SEARCH

SEARCH BY CITATION

Keywords:

  • Biodiversity;
  • community assembly;
  • functional trait;
  • neutral theory;
  • species niche;
  • species pool

ABSTRACT

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. EXAMPLES
  5. CONCLUSIONS
  6. ACKNOWLEDGEMENTS
  7. REFERENCES
  8. BIOSKETCH
  9. Supporting Information

The relevance of neutral versus niche-based community assembly rules (i.e. the processes sorting species present in a larger geographical region into local communities) remains to be demonstrated in ecology and biogeography. To attempt to do this, a number of complex null models are increasingly being used that compare observed community functional diversity (FD, i.e. the extent of trait dissimilarity between coexisting species) with randomly simulated FD. However, little is known about the performance of these null models in detecting non-neutral community assembly rules such as trait convergence and divergence of communities (supposedly revealing habitat selection and limiting similarity, respectively). Here, using both simulated and field communities, I show that assembly rule detection varies systematically with the magnitude of the observed FD, so that these null models do not really succeed in breaking down the observed functional relationships between species. This is a particular concern, making detection of community assembly dependent on: (1) the pool of samples considered, and (2) the capacity of observed FD to correctly discriminate these rules. Null models should be more thoroughly described and validated before being considered as a magic wand to reveal assembly patterns.


INTRODUCTION

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. EXAMPLES
  5. CONCLUSIONS
  6. ACKNOWLEDGEMENTS
  7. REFERENCES
  8. BIOSKETCH
  9. Supporting Information

Species diversity is promoted and maintained by ecological processes operating on species attributes through space and time (Weiher & Keddy, 1995; Mayfield et al., 2010). Understanding the fundamental mechanisms by which species, present in larger geographical regions, will be found at a local site requires an assessment of the importance of different abiotic and biotic processes for community structure (Swenson & Enquist, 2009; Shipley, 2010). Ecological theory suggests that two main processes, habitat selection and biotic interactions, drive the non-neutral assembly of biological communities from broad to local scales. While habitat selection (or ‘habitat filtering’) should reduce the spread of ecological strategies reflecting shared ecological tolerances (Cornwell et al., 2006; Bruzgul & Hadly, 2007), biotic interactions within communities should maximize the differentiation between coexisting species (leading to a limited similarity between co-occuring species; Petchey et al., 2007; de Bello et al., 2009; but see Mayfield & Levine, 2010). Despite conceptual and empirical attempts to reconcile opposing views, the quest to demonstrate the predominant effect of different ecological processes on community structure remains active (Mason et al., 2008; Podani, 2009; Shipley, 2010), thus inevitably producing great uncertainties in modelling diversity patterns for biogeography.

Functional trait diversity (FD), a measure of the ecological differentiation between species, reflects one of the most relevant components of biodiversity underlying patterns in community assembly (Thompson et al., 2010). Hence, in recent years an increasing number of studies have implemented community FD (the trait dissimilarity between coexisting species) together with complex null models, to investigate the extent of different community assembly processes (Cornwell et al., 2006; Mason et al., 2007; de Bello et al., 2009). In essence, null communities are generated to break down the relationships between coexisting species and to produce randomly simulated FD values. This largely follows the approach already applied for species and phylogenetic diversity studies (Gotelli & McCabe, 2002; Miklos & Podani, 2004; Thuiller et al., 2010). Theoretically, differences between observed and simulated FD reveal the operation of non-neutral assembly rules (Petchey et al., 2007). The decrease (often defined as ‘trait convergence’) and increase (‘trait divergence’) of community FD compared with null models is supposedly due to the extent of habitat selection versus limiting similarity due to biotic interactions.

Knowledge is accumulating on which indices of FD should be preferred for testing community assembly (Mouchet et al., 2010). However, little is known about the performances of different null models in detecting realistic assembly rules. Moreover, whilst reading null-model methods applied in the literature (indeed including my work), one may have the impression of reading a book of magic spells. Randomization schemes under null models are definitely not easy, but the general impression that they are produced by waving some kind of magic wand should be avoided. The null-model definition determines which ecological mechanisms are allowed and which are excluded; only those that are excluded can generate deviation from the null model. However, the criteria for inclusion and exclusion of such mechanisms can be, in practice, very subtle (Zobel et al., 1993; Gotelli, 2000). Unfortunately, these criteria are not often discernible, leaving uncertainties about the real implications of results and when comparing results across studies. Here, using both field data and simple simulations, I show that these null models could largely fail to break down the functional relationship between species, leading to great uncertainty in assembly rule detection. In essence it is argued that these randomizations should be more thoroughly described and validated before being indiscriminately used to assess community assembly.

EXAMPLES

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. EXAMPLES
  5. CONCLUSIONS
  6. ACKNOWLEDGEMENTS
  7. REFERENCES
  8. BIOSKETCH
  9. Supporting Information

Whatever null model is applied, results concerning the importance of community assembly patterns are often obtained by comparing the deviation of observed community FD from randomly simulated FD. Given a specific community (‘sample’ or ‘plot’ including a list of coexisting species), FD indices can be computed according to different algorithms (Villéger et al., 2008; de Bello et al., 2010). It is often believed that FD values become comparable across communities after removing the effects due only by chance. Commonly, a simulated FD is obtained by randomizing community composition (say, e.g., 500 or 1000 times) across a set of sampled communities. How these random communities are assembled varies, and it is still widely debated (Mason et al., 2007; de Bello et al., 2009; Thompson et al., 2010), but they usually include all species found within a combination of samples in a region. Subsequently, different standardized indices are generally considered to infer community assembly patterns. The standardized effect size index (‘SES’; Gotelli & McCabe, 2002) is often calculated: SES = (observed FD – mean of simulated FD)/standard deviation of simulated FD). SES > 0 suggests the prevalence of trait divergence patterns and SES < 0 suggests trait convergence. The statistical significance of the detected assembly patterns is then considered, for example, as the number of times that the observed FD values are lower or higher than the simulated FD values. Although other approaches exist, this set of indices represents the emblematic toolbox applied to test community assembly.

Using different field data, and different randomization types, I consistently detected a strong linear dependence of indices such as SES, and the relative P-values, on observed FD indices. For example (Fig. 1), FD values were calculated for plant communities (10 m × 10 m quadrats) sampled from pastures along a broad altitudinal and climatic gradient in north-east Spain (with 12 communities for each of 5 different vegetation belts; see de Bello et al., 2009). Other data sets on plant communities were considered, all producing this strong linear dependence of SES on observed FD indices (not shown). Results shown here refer to the Rao index used for FD (Lepšet al., 2006; http://botanika.bf.jcu.cz/suspa/FunctDiv.php). Although different FD indices were calculated (Villéger et al., 2008), all showed the dependence of SES on observed FD values. The Rao index expresses the average dissimilarity between species pairs in a community and does not trivially correlate with the species richness (de Bello et al., 2010). This is one of the key indices recommended to test the relevance of community assembly (Mouchet et al., 2010). Here, the Gower distance (a standardized measure of trait dissimilarity appropriate for the Rao index; de Bello et al., 2010) was considered.

image

Figure 1. Relationship between observed functional diversity (FD) values and standardized effect size (SES) resulting from null-models (with SES > 0 suggesting trait divergence and SES < 0 suggesting trait convergence). Results are based on data from north-eastern Spanish pastures. Randomizations to create null communities were run across all vegetation belts or within given vegetation belts (results for each vegetation belt are shown in Appendix S1). Borders of significance (dotted lines) indicate cases when the observed FD values were higher, or lower, than the 5% of null communities. The Rao index used here to compute FD was 1/(1 – Rao), for which the index equates the number of species when all species are functionally different (de Bello et al., 2010).

Download figure to PowerPoint

The reason for this dependence of SES on observed values of FD is fairly simple: the mean of simulated FD values converges to a rather constant value for all communities. Most ecologists would agree that this is an intuitive and apparently trivial pattern. With randomization, the ‘average’ condition in a data set is produced so that extremes in given a data set are less likely to occur randomly under null models. In this way: (1) the mean of simulated FD values is often rather constant, and (2) the differences between observed FD values and mean simulated ones are very much driven by the observed FD (see also below and Fig. 2). This pattern, if considered carefully, should alert ecologists to the fact that these null models could ultimately lead to uncertain biological conclusions. For example, the mean simulated FD will vary considerably by changing the pool of samples among which randomizations are run, so that ultimately one plot considered as convergent in one data set will be considered divergent in another. The idea that null-model results will change with the pool of samples considered is definitely not novel (Weiher & Keddy, 1995), but seems to be often ignored.

image

Figure 2. Relationship between observed (‘obs’) functional diversity (FD) and standardized effect size (SES) and with the numerator of SES (i.e. observed – mean simulated FD). Simulated communities (n= 150) were assembled with different species richness (from 8 to 20). The denominator of the SES index (i.e. the standard deviation of simulated FD) varies with the number of species in a community (right panel).

Download figure to PowerPoint

For instance, two different types of randomizations were run for this example, i.e. using different sampling pools in the creation of null communities (Fig. 1). Firstly (1) across all 60 plots, i.e. including all communities from all vegetation belts, and then (2) only including plots within a given vegetation belt. The results of these two approaches (Fig. 1) are, interestingly, rather different. Randomizations only within a vegetation belt gave higher importance to trait divergence than randomizations across vegetation belts and failed to detect trait divergence. This is the result of lower mean simulated FD values within randomizations within a vegetation belt compared with randomizations across belts (P < 0.001). It indicates that reshuffling species that have, in reality, few chances to occur together (i.e. having rather different environmental requirements) will most likely lead to a decrease in the detection of limiting similarity patterns. In fact, species from different environments tend to have rather different trait values (de Bello et al., 2009). Therefore, shuffling them across all vegetation belts produces higher simulated FD than across a more environmentally homogeneous pool of species. When assessing the effect of biotic interactions randomization tests should be constrained within rather homogeneous environmental conditions (Weiher & Keddy, 1995; Mason et al., 2011). This is, unfortunately, still not the case for many existing studies. Overall, it should be evident that randomizations across species from different environments cannot disentangle habitat selection from limiting similarity (Mayfield & Levine, 2010). Random results could also be produced simply from the opposing effect of these two forces (Swenson & Enquist, 2009) while the bigger the spatial scale considered, the more likely that habitat effects will be detected (Weiher & Keddy, 1995; Willis et al., 2010).

To better assess the dependence of SES values on FD, and to show that it is not only the result of a given data set, I used simple simulated communities with varying FD values. From a species pool of 100 species, communities were assembled with a fixed or variable number of species (from 8 to 20) generating 150 communities. Each of the species had a hypothetical trait value ranging from 1 to 100. Species were selected randomly within the possible trait values. Results from null models on this community confirm that SES values vary systematically with the magnitude of the observed FD (Fig. 2), mostly because the mean of simulated FD values is rather constant for all communities (i.e. resulting in a linear relationship between the observed FD and the differences between observed and mean simulated FD). The relationship was stronger when the range of species richness was lower and/or fixed (not shown), probably because the standard deviation of simulated FD (included in the SES calculation; see above) also varies with the number of species in a plot (Fig. 2). With the Rao index, these patterns imply a less linear relationship between observed FD and SES at lower values of FD.

The approach of using simulated communities was then applied to illustrate the risk that plots considered as convergent in one data set could be potentially considered divergent in another (Fig. 3). Communities were assembled from a species pool of 100 or 50 species (i.e. one half of the species pool of 100 species used above). Null models were run to produce SES values for each community. Analysing both scenarios separately showed a rather equal number of samples departing from the expectations of random assembly, i.e. supposedly indicating either traits of divergence or convergence. As a consequence, this suggests that studies reporting fewer than 5% of samples with a given non-random assembly process should derive biological conclusions carefully. Most importantly, when mixed together, a considerable number of samples previously identified according to a given process (trait convergence, divergence or random assembly) were no longer recognized as being formed under that process. This result should alert again ecologists to the fact that detection of community assembly is strongly dependent on the pool of samples considered.

image

Figure 3. Relationship between observed functional diversity (FD) values and standardized effect size (SES) values using simulated communities (n= 150) assembled randomly from a species pool of 100 species and 50 species (white and black circles, respectively), with 8 to 20 species per community. Borders of significance reflect cases when the observed FD values were higher, or lower, than the 5% of null communities. Mixing scenarios together (right panel) led to classification of samples according to different assembly rules with respect to analyses on single scenarios (left and central panels).

Download figure to PowerPoint

Finally note that different randomization tests were used to calculate null models and the simulated FD values. The results presented refer to the quasiswap method of Miklos & Podani (2004), with 499 randomizations, which satisfies the requirements for equidistribution of species. The commsimulator function was used to generate null communities (vegan library; R Development Core Team, 2009). The quasiswap method randomizes species composition while keeping the number of species per plot in the randomized data fixed (see de Bello et al., 2009, for details). Other randomization approaches (i.e. without keeping the number of species or randomizing trait values across species instead of species composition) gave very similar results. Even using the approach by Shipley (2010), based on maximum entropy, deviations from the simulated FD were linearly dependent on the observed FD (using the function ‘maxent’ in the FD library in R).

CONCLUSIONS

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. EXAMPLES
  5. CONCLUSIONS
  6. ACKNOWLEDGEMENTS
  7. REFERENCES
  8. BIOSKETCH
  9. Supporting Information

Null models have a long tradition in ecology in testing patterns in community assembly. The debate over null models in Strong et al. (1984) nearly led to an abandonment of the method. Recently null models were resurrected and are increasingly combined with functional traits. Here I showed that the results from both simulated and real communities demonstrate, unequivocally, a strong linear dependence of SES on observed FD. This apparently trivial result implies, for instance, that comparing SES (or similar indices) across communities might correspond largely to using simple observed FD values (with the disadvantage of needing complex and time-consuming randomizations). Often, indices such as SES are preferred to observed FD values (and often demanded by referees), for example when assessing the relationship between species diversity or environmental variables with FD. This approach assumes independence between FD values and results under null models but should be carefully applied. The dependence of SES on observed FD, and therefore on the set of samples considered, might ultimately lead to the fact that one plot considered as being under divergence selection in one data set might be considered as being under convergence selection in another. Also, null models do not resolve problems connected with the fact that observed FD values could not always properly distinguish assembly rules (Mouchet et al., 2010). Therefore, if indices do not discriminate ecological processes well, running elegant null models will still show uncertain results. Validating null models, e.g. with the probability of Type I and II errors using simple model data with a known structure, as done by Zobel et al. (1993) and Gotelli (2000), remains the exception rather than the rule. In conclusion, the null model applied for FD should be more thoroughly described and validated before being considered as an essential tool for assessing community assembly. The results concerning potential underlying assembly rules must always be interpreted relative to the species pool and null model considered.

ACKNOWLEDGEMENTS

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. EXAMPLES
  5. CONCLUSIONS
  6. ACKNOWLEDGEMENTS
  7. REFERENCES
  8. BIOSKETCH
  9. Supporting Information

This study was supported by the grants APIC-RT-PICs-4876 from CNRS and AV0Z60050516 from the Czech Academy of Sciences. I thank Tamara Muenkemueller, Wilfried Thuiller, Sébastien Lavergne, Norman Mason, Meelis Partel and Martin Zobel for their invaluable suggestions and Rebecca Brown for the English editing and the ‘magic’ touch. One anonymous reviewer and Evan Weiher helped considerably to raise the interest of this work.

REFERENCES

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. EXAMPLES
  5. CONCLUSIONS
  6. ACKNOWLEDGEMENTS
  7. REFERENCES
  8. BIOSKETCH
  9. Supporting Information

BIOSKETCH

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. EXAMPLES
  5. CONCLUSIONS
  6. ACKNOWLEDGEMENTS
  7. REFERENCES
  8. BIOSKETCH
  9. Supporting Information

Francesco de Bello is a researcher at the Institute of Botany, Czech Academy of Sciences. Using meadows and alpine vegetation as a study framework, he assesses the role of functional diversity in the interface between community assembly and ecosystem service delivery.

Editor: Tim Blackburn

Supporting Information

  1. Top of page
  2. ABSTRACT
  3. INTRODUCTION
  4. EXAMPLES
  5. CONCLUSIONS
  6. ACKNOWLEDGEMENTS
  7. REFERENCES
  8. BIOSKETCH
  9. Supporting Information

Appendix S1 Detailed relationship between observed functional diversity (FD) and standardized effect size (SES) for the five vegetation belts considered. Randomizations under null-models were run within each vegetation belt (as for the filled circles in Figure 1). Vegetation belts follow a gradient from drier at lower altitudes to more humid at higher altitudes (see de Bello et al. 2009 for more details).

FilenameFormatSizeDescription
GEB_682_sm_AppendixS1.pdf63KSupporting info item

Please note: Wiley Blackwell is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.