This research was funded by Czech Science Foundation Grant GACR16-15012S and Czech Academy of Sciences Grant RVO 67985939 and by the Spanish Plan Nacional de I + D + i (project PGC2018-099027-B-I00). R.J.P. was supported by the Scottish Government's Rural and Environmental Sciences and Analytical Services division. M.P. and C.P.C. were supported by the Estonian Research Council grant (grant nos. PRG609 and PSG293). M.P. and M.Z. were supported by the European Regional Development Fund (Centre of Excellence EcolChange). S.K.W. was supported by the Strategic Science Investment Fund of the New Zealand Ministry of Business, Innovation and Employment. E.V. was funded by the 2017 program for attracting and retaining talent of Comunidad de Madrid (grant no. 2017-T2/AMB-5406). B.A.W. is funded by NERC under AgZero + NE/W005050/1 and RestREco NE/V006444/1. R.M. was supported by Defra and the Leverhulme Trust. T.P.Y. was supported by the U.S. National Science Foundation (19-31224). J.P. was supported by the Fundación Ramón Areces grant CIVP20A6621.