Little evidence for limiting similarity in a long-term study of a roadside plant community


Correspondence author. E-mail:


1. Understanding the assembly of ecological communities is central to managing such communities for conservation, restoration and invasion resistance. A central plank of modern theory is limiting similarity, i.e. a finite limit to the similarity of coexisting species.

2. Here, we test the theory of limiting similarity in a roadside plant community that has been monitored for almost 50 years. We measure differences between observed trait distributions (‘functional diversity, FD’) and those expected on the basis of three different null models, several trait combinations and four different methods of measuring FD.

3. The general pattern was of lower observed than expected FD, or no difference between observed and expected FD, independent of FD measure and trait set. Some rare species had ‘unusual’ traits; thus if the null model assumed all species had equal chance of occurring, observed FD was generally lower than expected by chance, but if the null model assumed the probability of occurrence of rare species was proportional to the observed frequency of occurrence or excluded rare species, observed FD was not significantly different from expected.

4.Synthesis. The results do not support the idea that niche differences are central to the assembly of this plant community. However, they do suggest that rare species may be functionally unusual, and that the patterns revealed by this kind of analysis can be strongly influenced by the choice of null model, and much less so by the choice of FD measure.


A long-term aim of plant ecology has been to understand how local communities are assembled from the available species pool (Diamond 1975; Weiher & Keddy 1999). Such assembly is assumed to represent the outcome of two opposing forces: abiotic controls (environmental filtering) that tend to constrain permitted traits and trait combinations within certain limits, and internal forces (niche partitioning, limiting similarity or complementarity) that prevent coexisting species from being too similar (Grime 2006). Complementarity has assumed particular importance in the last 15 years, as the frequently observed superior yield of mixtures of species compared to monocultures is often attributed to complementarity, i.e. reduced competition and fuller use of available resources by species mixtures (Cardinale et al. 2007). However, in the context of complementarity, much of this recent literature has two rather curious features. First, much of the evidence for the role of complementarity in enhancing the performance of species mixtures comes from post hoc attribution of yield changes either to complementarity or to the ‘selection effect’, i.e. the tendency of mixtures to be dominated by the most productive species in monoculture (Huston 1997; Loreau & Hector 2001). There have been rather few direct attempts to detect the existence of complementarity in plant communities (Weiher, Clarke & Keddy 1998; Stubbs & Wilson 2004; Fukami et al. 2005; Mayfield et al. 2005; Cornwell, Schwilk & Ackerly 2006). Second, much of the evidence derives not from natural communities, but from synthesized experimental communities (Cardinale et al. 2007). Yet, if niche partitioning is central to community assembly, there is every reason to expect that natural communities will show more evidence of complementarity than artificial communities. The limited evidence suggests that they do (Flombaum & Sala 2008) and also shows that evidence of complementarity in experimental communities increases with age, interpreted as an increasing impact of species interactions on their structure (Cardinale et al. 2007).

One method for detecting the signature of complementarity or environmental filtering in species composition is to measure differences between observed trait distributions and those expected on the basis of a null model. When considering multiple trait dimensions simultaneously, this can be achieved by comparing measures of trait dispersion in multivariate space. Very dispersed observed multivariate trait distributions compared to expected (null) trait distributions would indicate greater than expected differences among species with respect to their trait values. This would be consistent with limiting similarity or niche partitioning. Conversely, observed multivariate trait distributions similar to expected (null) trait distributions would be consistent with random community assembly. Here, the dispersion of species in trait space is compared using measures of the functional diversity (FD) of communities (Petchey & Gaston 2007). Furthermore, we use a range of possible null models, each of which gives a different insight into the nature of any non-randomness in observed communities.

We present the results of a search for the ‘signature’ of complementarity in the assembly of one very common ecosystem type: the plant community (NVC MG1: Rodwell 1992) on a road verge at Bibury in Gloucestershire, UK. Community composition has been recorded here nearly every year since 1958, providing a unique opportunity to examine long-term patterns in community trait distributions. We focused on four primary questions: (i) were observed (O) and expected (E) trait distributions different from each other, i.e. did O minus E (O−E) differ from zero, (ii) were there differences in O−E among quadrats, (iii) did O−E change through time and (iv) were there any differences among quadrats in how O−E changed through time? In order to arrive at a consensus-based conclusion, we asked this set of four questions for three different null models, a range of trait combinations and a variety of different methods of measuring FD.

Materials and methods

Study system

The 700-m length of experimental road verge at Bibury (UK National Grid Reference SP 119048) is unusually wide (3.3–5.5 m), with relatively uniform vegetation, and the accompanying road bears only very light traffic. The mature vegetation at Bibury has experienced the same management (mowing annually in late autumn) and maintained a very similar species composition over 50 years of monitoring. Since 1958, the vascular plant composition of eight 1 × 0.5 m quadrats have been determined in most years. Monitoring continues, but we use data up to and including 2003 (325 quadrats × recording years). Over 46 years, 75 species have been recorded, although the mean species richness of individual quadrats in any 1 year is 14.9, the minimum was seven and the maximum 26.

Observed functional diversity and uniqueness

We define the 75 recorded species as the pool from which the communities in quadrats are assembled. For each of 325 data points (quadrats × recording years), we calculate FD (Petchey & Gaston 2002) of the actual community (observed FD) and FD of a community of the same richness assembled at random from the species pool (expected FD).

To calculate FD we used 23 traits (Table 1) that have been suggested to play a role in niche partitioning (Hooper 1998; Westoby et al. 2002; Lambers et al. 2004; van Ruijven & Berendse 2005; Vojtech et al. 2008). Since the actual traits responsible for complementarity are unknown, we use four different sets of traits: (i) all traits, (ii) traits of the adult plant (16), (iii) seed traits (7), and (iv) the set of four traits whose FD was found in a separate study (using different data from those analysed here) to have the greatest power to explain differences in biomass at Bibury (Thompson et al. 2005). These four traits are legume, individual leaf area, leaf dry matter content and canopy structure (‘bib maxr2’ traits).

Table 1.   Plant traits used in the analysis. Figures in brackets are the number of classes or categories for that trait. Trait data were obtained from Grime, Hodgson & Hunt (2007), PLANTATT (Hill, Preston & Roy 2004), Thompson, Bakker & Bekker (1997) and the LEDA data base (Kleyer et al. 2008). Seed dispersal was based on seed morphology and was a binary classification, i.e. species were classed as dispersed by a particular vector or not, and classes were not mutually exclusive
Life-formCategorical (5)Adult
Life historyCategorical (3)Adult
ClonalityCategorical (3)Adult
Specific leaf areaContinuousAdult
Individual leaf areaContinuousAdult
Leaf thicknessContinuousAdult
Leaf dry matter contentContinuousAdult
Canopy structureOrdinal (3)Adult
Canopy heightContinuousAdult
Lateral spreadOrdinal (5)Adult
MycorrhizasOrdinal (3)Adult
Flowering startOrdinal (7)Adult
Flowering periodOrdinal (8)Adult
Leaf phenologyCategorical (5)Adult
Seed weightContinuousSeed
Seed dormancy breakingCategorical (3)Seed
Seed bank persistenceOrdinal (3)Seed
Seed terminal velocityContinuousSeed

The species × trait matrix was converted into a distance matrix and this was clustered to produce a dendrogram that depicts the functional relationships within the plant assemblage (Petchey & Gaston 2002). Gower distance (Pielou 1984) was used throughout because Gower distance is appropriate for mixed types of data (ordinal, categorical, continuous; Sneath & Sokal 1973). UPGMA clustering was used as it most often led to the highest cophenetic correlation (Blackburn et al. 2005; Petchey & Gaston 2007; Petchey et al. 2007).

The FD of a notional assemblage of all 75 species is the total branch length of this dendrogram, and all measures of FD are standardized by this value, so that variation in FD ranges from 0 to 1, where a value of zero occurs for single-species communities (Petchey & Gaston 2006). The FD of each assemblage is the total length of the branches required to connect all of the species in the assemblage (Petchey & Gaston 2006). Standardization of FD and species richness between 0 and 1 has no quantitative effect on our results.

In addition to FD (Petchey & Gaston 2002, 2006), we also calculated mean dissimilarity (Heemsbergen et al. 2004), Q (Botta-Dukát 2005), and a measure (PS) by Podani & Schmera (2006). Mean dissimilarity is the mean of the pair-wise distances between species in trait space, Q is the mean of the pair-wise distances weighted by relative abundance and PS is a version of FD calculated from the local rather than regional dendrogram (Petchey & Gaston 2009; Petchey, O’Gorman & Flynn 2009). These indices have numerous statistical and mathematical properties that have been much debated; here we were interested only in whether they lead to qualitatively different answers to our ecological question.

We also calculated the functional uniqueness of each species as a tool to understand some of the patterns revealed. The measure of functional uniqueness used was the functional analogue of Pavoine, Ollier & Dufour (2005) measure of phylogenetic uniqueness. Instead of calculating it from the phylogeny, we calculated it from the functional dendrogram, which is also used to measure FD.

Expected functional diversity

Three null models provided three expected measures of FD. First, we constructed the random communities assuming that each species had an equal chance of occurring. Second, a species’ probability of occurring in a random community was proportional to the species’ occurrence in the 325 quadrat × year samples. The latter represents a null model that accounts for the rarity and commonness of different species. We also performed all analyses including or excluding the half of the species with lowest occurrence. This completely excludes any effect of the rare species on O−E. In all cases, 50 independent randomizations were made, to give 50 expected FD values, of which the mean was taken for comparison with observed FD.

Analysing variation in observed and expected FD

If the community is assembled at random, we expect observed minus expected FD (O−E) to be not significantly different from zero. We hypothesize that complementarity (or limiting similarity) will be manifest as the FD of real communities exceeding that of random communities, i.e. coexisting species are more different from each other in trait space than one would expect by chance, and therefore O−E should be significantly >0 (see Fig. 1 in Petchey & Gaston 2007). Therefore, the response variable in statistical models was O−E FD.

Figure 1.

 Deviation of observed minus expected functional diversity (FD) from zero in a road verge plant community. Patterns through time of eight quadrats are plotted separately. (a) Expected FD calculated using a null model where all species have equal chance of occurring. (b) Expected FD calculated using a null model where species’ chance of occurring is proportional to their frequency in the species pool. (c) Both observed and expected FD patterns calculated for only the 50% most common species.

Mixed effects models were used to estimate if O−E (i) was different from zero, (ii) differed between quadrats, (iii) changed through time and (iv) changed through time differently in different quadrats. Two statistical models were used: M1, a nested mixed effect model with the explanatory variables year and quadrat, where year was a continuous variable and quadrat a random variable and year was nested within quadrat; and M2, a mixed effect model in which the explanatory variables were year and random effect of quadrat. Both statistical models account for temporal autocorrelation and/or non-independence caused by repeated sampling of the same quadrats.

The overall intercept of the model is the mean of the intercept values for each of the quadrats. In all models, year was transformed by subtracting the mean and rescaling to the interval [−1, 1]. Hence, intercept estimates were half way through the duration of the observations, and estimates of slopes (year effects) were changes in O−E expected over half of the duration of the observations.

Question 1 was addressed by the intercept of the model M1, and significance of difference from zero was tested with Markov chain Monte Carlo simulation. Question 2 was addressed by comparing the explanatory power of M2 and M1 using a chi-squared statistic. Question 3 was addressed by the year effect in M2. Significance of the year effect was estimated by Markov chain Monte Carlo simulation. Question 4 was addressed by comparing the explanatory power of M1 and M2 using a chi-squared statistic. In no case was there a significant time × quadrat interaction, so Question 4 is not considered any further.

We answered each of the three remaining questions separately for every possible combination of trait set and FD measure, and used classification tree analyses to detect significant effects of these parameters on the answers to questions 1, 2 and 3.


Differences from expected functional diversity

The general pattern was of lower observed than expected FD, or no difference between observed and expected FD (Fig. 1). This qualitative pattern is consistent across differences in null model used (Fig. 2a), FD measure used (Fig. 2b) and trait set used (Fig. 2c). These parameters did, however, affect the magnitude and statistical significance of the difference between observed and expected FD (Fig. 2a–c). There was also evidence of significant effects of both quadrat (Fig. 2d–f) and year (Fig. 2g–i). While the magnitude of the year effect was very small relative to the overall intercept, quadrats often differed considerably. Thus, while the overall pattern is for observed FD to be less than or equal to expected, there can be significant differences between quadrats.

Figure 2.

 Boxplots of estimates of intercept, maximum quadrat effect and the year effect for each of the null models, functional diversity (FD) measures and trait sets. For ease of interpretation, the intercept is the mean of the quadrat effect at the midpoint of the time series in Fig. 1, and the slope is the effect size for half the duration of the time series. For explanation of trait lists and FD measures, see text. Null model All list: all species in pool have equal probability of occurring in the expected community; All relab: probability of occurring in expected community weighted by frequency in the pool; Common list: expected community based only on the 50% most frequent species in the pool.

The magnitude and statistical significance of the overall difference, quadrat effect and year effect depended on null model, FD measure and trait set. The influence of the null model on the three effects (overall intercept, quadrat effect and year effect) is shown in Fig. 2a,d,g. Regression tree analyses indicate a strong and significant effect of the null model on the overall intercept (Fig 2, Fig. 2a). If the null model assumed all species had equal chance of occurring, observed FD was generally lower than expected by chance (Figs 1a and 2a). The other two models either assumed a species’ probability of occurring was proportional to the observed frequency of occurrence or they excluded rare species; both null models resulted in observed FD that was not significantly different from the expected (Figs 1b,c and 2a).

Table 2.   Regression and classification tree effect sizes among different analyses. For explanation of trait lists and functional diversity (FD) measures, see text. Null model all: all species in pool have equal probability of occurring in the expected community; relab: probability of occurring in expected community weighted by frequency in the pool; common: the 50% most frequent species in the pool. Trait set bib maxr2: the four traits whose FD was found in a separate study to have the greatest power to explain differences in biomass at Bibury O−E: observed minus expected functional diversity. For explanation of different measures of functional diversity (FD, MD, PS and Q), see text
O−E intercept
 (1) Root48−0.038
 (2) Null model = all16−0.082
 (3) Null model = relab, common32−0.016
 (6) FD measure = Q8−0.034
 (7) FD measure = FD,MD,PS24−0.011
Maximum absolute quadrat effect
 (1) Root480.042
 (2) FD measure = FD, MD, PS360.019
 (4) FD measure = FD,P S240.015
 (5) FD measure = MD120.027
 (3) FD measure = Q120.109
Year effect
 (1) root480.0005
 (2) Trait = seed traits12−0.0003
 (3) Trait = adult traits, all traits, bib maxr2360.0007
 (6) FD measure = FD, MD, PS270.0004
 (12) trait = all traits90.0002
 (13) trait = adult traits, bib maxr2180.0005
 (7) FD measure = Q90.0017

The measure of FD used influenced each of the three effects, but did not change the overall conclusion that observed FD is less than or equal to expected FD (Fig. 2b,e,h). Regression tree analyses confirmed that FD, MS and PS resulted in very similar effect sizes, which were different from the effect sizes given by Q (Table 2). The measure Q generally suggested an effect size larger than the other three FD measures, although of the same sign (see Table S1 in Supporting Information). Consequently, the finding of lower observed than expected FD was robust to the diversity measure used.

The trait set used had a small but significant influence on the year effect (Fig. 2i; regression tree analysis, Table 2), but not on the overall intercept or the quadrat effect (Fig. 2c,f; see Table S1).

Occurrence and functional uniqueness

Relationships between the relative frequency of occurrence and functional uniqueness are displayed in Fig. 3 separately for each of the four trait sets. In each case, the functional uniqueness of rare species ranges from low to high values – some rare species were relatively unique. In contrast, there were no functionally unique common species.

Figure 3.

 Relationships between occurrence and functional uniqueness of species, given for four different combinations of traits: All traits, Adult traits, Seed traits, and Bibury max r2 traits: the four traits with the greatest power to explain differences in biomass at Bibury in a previous study: legume, individual leaf area, leaf dry matter content and canopy structure.


These results provide little support for the idea that complementarity or niche differences are central to the assembly of this plant community. At the small scale, common species appear to be assembled into communities at random. In other words, once species have passed through a dispersal filter (i.e. they have arrived at the site) and an abiotic filter (i.e. they are able to survive and grow under the prevailing conditions), niche-based interactions appear to play little further part in local community assembly.

Failing to account for the rarity of some species fundamentally changed our conclusions. In particular, the null model that did not account for species’ rarity (or commonness) created random communities with a high occurrence of species that are rare in the observed communities. As these rare species can also be very functionally unique (Fig. 3), the resulting random communities often have high FD. This gives the impression that the observed levels of FD are low and may explain observations of lower than expected FD. Note that given the close proximity of the study quadrats, in a single, apparently homogeneous road verge, any real evidence of environmental filtering at this scale (i.e. within the verge) would be hard to explain. In other words, we did not expect to find evidence of environmental filtering in our data set, and we did not find any such evidence, provided we used an appropriate null model.

Consistent with the effect on the null model of failing to account for species’ rarity, four recent studies that do not appear to control for species’ relative abundance in their null model find lower than expected FD (Fukami et al. 2005; Mayfield et al. 2005; Cornwell, Schwilk & Ackerly 2006; Petchey et al. 2007). In contrast, two studies that controlled for relative abundances found no difference from random (Weiher, Clarke & Keddy 1998; Stubbs & Wilson 2004). Some of these studies have suggested that their finding of lower than expected FD represents evidence against neutral theory. These conclusions need to be reconsidered in the light of our findings.

The difference between null models resulted from the intriguing finding that rare species have ‘unusual’ traits, or at least traits that are different from the common species’. Examination of the values of species’ functional originality reveals a suite of rare, functionally unusual species consisting of rare growth forms (e.g. geophytes such as Arum maculatum, Allium vineale, A. oleraceum; climbers such as Tamus communis), life history (e.g. annuals such as Lolium multiflorum) or nutrition (e.g. the legume Lathyrus pratensis). Nomenclature follows Stace (1997). Some of these species are always likely to be minority components of this kind of tall-herb community, but not all; a few, such as Calystegia silvatica, have the clear potential to become dominants themselves under different circumstances. Nevertheless, these rare species currently have little or no impact on the overall properties of the community, such as biomass; the least frequent 50% of species in the pool account for only 3% of the biomass (J.P. Grime, K. Thompson & A.P. Askew, unpublished data).

Two methodological concerns require some comment. The first is whether it is legitimate to use a single species pool to generate the expected communities, as not all species occurred in every quadrat; indeed, the very rarest species could not do so by definition. The fundamental question is whether some species are actually unable to grow in one or more quadrats, i.e. they are not part of the potential species pool for those quadrats. As all eight quadrats are located in the same road verge, with identical geology and management, and separated by only a few hundred metres, this seems highly unlikely. An examination of the data supports this view; for example Poa trivialis occurred in only seven quadrats, yet is one of the least specialized species in the British flora, abundant in a wide variety of habitats including wetlands, woods, grassland and arable fields (Grime, Hodgson & Hunt 2007). Plantago lanceolata, another species of wide ecological amplitude, also failed to occur in one quadrat, strongly suggesting that such absences are stochastic, and not attributable to environmental variation. Sixteen species, including all the dominants responsible for the overwhelming majority of the biomass, occurred in all eight quadrats.

A second question is whether the spatial scale employed (0.5 m2) is the right one for examining the assembly of this community. Clearly quadrats of this size would be too large to investigate coexistence in very small plants, e.g. in lawns (Wilson & Roxburgh 1994; Wilson & Watkins 1994), where species could easily occur in the same quadrat, yet never actually encounter each other. In the Bibury community, however, the dominant species have heights and spreads of the order of large fractions of a metre (Grime, Hodgson & Hunt 2007), and there seems little doubt that plants of such a size could not occur in the same quadrat without interacting, or that the biotic environment of all species is strongly conditioned by the characteristics of their neighbours.

Although both year and quadrat had small but significant effects, neither alters the overall conclusion that observed FD is less than or equal to expected. The year effect reflects a slight but significant tendency for O−E to become less negative (closer to zero) over time. This could be interpreted as a trend towards increasing (but still small) biotic interactions through time. Nevertheless, this effect seems to consist largely of the occurrence and frequency of rare species becoming increasingly constrained by the presence of the common community dominants; note that the year effect is least when analysis is confined to the common species (Fig. 2g). Quadrats varied in mean species richness, and the most diverse quadrats had the largest negative O−E values, while the least diverse had O−E closest to zero (Fig. 4); thus the quadrat effect is consistent with the overall conclusion that increased FD is strongly dependent on the presence of rare species.

Figure 4.

 Boxplots of estimated values of the difference between observed and expected functional diversity (FD) in 1980, the year to which the model intercept was set. The distribution of intercepts for each quadrat is given by modelling four trait sets, four measures of FD and three null models. Numbers above quadrat labels are mean species richness of each quadrat, arranged in order of increasing species richness.

The Bibury community has much in common, in terms of biomass, diversity and species composition, with the five northern BIODEPTH communities (Hector et al. 1999; Thompson et al. 2005). The latter were synthesized from seed, with the spatial arrangement and relative abundance of the different species determined, at least initially, by the experimenters. The former, in complete contrast, has been allowed to assemble itself, so that the present spatial arrangement of the plants, and their strongly skewed distribution of abundance, is a product of interactions between the component species over many decades. Theory suggests that we would expect Bibury to show the strongest evidence of interspecific interactions, and thus of complementarity, yet only BIODEPTH seems to provide such evidence.

This result suggests we still have much to learn about the importance or otherwise of niche differentiation in the assembly of plant communities, and in particular we cannot assume, as some concerned with managing invasions and restoration are beginning to (Shea & Chesson 2002; Funk et al. 2008; Moles, Gruber & Bonser 2008), that all (or even most) communities are strongly structured by limiting similarity or niche partitioning. While observational studies of plant communities and phenomenological analyses of their dynamics will remain important, experimental examinations of the mechanisms responsible for dynamics will also be important, particular for understanding the influence of trait choices and environmental heterogeneity.


K.T. was supported by NERC grant NE/D52222X/1 as part of the European Science Foundation Eurocores ASSEMBLE project. O.L.P. is a Royal Society Research Fellow. A.P.A. was supported by the Leverhulme Trust. Long-term monitoring at Bibury has at various times been supported by the British Ecological Society.