Ecology Letters (2011) 14: 19–28
A recent increase in studies of β diversity has yielded a confusing array of concepts, measures and methods. Here, we provide a roadmap of the most widely used and ecologically relevant approaches for analysis through a series of mission statements. We distinguish two types of β diversity: directional turnover along a gradient vs. non-directional variation. Different measures emphasize different properties of ecological data. Such properties include the degree of emphasis on presence/absence vs. relative abundance information and the inclusion vs. exclusion of joint absences. Judicious use of multiple measures in concert can uncover the underlying nature of patterns in β diversity for a given dataset. A case study of Indonesian coral assemblages shows the utility of a multi-faceted approach. We advocate careful consideration of relevant questions, matched by appropriate analyses. The rigorous application of null models will also help to reveal potential processes driving observed patterns in β diversity.
β diversity, generally defined as variation in the identities of species among sites, provides a direct link between biodiversity at local scales (α diversity) and the broader regional species pool (γ diversity) (Whittaker 1960, 1972). The past decade has witnessed an especially marked increase in studies under the name of β diversity (Fig. 1). Indeed, the study of β diversity is genuinely at the heart of community ecology – what makes assemblages of species more or less similar to one another at different places and times (Vellend 2010)?
Many different measures of β diversity have been introduced, but there is no overall consensus about which ones are most appropriate for addressing particular ecological questions (Vellend 2001; Koleff et al. 2003; Jost 2007; Jurasinski et al. 2009; Tuomisto 2010a,b). Debates persist regarding whether the measures used for partitioning γ diversity in terms of α and β components should be additive or multiplicative (Lande 1996; Crist & Veech 2006; Jost 2007). Many ecologists also now use β diversity to describe measures that incorporate additional information, such as the relative abundances of species (Legendre et al. 2005), or the taxonomic, phylogenetic or functional relationships among species (Izsak & Price 2001; Clarke et al. 2006; Graham & Fine 2008; Swenson et al. 2010). The intrinsic relationship between β and α diversity, including dependence on scale and sample size (Loreau 2000), has also prompted a variety of proposed corrections to classical β diversity measures (e.g. Harrison et al. 1992; Chao et al. 2005; Chase 2007; Vellend et al. 2007).
Added to the perplexing array of potential measures (e.g. Tuomisto 2010a) are a variety of statistical approaches for analysing patterns in β diversity (Legendre et al. 2005; Anderson et al. 2006; Tuomisto & Ruokolainen 2006; Qian & Ricklefs 2007; Legendre 2008). There are strongly divergent opinions regarding these methods (Legendre et al. 2008; Tuomisto & Ruokolainen 2008), and how statistical dependence among α, β and γ influences tests of hypotheses (Baselga 2010; Jost 2010; Veech & Crist 2010a,b). The use of different measures or analytical approaches on a single set of data can naturally result in quite different outcomes and interpretations (e.g. Smith & Lundholm 2010). In addition, most measures of β diversity are applied without incorporating statistical null models, even though they might be appropriate, given known interrelationships between α, β and γ diversity.
Tuomisto (2010a,b) has provided an extensive review of existing measures of β diversity and their mathematical interrelationships. Moreover, in an effort to diminish growing confusion, Tuomisto (2010a,b) proposed that ‘β diversity’ be used exclusively to refer to one specific measure (called ‘true β diversity’, denoted by βMd therein). However, this belies the fact that Whittaker’s original concept of β diversity (indeed, as nicely summarized by Tuomisto 2010a) was much more general; several different measures of β diversity were proposed in Whittaker’s (1960, 1972) seminal work. Some plurality of concept is evident in the framework of Jurasinski et al. (2009), who identified ‘inventory’, ‘differentiation’ and ‘proportional’β diversity. However, this places certain measures in different categories (such as Whittaker’s βW and the Jaccard resemblance measure), even though they are, in practice, intimately related.
The purpose of this article is to provide a practical and hypothesis-driven roadmap for ecologists in the analysis of β diversity. Multivariate species data are complex and hold much information. We consider that ecologists need a framework that both simplifies the enormous list of existing methods (by pointing out relevant congruencies that will occur in practice), while nevertheless maximizing the utility of having more than one concept and measure for β diversity. First, we distinguish two essential concepts: turnover (directional) and variation (non-directional). Second, we outline a series of core ecological mission statements regarding β diversity and connect these directly with appropriate analyses. Third, we describe the key ecologically relevant properties of commonly used resemblance measures, indicating also direct links between these and classical measures of β diversity. Fourth, we provide a case study (the response of coral assemblages in Indonesia to an El Niño weather event) which illustrates these properties and exemplifies the strategy of using a suite of measures in concert to yield an informative holistic analysis of β diversity for community data.
Turnover vs. variation
We distinguish two types of β diversity: turnover and variation (Fig. 2; see also Vellend 2001). Both have clear historical roots in Whittaker’s (1960, 1972) original conceptualization. The first is the notion of β diversity as turnover (Fig. 2a). The essential idea here is to measure the change in community structure from one sampling unit to another along a spatial, temporal or environmental gradient. By ‘change in community structure’, we mean a change in the identity, relative abundance, biomass and/or cover of individual species. Questions associated with turnover include: How many new species are encountered along a gradient and how many that were initially present are now lost? What proportion of the species encountered is not shared when we move from one unit to the next along this gradient? Turnover can be expressed as a rate, as in a distance–decay plot (e.g. Nekola & White 1999; Qian & Ricklefs 2007). Turnover, by its very nature, requires one to define a specific gradient of interest with directionality. For example, the rate of turnover in an east–west direction might differ from that in a north–south direction (e.g. Harrison et al. 1992).
The second type of β diversity is the notion of variation in community structure among a set of sample units (Fig. 2b) within a given spatial or temporal extent, or within a given category of a factor (such as a habitat type or experimental treatment). This is captured by Whittaker’s original measures of β diversity as variation in the identities of species among units (see βW below) or the mean Jaccard dissimilarity among communities (see below). Here, the essential questions are: Do we see the same species over and over again among different units? By how much does the number of species in the region exceed the average number of species per sampling unit? What is the expected proportion of unshared species among all sampling units? Variation is measured among all possible pairs of units, without reference to any particular gradient or direction, and has a direct correspondence with multivariate dispersion or variance in community structure (Legendre et al. 2005; Anderson et al. 2006).
Measures of β diversity
The two most commonly used classes of measures of β diversity used in studies of either turnover or variation are: (1) the classical metrics, calculated directly from measures of γ (regional) and α (local) diversity and (2) multivariate measures, based on pairwise resemblances (similarity, dissimilarity or distance) among sample units.
Let αi be the number of species (richness) in sample unit i, let be the average number of species per unit obtained from a sample of N units within a larger area or region, and let γ be the total number of species for this region. One of the original measures described as β diversity by Whittaker (1960) was . It focuses on species’ identities alone and is the number of times by which the richness in a region is greater than the average richness in the smaller-scale units. It thus provides a multiplicative model which, being additive on a log scale (Jost 2007), can also be used to calculate additive partitions of β diversity at multiple scales (Crist et al. 2003).
An additive rather than multiplicative model is given by (Lande 1996; Crist & Veech 2006). βAdd, like βW, can be partitioned across multiple scales (Veech & Crist 2009). βAdd is in the same units as and γ, so is easy to communicate in applied contexts (Gering et al. 2003) and can be compared across multiple studies, when and βAdd are expressed as proportions of γ (Veech et al. 2003; Tuomisto 2010a).
More recently, Jost (2007) has defined a measure that also includes relative abundance information: βShannon = Hγ /Hα, where is an exponentiated Shannon–Wiener index (i.e. effective diversity) for the γ-level sample unit (obtained by pooling abundances for each species across all α-level units) and is the average of the exponentiated indices calculated for each α-level sample unit. βShannon shares the property with βW of being multiplicative, and thus additive on a log scale, (MacArthur et al. 1966). It can also be partitioned for a hierarchy of spatial scales (Ricotta 2005; Jost 2007).
We first define a sampled community as a row vector y of length p containing values for each of p species within a given sample unit (a plot, core, quadrat, transect, tow, etc.). The values in the vector may be presence/absence data, counts of species’ abundances or some other quantitative or ordinal values (biomass, cover, etc.). A set of N such vectors (sampled communities) generates a matrix Y, with N rows and p columns. We shall use Δy (or dij) to denote a change in community structure from one unit to another , as would be measured by a given pairwise dissimilarity measure [Jaccard (dJ), Bray–Curtis (dBC), etc.]. Multivariate measures of β diversity begin from a matrix D containing all pairwise dissimilarities (dij or Δy) among the sample units. For N units, there will be m = N(N − 1)/2 pairwise dissimilarity values.
β diversity as turnover can be estimated as the rate of change in community structure along a given gradient x, which we shall denote as ∂y/∂x. For example, the similarity between pairs of samples [denoted here as (1 − Δy) for measures like Jaccard, where 0 ≤ Δy ≤ 1] is expected to decrease with increasing geographical distance. Given a series of sample units along a spatial gradient (as in Fig. 2a), we can fit, for example, an exponential decay model as: (1 − Δyk) = exp(μ + βΔxk + ɛk), where (1 − Δyk) is the similarity between the kth pair of sample units and Δxk is the geographic distance (the difference in latitude, say) between the kth pair, for all unique pairs . This is visualized by a distance–decay plot of (1 − Δyk) vs. Δx. The estimated slope, in absolute value, is a direct measure (on a log scale) of turnover (∂y/∂x; Fig. 2a; Nekola & White 1999; Vellend 2001; Qian et al. 2005; Qian & Ricklefs 2007): the steeper the slope (larger negative values in the exponential decay), the more rapid the turnover. Note that Δx might also denote environmental change along a gradient, such as altitude, soil moisture, temperature or depth; it need not necessarily be a spatial distance.
β diversity as variation in community structure among N sample units shall be denoted by . This idea is captured by the notion of the dispersion of sample units in multivariate space (Anderson et al. 2006) and can be measured directly using the sum of squared interpoint dissimilarities: (e.g. Legendre & Anderson 1999; Anderson 2001; McArdle & Anderson 2001), the average interpoint dissimilarities (e.g. Whittaker 1960, 1972; Vellend et al. 2007), or the average distance-to-centroid of the N points in the space defined by the resemblance measure (here referred to as ; see Anderson 2006; Anderson et al. 2006 for further details). For Euclidean distances and one species (p = 1), is the classical unbiased estimate of the univariate sample variance. Legendre et al. (2005) suggested SS(Y), the sum of the individual sum of squares across all species, as a measure of β diversity. For Euclidean distances, ; that is, , the sum of the estimated variances across all species, where is the estimated variance–covariance matrix of dimension p × p. More generally, for non-Euclidean measures (Jaccard, Sørensen, Bray–Curtis, etc.), , where G is Gower’s centred matrix obtained directly from matrix D (McArdle & Anderson 2001). Taking the square root yields a measure of variation expressed in the same units as the chosen resemblance measure.
Mission statements and associated analyses
We articulate a series of mission statements regarding the analysis of β diversity. These are numbered and discussed in two separate groups by reference to the two conceptual types of β diversity: turnover and variation (Fig. 2). Figures 3 and 4 show schematic representations of the relevant sampling designs and associated analyses in each case.
T1. Measure the turnover in community structure between two communities. The focus here is on simply estimating Δy between two communities.
T2. Measure the turnover in community structure between two communities and model this along an environmental gradient or other factor. For example, if we estimate the turnover in community structure between serpentine and non-serpentine soils (Δy), does this change with latitude (x)? Interest lies in modelling Δy vs. x directly, which can be done by fitting a linear or nonlinear model. Note that each Δy value is obtained independently in this scenario: with one (or more, if one has independent replicates of such pairs) for each value of x. Note these are not all possible pairwise values in a distance matrix.
T3. Explore and model the relationship between pairwise dissimilarities in community structure and pairwise differences in space, time or environment. Here, interest lies in modelling all pairwise Δy values as a function of Δx. For example, are differences in insect community structure related to differences in precipitation? The Mantel test (Mantel 1967) may be used to test statistical significance of such relationships. This approach ‘unwinds’ the dissimilarity matrix into a single vector of values, but permutations are done correctly by treating the N sample units (not the m pairs) as exchangeable under the null hypothesis of no relationship between Δy and Δx (Legendre & Legendre 1998). These m values are not independent of one another, so one cannot use classical regression methods (partitioning and associated tests) directly on the Δy values (Manly 2007). Note that the Mantel test, which may be useful for analysing a single gradient, is not recommended for investigating more than one gradient at a time (such as spatial gradients in two dimensions), due to the omni-directional nature of dissimilarities and lack of power (Legendre & Fortin 2010).
T4. Estimate the rate of turnover in community structure along a spatial, temporal or environmental gradient. Interest lies in modelling. Interest lies in modelling Δy vs. Δx, which is essentially the same thing as T3, but specifically now with the goal of estimating the rate of turnover (∂y/∂x), as in a distance–decay model (Qian & Ricklefs 2007). In most cases, similarity [(1 − Δy), where 0 ≤ Δy ≤ 1] is modelled as a linear or nonlinear function of Δx (usually an exponential decay); for simplicity we shall refer to the estimated slope as ∂y/∂x. One might also consider the relative strength of the relationship (r2), which is not necessarily monotonic on the estimated slope. Thus, we recommend that both the r2 and slope values be reported in comparative studies of distance–decay models. A potential issue arises when there is no evidence against the null hypothesis of the slope being zero (tested using the Mantel test, as in T3 above). It is unlikely that this corresponds to there being zero β diversity. Rather, there may well be variation that is simply unrelated to the measured gradient.
T5. Compare turnover along a specific gradient for two different sets of species or taxonomic groups. For example, is the rate of species turnover along a gradient in soil type different for native species than it is for exotic species? Here, one examines two rates along a common gradient. Interest lies in comparing (say) ∂ynative/∂x with ∂yexotic/∂x. This can be done visually by looking at plots of the models, but note that lack of independence among the Δy values precludes the use of a classical ancova. A test of the null hypothesis of no difference in the slopes may be done, however, by randomly re-allocating the species into the groups (native vs. exotic), but leaving x fixed, to generate a null distribution for the difference in slopes. The concept of halving distance (Soininen et al. 2007) might also be considered here.
T6. Explore and model the rate of turnover along a gradient across different levels of another factor or along another gradient. For example, is the rate of turnover in marine benthic invertebrates along a depth gradient different for different latitudes or through time? Here, the response variable is turnover (∂y/∂x) along a chosen gradient, and one may model this in response to a complex experimental design (e.g. with several factors and their interactions) or sets of other continuous predictor variables (e.g. temperature, salinity, nutrients, etc.). There are no limitations on the types of models that could be used here (linear or nonlinear, classical or nonparametric), provided independent (and preferably replicated) values of ∂y/∂x are estimated at each point within the sampling design. For example, separate independent estimates of turnover along a depth gradient ∂y/∂xdepth may be modelled as a function of latitude, substratum type and/or nutrients.
V1. Measure the variation in community structure among a set of samples. Here, the focus is simply on estimating variation, which can be achieved by calculating one or more of the classical (βW, βAdd) or multivariate measures discussed above (, or , on the basis of a chosen resemblance measure).
V2. Explore the relationship between community structure and some factor(s) or variable(s) of interest. Here, interest lies in visualizing the potential relationship of Y vs. x (a single variable) or X (several continuous variables or indicators of factors). The factor(s) or variable(s) of interest may be temporal, spatial or environmental from an observational survey, or they may be experimentally manipulated treatments. Unconstrained ordination such as principal coordinates analysis or non-metric multi-dimensional scaling (MDS) can be used to examine patterns in a multivariate data cloud on the basis of a chosen resemblance measure. Potential relationships with X are explored by superimposing labels on points (for groups), bubbles (for quantitative variables) or vectors (showing multiple linear relationships with axes). This is called indirect gradient analysis (e.g. ter Braak 1987) and covers a plethora of methods. Importantly, differences in the relative sizes of multivariate dispersions for different groups can be visualized on unconstrained ordination plots (e.g. Anderson 2006; Chase 2007, 2010).
V3. Partition the variation in community structure in response to some quantitative variables or factors (spatial, temporal, environmental, experimental). This is achieved by modelling Y in terms of x (or X). The total variation in Y is , but interest lies here specifically in determining how much of this variation is explained by functions of other variables (and their overlap if they are non-independent). For example, how much of the spatial variation in communities of herbs is explained by the factors of fire frequency, fencing to prevent grazing, and their interaction? If the partitioning involves a fixed factor (e.g. disturbed vs. undisturbed treatments in an experiment), then the component of variation for that factor is interpreted as an effect size. can be partitioned directly according to multi-factor experimental or hierarchical sampling designs (using permanova; Anderson 2001; Anderson et al. 2008) or continuous environmental or spatial gradients [using redundancy analysis (RDA), canonical correspondence analysis (CCA) or distance-based redundancy analysis (dbRDA); Borcard et al. 1992; Legendre & Anderson 1999; McArdle & Anderson 2001; Anderson et al. 2008]. dbRDA on Euclidean distances yields a classical RDA, as tr(G) = SS(Y) in that case, while dbRDA on chi-squared distances yields results very close to CCA (ter Braak 1986). Partitioning in the space of the chi-squared, Hellinger or chord measures can also be obtained by RDA on a simple transformation of the values in matrix Y (Legendre & Gallagher 2001). Advantages to using RDA, thus working with SS(Y), on either raw or transformed data include the direct interpretability of ordination axes in terms of the original variables and the computational speed of partitioning a p × p matrix of sums of squares and cross products (SSCP) rather than the N × N matrix (G) if . Although the direct link to original Y variables is broken once a dissimilarity matrix D has been formed, dbRDA allows much more flexibility in the choice of resemblance measure (Jaccard, Sørensen, Bray–Curtis, etc.), and yields a faster core algorithm when p > N. Also, dbRDA does not require calculation of principal coordinates or corrections for negative eigenvalues, but directly partitions matrix G (see Fig. 4; McArdle & Anderson 2001; Anderson et al. 2008).
V4. Compare variation in community structure among several levels of a factor (categorical) or along a gradient (continuous). For example, does the degree of variation in species’ identities change with depth? If one has n replicate sample units within each of g levels of a factor (N = g × n) then we can formally test the null hypothesis of homogeneity of multivariate dispersions (Anderson 2006; Anderson et al. 2006, 2008). For example, we can compare for shallow vs. deep sites. A statistical comparison of values among groups could also be performed using a separate-sample bootstrap, as described by Manly (2007) for univariate data. Furthermore, if groups occur along a gradient (e.g. in a series of depth strata), then we may model values of , or vs. depth (x). More complex designs are also possible where multiple values of have been obtained along more than one gradient or factor.
V5. Partition the variation in community structure according to a series of additive hierarchical spatial scales. When there is more than one spatial scale of interest, a relevant sampling design would have hierarchical random factors at a number of scales within a region, such as locations, sites within locations and replicates within sites. Here, one would calculate (for example): . This yields additive components of variation. Estimators for these can be calculated from mean squares and tested using permutation methods as pseudo multivariate variance components (Anderson et al. 2005), direct analogues to the unbiased univariate anova estimators (Searle et al. 1992). Although partitioning to obtain sums of squares for each factor is calculated from SS(Y) (in RDA) or tr(G) (in dbRDA or permanova), the actual components of variation (, which take into account degrees of freedom), are required for making valid comparisons. For analyses of one variable using Euclidean distances, these are the classical univariate variance components (Searle et al. 1992). Notably, unbiased estimators for these components are derived from expectations of mean squares, which will be specific not just to the individual component being estimated, but also to the particular model in which they are found; they will depend especially on the nature of any nested structures and whether factors included in the model are to be treated as fixed or random (Searle et al. 1992; Anderson et al. 2008). Partitioning might also be done as γ = α +βreplicates + βsites + βlocations (Crist et al. 2003; Crist & Veech 2006). Note that α is a measure of diversity within a sample, which is not discussed explicitly here in the form of a variance component, but see Pélissier & Couteron (2007). Thus, γ is not the same as because γ includes α. Similarly, a multiplicative hierarchical partition is: γ = α × βreplicates ×βsites × βlocations. Either an additive or multiplicative partitioning of these classical measures can be calculated, with statistical tests of null hypotheses (Veech & Crist 2009). Finally, for modelling scales of variation along a continuum, rather than hierarchically, one may consider doing an analysis using principal coordinate analysis of neighbour matrices (PCNM) (Dray et al. 2006; Legendre et al. 2009).
V6. Compare individual components of variation in community structure from a partitioning across some other factor or variable of interest. For example, how does the partitioning of change when we look at disturbed vs. undisturbed environments? Specifically, we may wish to test for a difference in the sizes of individual components; is β diversity at the scale of sites, , significantly larger (or smaller) in disturbed than in undisturbed environments? A direct multivariate analogue to the univariate two-tailed F-ratios (Underwood 1991) could be used to compare such components, but with P-values obtained using bootstrapping (Davison & Hinkley 1997; Manly 2007). Components could also be compared across multiple levels of other factors, with formal tests for differences obtained using bootstrapping, as has been done to compare univariate variance components (Terlizzi et al. 2005).
V7. Compare components of variation in community structure for different sets of species or taxonomic groups. An example here might be: is β diversity for annelids at the scale of sites, , larger (or smaller) than that for molluscs? This is a bit like comparing components across levels of another factor (V6). Components of variation for different groups of organisms can be calculated and compared directly (Anderson et al. 2005). Care is needed, however, in designing formal tests; if components for different groups are calculated from the same dataset they may not be independent.
Key properties of pairwise resemblance measures for ecological interpretation
Pairwise dissimilarities form the basis of multivariate analyses of β diversity. Different measures have different properties. They emphasize different aspects of community data and therefore can yield very different results. Rather than being a handicap, we advocate that this plurality be used as an advantage. Comparing and contrasting the results obtained from judicious use of a suite of directly interpretable measures can yield important ecological insights into the actual nature of patterns in β diversity. Analyses performed using different measures correspond to different underlying ecological hypotheses.
We provide here a key to the essential properties associated with the most commonly used measures in the ecological analysis of community data (Table 1; Fig. 5). See Legendre & Legendre (1998), Koleff et al. (2003) and Tuomisto (2010a,b) for more.
|Exclude joint absences||Jaccard||Bray–Curtis|
|Include joint absences||Simple matching||Euclidean|
|Baroni-Urbani and Buser||Manhattan|
Presence/absence vs. relative abundance information
The first important conceptual distinction (Table 1; Fig. 5) is between measures that use identities of species only (presence/absence data), vs. those that include abundance (or relative abundance or biomass or other) information as well. The classical measures of βW and βAdd do not include relative abundance information, but βShannon does. This distinction is fundamental and dramatically different results can be obtained when relative abundance information is included. There may be good reasons to focus on identities of species alone for some applications, as species (rather than individuals) are often the units of interest in conservation and biodiversity studies.
Abundance information is, however, an important aspect of community structure and there is no reason not to include it in analyses of variation in communities. Indeed, comparing analyses of β diversity that emphasize species identities alone (with a strong role for rare species) to those that emphasize differences in relative abundances (where common and numerically dominant species play a strong role) can yield useful insights into the specific nature of community-level changes (Olsgard et al. 1997; Anderson et al. 2006).
Inclusion vs. exclusion of joint-absence information
The next important distinction is between measures that exclude joint-absence information and those that do not (Table 1; Fig. 5). Measures based on presence/absence data generally use the following quantities for their calculation: a is the number of species shared between the two units, b is the number of species occurring in unit i but not unit j; c is the number of species occurring in unit j but not unit i; e is the number of species absent from both units. Neither Jaccard: dJ = [1 − a/(a + b + c)], nor Sørensen: dS = [1 − 2a/(2a + b + c)] use the quantity e in their calculation. dJ has a direct interpretation as the proportion of unshared species observed in the two sample units. dS (equivalent to Bray–Curtis on presence/absence data) is monotonic on dJ, so these two will yield highly similar results.
For many applications, the exclusion of joint absences is appropriate: two sites are not considered more similar if they both lack certain species. In analyses of communities along environmental gradients, such as altitude, high and low-altitude communities are not considered more similar because they both lack species from middle altitudes. Importantly, there is an intimate link between dJ (or dS) and Whittaker’s βW. Specifically, in the case of N = 2, βW = 1 + dS = 2/(2 − dJ) (Tuomisto 2010a). Thus, βW is classified here as a measure that uses identities of species only and excludes joint absences. It is expected to give results similar to those obtained from multivariate analyses based on either dJ or dS (Fig. 5).
In some cases, however, joint absences are informative. They can be relevant, for example, when hypotheses relate to the disappearance of species, like in studies examining the effects of environmental impact, predation or biological invasions. Similarly, at broader scales, species absences from suitable habitats may occur due to stochastic extinction or dispersal limitation. A measure based on presence/absence data that includes joint absences (quantity e) is the simple matching coefficient: dSM = 1 − (a + e)/(a + b + c + e).
Interestingly, βAdd also includes joint-absence information. βAdd can be defined as the average number of unseen species per α-level sample unit that are present in the larger γ-level unit. Although βAdd = ½(b + c) when N = 2, the inclusion of joint-absence information (e) in βAdd is explicit when one considers the contribution (β*) of any two sample units towards βAdd when N > 2, namely, β* = e + ½ (b + c). In addition, βAdd when N = 2 is also a function of Euclidean distance (dEuc) when calculated on presence/absence data, namely βAdd = ½(dEuc)2 (Tuomisto 2010a). Thus, results obtained using βAdd are expected to give similar results to multivariate analyses based on dSM.
There are many ecological dissimilarity measures that include relative abundance information (Legendre & Legendre 1998; Chao et al. 2005; Anderson et al. 2006; Clarke et al. 2006), and most of these exclude joint absences (Table 1; Fig. 5). Measures in this class include Bray–Curtis, one of the most popular abundance-based metrics (Bray & Curtis 1957; Clarke et al. 2006), along with modified Gower (Anderson et al. 2006), chi squared (having a kinship with correspondence analysis, ter Braak 1985; Legendre & Legendre 1998) and Hellinger (Rao 1995; Legendre & Gallagher 2001).
Joint-absence information may be relevant to include, however, if hypotheses focus on phenomena that can cause changes in total (rather than proportional) abundances, biomass or cover, such as in studies of productivity, upwelling, disturbance or predation. Measures in this category include Euclidean distance and the Manhattan measure. When analysing counts of abundances (which are often overdispersed), such distances are usually calculated on log( y + 1)-transformed data. Figure 5 shows schematically how changes in the choice of measure, as well as the transformation used, will alter the relative importance of composition, relative or raw abundance information in terms of their contribution towards the results obtained, as a continuum.
Probabilistic measures under a null model: accounting for differences in α
Pairwise measures of dissimilarity, such as dJ or dS, will depend to some extent on the number of species in the sample units. When there is a large difference in richness between two samples, the corresponding dissimilarity should automatically increase, as the potential for overlap (quantity a) is reduced (Koleff et al. 2003). This issue has led to various attempts to remove the effect of differences in α from measures of β diversity (Lennon et al. 2001).
One way to remove effects of α on β is to use a null-modelling approach. For example, Raup & Crick (1979) proposed a probabilistic resemblance measure, dRC, which is interpretable as the probability that two sample units share fewer species than expected for samples drawn randomly from the species pool, given their existing differences in richness (see also Chase 2007, 2010; Vellend et al. 2007). More specifically, let α1 and α2 be the respective number of species in each of two sample units. One generates a null distribution of dJ from repeated random draws of α1 and α2 species from the species pool (γ), with the probability of drawing each species being its proportional occurrence in all sample units. dRC is the proportion of pairs of communities generated under the null model that share the same number or more species in common than the original sample units. Thus, dRC measures β diversity while conditioning on α.
Although dRC still depends on γ (a topic for further research), analyses based on dRC allow one to identify changes in β diversity (increases in variation as measured by dJ) that are driven by changes in α alone (Vellend et al. 2007). By teasing out the α-driven component of β diversity for presence/absence data, the probabilistic null model implemented by dRC yields a very useful tool that, especially when coupled with well-designed experiments, can help to unravel the underlying mechanisms generating variation in ecological communities (Chase 2010).
Revealing the nature of changes in β diversity using different measures: a case study
A full set of analyses for all of the mission statements is beyond the scope of this article, hence we focus here on an illustration of how multiple analyses of a given dataset can yield deeper insights than any one analysis. Rather than choosing a single measure of β diversity, we recognize that communities have a variety of ecological properties of interest, and we advocate using a suite of measures, each driven by specific hypotheses. This approach can directly reveal the nature of changes in community structure. This is not to suggest that all available measures should always be used. Rather it is to compare results obtained using a subset of contrasting measures that focus on different properties, so meaningful interpretations can follow.
We illustrate this approach by analysing observational data where the mission is to compare variation among several groups of samples (V4a) in response to a disturbance. This is one of the most common and general mission statements in ecological studies and this case study purposefully exemplifies strong contrasts in results with choice of resemblance measure. The percentage cover of 75 species of coral was measured along each of 10 transects on reefs in the Tikus Islands, Indonesia, in each of several years from 1981 to 1988 (Warwick et al. 1990; data are provided in Appendix S1 in Supporting Information). In 1982, there was a dramatic bleaching of the corals (disturbance), triggered by El Niño. We examined community variation for n = 10 transects in three years: 1981, 1983 and 1985. Results differed dramatically for different measures (Fig. 6; Table 2), but several classes of outcomes were apparent. β diversity (as variation) in communities of coral species following the El Niño (1983) significantly increased (using βW, Jaccard or Bray–Curtis), or decreased (using βAdd, simple matching or Euclidean) or showed no significant change (using modified Gower log base 5, or Raup–Crick). There was a clear dichotomy between the multivariate results obtained when joint absences were included vs. excluded. This was paralleled directly by the classical metrics: βAdd reflected results obtained by including joint absences, while βW reflected results obtained by excluding joint absences.
|Diversity metrics||1981||1983||1985||F||P||Pairwise test results|
|18.00||3.60||9.50||17.71||0.0001||81 > (85, 83)|
|βW||3.00||5.83||3.47||6.06||0.0011||83 > (85, 81)|
|βAdd||36.0||17.40||23.50||30.38||0.0024||81 > (85, 83)|
|H||12.92||3.33||7.63||15.78||0.0003||81 > (85, 83)|
|βShannon||2.50||5.35||1.79||7.37||0.0001||83 > (81, 85)|
|Multivariate measures||1981||1983||1985||F||P||Pairwise test results|
|Euclidean, proportions||27.82||62.51||28.27||14.97||0.0002||83 > (81, 85)|
|Chi squared||1.56||4.70||1.50||71.17||0.0001||83 > (81, 85)|
|Jaccard||47.91||63.82||44.97||22.60||0.0001||83 > (81, 85)|
|Sørensen||38.11||61.78||35.66||28.27||0.0001||83 > (81, 85)|
|Bray–Curtis||47.50||62.41||39.79||16.38||0.0002||83 > (81, 85)|
|Hellinger||0.72||0.90||0.62||23.62||0.0001||83 > (81, 85)|
|Modified Gower (log10)||0.76||0.80||0.72||1.43||0.3089||n.s.|
|Bray–Curtis, adjusted||46.83||52.74||38.95||6.92||0.0089||83 > (81, 85)|
|Modified Gower (log5)||0.89||0.87||0.84||0.37||0.7275||n.s.|
|Modified Gower (log2)||1.43||1.16||1.36||3.81||0.0560||n.s.|
|Gower excluding 0-0||1.15||1.08||0.30||0.89||0.5329||n.s.|
|Euc log(x + 1)||4.75||1.71||3.22||38.24||0.0001||81 > 85 > 83|
|Binomial deviance||45.46||5.32||20.59||25.98||0.0001||81 > 85 > 83|
|Simple matching||18.33||5.60||9.54||17.71||0.0001||81 > (85, 83)|
|Binomial deviance (scaled)||9.85||2.91||5.15||18.60||0.0001||81 > (85, 83)|
|Euclidean||20.99||3.26||14.16||16.06||0.0003||(81, 85) > 83|
|Canberra metric||15.13||4.21||7.90||19.89||0.0001||81 > (85, 83)|
|Manhattan, log(x + 1)||22.55||4.23||11.10||27.97||0.0001||81 > 85 > 83|
|Manhattan||72.94||7.73||37.46||27.57||0.0001||81 > 85 > 83|
These classes of outcomes can be visualized in a second-stage MDS plot based on Spearman rank correlations (Somerfield & Clarke 1995) among the dissimilarity matrices (Fig. 6e). The strongest contrast in results for these data is in the exclusion vs. the inclusion of joint absences, exemplified by dJ on the left and dSM on the right (along MDS axis 1), which differ only in this respect. Exclusion of joint absences led to significantly greater observed variability post-disturbance. Measures emphasizing proportional composition (Euclidean on proportions, or chi squared, which tends to heavily emphasize rare species, Legendre & Gallagher 2001) are shown even further to the left. The second MDS axis shows a gradient of measures emphasizing species’ identities or composition (towards the bottom) vs. abundances (towards the top) (Fig. 6e; see also Fig. 5).
One of the reasons that including or excluding joint absences yielded such differing results is that the bleaching event dramatically reduced the total cover and the average richness () of corals (Table 2). The loss of species in 1983 led to sparse samples (many zeros) and fewer matched species among samples. Such sparseness tends to inflate measures that exclude joint absences (Clarke et al. 2006). This is the major reason why βW and dispersions based on dJ or dBC (Fig. 6a,b) increased, while βAdd and dispersions based on dSM or dEuc (Fig. 6c) decreased in 1983 compared to 1981 (Table 2). Inclusion of joint absences (as in βAdd or dSM) can provide greater resolution for measuring changes in communities where many species are either rare or narrowly distributed.
Differences in β diversity are often accompanied by changes in richness (α). An analysis based on Raup–Crick, which explicitly takes into account differences in richness by conditioning on a null model, yielded no statistically significant differences in multivariate dispersion among the three groups (Table 2). This indicated that the effects of the El Niño on β diversity as measured by dJ (or βW) were confined to effects on richness. In other words, the increase in multivariate dispersion based on dJ after the El Niño appears to have occurred because of a non-selective reduction in richness, consistent with the expected increase in dJ that accompanies reduced numbers of species under the null model. In general, a null model is needed when testing hypotheses about βAdd or βW, because observed α influences the expression of β diversity differently with these metrics (Veech & Crist 2010a). In addition, although the coral data did not show a particularly strong effect of abundance information on results [viz. the relative proximity in Fig. 6e of dSM and dEuc on log( y + 1)-transformed data], this is not always the case, and interpretations of results for a given dataset must allow for a variety of classes of outcomes, all of which can inform the nature of changes in β diversity.
When discussing β diversity, clarity is needed regarding the type of β diversity of interest: either turnover by reference to a specific gradient or variation (Fig. 2). The sampling design and ensuing analysis should reflect this (Figs 3 and 4). In addition, while a variety of measures may be used advantageously in concert, results must be interpreted in accordance with the ecological properties emphasized by those measures.
Recently, β diversity was described as a ‘level 3 abstraction’, where one examines ‘variation in variation in raw data’ (Tuomisto & Ruokolainen 2006). Analyses of quantities such as or models of vs. xℓ, as in missions V4(a) or V4(b) above, are indeed analyses of how variation, itself, is changing along a gradient or among regions. However, the estimated variance of the dissimilarity values, namely, (Tuomisto & Ruokolainen 2006) is not an interpretable measure of β diversity. Simply put, the m = N(N − 1)/2 pairs of dissimilarity values calculated from a set of N sample units are not independent of one another. There are not (m − 1) degrees of freedom in the system as implied by this approach. Unfortunately, the nature of the non-independence among dij values for a given system of N points is not easy to unravel for direct modelling purposes.
It has been suggested, furthermore, that multiple regression directly on dissimilarities can be used to ‘partition’ variation in Δy, such as Δyk = μ + β1Δxk + β2Δzk + εk, where Δxk are (say) spatial distances and Δzk are environmental distances. It is certainly possible to estimate coefficients for this model, despite obvious violation of the assumption of independence. However, the usual agenda here is to see how much of the ‘variation’ in Δy is ‘explained’ by space vs. the environment (Duivenvoorden et al. 2002; Tuomisto et al. 2003). Unfortunately, there is no clear sense in which the fitting of Δx to Δy has ‘controlled’ for spatial relationships among the sample units, and the meaning of the partial relationship of Δy with Δz, given Δx, in terms of actual underlying spatial variation, is quite unclear. Dutilleul et al. (2000) have demonstrated how the Mantel correlation between distance matrices does not accurately reflect known true correlations in underlying variables, even for Euclidean distances on univariate normal variables. Furthermore, Manly (2007, p. 215) has shown how the effects of even simple spatial autocorrelation are not removed by the regression of Δy on spatial distances Δx. Thus, we do not advocate the use of multiple regression directly on dissimilarity values (where the Δy values are treated as a univariate response variable). The partial Mantel approach (Smouse et al. 1986) has been known for some time to be problematic for interpretation (e.g. Dutilleul et al. 2000; Legendre 2000; Legendre et al. 2005; Legendre & Fortin 2010), in contrast with the simple Mantel test which is a valid approach to relate two distance matrices (Fig. 3, T3).
What is even more problematic is the use of partitioning methods to make direct inferences regarding the relative importance of underlying processes driving patterns in β diversity (e.g. Duivenvoorden et al. 2002; Tuomisto et al. 2003). For example, researchers partition into a portion explained by a set of spatial variables (X), a portion explained by a set of measured environmental variables (Z), some overlap in what these two sets explain, and a residual (unexplained) portion using RDA or dbRDA (Borcard et al. 1992; Legendre et al. 2009). This, in and of itself, is fine (see V3 above). However, extreme caution is required when interpreting results; it is sometimes claimed that the portion attributable to ‘space’ directly represents the relative importance of ‘neutral processes’ (sensuHubbell 2001), while that portion attributable to ‘environment’ represents the relative importance of ‘niche-based processes’. Unfortunately, such conclusions cannot logically be inferred from observational data (Underwood 1990). First, spatial structure in measured environmental variables (leading to overlap in explained variation), precludes any logical inference about whether processes were niche-based or neutral. Also, apparently ‘spatial’ portions interpreted as ‘neutral’ could simply have been due to unmeasured environmental variables. For example, researchers often neglect small-scale environmental variables when studying ecological systems at large scales. This does not necessarily mean that small-scale variation (appearing in the ‘spatial’ portion when using the rather powerful method of PCNM, for example, Dray et al. 2006) is driven by neutral processes. Even variation attributed to environmental variables alone might co-incidentally mirror patterns in species that actually arose from neutral processes. Finally, individual species vary in their degree of aggregation (McArdle & Anderson 2004), so neutral processes should yield spatial patterns at different scales for different species. Thus, patterns in multivariate data are not easily interpreted regarding the actual mechanisms at work for individual species.
Methods of partitioning are helpful for uncovering patterns and generating hypotheses (Underwood et al. 2000). To test potential mechanisms underlying observed patterns, manipulative experiments isolating the factor(s) of interest are required (Chase 2007, 2010). When controlled experiments are not possible, a ‘toe-in-the-door’ regarding mechanisms might be obtained from ever-more-specific null models (Chase 2007; Vellend et al. 2007), incorporating explicit hypotheses regarding species pools (γ) at a variety of scales, or expectations of occupancies or relative abundances. Insights can also be achieved through contrasting simultaneous analyses of observational data using taxonomic, phylogenetic and functional β diversity (e.g. Graham & Fine 2008; Swenson et al. 2010). In addition, simulations of ecological processes under a variety of stochastic and deterministic forces might be used to identify plausible hypotheses for the mechanisms governing β diversity.
We agree that researchers should be explicit about which ‘β diversity’ they are referring to, but we disagree that there is only one ‘true β diversity’sensuTuomisto (2010a). Multivariate species data are information rich. Plurality in the concept of β diversity can yield important ecological insights when navigated well. By knowing the properties of the measures being used and applying more than one, the underlying ecological structures in the data generating patterns in β diversity can be revealed, such as a selective vs. non-selective loss of shared species, or an increase in the variance of log-abundances.
We highlight the special utility of null models for studying β diversity, which can eliminate the dependence of β diversity on α diversity (e.g. the Raup–Crick measure) and/or γ diversity. Using the Raup–Crick measure also goes some way (though not entirely) towards alleviating the well-known problem of the classical (and most often used) measures of β diversity (Whittaker, Jaccard and Sørensen), which lose resolution for datasets having many samples that share few species. This occurs in sparse datasets (often encountered in studies of disturbance or predation) and also in datasets spanning large spatial or temporal scales (as in studies of latitudinal gradients or biogeography). The appropriate species pool (γ) to use in null models, especially at broad scales, remains an important topic for future research.
We consider that future studies of β diversity will lie not just in the meaningful use of multiple approaches for examining patterns, but also in the development of stronger frameworks for assessing underlying processes. More experiments directly testing mechanisms which generate β diversity are needed. Where manipulative experiments are not feasible (at large spatial or temporal scales), simulations, multi-scale null models and con-joint analyses of abundance, taxonomic, phylogenetic and functional information might be used to narrow down potential instrumental models of the mechanisms driving β diversity.
This work was made possible by support from the National Center for Ecological Analysis and Synthesis (NCEAS), Santa Barbara, USA, through the activities of the working group entitled ‘A synthesis of patterns, analyses, and mechanisms of β diversity along ecological gradients’. M.J. Anderson was also supported by a Royal Society of New Zealand Marsden Grant (MAU0713). J.C. Stegen was supported by an NSF Postdoctoral Fellowship in Bioinformatics (DBI-0906005). N.J. Sanders was supported by grant DOE-PER DE-FG-02-08ER64510. N.J.B. Kraft was supported by the NSERC CREATE Training Program in Biodiversity Research. We thank P. Legendre and an anonymous referee for their comments on the manuscript.