The imprint of the geographical, evolutionary and ecological context on species–area relationships



Species–area relationships (SAR) are fundamental in the understanding of biodiversity patterns and of critical importance for predicting species extinction risk worldwide. Despite the enormous attention given to SAR in the form of many individual analyses, little attempt has been made to synthesize these studies. We conducted a quantitative meta-analysis of 794 SAR, comprising a wide span of organisms, habitats and locations. We identified factors reflecting both pattern-based and dynamic approaches to SAR and tested whether these factors leave significant imprints on the slope and strength of SAR. Our analysis revealed that SAR are significantly affected by variables characterizing the sampling scheme, the spatial scale, and the types of organisms or habitats involved. We found that steeper SAR are generated at lower latitudes and by larger organisms. SAR varied significantly between nested and independent sampling schemes and between major ecosystem types, but not generally between the terrestrial and the aquatic realm. Both the fit and the slope of the SAR were scale-dependent. We conclude that factors dynamically regulating species richness at different spatial scales strongly affect the shape of SAR. We highlight important consequences of this systematic variation in SAR for ecological theory, conservation management and extinction risk predictions.


The diversity of life is perhaps the most stunning feature of our planet and has consequently stimulated ever-increasing interest in understanding how regional and local differences in the number of species arise and are maintained. Recent decades have seen a paradigm shift from a more local species interaction-based explanation of diversity patterns to the acknowledgement of the importance of large-scale processes (Ricklefs 1987, 2004). Differences in species richness at local spatial scales are no longer linked to competition, predation and population dynamics alone, but also to differences in the number of species on larger – regional – spatial scales originating from speciation, colonization and extinction dynamics. The relative importance of local and regional processes in shaping diversity patterns is still under debate (Cornell & Lawton 1992; Lawton 1999; Shurin et al. 2000; Hillebrand & Blenckner 2002). Metacommunity theory (Amarasekare & Nisbet 2001; Mouquet & Loreau 2002; Leibold et al. 2004) and experiments (Cadotte & Fukami 2005) have addressed the regulation of species coexistence in a spatial context, highlighting the importance of connectivity and species turnover between local habitats for the maintenance of regional species richness. Moreover, it has become obvious that the relationship of diversity to major environmental gradients such as productivity changes with spatial scale (Chase & Leibold 2002). The acknowledgement of metacommunity dynamics in community ecology concurs with an increasing awareness of the importance of scale in the macroecological assessment of diversity (Hillebrand 2004; Rahbek 2005). Including spatial scale explicitly into macroecological approaches yields new and important theoretical advances (Adler et al. 2005; Storch et al. 2005).

The differences in the number of species at different spatial scales and the species turnover between small and large areas are captured by species–area relationships (SAR), which are among the most widely studied phenomena and robust generalizations in ecology. SAR reflect the fact that species richness (S) tends to increase with increasing sampling area (A), a relationship recognized since the beginning of quantitative ecology (Arrhenius 1921; Gleason 1922). While finding more species in larger areas is inevitable (in spatially nested areas at least) because species are not distributed identically in space, the systematic nature of this relationship is the key to its importance for ecologists. In particular, the empirical evidence that many assemblages have SAR well summarized by simple mathematical relationships between S and A and the likelihood that key ecological processes are reflected in the shape of SAR means that they remain of great interest (Rosenzweig 1995).

Moreover, current ecological research increasingly addresses issues related to habitat fragmentation, global environmental change and loss of biodiversity. Against this background, understanding the relationship between species richness and area is particularly important. SAR are central in the prediction of species loss in response to global environmental change (Thomas et al. 2004) and regional habitat loss (Ney-Nifle & Mangel 2000) as well as the risk of future diversity loss because of reduced speciation (Rosenzweig 2001). Additionally, SAR have a fundamental role in theoretical ecology, e.g. in neutral biodiversity models (Hubbell 2001) and the spatial scaling of trophic interactions (Brose et al. 2004).

Despite the importance of SAR, there have been surprisingly few attempts to empirically quantify systematic variation in their key parameters, such as the exponent z of the power law model. Reviews of SAR are rare and although there are many individual studies, these typically lack generality by focusing on particular taxa and restricted geographical locations. One notable exception is the highly influential review by Connor & McCoy (1979). They analysed 100 SAR in order to see if there is a unique theoretical basis to SAR and whether the parameters, especially the exponent z of the power-law, reveal any deeper biological meaning. They remained sceptical about both aspects.

Connor & McCoy (1979) also discussed the two major hypotheses explaining why richness should increase with area: (i) the environmental heterogeneity (habitat diversity) hypothesis; and (ii) the demographic process hypothesis. The first explanation is a pattern-based view of SAR, such that larger areas have a higher probability of containing more habitat types; this increase of habitat diversity with area generates the SAR because of species’ habitat associations. In contrast the ‘demographic’ explanation is process-based, incorporating the dynamic processes of dispersal, colonization, speciation and extinction at multiple spatial scales. Larger areas have higher probabilities of colonization and speciation and lower probabilities of extinction (MacArthur & Wilson 1967; Hubbell 2001), fostering higher diversity. However, very high colonization success by dispersal via strong connectivity may in turn flatten SAR at coarser scales by reducing species turnover between them (Preston 1962; Mouquet & Loreau 2002; Leibold et al. 2004).

With the growing importance of space in ecology (Rosenzweig 1995; Tilman & Kareiva 1997), further insights into the theoretical basis of SAR have been sought. These analyses have targeted both the pattern-driven (e.g. Leitner & Rosenzweig 1997; Harte et al. 1999) and process-driven (Hanski & Gyllenberg 1997; Hubbell 2001) view of SAR.

In the pattern-based approach, one major discussion has centred on the utility of self-similarity concepts (Harte et al. 1999; Maddux & Athreya 1999; Harte et al. 2001; Lennon et al. 2002). ‘Community self-similarity’ implies that dividing area into constant ratios produces also constant ratios of species richness (Harte et al. 2001). Whilst the understanding of these intrinsic properties of SAR models advances (Hubbell 2001; Harte et al. 2005), it remains unresolved at present exactly how and under what circumstances power-law SAR (or any other SAR) arise from species’ spatial distributions, Preston's (1962) ground-breaking theoretical work notwithstanding.

Tightly linked to the issue of self-similarity is the question as to whether SAR are scale-independent (Harte et al. 1999; Crawley & Harral 2001). There are several properties of different SAR that are in some respects scale-invariant, depending on their shapes (Lennon et al. 2001). However, it is now widely accepted that SAR generated at very small (within communities) and at very large (across evolutionary provinces) spatial scales differ from SAR generated at intermediate scales because the importance of different processes underlying SAR also vary with scale (Rosenzweig 1995; Hubbell 2001, and references therein). At local scales (within communities), SAR are driven by rank-abundance distributions and are nonlinear in log–log space. At intermediate scales, SAR depend less on abundance patterns and more on speciation, dispersal and extinction, and give a linear relationship in log–log space. At even larger geographical scales (crossing evolutionary provinces), assemblages do not share evolutionary history and are not connected by dispersal, leading to an increase in steepness of slopes as species-assemblages are unrelated and species turnover maximized.

From a more dynamic perspective on SAR, metacommunity models have shown how increasing dispersal ability alters the parameters of SAR by changing colonization-extinction dynamics and homogenizing local communities (Amarasekare & Nisbet 2001; Hubbell 2001; Shurin & Allen 2001; Mouquet & Loreau 2002; Leibold et al. 2004; Hovestadt & Poethke 2005). Increasing dispersal in these models reduces the slope of SAR by making local assemblages more similar to each other.

Twenty-five years after Connor and McCoy's study – and despite the conceptual advances made – a more general quantitative analysis of SAR across habitats and organisms is still lacking, but particularly warranted in order to provide a differentiated baseline of SAR parameters for predictive ecology and to identify factors resulting in significant variation between SAR. Here we report the results of such a general analysis using 794 SAR obtained from the literature. These SAR encompass terrestrial and aquatic organisms ranging from unicellular eukaryotes to vertebrates and higher plants and include habitats from all over the globe. We identify changes in SAR with sampling method and characteristics of both the organisms and habitats. These characteristics were chosen to reflect processes highlighted by the hypotheses outlined above (see Factors and hypotheses for details). Each SAR was described by its exponent z of the power law relationship and by a transformed correlation coefficient rZ, a measure of the goodness-of-fit of the SAR model. We employed quantitative meta-analysis techniques to compare the fitted parameters of SAR across the published studies. We also put major effort into excluding artefacts in our analyses by considering spatial autocorrelation and by repeating the analysis using an alternative SAR model.

Materials and methods


Six abstracting services were searched: JSTOR (1913–1999, six ecological journals), BIOSIS Biological Abstracts (1980–2003), BIOSIS Zoological Record Plus (1978–2003), CSA Entomology Abstracts (1982–2003), ISI Web of Knowledge (1987–2003) and Aquatic Sciences and Fisheries Abstracts (1978–2003). The search strings were ‘species area relationship*’ and ‘species area curve*’, with and without hyphens. More recent papers were included if present, and also papers derived from the bibliographies of the papers screened.

The mathematical formulation of SAR has been widely discussed in the ecological literature. The most common form is the power-law, S = c Az, describing a linear increase of log S with log A with a slope of z (Arrhenius 1921; Rosenzweig 1995; Hubbell 2001). Alternatively, the semi-log SAR (Gleason 1922) defines a linear increase of S with log A, such that S = a + b log (A). Other formulations – including those with an asymptote and with sigmoid (triphasic) shapes – have been published, but are much less frequently used (Tjörve 2003). While focusing mainly on the power law here, we also present results for the semi-log SAR in order to show that our results can be generalized beyond one specific (albeit long-established and near-ubiquitous) equation (see supporting online material, SOM). We did not include asymptotic or sigmoid equations (Tjörve 2003); their parameters are not analogous to parameters in the power-law and semi-log SAR and so the results from the meta-analysis would not be comparable.

We obtained slopes (z) and correlation coefficients (r) from 794 SAR published in the literature. These were obtained either from the equations given in the original paper (379 studies) or from our own calculations on the source data as published (415 studies). From this total of 794 studies, 770 gave the correlation coefficient, whereas 553 analyses gave the slope and its standard error, an input needed for the weighted meta-analyses on z (see below). We electronically append a table with references included in our database (see Table S1 in the SOM). For the semi-log SAR, we obtain parameters from 506 studies (SOM).

It has become common to term the fitted parameter z the slope of the power-law SAR, because z represents the slope of the linearized function in log–log space, but the slope of the SAR (as the tangent to the curve) in untransformed axes is determined by both fitted parameters, z and c (Connor & McCoy 2001). The exponent z can be understood as the proportionality constant in the rate-equation form of the power law: dS/dA = zS/A. The increase in richness per increase in unit area is directly proportional to species density, measured as the number of species per unit area. We use z as the parameter reflecting this proportionality in our meta-analysis. In order to avoid lengthy descriptions of the conceptual definition of z, we adhere to the term ‘slope’ throughout the manuscript.

In contrast to z, the second parameter c is not independent of spatial unit used (Rosenzweig 1995) and a comparison of c between studies is not straightforward.

In addition to z, we use a transformed correlation coefficient as a measure of the goodness-of-fit of the SAR model to the data. With this measure, we are able to analyse the predictive power of SAR models, and see how differences in organism and habitat characteristics affect how closely species richness tracks increases in area.

Factors and hypotheses

In addition to the two parameters z and r, each SAR was classified with respect to 10 variables. These variables presented a broad range of organism and habitat characteristics related to the proposed ultimate factors controlling SAR, i.e. habitat diversity, dispersal, speciation/extinction. A second pre-requisite was that the classification within these factors could be made objectively for all studies involved. That is, we opted for broad but reliable categories rather than detailed distinctions.

Species–area relationships can be produced by different sampling regimes: nested sampling means that a larger area contains all smaller areas, whereas independent sampling uses spatially non-overlapping areas. We used the factor method to separate SAR based on independent areas from those using nested areas. Others have proposed an even finer division of SAR based on methodology (Scheiner 2003; Gray et al. 2004a), but we restrict ourselves to this broad distinction because (i) it is well applicable across the breadth of studies included; and (ii) we investigate additional discriminatory criteria (island vs. mainland) using other factors. We expected nested SAR to be more predictable (have higher goodness-of-fit) than independent SAR because of the non-independence of data in the former. We expected nested and independent SAR not to differ in slopes because the expected number of species in a given area is independent of sampling method. Some studies used a mixture of nested and independent sampling regimes. They were included only in the analysis comprising the entire data set (all), but were excluded in the separate analyses for nested and independent data.

It has often been argued that substantial differences in SAR exist between mainland and island areas as well as between habitats of different habitat diversity and connectivity (Preston 1962; Rosenzweig 1995; Hanski & Gyllenberg 1997). We tested these predictions regarding habitat characteristics on two organizational levels. We contrasted aquatic and terrestrial ecosystems (the factor realm) and we tested for differences between broad and ubiquitous ecosystem types (the factor habitat, contrasting islands, forested mainland, unforested mainland, lakes, streams and the ocean). ‘Realm’ reflects the different physiological and evolutionary backgrounds of species in aquatic and terrestrial environments as well as any contrasts in spatial complexity between these major ecosystems. ‘Habitat’ focuses on potential differences arising from different spatial complexity and spatial isolation. We expected SAR from islands to be steeper than SAR on terrestrial mainlands (higher isolation, less dispersal) and likewise we expected lakes to differ from the more continuous and less isolated ocean.

Another question with major implications is whether or not SAR change systematically with spatial scale (Rosenzweig 1995; Harte et al. 1999; Crawley & Harral 2001; Hubbell 2001). We address two key aspects of scale, the grain (Rahbek 2005) and the spatial range. The sampling grain was given by two measures, the minimum and maximum area used for each SAR (log-transformed m2). From consideration of triphasic SAR, we expected a power-law SAR to fit best at intermediate scales, and expected highest slopes for SAR at large grains. As a measure of range of the SAR, we used the log-transformed ratio of maximum to minimum area. The range was only weakly correlated with minimum (r = −0.23) and maximum (r = 0.26) sampled area. We expected lower goodness-of-fit for SAR with large ranges, because these probably encompass greater environmental heterogeneity. Predictions for z are less straightforward, as several models predict non-linear responses of z to shifts in grain (see above); thus, integrating over a larger range of this non-linear spectrum may, therefore, increase or decrease z depending on the grain.

Dispersal is a key mechanism in neutral meta-community models (Hubbell 2001), and high dispersal lowers z-values because of the greater homogeneity produced through reduced species turnover and fewer local endemics. Reliable estimates of dispersal distances are not available for all the organisms included here, but we hypothesized that if dispersal matters it should result in distinctive parameter estimates for islands (see above) and for small organisms. It has been argued that small organisms (unicellular eukaryotes) have overall greater dispersal, which results in a flatter SAR (Finlay et al. 1998; Hillebrand et al. 2001). Therefore, organism body mass was derived from published sources (Peters 1983; Hillebrand 2004) as log-transformed wet weight (g). Some of the organism groups are characterized by large ranges of body masses. In these cases, we obtained geometric means or used the geometric mean of minimum and maximum estimates (see also Hillebrand 2004), reflecting the higher abundance of small individuals. While the size range within groups may be large (up to five orders of magnitude), it is small compared with the overall size range (> 13 orders of magnitude) obtained across organism groups. In order to accommodate concerns about the validity of body mass estimates for modular organisms such as plants, we repeated the analysis without these. To ensure that body size effects did not reflect scale differences (Azovsky 2002), we also used relative area as a body-size equivalent sampling grain (see SOM).

Modelling simple-structured communities, Holt et al. (1999) predicted that the slope of SAR should be larger for organisms at higher trophic levels. We defined trophic level as carnivorous, omnivorous, herbivorous, autotrophic, detritivorous, microbivorous and parasitic. Organism groups containing species differing in their trophic level were classified as omnivores.

Species–area relationships are expected to strongly reflect differences in overall diversity and in species spatial turnover (Lennon et al. 2001). Both diversity and turnover are supposed to differ with latitude (Hillebrand 2004), therefore, we hypothesized that the slope of the SAR should show a latitudinal decline. We characterized the geographic position of the sampled areas by the mid-latitude of the area(s) sampled, irrespective of northern and southern hemisphere. A separation of the hemispheres was not possible because of the relative paucity of Southern hemisphere studies. For this analysis, we focused on studies with a confined latitudinal extent, i.e. covering ± 5 ° N or S of the mid-latitude in order to exclude global studies where the calculation of mid-latitude would be meaningless. Analysing data from many geographically scattered sites raises the issue of spatial autocorrelation, which however, played a minor role (see SOM).

Effect sizes and analyses

The choice of effect sizes is a critical step in any meta-analysis. We used the parameter z of the log–log SAR as effect size reflecting the ‘slope’ (Hillebrand et al. 2001). We used the Fischer-Z-transformed correlation coefficients (rZ) as an effect size for the goodness-of-fit of the association between richness and area (Rosenberg et al. 2000), where rZ = 0.5 ln[(1 + r)/(1 –r)]. The use of this transformed correlation coefficient is recommended to normalize the variable (Rosenberg et al. 2000). For both measures, we obtained sampling variances to conduct weighted meta-analysis. Weighted analyses are much more powerful than unweighted ones, because the impact of single studies on the overall results depends on their reliability (Gurevitch & Hedges 1993). For z, the sampling variance is its squared standard error. For rZ, it is calculated from the number of observations (N) as 1/(– 3).

We used weighted mixed-model meta-analysis to obtain grand means of z and rZ and to test the effect of the variables described above. For overall and group-wise average effect sizes, 95% confidence intervals were calculated by bootstrapping (Rosenberg et al. 2000). As we used data in multiple tests, we adjusted significance levels by sequential Bonferroni adjustment, which is recommended over the overly conservative traditional Bonferroni adjustment (Quinn & Keough 2002). In the sequential procedure, significance levels are adjusted by ranking the analyses from most to least significant and by dividing the nominal significance level by t, the number of tests, for the most significant original analysis, by t – 1 for the second one, t – 2 for the third, etc.


The average z for the power-law relationship across all data was 0.27, but ranged very widely, from 0 to c. 1. The average slope b of the semi-log relationship was 9.11. Average goodness-of-fit (rZ) was 0.726 for the power-law SAR, and only slightly smaller for the semi-log SAR (rZ = 0.718). The power law was a better fit than the semi-log model in 52% of SAR, the semi-log model was a better fit than the power-law in 46%, and they were equal for the remaining 2% of cases.

Species–area relationships varied according to the two major census methods; that is, whether nested or independent areas were used (Table 1, Fig. 1a,b). Nested SAR were on average both steeper (z = 0.36) and of better fit (rZ = 2.16) than their independent counterparts (z = 0.24, rZ = 0.97). The same differences were observed for semi-log SAR (SOM).

Table 1. Results of weighted mixed-model meta-analysis on the fit rz and slope z of the power-law species–area relationship
FactorDataFit (rz)Slope (z)
k P adj B k P adj B
  1. The table denotes the factor, the data set (all, unrestricted; nes, nested; ind, independent), the number of studies for each test (k), and the adjusted significance level. Adjustment was by sequential Bonferroni correction (see Materials and methods). For significant continuous predictors, the slope B of the weighted regression model is given. For a more detailed table of results, see Table S2 in the Supporting Online Material.

MethodAll713< 0.01 524< 0.01 
RealmAll770ns 553ns 
Nes158ns 130< 0.1 
Ind555ns 394ns 
HabitatAll682< 0.01 488< 0.01 
Nes148< 0.01 128< 0.01 
Ind492< 0.01 339< 0.01 
RangeAll728ns 535< 0.01−0.024
Nes153ns 130< 0.001−0.025
Ind530ns 386ns 
Minimum areaAll633< 0.001−0.086535ns 
Nes147ns 130< 0.0010.032
Ind454< 0.001−0.032388ns 
Maximum areaAll633< 0.001−0.071535ns 
Nes147ns 130< 0.10.016
Ind454ns 388ns 
Relative areaAll620< 0.001−0.093526ns 
Nes142ns 125< 0.0010.025
Ind447< 0.001−0.040385ns 
Body massAll749< 0.010.054542< 0.010.020
Nes153ns 125< 0.0010.040
Ind541< 0.0010.043389ns 
Trophic levelAll769< 0.01 553ns 
Nes158ns 130ns 
Ind555ns 394ns 
LatitudeAll579ns 383< 0.05−0.003
Nes153< 0.1−0.006127< 0.001−0.007
Ind402ns 254ns 
Figure 1.

Effects of categorical variables on the slope z (left column) and the fit rZ (right column) of the power-law SAR. Black circles represent all SAR, blue squares nested and red diamonds independent SAR. The average slope or fit is given ± 95% confidence intervals (CI) for the entire data set (a, b), different realms (c, d), and different habitats (e, f). Realms are terrestrial (terr) or aquatic (aqu); habitats are islands (isl), forests (for), non-forestal mainland (non-f), lakes (lake), streams (str) and marine ecosystems (mar).

We found that differences in SAR between the aquatic and terrestrial realm were minor. For nested power-law SAR, we found a marginally significant trend of terrestrial systems having larger z's than aquatic systems (Fig. 1c), whereas they did not differ in their average fit (Table 1, Fig. 1d). Independent semi-log SAR showed significantly higher fit in terrestrial than in aquatic systems.

We found highly significant variation in SAR between different habitats for both parameters (Table 1, Fig. 1e,f) and in both SAR models (SOM). From the number of significant differences in fit and slope, we highlight two major contrasts.

First, our analysis does not support the generally accepted view that SAR generated from islands are distinctively different from those from mainlands. For the entire data set, island SAR have intermediate average z-values. The average slope of nested island SAR is higher than most other habitats, but they are indistinguishable from mainland forest habitats (Fig. 1e). This lack of contrast is corroborated by the semi-log SAR model (SOM) and by aquatic ‘habitat islands’ such as lakes, which have slopes similar to (nested SAR) or lower than (independent SAR) the more continuous ocean.

Second, forested mainland habitats have steeper SAR than non-forested habitats across all SAR and across nested SAR, for both power-law and semi-log models. Likewise, marine habitats have significantly steeper power-law SAR than both freshwater habitats (lakes and streams).

Spatial scale measured as grain played a small role for variation in z. When considering the entire data set, we found that z was unaffected by minimum or maximum sampling area (Table 1, Fig. 2a,c). Only when considering nested SAR alone did we find that z increased significantly as minimum area increased (Fig. 2a) and showed a similar, but marginally non-significant, trend for maximum sampling area (Fig. 2c). Similarly, the slope b of nested semi-log SAR significantly increased as both the minimum and the maximum sampling area increased (SOM). Slopes of nested SAR remained positively correlated to the grain in both SAR models when we used relative area as a body-size adjusted measure of grain (SOM). Only for the entire semi-log SAR data did we observe a decrease in the slope b with increasing maximum area and increasing relative area (SOM).

Figure 2.

Effects of scaling variables on the slope z (left column) and the fit rZ (right column) of the power-law SAR. Blue squares represent nested SAR, red diamonds independent SAR. Each data point represents one SAR. Slope and fit are plotted over the minimum (a, b) or maximum (c, d) area used in the SAR as well as over the range of areas used (e, f). For visual clarity, six very low or high slopes were omitted from (a), (c), and (e), but not from the statistical analysis. Note that diagrams display raw data points of single gradients, whereas the statistical analysis (Table 1) weighs each data point with the inverse of the sampling variance (see Materials and methods).

The fit of both SAR models significantly decreased with both minimum and maximum sampling area (Table 1, Fig. 2b,d, SOM). At coarser scales, SAR models are noisier and poorer descriptions of diversity scaling. This relationship was significant for the entire data set and – in the case of power-law model and minimum area – also for the independent data set (Fig. 2b).

The range of the sampled scale significantly affected the slope of both the power-law (Table 1, Fig. 2e) and semi-log SAR (SOM). The estimates for z and b decreased significantly with increasing range for the entire data set and for the nested subset, whereas the same trend was present but was not significant for the independent subset. The goodness-of-fit did not respond significantly to variation in the range of the sampling (Table 1, Fig. 2f) except for a decrease in rZ of the nested semi-log model with increasing range.

Species–area relationships became steeper with increasing species’ body size (Table 1, Fig. 3a). Depending on the SAR model used, this was significant for all data subsets (semi-log, SOM) or only for the entire and nested data partitions (power-law, Table 1, Fig. 3a). The fit of the power-law model was also better for larger organisms, which was significant for both the entire and the independent-area methods (Table 1, Fig. 3b). Deleting modular organisms (plants) from this analysis of body mass effects significantly reduced the size of the database and resulted in non-significant relationships for nested SAR. For the independent data set, the effect of body mass on the fit of the SAR was significantly positive, whereas the effect of body mass on the slope remained positive but became non-significant.

Figure 3.

Effects of continuous variables on the slope z (left column) and the fit rZ (right column) of the power-law SAR. Blue squares represent nested SAR, red diamonds independent SAR. Each data point represents one SAR. Slope and fit are plotted over the body weight of the organisms (a, b) and the mid-latitude of the area used in the SAR (c, d). For visual clarity, six very low or high slopes were omitted from (a) and (c), but not from the statistical analysis. Note that diagrams display raw data points of single gradients, whereas the statistical analysis (Table 1) weighs each data point with the inverse of the sampling variance (see Materials and methods).

Although the goodness-of-fit for the entire database differed between some trophic levels, the slope of the SAR did not – we found no relationship between z and the trophic level of organisms (Table 1).

Species–area relationships slopes strongly reflected the latitudinal gradient of species diversity: with increasing latitude, z and b significantly decreased (Table 1, Fig. 3c, SOM). This pattern could not be attributed to spatial autocorrelation strongly affecting the statistical independence between sites (SOM). The reduction of z from tropical to temperate regions was significant for the entire data set and for the nested subset of SAR (Table 1), whereas for b this effect was significant in both nested and independent SAR. The fit of the SAR was not significantly affected by latitude (Table 1, Fig. 3d).


The results of our meta-analysis highlight two important aspects of SAR. First, SAR show extensive, systematic variation – they are far from being an invariant baseline for the analysis of diversity or extrapolation of species ‘carrying capacities’. Instead, there are important differences between parameters of the SAR based on the sampling scheme (nested vs. independent), across latitudes and sizes of organisms, and between different habitats. Second, this variation in SAR is not stochastic, but instead strongly reflects the geographical, evolutionary and ecological context of the species considered. Conspicuous large-scale spatial patterns such as the global latitudinal diversity gradient, fundamental organism features such as body size, and the different spatial organization and evolutionary backgrounds of habitats leave general imprints on SAR.

There is a major dichotomy in SAR between nested and independent sampling methods, the former being steeper and of higher goodness-of-fit than the latter. Interpreting these results is relatively straightforward for rZ. A better model fit is expected because data points are non-independent in a nested design and species richness always has to increase with increasing area, whereas it is free to increase or decrease with area in an independent design.

Our finding of significantly larger z's in nested relative to independent areas is contrary to our expectations. Whereas the average estimates of z for all data (0.27) and for independent data (0.24) conform to the ‘canonically’ predicted value (0.262) by Preston (1962), our observed average estimate of z for nested SAR is much higher (0.36). Increases in z in nested sampling schemes could be artificial if they are compensated for by lower estimates of the parameter c. We tested for significant differences in log c between nested and independent subsets and found that log c was also significantly higher in nested than in independent SAR. Thus, the higher z in nested SAR represents a real increase in the rate with which species richness increases with area. In part, this may be a consequence of the different spatial scales of these studies. Over the entire data set, the slope tended to decrease with increasing range and increasing grain, both of which were smaller for nested SAR. However, there is considerable overlap between the scales of the nested and independent studies and the latter tend to have steeper slopes than the former in these areas of overlap (cf. nested and independent data in Fig. 2a–d). That is, other factors have to contribute to higher z in nested SAR; what these factors are remains unclear. Based on the strong differences shown by nested vs. independent sampling methods, we strongly concur with those advocating a careful separation of sampling regimes when discussing SAR (Scheiner 2003; Tjörve 2003; Gray et al. 2004a,b; Scheiner 2004).

We found that aquatic systems are characterized by a marginally slower accumulation of species with area than terrestrial systems in nested SAR. However, differences between realms were small compared with differences between habitats.

Previous studies have suggested that SAR from islands should on average have higher z than those constructed from mainland distributions (Preston 1962; Rosenzweig 1995; Hanski & Gyllenberg 1997). Our finding that island SAR are generally no steeper than mainland SAR clearly does not support this proposition. Whereas average estimates of z were high for island SAR, they were not significantly different, at least from SAR in forested mainland habitats. This result may depend on the fact that the isolation of islands may either increase or decrease the SAR slopes. On the one hand, islands represent disconnected habitats with lower dispersal than similar areas on mainland, which may increase z by increasing species turnover (Preston 1962; Rosenzweig 1995; Hovestadt & Poethke 2005). On the other hand, in very isolated archipelagos, colonization is not from the mainland but from neighbouring islands, which may reduce overall species richness and species spatial turnover and therefore also z (Hanski & Gyllenberg 1997; Hovestadt & Poethke 2005).

Another important difference between habitats occurred between forested (high slope) and non-forested (low slope) mainland terrestrial habitats on the one hand and between marine (high slope) and lake (low slope) habitats on the other. Our explanations for this pattern have to remain speculative, as we lack direct measures to test them. Nevertheless, these habitats exhibit a number of contrasts potentially affecting species turnover. These contrasts include differences in evolutionary history (oceans being phylogenetically richer than freshwater habitats), their different vertical dimension (forests being higher than non-forests, oceans being deeper than freshwaters), and differences in spatial structure (forests being more spatially structured than grasslands, deserts or wetlands). With the addition of a stronger vertical spatial dimension to area comes the opportunity for stacking additional habitats/environments within a given area. Expanding areas with a strong vertical component opens up the opportunity for turnover of habitats in different strata; essentially, adding a significant vertical dimension to area allows more rapid species turnover between areas and so steeper SAR.

Our results shed new light on the much discussed aspect of scale-dependence of SAR. Considering the sampling grain, we present evidence that the slope of the nested SAR is scale-dependent. Increasing the sampling grain of nested SAR produced higher estimates of z. Thus, our results support the argument that coarse-grained SAR should have steeper slopes as predicted by Rosenzweig (1995) and Hubbell (2001). This conclusion, however, was true only for the nested subset of SAR. Across the multitude of habitats and organisms included in our analysis, we found no evidence for a unimodal pattern of z across scales, as reported for plant SAR (Crawley & Harral 2001). However, we observed a decrease in the slope b of the semi-log SAR with increasing maximum area and increasing relative area when using all data (SOM).

The goodness-of-fit of SAR models was significantly diminished at coarser minimum and maximum area in both SAR models. As spatial scale increases, there is both a decrease in the number of data points available and an increase in the physico-geographical variability of the areas analysed. Clearly, both trends have the potential to reduce the predictability of the SAR.

The range of areas sampled affected the slope of the SAR (but not the fit). In contrast to predictions of triphasic SAR (Rosenzweig 1995), the fit was not reduced when the SAR covered a large spatial range. The slope, however, decreased in wide-ranged SAR, an effect prominent in the entire dataset and in nested SAR. Similar results have been found in nested SAR for bird data reanalysed by He & Legendre (1996). This decrease was mainly observed for very small ranges (Fig. 2e), indicating that species packaging at smallest scales leads to an initially steep increase of S with area (Rosenzweig 1995; Hubbell 2001). Partly, this decrease in z can be explained by the significant negative correlation between range and minimum area (see Materials and methods). Wide-ranged studies tend to have lower minimum areas, and lower minimum areas tend to produce lower z in the nested subset of data (see Fig. 2a).

We found a strong positive relationship between body size and both the slope and the fit of the SAR. The increase in z was significant for both sampling methods in the semi-log model, but only for the entire and nested partition in the power-law model. The independent subset showed a similar trend. Our data support the notion that small organisms have flatter SAR (Finlay et al. 1998; Hillebrand et al. 2001), which has been explained by their higher dispersal abilities. Theoretical models show that increased dispersal leads to lower species turnover and higher homogenization of species compositions (Amarasekare & Nisbet 2001; Hubbell 2001; Mouquet & Loreau 2002; Hovestadt & Poethke 2005). The relationship between body size and dispersal has frequently been highlighted (Finlay et al. 1996; Fenchel et al. 1997; Finlay & Fenchel 1999). One pattern corroborating this proposition is the slower decay of assemblage similarity along geographic distance for small vs. large organisms (Hillebrand et al. 2001). Using published data on species compositions for a number of taxa, Hillebrand et al. (2001) observed that the similarity of assemblages decreased with increasing distance for all organisms, but the rate of this decay of similarity (and thus the increase in species turnover) was much higher for large organisms.

Recent evidence suggests that the trend of increasing z with increasing body mass that we found also holds when considering genetic diversity–area relationships for eukaryotic (Green et al. 2004) and prokaryotic (Horner-Devine et al. 2004) microorganisms. Both studies produced very small estimates for z (0.07 for fungi and 0.04 for bacteria, respectively) using molecular estimates of diversity. These studies show that ‘flat’ SAR for small organisms cannot be explained by a bias towards lower taxonomic resolution in these groups because of smaller number of identifiable morphospecies. However, a consensus as to the nature of the link between organism size and dispersal rate has yet to be established. Much larger organisms have larger estimates of z in our analysis, even in studies where they probably are not dispersal-limited. Other factors that may contribute to the body size gradient observed here are: (i) differences in methods from complete census of the habitat for large organisms compared with selected sub-samples for small organisms; and (ii) differences in the spatial distribution patterns of large vs. small organisms. Both aspects can affect the shape of SAR and both differ between organisms of different body size (He & Legendre 1996; Finlay 2002; He & Legendre 2002).

Modelling simple structured communities, Holt et al. (1999) predicted that the slope of SAR should increase for organisms at higher trophic levels. We did not find such an increase when comparing SAR slopes across autotrophs, herbivores, omnivores, carnivores, microbivores, parasites and decomposers. This may be at least partly explained by the nature of Holt et al.'s (1999) trophic model – the model assumptions are restrictive as the areas were closed for immigration and the food web was built using specialist consumers alone.

We observed a strong latitudinal gradient in z, decreasing towards the poles, which never has been shown before on this general level, but has only been reported for some particular groups (Koleff et al. 2003b; Rodriguez & Arita 2004). In contrast to the increase in z found here, others have even found a decrease in z towards the equator within single groups (Lyons & Willig 2002). In the present study, we found a significant negative trend of z and b with latitude across a wide range of habitats and organisms. This result mirrors a previously observed effect of sampling grain on the steepness of the latitudinal species richness cline (Hillebrand 2004). Both local assemblage and regional pool richness were observed to increase significantly from the poles to the equator, but on average, the slope of this increase was three-times steeper for richness measured on large areas (Hillebrand 2004). This finding in turns means that the increase of richness from local to regional areas has to be steeper in the tropics. The data set used for the study on latitudinal gradients (Hillebrand 2004) and for this analysis of SAR show very little overlap in the primary literature used. Thus, these two analyses strongly indicate that not only species richness increases towards the equator, but species turnover (measured as z of the power-law SAR) does so too.

Conclusion and consequences

Species–area relationships play a central role in ecology, and the trends we identified have potentially important consequences. Applying different slopes of SAR appropriately may improve the outcome of management tasks such as planning nature reserves and species conservation prioritization (Zurlini et al. 2002). More generally, we identified significant heterogeneity between species and habitats in SAR, which identify clearly matters when assessing world-wide extinction risks (Thomas et al. 2004) and speciation deficits (Rosenzweig 2001). The future of global diversity is tightly linked to the number and kinds of species an area can support, and as this study highlights, responses to changes in habitat area vary between species and habitats systematically. Species groups with large z-values increase in diversity rapidly with area, but also decline just as rapidly with shrinking area. Thus differences in the slope of the SAR between species groups are a strong indicator of their sensitivity to habitat and climate-space loss. Larger species, species living at lower latitudes, in forests or in oceans, for example, may be most sensitive to reduction in habitat area and such reductions may result in greater numbers of such species ‘committed to extinction’ (Thomas et al. 2004). In the face of global change and large-scale habitat destruction, our analysis provides the means to more accurately predict how loss of habitat transfers into loss of diversity.


This synthesis of SAR data has revealed some general insights into the spatial scaling of diversity, but has only touched the surface of many areas where our understanding is incomplete. From the many exciting research fronts, we will mention just three that seem to us particularly interesting.

First, with respect to SAR there are clearly various lacunae to be filled in the species and ecosystems studied. The strong taxonomic and body-size bias towards vascular plants, birds and mammals and the neglect of other taxa, e.g. arthropods and microorganisms is problematical for comprehensive generalization. Particularly, it also highlights yet another bias, namely that towards spatial scales (areas) much coarser than the scales directly experienced by many smaller organisms. This is slowly changing, for microorganisms at least, as exciting new molecular methods are developed and applied to diversity scaling (Green et al. 2004; Horner-Devine et al. 2004; van der Gast et al. 2005), but sampling down to the scales experienced by individual soil bacteria, for example, remains a formidable technical challenge when there may be 3 × 104 species g−1 of soil (Curtis et al. 2002). It is also easy to identify major ecosystem types greatly under-represented in terms of SAR studies, the open ocean perhaps being the most conspicuous of these.

Second, integrating the various ideas about species turnover (beta diversity) with those about SAR is very much worthwhile. Species turnover is the key, or at least part of the key, to understanding how species diversity changes with scale and how, in turn, this scale of observation changes patterns of species richness along spatial (Hillebrand 2004; Rahbek 2005) or environmental (Chase & Leibold 2002) gradients. Unifying within a common theoretical framework the many analyses of the frequency distributions of species range sizes, the scale-dependency of individual species distribution patterns (fractal or otherwise), species turnover patterns, SAR and species–time relationships is particularly useful (Preston 1962; He & Legendre 1996; Leitner & Rosenzweig 1997; Harte et al. 1999, 2001; He & Legendre 2002; Lennon et al. 2002; Koleff et al. 2003a,b; Adler et al. 2005; Harte et al. 2005; Rahbek 2005). This unification is likely to come about in part by expressing different measures of patterns in common terms (e.g. SAR and spatial turnover, Koleff et al. 2003a) and in part by developing pattern-analysis theory and analytical methods that explicitly involve variable spatial scales; most analyses of diversity patterns, for example, use only one arbitrary scale. Moreover, relating connectivity (rather than area per se) to the spatial structure of diversity is clearly tremendously important, both for theory and for practical conservation planning.

Third, another challenge holding great promise is to extend SAR and related pattern-based approaches to incorporate and so understand key processes such as dispersal, local extinction/recolonization, species extinction and speciation, and last but not least species interactions. Dispersal rates between local habitats have been identified as a major driver of species coexistence on the one hand and homogenizing species composition between habitats on the other. Practically, actual measurements of dispersal in nature are notoriously difficult and the dependency of dispersal on proximate variables such as body size is far from being settled. On the theoretical front, a great beginning has been made in the form of new ideas in metacommunity theory (both neutral and non-neutral), focusing our thinking about community ecology in general and especially about processes acting on different spatial scales and affecting coexistence of species (e.g. Amarasekare & Nisbet 2001; Mouquet & Loreau 2002; Leibold et al. 2004).

Species–area relationships and allied approaches are likely to remain central in our understanding biodiversity patterns, and, as we have shown in the present study, are powerful enough to reveal trends indicative of important underlying ecological mechanisms.


We acknowledge the helpful comments from S. Flöder, B. J. Hawkins, C. Rahbek, B. Worm, M. Xenopolous and the Aquatic Ecology workgroup at the Botanical Institute in Cologne on earlier drafts of this manuscript, which was also greatly improved by Jon Chase and three anonymous referees. Financial support was granted to HH by the Swedish council for basic research (Vetenskapsrådet).