Which plant traits determine abundance under long-term shifts in soil resource availability and grazing intensity?


Correspondence author. E-mail: etienne.laliberte@uwa.edu.au


1. Soil resource availability and disturbance are widely recognized as key drivers of plant community structure. However, the relative importance of different plant traits in determining species abundance following shifts in soil resource availability and disturbance remains little studied, particularly in long-term experiments.

2. We studied trait-based plant community assembly in a 27-year grassland experiment where 25 plant species were sown into resident vegetation, after which annual manipulations of soil resource availability (five levels of superphosphate fertilizer; the highest level was also irrigated) and disturbance (three ‘mob-grazed’ sheep grazing intensity levels: lax, moderate, hard) were applied. We used community assembly through trait selection (CATS) models based on entropy maximization to predict species relative abundances and to quantify the relative importance of each trait in determining abundance.

3. Plant species were primarily differentiated along a trade-off axis corresponding to traits promoting rapid growth (e.g. high leaf [N] and specific leaf area [SLA]) vs. those promoting long leaf life span. Using 12 traits, the CATS model predicted >80% of the variation in the relative abundances of 51 species, suggesting that trait-based filtering was important.

4. Species with leaf attributes that reduce nutrient losses held a long-term advantage under the lowest soil resource availability, whereas those associated with a rapid growth rate became dominant under soil resource addition. Species with thinner leaves were also favoured under greater soil resource availability, which may reflect a strategy to maximize SLA without sacrificing leaf density and thus maintain leaf structural defences under grazing disturbance. Greater leaf [S] and the ability to symbiotically fix atmospheric N were favoured under greater soil resource availability. Greater plant height, thinner leaves and higher leaf [N] were favoured under lower grazing intensity.

5.Synthesis. Our results highlight the importance of species functional differences to understand how plant communities react to increases in soil resource availability and disturbance, two important and inseparable components of land-use change in grasslands world-wide.


Given a common regional species pool, what determines the composition of local communities? Answering this fundamental question has been a major goal of community ecology since its beginnings (Clements 1916; Gleason 1926). One approach is to focus on species functional traits, which can be defined as morphological or physiological characters that influence species responses to different environmental conditions (Violle et al. 2007; Shipley 2010a). It has been argued that trait-based approaches are essential to make community ecology a more general, predictive science (Keddy 1990; McGill et al. 2006; Webb et al. 2010).

Trait-based community assembly was initially viewed as a series of successive abiotic or biotic filters that exclude unsuitable strategies from those found across the entire available species pool (Keddy 1992; Díaz, Cabido & Casanoves 1998). The successive set of filters thus represents the trait-based rules that govern community assembly at a local site (Keddy 1992). Despite its conceptual appeal, the strict notion of filters invokes a series of binary decisions that dictate whether a species will be present or not at a site, yet says nothing about its expected abundance (Cornwell & Ackerly 2010; Shipley 2010a). For instance, both soil resource availability and disturbance intensity strongly influence the structure of plant communities (Grime 2002), yet these continuous factors do not simply allow or prevent particular ecological strategies from occurring.

In an effort to modify Keddy’s (1992) original concept into a more probabilistic view of trait-based community assembly, Shipley (2010a) used the analogy of a stochastic filter that specifies the expected relative abundance of a species rather than its presence or absence. Under this new analogy, species relative abundances are constrained by their functional attributes, such that species possessing favourable attributes become more abundant on average, but with a strong stochastic component that recognizes the importance of chance events affecting community composition. This trait-based community assembly conceptual framework has recently been translated into a quantitative framework to predict species relative abundances at local sites (Shipley, Vile & Garnier 2006; Shipley 2010a). We refer to this model as the ‘community assembly through trait selection’ (CATS) model.

The general concept of the CATS model is the following: if species sorting processes constrain the composition of a local community in such a way that the individuals of species possessing favourable functional attributes in a particular environment become more abundant (Webb et al. 2010), then predictable community-level properties should emerge (Shipley 2010b). Shipley, Vile & Garnier (2006) defined these community-level properties, or model constraints, as community-weighted traits (Garnier et al. 2004; Díaz et al. 2007a). Community-weighted traits are best understood as the trait values of an average individual or biomass unit chosen at random from a community (Shipley 2010a). However, given a known species pool, a very large number of different community compositions (i.e. combinations of varying species relative abundances) can adequately meet the criteria of any set of model constraints expressed as community-weighted traits (Shipley 2009). One solution to this problem provided by Information Theory and Bayesian statistics is to find, among the very large set of possible community states, the one that maximizes Shannon’s index of information entropy or relative entropy. This decision is justified by the fact that this particular community composition is the only one that is both consistent with all stated constraints and that does not imply any additional constraints (Jaynes 2003; Shipley, Vile & Garnier 2007).

To identify trait-based assembly rules, the relative importance of different traits in determining relative abundance needs to be measured, before testing their significance. Provided that good predictions of the observed relative abundances can be derived from CATS model, then particular parameters (i.e. the λ-values, or weights on the traits) can be used to assess the relative importance of different traits in determining species relative abundances and to enable us to find a parsimonious set of the most important traits (Shipley 2010a). Importantly, in doing so, the relative importance of different traits in determining abundance can be compared across different environments.

Grime (2002) argued that trait-based community assembly would be best studied by constructing synthetic communities from a common species pool, experimentally altering environmental conditions and/or disturbance regimes and documenting changes in composition as communities adjusted to the altered conditions. Ideally, such experiments should be of a sufficiently long duration to allow time for communities to adjust to their new, experimentally altered conditions. Here, we study trait-based plant community assembly in a 27-year experiment conducted in New Zealand grasslands, where a common species pool of 25 plant species was sown into the resident vegetation within a 3-ha area in 1982, after which soil resource availability (five levels of superphosphate fertilizer; the highest level was also irrigated) and grazing intensity (three ‘mob-grazed’ sheep grazing intensity levels: lax, moderate, hard) were manipulated (Scott 1999). This experiment allowed us to explore in a controlled setting how these two factors together drive long-term plant community assembly.

Specifically, we asked the following questions:

  • 1 How are trait distributions altered following long-term shifts in soil resource availability and grazing intensity?
  • 2 What are the most important plant functional traits in determining abundance in these grasslands, and how does the importance of different traits vary across the experimental treatments?

Materials and methods

Study Area and Site

The study was conducted on the AgResearch trial site at Mount John, west of Lake Tekapo in the Mackenzie Basin of New Zealand’s South Island (43°59′ S, 170°27′ E, 820 m a.s.l.). A detailed description of the study area and study site is given elsewhere (Scott 1999). The dominant vegetation type prior to human settlement in the area is likely to have been short-tussock grassland with a variable woody component, probably near the tall tussock (Chionochloa sp.)/short tussock (Festuca novae-zelandiae) transition (McGlone 2001). Extensive grazing by sheep began in the area in the 1850–60s and remains the most important land use today.

Experimental Design

The experiment is described in detail by Scott (1999) and is only summarized here. In 1982, a mixture of 25 agricultural grass and legume pasture species (Scott 1999) was sown using a rotary hoe drill into a 3-ha area of depleted fescue tussock (F. novae-zelandiae) grassland dominated by the exotic mouse-ear hawkweed (Hieracium pilosella) and an estimated 30–40 other vascular plant species. This vegetation type is representative of large areas of New Zealand rangelands (Wardle 1991). The trial followed a split-plot design consisting of two spatial replications (blocks), each split into five whole plots receiving one of the following five nominal fertilizer treatments: 0, 50, 100, 250 and 500 kg ha−1 year−1 of sulfur (S)-fortified superphosphate (i.e. a P/S fertilizer), which is typical in Australian and New Zealand legume-based pasture systems. The whole plots receiving 500 kg ha−1 year−1 were also irrigated fortnightly from November to May of each year. Fertilizer was applied each year for the first 19 years of the experiment only. The fertilizer was applied annually to plots in early spring. Actual mean P and S rates applied over the first 10 years of the experiment are given by Scott (1999).

Each whole plot was further split into six 8 × 50 m subplots (thereafter simply referred to as ‘plots’) corresponding to a two-way factorial design involving sheep grazing intensity (lax, moderate and hard) and stocking method (mob vs. sustained). In mob-grazing plots, a larger number of sheep (with actual numbers depending on available feed-on-offer of the moderate plot) were introduced to plots for 3–4 days, while sustained grazing plots received a lower number of sheep for a longer period (i.e. several weeks). However, only plots corresponding to the mob-stocking type were considered in this study. Lax, moderate and hard grazing intensity levels corresponded to a ratio of 1 : 2 : 4 sheep grazing days, respectively, in the years 2–4 of the experiment and a ratio of 2 : 3 : 4 in subsequent years.

Plots were grazed in groups of three for each fertilizer level, corresponding to the three mob-stocked grazing intensity levels per whole plot. For each grazing event (i.e. when vegetation had reached c. 30 cm in height), sheep numbers were adjusted based on available feed-on-offer of the moderate grazing treatment. The duration of grazing was the same for all three plots but was based on residual bulk of the moderate grazing treatment (i.e. height of 1–2 cm). The grazing intensity treatment is relative (i.e. within each whole plot), not absolute, as the annual sheep grazing days achieved depended on the forage growth of the different fertilizer treatments. This is a key component of the experimental design that increases realism and relevance, since grazer density and production depend on primary production in grasslands (McNaughton et al. 1989). Grazing occurs in the period November–May.

Two 8 × 50 m plots were added in an area directly adjacent to the experiment. These two plots had not been sown, fertilized, irrigated and had not been grazed since at least 1981 and only lightly grazed before then, at the same intensity as the rest of the experimental site prior to this experiment being initiated.

Vegetation Sampling

Visual ranking

Every year since 1983, 10 most abundant plant species in each plot were visually ranked according to standing biomass (Scott 1989). The abundance ratio between the fifth- and first-ranked species (R5,1) was also estimated (Scott 1989). To derive species relative abundances pi from ranks ri, we followed Scott’s (1989) suggestion to use the geometric series description of the abundance–rank relationship:


where inline image. Plots were sampled at least three times on different days, and abundances from the three rounds were averaged (Lalibertéet al. 2010). These data were used to explore temporal patterns in community-weighted traits but were not used with CATS models.

Percentage cover

Sampling of all vascular plant species present within each plot (see Table S1 in Supporting Information) was also undertaken in November 2007. Twenty 1 × 1 m quadrats were randomly positioned along two longitudinal transects (10 quadrats per transect) in each plot, and percentage cover of all vascular plant species present in each 1 × 1 m quadrat was recorded (see Laliberté & Tylianakis 2011). While the results obtained from the visual ranking and percentage cover sampling methods are only moderately correlated, both methods detect consistent differences in plant community structure among fertilizer and grazing treatments (Lalibertéet al. 2010). These data were used to test for differences in community-weighted traits and for CATS models.

Plant Functional Traits

Morphological and chemical traits

We selected a set of 14 functional traits (Table 1) from the standard lists of Cornelissen et al. (2003) and Garnier et al. (2007) that have been identified as traits that determine plant responses to environmental change, while still being easily measurable across a wide range of species (Weiher et al. 1999; Cornelissen et al. 2003; Garnier et al. 2007). Specific leaf area (SLA), leaf area (LA), leaf dry matter content (LDMC) and leaf nutrient concentrations [leaf nitrogen concentration (LNC), leaf carbon concentration (LCC), leaf phosphorus concentration (LPC) and leaf sulfur concentration (LSC); Table 1] were measured as described in Laliberté & Tylianakis (2011). Because leaf nutrient concentration can vary along soil resource availability gradients, we used individual samples from each fertilizer level for the six species that together accounted for more than 80% of total cover among all plots (Table S1), following Garnier et al. (2007). Field-measured traits were measured on all vascular plant species in November 2007, just prior to the start of the grazing treatments for that year.

Table 1.   List of the functional traits measured on the species
CodeTraitTypeUnits/categoriesRangeNumber of species
  1. *Height of highest photosynthetic organ at reproductive stage.

  2. †LT was not measured but was instead estimated as (SLA × LDMC)−1.

ASAcceptability to sheepOrdinal(1) low; (2) medium; (3) high1–354
H*Plant height*Continuousmm4–56454
LALeaf areaContinuousmm21–838252
LCCLeaf carbon concentrationContinuous% (dry weight)40.2–47.451
LDMCLeaf dry matter contentContinuousmg g−1126–49852
LNCLeaf nitrogen concentrationContinuous% (dry weight)1.1–5.251
LPCLeaf phosphorus concentrationContinuous% (dry weight)0.09–0.6751
LSPlant life spanOrdinal(1) annual; (2) biennial; (3) perennial1–354
LSCLeaf sulfur concentrationContinuous% (dry weight)0.06–0.5851
LT†Leaf thickness (estimate)Continuousmm0.14–1.2052
NFAbility to fix nitrogenBinary(0) no; (1) yes0–154
OFOnset of floweringContinuousDay of year (from July 1)124–21049
SLASpecific leaf areaContinuousm2 kg−11.8–40.952
SMSeed massContinuousmg0.03–21.242

Height of mature plants (H) was measured from ground level to the tip of the highest photosynthetic organ on at least 10 individuals from each species among the different fertilizer treatments (Garnier et al. 2007), prior to the first grazing in November. Seed mass (SM) and onset of flowering (OF) were measured following Cornelissen et al. (2003) and Garnier et al. (2007), but could not be measured on all species (Table 1). Life span (LS) and ability to fix atmospheric nitrogen (NF) were assessed using floras (Allan 1982; Webb, Sykes & Garnock-Jones 1988; Edgar & Connor 2000) and electronic data bases (Landcare Research 2008; Peat, Fitter & Ford 2008).

Finally, we estimated leaf thickness (LT) as (SLA × LDMC)−1 (Vile et al. 2005). This assumes an average leaf density (fresh mass) ρF ≈ 1 kg m−3, which has been shown to be reasonably accurate (Sims, Seemann & Luo 1998; Garnier et al. 1999). This estimate is likely to be biased for species with non-laminar leaves (e.g. the tightly inrolled leaves of F. novae-zelandiae). Therefore, we only used our LT estimate when exploring temporal patterns in community-weighted traits and in the CATS models, because LT and LDMC can vary independently along environmental gradients (Witkowski & Lamont 1991).

Acceptability to sheep

Since all plots were grazed by sheep (except for the two additional plots that had not been grazed since 1982), the relative acceptability of plant species as forage for sheep (AS; Table 1) was considered as a potential determinant of plant community assembly. We equate acceptability with preference (Johnson 1980), such that the acceptability of a plant species is assessed as its grazing removal against its abundance or availability. For instance, a species would rate as highly acceptable if it was grazed disproportionally to its availability and vice versa. We used a restricted ordinal scale: 1 (low acceptability), 2 (medium acceptability) and 3 (high acceptability), based on existing data from the literature (Cockayne 1920; Hugues 1975; Covacevich 1991) and from previous experience over the trial.

Statistical Analyses

Relationships among species traits

We used principal component analysis (PCA) to visualize the interrelationships among all continuous plant functional traits (Table 1) other than OF, SM (because of missing values) and LT (because of reasons mentioned in the Materials and methods section). Correlation biplots (type-II scaling) were used such that the angles between vectors (traits) and PCA axes reflected their correlations, with a small angle difference indicating a high correlation (Legendre & Legendre 1998). Prior to analyses, some traits (H, LCC, LNC, LPC, LSC and LA) were first log-transformed to make their distributions more symmetric and thus reduce the influence of species with extreme trait values. Pearson correlation coefficients r were computed between traits and tested for statistical significance. These analyses, and all others in sections below, were conducted in the R environment (R Development Core Team 2011).

Shifts in community-weighted traits

We used community-weighted traits as a measure of functional composition (Garnier et al. 2004). Community-weighted traits were computed as the weighted trait means for each plot, where weights are species relative abundances. Community-weighted traits for all leaf chemical traits (LCCcw, LNCcw, LPCcw and LSCcw) were computed using the trait values measured from each fertilizer level (0, 50, 100, 250 or 500 kg ha−1 year−1) because these data were available for the six most abundant species that together comprised >80% of total abundance from all plots. Doing so partially takes into account intraspecific variation for these traits, which can be important (Garnier et al. 2007; Albert et al. 2010).

Temporal trends (1981–2008).  First, we explored temporal trends in all continuous community-weighted traits (Table 1) for the first 27 years of the experiment (i.e. 1981–2008) from the visual ranking vegetation data and the trait data from year 26 (cover data). For simplicity, in presenting the main trends, we divided the 27-year time series into three distinct periods: 0–5 years (the ‘adjustment’ period), 5–19 years (the ‘middle’ period) and 19–27 years (the ‘post-fertilizer’ period, because fertilizer application was ceased at year 19). To highlight major temporal trends in the data, we used generalized additive models (GAMs; Wood 2006) on community-weighted traits between fertilizer and grazing intensity levels.

Because vegetation composition was not assessed prior to the start of the experiment (i.e. in 1981), but only from 1983 onwards, we used the two additional plots from the 2007 vegetation cover data as estimates of the starting conditions for all plots. These estimates are necessarily imprecise, because they assume that all plots had exactly the same vegetation composition and because the data were obtained from a different vegetation sampling method (i.e. visual ranking vs. vegetation cover) whose results are only moderately correlated with each other (Lalibertéet al. 2010). However, community-weighted traits have little sensitivity to vegetation sampling method (Lavorel et al. 2008). Moreover, the two additional plots should represent reasonable estimates of starting conditions since sampling records show that these have changed little since 1981, while fertilizer treatments led to large and rapid changes in vegetation composition (Scott 2007). GAMs were computed in R, using the gam function in the mgcv package (Wood 2006).

Differences in community-weighted traits at year 26  We performed statistical tests of differences in community-weighted traits among fertilizer and grazing intensity treatments on the vegetation cover data at year 26 (i.e. 2007). This was performed because: (i) these data encompassed all plant species (Table S1), not just the most abundant ones in each plot (as in the visual ranking data), and (ii) the community-weighted traits from that year were used as the constraints in the CATS models (see ‘CATS models’ section). For these analyses, H, LCC, LNC, LPC, LSC and LA were log-transformed prior to computation of community-weighted traits to reduce the influence of species with large trait values. SMcw and OFcw were not computed because of missing trait values for a number of species, some of which were relatively abundant (Table 1). PCA correlation biplots were used to visualize the interrelationships among community-weighted traits in multivariate space. Differences in individual community-weighted traits among the different experimental treatments were tested using a split-plot anova model. Post hoc Tukey’s HSD tests were conducted when a significant interaction or main treatment effect was detected.

CATS models

A general description of the CATS model (Shipley, Vile & Garnier 2006; Shipley 2010a) is given in Appendix S1. Our main interest was the λ-values for each trait that are estimated by the model. The λ-values measure the degree to which the relative abundance of a species changes as the value of the trait changes, holding other traits constant.

When traits are standardized to unit variance, these λ-values can be directly compared to estimate the relative importance of different traits in determining community structure (Shipley 2010a). A positive λ-value for a given trait indicates that when other traits are held constant, species with greater values for this trait are more abundant in the local community than could be expected from the regional abundances, while negative λ-values indicate the opposite. A λ-value of zero suggests that a trait does not influence abundance. Hence, provided that the predicted relative abundances obtained from the CATS model match well the observed ones, λ-values can provide information about the importance of different traits during community assembly (Shipley 2010a).

We used different sets of traits in our CATS models. In addition, we also ran the model by excluding rare species. We also used more uninformative as well as more informative priors (Shipley 2010a). Model fit was evaluated, and the ability of different trait constraints to improve predictive capacity was determined using permutation tests (Shipley 2010c). Finally, λ-values were compared using anova and post hoc Tukey’s HSD tests. Doing so allowed us to assess the relative importance of different traits in determining abundance across the different experimental treatments. Details on all of these analyses are given in Appendix S1.


Relationships Among Species Traits

There were moderately strong (|r| ≤ 0.6) and significant (≤ 0.05) correlations between several of the functional traits measured (Fig. S1). A PCA conducted on all continuous traits (other than SM and OF, because of missing values for these two traits; Table 1) highlighted the intercorrelation between several traits along the two first principal components (Fig. 1a). Together, the first two principal components represented 74% of the variance in these traits. LNC, LPC, LSC, LDMC and SLA all had strong loadings on the first axis, which explained 54.3% of the variance in traits (Fig. 1a). The second axis, which explained an additional 19.7% of the variance in traits, was primarily driven by plant height and LCC (Fig. 1a), although these two traits were not significantly correlated with each other (Fig. S1). LCC had the strongest loading on the third axis, which explained 9.9% of the variance in traits. Finally, there was a sharp distinction between native and exotic species along the first PCA axis, with native species having distinctly greater than average LDMC values and lower than average LNC, LSC, SLA, LSC and LPC values than exotic species (Fig. 1a).

Figure 1.

 Principal component analysis correlation biplots (type-II scaling) showing the relationships among (a) nine traits (black vectors) for all species and (b) the same nine community-weighted traits (black vectors) for all 32 plots (2007 vegetation cover data). Grey dots represent species (see Table S1 for species codes definitions). Traits that were log-transformed in (a) were also log-transformed in (b) before computing the community-weighted traits. See Table 1 for a description of the trait codings.

Shifts in Community-Weighted Traits

Temporal trends: 0- to 5-year period

In the 0- to 5-year adjustment period, there were rapid changes in several community-weighted traits, such that distinct peaks (positive or negative) were present (Fig. 2). SLAcw initially increased rapidly at all fertilizer levels, but became highest (c. 26 m2 kg−1) in the high fertilizer/irrigated treatment. A similar trend was observed for AScw. LTcw decreased in the high fertilizer/irrigated treatment (c. 0.20 mm) but increased in all other treatments (c. 0.35 mm).

Figure 2.

 Shifts in community-weighted traits among fertilizer treatments for the 27 years of the experiment (1981–2008). See Table 1 for units. Smoothed lines are fitted values from generalized additive models (GAMs), with grey shaded areas representing standard errors. The vertical dashed lines indicate year 5 (the end of the ‘adjustment’ period) and year 19, when fertilizer treatments were ceased.

LNCcw and LSCcw increased in all treatments, but more so in fertilized than in unfertilized ones; LDMCcw showed a weak trend in the opposite direction. The trends in LNCcw and LSCcw matched the proportional increases in N-fixing species (NFcw) among the different fertilizer treatments. LPCcw showed a different pattern in that it initially increased to a greater level in unfertilized than in fertilized treatments. LCCcw increased in all treatments, but less so in high fertilizer/irrigated ones. SMcw and LAcw strongly increased in dryland fertilized treatments, increased moderately in unfertilized treatments and remained stable in the high fertilizer/irrigated treatment. Hcw increased in the dryland fertilized treatments but remained stable in high fertilizer/irrigated and unfertilized treatments. OFcw was delayed in high fertilizer/irrigated treatments but appeared to vary little in the other treatments.

The changes in community-weighted traits during the adjustment period reflected marked increases in the relative abundance of N-fixing species (Fig. S2). Lupinus polyphyllus, a tall legume (c. 40 cm), rapidly increased in relative abundance in all but the highest fertilizer treatments (Fig. S2). Clovers (Trifolium spp.; see Table S1) increased in abundance in all dryland fertilizer treatments and particularly so at the highest fertilizer level, where they became dominant (Fig. S2).

Temporal trends: 5- to 19-year period

Traits that showed a rapid initial increase (or decrease) in the initial adjustment period shifted back to a value closer to the estimated starting value during the 5- to 19-year period (Fig. 2). For example, after its initial peak, SLAcw decreased strongly in the high fertilizer treatment, while it remained stable (after a temporary decrease) in the dryland (i.e. non-irrigated) fertilizer treatments and progressively decreased in the unfertilized treatment. Consequently, all fertilized treatments converged towards a similar SLAcw (c. 20 m2 kg−1) around year 15, while SLAcw in the unfertilized treatment was still distinctly lower (c. 15 m2 kg−1). Similar trends were observed for AScw, LNCcw and NFcw.

LSCcw remained stable after its initial peak for all fertilized treatments, while it consistently decreased in the unfertilized treatment. LDMCcw increased in all treatments, but more so in the unfertilized treatment, which had the highest LDMCcw at year 19. SMcw and LAcw (and to a lesser extent LTcw) remained stable in the high fertilizer/irrigated treatment but decreased sharply in dryland fertilized treatments around year 15. OFcw remained stable in the high fertilizer/irrigated treatment and increased progressively in all other treatments. Hcw remained relatively stable (showing only a slight increase) among treatments during the 5- to 19-year period.

The 5- to 19-year period was marked by decreases in the abundance of L. polyphyllus in unfertilized and dryland fertilized treatments and increases in exotic grasses in all treatments (Fig. S2).

Temporal trends: 19- to 27-year post-fertilizer period

The post-fertilizer 19- to 27-year period was characterized by less temporal variability within treatments for most community-weighted traits. The main exception to this general trend was in the unfertilized treatment, where SLAcw, LNCcw, LSCcw, LPCcw and NFcw steadily decreased, while LDMCcw kept increasing (Fig. 2). Similarly, there were smaller fluctuations in the relative abundance of different groups of species than in the previous two periods (Fig. S2).

Temporal trends: grazing intensity

In contrast to the large differences that were observed among fertilizer treatments in the temporal trends of community-weighted traits, the differences among grazing intensity treatments were much subtler and all three levels followed similar trajectories (Fig. S3). In general, differences among treatments became more apparent with time (particularly after around year 15), but such differences were always small and only noticeable for a few traits (e.g. AScw, LNCcw, LSCcw, LDMCcw, NFcw and Hcw). Most notably, there was a trend for Hcw to increase under lower grazing intensity compared to the other two levels (Fig. S3).

Differences in community-weighted traits at year 26

The pattern of trait variation observed at the species level (Fig. 1a) was also reflected at the community (plot) level (Fig. 1b). Although qualitatively similar, correlations between community-weighted traits were generally stronger than those between species-level traits (Fig. S4). A PCA of community-weighted traits showed that plots receiving no fertilizer had greater than average LDMCcw and lower than average SLAcw, LNCcw, LSCcw and LPCcw, than all fertilized plots (Fig. 1b). There was also a tendency for plots under low grazing to have greater Hcw than plots under moderate or high grazing (Fig. 1b).

There were significant (≤ 0.05) differences between unfertilized and fertilized treatments for SLAcw, LAcw, LSCcw and LNCcw, which were all greater in the fertilized treatments, yet there were no differences within the different fertilized treatments (Fig. 3a). LPCcw, Hcw, LDMCcw and LCCcw did not significantly differ among fertilizer treatments (Fig. 3a). LSCcw increased with grazing intensity and differed significantly among the three grazing intensity levels (Fig. 3b). LNCcw was significantly lower under lax grazing than under hard grazing (Fig. 3b). LPCcw was significantly higher at the highest grazing intensity, while the opposite was true for LDMCcw (Fig. 3b). Hcw was higher in the lax-grazing treatment than in the other two treatments (Fig. 3b). Two traits (LSCcw and LPCcw) showed significant fertilizer × grazing interactions. LSCcw of high fertilizer/irrigated plots was significantly greater under hard grazing than under lax or moderate grazing (Fig. S5). For LPCcw, the significant interaction was due to significant differences in the high fertilizer/irrigated treatments between lax and hard grazing and similarly for the 100 kg ha−1 year−1 fertilizer treatment (Fig. S5).

Figure 3.

 Boxplots showing the differences in community-weighted traits among (a) fertilizer and (b) grazing intensity treatments at year 26. Different letters indicate significant differences (α = 0.05) based on post hoc Tukey’s HSD tests. The central bar shows the median, the box represents the interquartile range (IQR), the whiskers show the location of the most extreme data points still within 1.5 × IQR ± the upper or lower quartiles, and the grey points are outliers. See Table 1 for a description of the trait codings and units. Statistical tests were based on community-weighted traits computed from log-transformed traits for Hcw, LCCcw, LNCcw, LPCcw, LSCcw and LAcw, but the figure shows untransformed community-weighted trait values. mod, moderate.

CATS models

Prediction of relative abundances

Starting with each of the community-weighted traits except LT, SM and OF (Table 1) with a maximally uninformative uniform prior, we were able to retrospectively predict 66% of the variation (≤ 0.05) in the observed species relative abundances of all 51 species from the regional pool overall plots (Figs 4a and 5a). Plotting the predicted relative abundances against observed relative abundances on a logarithmic scale showed that rare species (those forming <5% of total abundance per plot) were less well predicted than the most common ones (Fig. 5b). In this model, progressive removal of traits with the smallest absolute λ-values suggested that LNC, LDMC, LA and SLA were the most important traits, while LCC, plant height, LPC and NF were the least important ones (Fig. 4a). The predictive capacity of the CATS model was significant once two traits (LNC and LDMC) were included, although the resulting r2 was low when only these two traits were considered (Fig. 4a).

Figure 4.

 Predictive capacity (Pearson r2) of the CATS model over all plots as the number of traits used increases. (a) All traits except LT (Table 1), uniform prior. (b) All traits including LT, using either a uniform or a more informative neutral prior. The black lines and trait labels show where the CATS models start becoming statistically significant (≤ 0.05).

Figure 5.

 Results of CATS model for the 2007 vegetation cover data, using (a) all traits except LT and a uniform prior on arithmetic or (b) logarithmic scales (i.e. log10 [+ 0.0001]) and using (c) all traits including LT with an informative neutral prior on arithmetic or (d) logarithmic scales. Analyses were performed using all 32 plots and all species. *≤ 0.05. The dashed lines cut both axes at a relative abundance of 0.05 (5%).

Using a more informative neutral prior in the model gave even better results (r2 = 0.784, ≤ 0.05; Fig. S6), although some traits that were previously important (e.g. SLA) became unimportant in the new model following our backward selection procedure (Fig. S7). Using the uniform prior but considering only the most abundant species accounting for 99% (35 species in total), 90% (21 species) or 80% (16 species) of the total abundance in each plot yielded much better predictions (Fig. S8).

Compared with our initial model (Figs 4a and 5a), adding LT as a trait in the CATS model greatly increased predictive capacity, particularly with the more informative prior (r2 = 0.929, ≤ 0.05; Figs 4b and 5c). In these new models, LT, SLA and LSC were important traits, while LDMC and LNC became much less important (Fig. 4b). In the model with the more informative prior, four traits (LT, SLA, LSC and NF) were required to contribute significant predictive capacity beyond the information already contained in the prior. The model with the uniform prior attained significant predictive capacity when only one trait (LA) was added (Fig. 4b).

Comparison of λ-values

For the CATS model with the highest predictive capacity (i.e. the one that included LT and a more informative neutral prior; Fig. 5c), species with lower LT were advantaged at all fertilizer levels (i.e. negative λ-values for LT), and there was a non-significant trend for this advantage to increase under greater fertilizer application (Fig. 6a). A similar pattern was obtained for SLA, except that the trend was significant: lower SLA was, surprisingly, more advantageous in the high fertilizer/irrigated treatment compared with the unfertilized treatment (Fig. 6a). Species with greater LSC and NF were favoured at all fertilizer levels, although this pattern was more important in the fertilized treatments compared to the unfertilized one (Fig. 6a). There was also a significant difference in λ-values for LDMC between 50 and 500 kg ha−1 year−1 treatments (Fig. 6a); λ-values for LDMC were positive in the 50 kg ha−1 year−1 treatment level, whereas they were negative at the highest fertilizer level.

Figure 6.

 Boxplots showing the differences in λ-values among (a) fertilizer and (b) grazing intensity treatments for the 12 traits (including LT) used in the CATS model with the informative neutral prior (Fig. 5c). Different letters indicate significant differences (α = 0.05) based on post hoc Tukey’s HSD tests. Traits are ordered according to their importance, based on Fig. 4b. The central bar shows the median, the box represents the interquartile range (IQR), the whiskers show the location of the most extreme data points still within 1.5 × IQR ± the upper or lower quartiles, and the grey points are outliers. See Table 1 for a description of the trait codings. mod, moderate.

With regard to grazing intensity, species with greater LT were more strongly disadvantaged under lax grazing than under hard grazing (Fig. 6b). Significant differences in λ-values for plant height were also observed, such that species with greater height were favoured under lax grazing, but were disadvantaged under moderate and hard grazing (Fig. 6b). Higher LNC was favourable under lax grazing but unfavourable under hard grazing, and this difference was significant (Fig. 6b).

Significant fertilizer × grazing interactions on λ-values were detected for LT, SLA and LNC (Fig. S9). In the 500 kg ha−1 year−1 fertilizer treatment, both greater LT and SLA were more unfavourable under lax grazing than under hard grazing. At the same fertilizer level, higher LNC was favourable under lax grazing, yet was unfavourable under hard grazing.


Shifts in Community-Weighted Traits

Responses to fertilisation

The first 5 years of the experiment were characterized by rapid shifts in community-weighted traits that reflected changes from more conservative nutrient-use strategies to less conservative ones across all fertilizer levels, with the trend most pronounced in the high fertilizer/irrigated treatment. These shifts were linked to marked increases in the relative abundance of species with N-fixing ability possessing leaf attributes associated with rapid growth, namely L. polyphyllus and clovers (Trifolium spp.). The increase in L. polyphyllus was particularly important in the dryland fertilized treatments (50, 100 and 250 kg ha−1 year−1) and to a lesser extent in the unfertilized treatment, but not in the high fertilizer/irrigated treatment. Instead, clovers became largely dominant, although there was a shift in time from Trifolium repens to Trifolium ambiguum (Scott 2007). Since the fertilizer used in this study was sulfur-enriched superphosphate (i.e. a P/S fertilizer typical of legume-based pasture systems in Australia and New Zealand), the rapid increase in these N-fixing species in fertilized treatments can be explained by the strong dependency of N fixation on P (Israel 1987; Olivera et al. 2004) and S (Scherer & Lange 1996; Krusell et al. 2005; Varin et al. 2010), at least for crop legume species (Sprent 1999). On the other hand, the initial rapid increase in the relative abundance of L. polyphyllus in the unfertilized treatment may be partly due to this species exhibiting the same P acquisition strategy as congeneric Lupinus angustifolius, which can release large amounts of carboxylates through root exudation to solubilize P when it is deficient (Hocking & Jeffery 2004).

Five years after the start of the experiment, there were decreases in SLAcw and LNCcw and increases in LDMCcw, at all fertilizer levels, particularly at the highest fertilizer level. In the unfertilized treatment, all other leaf traits that were tightly related to leaf nutrient economy (i.e. LNCcw, LPCcw, LSCcw and LDMCcw) also reverted back to their original values after a rapid initial shift to larger values, although this was a gradual process that occurred over a period of about 20 years. This initial shift towards ‘fast-growing’ attributes in the unfertilized treatment is at odds with the view that slow-growing plant species are better competitors for soil resources than fast-growing ones in nutrient-limited environments (Tilman 1988). We hypothesize that the gradual return of the unfertilized treatment towards its initial community-weighted trait values (e.g. low SLAcw) is because nutrient limitation gives a long-term advantage to species with attributes that limit nutrient losses to herbivory and other stresses (Chapin 1980; Ryser 1996; Aerts & Chapin 2000; Grime 2002). Such attributes reduce growth rate (Lambers & Poorter 1992; Ryser & Lambers 1995), but can also lower acceptability to sheep (as evidenced by the gradual decrease in AS in the unfertilized treatment after its initial peak) and reduce nutrient losses by a slower tissue turnover rate. This interpretation is in line with microcosm experiments that showed how fast-growing plant species out-compete slow-growing ones under low-fertility conditions in the absence of generalist herbivores but that the reverse is true when herbivores are present (Fraser & Grime 1999; Buckland & Grime 2000).

Responses to grazing intensity

Compared to the large effects of soil resource availability on community-weighted traits, relative grazing intensity had only subtle effects, although these became more apparent with time. After 26 years, Hcw was greater under lax grazing than under moderate or hard grazing. This pattern agrees with a recent global analysis of plant trait responses to grazing (Díaz et al. 2007b).

Species with attributes associated with a rapid-growth strategy become more prevalent under higher grazing intensity. The prevalence of such a grazing-tolerant strategy under more intense grazing has also been reported elsewhere (Cingolani, Noy-Meir & Diaz 2005; Cruz et al. 2010). Westoby (1999) suggested that low-intensity, selective grazing should favour slower-growing, less palatable plants (grazing avoidance), whereas hard non-selective grazing should favour faster-growing, more palatable plants (grazing tolerance). Since all plots used in this study were ‘mob-grazed’ (i.e. a large number of sheep were introduced to plots for 3–4 days), our results support this idea. In addition, in the high fertilizer/irrigated treatment, LPCcw and LSCcw were significantly greater under hard grazing than under lax grazing. These significant fertilizer grazing interactions support the resource availability model (Coley, Bryant & Chapin 1985), which predicts that a rapid-growth, grazing-tolerant strategy becomes more advantageous under greater resource availability.

Importance of Traits in Determining Abundance: CATS Models

Despite providing some insights, testing for differences in community-weighted traits does not inform us on the relative importance of particular traits in determining abundance in different environments. This is because the importance of a trait in determining abundance must be defined by reference to the entire pool of species that could potentially colonize a site, including those that did not successfully establish (Shipley 2010a). For example, sites with higher soil resource availability could show higher SLAcw, yet greater SLA could still be unfavourable if the majority of species from the regional pool had very large SLA values and if these species were absent or present only at very low abundance in these sites. Moreover, SLAcw would be a constraint to community assembly, rather than simply a consequence of community assembly, only if the relative abundances that are predicted in its presence in the model differ from those that are predicted in its absence (Shipley 2010b). For instance, particular values of SLAcw may not reflect constraints on community assembly but instead simply arise from correlations with other important constraints, such as LNCcw. On the other hand, the parameters of the CATS model (i.e. the λ-values) can be directly used to quantify the relative importance of different traits during community assembly (Shipley 2010a).

Prediction of relative abundances

Using all traits except LT and a uniform prior, we were able to predict 66% of the variation in the relative abundances of the 51 species from the regional pool over 32 plots. This is substantially less than the first empirical application of the model, which predicted 96% of the variation in the relative abundances of 30 species from 12 plots, using eight traits and a uniform prior (Shipley, Vile & Garnier 2006). One potential explanation for this poorer fit is simply that the community-weighted traits used in our study reflect weaker constraints over community assembly. However, Shipley, Vile & Garnier (2006) used a subset of species from the regional pool, excluding very rare ones, whereas we used complete botanical surveys and associated trait data for all species. Reducing the number of rarer species greatly increased the predictive capacity of our CATS models. This suggests that the relative abundances of rarer species are only weakly determined by the traits included in the model. Recently, it has been pointed out that the influence of incomplete surveys and missing species on the predictive capacity of the CATS model is as yet unknown (He 2010); our results indicate that this effect can be large. Since our measure of relative abundance, based on visual cover estimates, is necessarily less precise than actual biomass estimates, this lower precision also likely contributed to the remaining lack of fit in the models.

Relative importance of traits across treatments

In an effort to increase the predictive capacity of the full CATS model that included all species, we used a more informative neutral prior and also added LT as a trait in the model because this trait can vary independently of LDMC or SLA along environmental gradients (Witkowski & Lamont 1991). This more complex model had much greater predictive capacity, explaining >92% of the variation in relative abundances in all species from all plots even when all species were included. Thus, this better model was used to compare its λ-values across experimental treatments.

In that model, greater SLA became increasingly unfavourable (i.e. species with larger SLA values were less abundant, holding other traits constant) as soil resource availability increased. This contrasted sharply with the marked and highly significant increase in SLAcw following fertilizer addition. However, because λ-values reflect the importance of single traits after taking into account other traits included in the model, the results for SLA need be interpreted accordingly, particularly since SLA is a product of LT and LDMC (Witkowski & Lamont 1991; Vile et al. 2005). First, there was a trend for thinner leaves to be more favourable under greater soil resource addition, although differences between fertilizer levels were not significant. However, given the similarity of the patterns found for LT and SLA (for which significant differences between fertilizer levels were found) and the clear monotonic decrease in λ-values for LT with fertilizer rate, our inability to reject the null hypothesis for LT likely reflects more the lower statistical power for tests of a whole plot factor (here, fertilizer level) in a split-plot design (Gotelli & Ellison 2004) than a true null hypothesis. Second, LDMC was unimportant in determining abundance across all fertilizer treatments with the exception of the 50 kg ha−1 year−1 treatment, where it had a positive effect. Therefore, one possible interpretation for these results was that although species with increasingly thinner leaves were favoured under greater soil resource availability, further increasing SLA of thinner leaves was disadvantaged because this can only arise through a lower LDMC. A lower LDMC may be disadvantageous under grazing, since a greater investment in cell wall material (and thus greater LDMC) can confer resistance against trampling (Lambers, Chapin & Pons 2008). While lower LT can increase the rate of CO2 diffusion inside the leaf (Syversten et al. 1995), thus potentially promoting photosynthesis, it seems unlikely that this alone can explain the advantage of lower LT under greater soil resource availability. Rather, it may be interpreted as a way to maximize SLA (and thus maximize light interception per unit leaf dry mass) without sacrificing LDMC (and thus maintain leaf structural defences). Whether this might represent a general strategy for plant survival in fertile grasslands under grazing requires further study.

The ability to symbiotically fix atmospheric N (NF) was favourable in all treatments that received fertilizer, whereas this trait was not important in the unfertilized treatment. Similarly, higher LSC was favourable in fertilized treatments, whereas it was unimportant in the unfertilized treatment. These results agree with the patterns that were found for NFcw and for LSCcw. Given that the fertilizer used was superphosphate (i.e. a P/S fertilizer), our interpretations thus remain similar to those put forward for the patterns in community-weighted traits. First, N is a crucial element for photosynthetic enzymes and thus strongly controls photosynthetic rates (Lambers, Chapin & Pons 2008) and, N enters natural ecosystems predominantly through fixation of atmospheric N (Chapin, Matson & Mooney 2002). However, since N fixation strongly depends on P (Israel 1987; Krusell et al. 2005; Varin et al. 2010) and S (Scherer & Lange 1996; Krusell et al. 2005; Varin et al. 2010), this may explain why this trait was favourable only in the presence of the P/S fertilizer. Second, higher soil resource availability favours a rapid-growth, nutrient-acquisitive strategy (Aerts & Chapin 2000), and greater LSC is associated with faster growth because of the importance of S in proteins and other organic compounds (Hawkesford 2007). Sulfur deficiency is widespread in many areas of New Zealand, especially in drier inland areas (Walker & Gregg 1975), and S is a limiting nutrient to pasture growth in our study area (Scott 2000). If greater LSC reflects a greater capacity for S uptake, then this could perhaps explain why LSC was favourable as S availability increases in fertilized treatments; however, this remains speculative. On the other hand, under low S availability, it may be more advantageous to conserve acquired S through greater constitutive defences (e.g. lower SLA and/or higher LDMC, which would ‘dilute’ LSC), and this may explain why LSC was unimportant in this treatment. There was only a weak advantage of higher LPC and no advantage of higher LNC at all fertilizer levels, but this probably reflected the fact that all three traits were highly positively correlated between species, such that greater LPC or LNC conferred little additional benefit once LSC (and other traits) was taken into account. As shown before with LT, SLA and LDMC, this highlights the importance of considering covariation between traits when interpreting the direction and strength of λ-values (Shipley 2010a).

Greater plant height was favourable under lower grazing intensity, but not under moderate or hard grazing. This was consistent with the results found for Hcw. A likely explanation is that under higher disturbance intensity/frequency, taller species lose a disproportionate amount of their above-ground biomass relative to shorter ones, whereas lower disturbance intensity/frequency can favour taller species because greater height is associated with greater competitive ability (Westoby 1998; Bullock et al. 2001; Grime 2002), at least when competition is primarily for light (Aerts 1999). We also found that greater LNC and thinner leaves were more favourable under lower than harder grazing intensity. Together, these results point towards an advantage to tall species with a rapid-growth strategy (through greater LNC and thinner leaves), which is consistent with the competitor (C) strategy of Grime (1974). However, it must be noted that this contradicts the results found for community-weighted traits, where higher grazing intensity led to significantly greater LNCcw, LPCcw, LSCcw and lower LDMCcw. This again illustrates that there is not always a direct link between community-weighted traits, which are computed only from the species that are actually present at a site, and the λ-values of the CATS model, which represent the importance of traits with regard to the entire available pool of species, including those that have already been excluded from the local community.

Model parsimony

Can we identify a parsimonious set of traits that are most important in driving community assembly? We attempted to answer this question through a simple backward selection procedure where traits with the smallest absolute average λ-values across all plots were progressively removed from the models. This approach suffers from high collinearity between traits, which can lead to unstable solutions (Shipley 2010a). However, no alternatives are currently available for selection of CATS models, and only forward or backward stepwise approaches have been used so far (Mokany & Roxburgh 2010; Shipley 2010a; Laughlin et al. 2011). Despite the limitations of our backward stepwise approach, it was clear that some traits such as LCC and LS were never important constraints to community assembly in these grasslands and could be omitted in future studies of grassland responses to fertilisation and/or grazing. Second, LSC was always more important than LPC, which may reflect the importance of S in these grasslands (Scott 2000). Third, the CATS model without LT increased the importance of LNC and LDMC. On the other hand, in the other models that included LT, we found that LNC and LDMC became unimportant, while correlated traits such as SLA and LSC became more important. This illustrates the difficulty of getting stable estimates λ-values when relatively high collinearity among traits is present (Shipley 2010a). Consequently, the development of more sophisticated procedures for trait selection in CATS models is an important area for future research.

Relationships Among Species Traits

Plant species were differentiated along two major axes of functional variation. The first and most important axis was mainly driven by LNC, LSC, SLA, LDMC and to a lesser extent LPC. This axis represents the fundamental trade-off between traits that promote rapid growth and those that promote persistence (Chapin 1980; Herms & Mattson 1992; Lambers & Poorter 1992; Aerts 1995; Reich, Walters & Ellsworth 1997; Aerts & Chapin 2000; Grime 2002). This nutrient acquisition–conservation trade-off has been identified as a primary axis of functional variation among species, both when several traits were compared across fewer species (Grime et al. 1997; Adler et al. 2004) and when fewer traits were compared across many species (Reich, Walters & Ellsworth 1997; Díaz et al. 2004; Wright et al. 2004).

We found that native species were functionally distinct from exotic species, with exotic species exhibiting a rapid-growth trait syndrome, while the reverse was true for natives. These results are consistent with those from a recent global meta-analysis that compared leaf traits among co-occurring exotic and native species, where exotic species were positioned further along the acquisition–conservation axis towards a faster-growth strategy (Leishman et al. 2007). The difference may be partly explained by the fact that all sown species at the start of the experiment were exotic species which had been selected based on their potential suitability as pasture species, of which high growth rate is a key characteristic. Other studies that compared co-occurring herbaceous native and exotic species in New Zealand grasslands also found that exotic species had faster growth rates compared with native ones (Scott 1970; King & Wilson 2006) or possessed leaf and root attributes associated with faster growth rates (Craine & Lee 2003). It may also reflect the particular evolutionary history of New Zealand grasslands, which prior to human settlement were confined to sites with marginal environmental conditions (McGlone 2001).

The second major axis of functional variation among species was best represented by plant height, a result similar to that of Díaz et al. (2004) and Adler et al. (2004). Plant height is an important aspect of competitive ability (Grime 1977; Gaudet & Keddy 1988) when competition is primarily for light (Aerts 1999). However, while taller species may be able to capture a greater proportion of light resources, frequent and/or intense disturbances can remove a disproportionate amount of their biomass relative to shorter species, putting them at a disadvantage (Grime 2002). Therefore, allocation of above-ground biomass to occupy vertical space reflects another trade-off in plant functional variation that is influenced by disturbance frequency/intensity, and which is largely independent of the nutrient acquisition–conservation trade-off (Westoby 1998; Grime 2002; Díaz et al. 2004).


By experimental additions to the plant species pool, manipulation of soil resource availability and grazing intensity and long-term (27 years) monitoring of shifts in trait distributions across experimental treatments, we were able to determine the importance of traits to community assembly. First, we found that the relative abundances of plant species from a common initial species pool were strongly influenced by functional traits across all experimental treatments, highlighting the importance of traits in determining relative abundance (Shipley, Vile & Garnier 2006; Cornwell & Ackerly 2010; Shipley 2010a). Second, we showed that the relative importance of particular traits shifted with soil resource availability and sheep grazing intensity. Third, our results provide long-term experimental support for the hypothesis that slow-growing species do not become dominant under nutrient-poor environments because they have greater competitive ability than fast-growing species (Tilman 1988), but instead because they hold a long-term advantage through leaf attributes that reduce nutrient losses (Chapin 1980; Ryser 1996; Aerts & Chapin 2000; Grime 2002). Finally, we provided further support for the importance of the nutrient acquisition–conservation trade-off as a primary axis of functional variation among plant species (Grime et al. 1997; Díaz et al. 2004; Wright et al. 2004) and for the presence of differences between native and exotic species along that axis (Leishman et al. 2007).

Some limitations of our study also highlight future challenges for trait-based community assembly research. First, we have largely focused on above-ground functional traits, partly because of their ease of measurement, yet below-ground traits (e.g. specific root length, average rooting depth, presence of specialized root structures) could be very important in determining community structure, especially in nutrient-poor soils (Aerts 1999; Lambers et al. 2010). Chemical traits involved in defence against herbivores could also be important (Funk & Throop 2010) but have been largely neglected in trait-based community assembly research, including this study. Second, several of the traits we used were strongly intercorrelated, which led to unstable solutions in our selection procedure for CATS models and also complicated interpretation. As a result, future studies should aim at measuring traits that reflect as many independent axes of functional variation as possible (Laughlin et al. 2011). Our study also clearly illustrates that there is a need to develop better model selection approaches for CATS models than the simple stepwise procedure we used (He 2010).

Despite these limitations, our study suggests a way forward to understand and describe changes in plant species composition and diversity under land-use change in grazing systems (Chapin et al. 2000). This is crucial because land-use change is expected to be the single most important driver of changes in biodiversity world-wide for this next century (Sala et al. 2000). In particular, pastures and rangelands cover 25% of the ice-free surface of the Earth (Asner et al. 2004) and are expected to undergo rapid intensification (Bouwman et al. 2005) to meet the forecasted doubling in global food demand by 2050 (Alexandratos 1999) and increased demand for meat from developing countries in particular [Food and Agriculture Organization of the United Nations (FAO) 2005]. In that regard, our results highlight the importance of considering species functional differences to understand how plant communities react to increases in soil resource availability and increasing intensity and frequency of biomass removal, two important and often inseparable components of land-use change.


We wish to thank P. Fortier for help with field work, L. Kirk, A. Leckie, N. Pink and J. M. Tylianakis for logistical and academic support and A. Simpson for the use of stock. We also thank R. K. Didham, N. Gross, H. Lambers, W. G Lee, M. M. Mayfield, E. Weiher and J. A. Wells, as well as the Editor and anonymous referees, for providing insightful comments on the manuscript. We acknowledge financial support from the Miss E. L. Hellaby Indigenous Grassland Research Trust and the School of Forestry, University of Canterbury. During the writing of this manuscript, E. L. was supported by awards from the University of Canterbury, the Fonds québécois de recherche sur la nature et les technologies (FQRNT), Education New Zealand and The University of Western Australia. This research is endorsed by the Global Land Project (http://www.globallandproject.org).