The adopted sampling design and statistical modelling (MEMs: Dray, Legendre & Peres-Neto 2006) allowed us to detect several spatial structures in the community from the habitat (500 m) to the subplot (1 m) scale. The most obvious of these spatial patterns is the broadest one: most of the variation observed in the community corresponds to the mere spatial separation of the beech forest from the grassland (see MEM1 in Figs S2 and S3 in Appendix S1, Supporting Information). Therefore, community patterns at this scale do not actually provide definitive statistical proof that the community is structured by the observed environmental differences between the two habitats (Legendre & Legendre 1998). In fact, variance partitioning results indicate that once we accounted for the fact that the correlation between community patterns and environmental variables is spatially structured at a scale corresponding to the geographical distance between the two habitats, a multitude of spatial patterns remained significant and were uncorrelated with the measured environmental variables. Overall, these spatial patterns accounted for 40% of community variation. This large amount of variation is consistent with that observed not only in other soil assemblages but also in different types of communities (see the meta-analysis by Cottenie 2005).
Neutral models predict that spatial patterns in community structure can be explained by limited dispersal only, independently of environmental variables. Given that only 4% of community variation was explained by environmental variables and 40% was explained by spatial autocorrelation independent of environmental variation, one might conclude that patterns of community structure in the studied assemblages suggest the prevalence of neutral forces. However, variance partitioning is limited by the fact that it is hardly possible to take into account all environmental variables relevant to niche dynamics. A recent study (Smith & Lundholm 2010) has shown that the relative proportions of variance explained by environmental and spatial components partially depend on the degree of interaction between the two. Several authors have instead proposed that an appropriate null hypothesis for testing the niche theory is paradoxically provided by the neutral models developed in recent years (Bell 2000; Hubbell 2001; Alonso, Etienne & McKane 2006; Etienne 2007, 2009) and explicitly tested in this study. Results clearly reject models based on neutral theories. Note that neutral models predicted that local communities in the beech forest were experiencing a higher rate of immigration than those in the grassland. According to the neutral theory (Hubbell 2001), this is consistent with the fact that the beech forest has much lower beta diversity (Fig. 2a). However, observed levels of dissimilarity were always significantly greater than those predicted by the neutral model: both the beech forest and the grassland have higher levels of dissimilarity, and their local communities diverge with respect to their neutral counterparts (Fig. 3). This important finding implies that patterns of beta diversity in Fig. 2a cannot be interpreted in the light of demographic stochasticity and limited dispersal alone, as models based on these mechanisms failed to predict patterns of ecological dissimilarity or beta diversity (Dornelas et al. 2006; Etienne 2007). Given that the data refuted the neutral hypothesis and despite the fact that variance partitioning indicated that environmentally independent spatial effects are prominent features of the communities, we accept the hypothesis that environmental filtering and/or niche-mediated competition are operating and, in particular, that demographic stochasticity and limited dispersal are not the only forces driving community structure. The divergence of communities suggests that high environmental heterogeneity is sorting species into relatively heterogeneous local assemblages (Dornelas et al. 2006). We exclude a disturbance regime, given that the study area is in a nature reserve showing no sign of physical alteration. Some environmental axes should account for the observed community patterns; however, the ones we measured had limited explanatory power. Despite the range of scales we could resolve with our sampling design and modelling strategy, and the apparently obvious ecological differences between a grassland and a beech forest, most (56%) of the community structure remained unexplained and only a small fraction (2%) of community variation was accounted for by the spatial structure in the environmental variables or a spatially independent environmental effect (2%). We found a total of 156 species, most of which were not shared by the two habitats. Our data therefore raise the question of the unsolved ‘enigma of soil diversity’ (Anderson 1975): how is it possible that such a large number of species coexist in a relatively homogenous environment such as soil? Wardle (2002) stated that soil actually provides both high niche dimensionality (owing to its three-dimensional structure) and the opportunity to partition resources considerably. Furthermore, differential rates of activity can enhance coexistence by avoiding competition (Chase & Leibold 2003). Differential rates of activity and local patchy variations in resource availability, soil structure or features affecting life cycles (e.g. temperature) may increase local species richness by favouring local niche partitioning and increasing beta diversity (Wardle 2002). The local variations in our environmental variables are not consistent with this hypothesis because, for example, organic matter and water content were more variable in the beech forest (Fig. 3b), which has lower beta diversity (Fig. 3a) but higher species richness (Table 1, estimates of θ). Nevertheless, the pattern of heterogeneity observed in the environmental variables (Fig. 2b) matches that observed in the neutral analysis (Fig. 3, compare the more divergent beech with the less divergent grass), which offers a novel key for interpreting results. Indeed, there thus remain three not mutually exclusive possibilities: some environmental variables driving local communities at relatively fine scales were not measured, although we detected the strong spatial structures they determine (environmental filtering); local species interaction determines a substantial amount of community structure (niche), such that the neutral model was rejected and beta diversity was highly variable (Figs 2b and 3); temporal variations in the spatial patterns of the environment not accounted for in this study explain spatiotemporal variations in the assemblages. We cannot exhaustively address this latter potentially critical point. Nevertheless, we have good temporal information on the seasonal variability of the arthropod community in the study area (Migliorini, Petrioli & Bernini 2002), confirming that the analysed data set is representative of the community in terms of diversity and species composition. The large number of singletons might suggest that several species have not been sampled. Estimates of actual richness based on rarefaction curves (Chao 1, Jackknife 1, and Bootstrap: Magurran 2004; : Table S2 in Appendix S1, Supporting Information) showed that observed richness values did indeed underestimate total richness, especially in the beech forest. This result may be due not only to the very high diversity of the studied system but also to the use of the Berlese–Tullgren apparatus, which exposes soil arthropods to high thermal stress. However, under the dry conditions of Mediterranean areas (including relatively moister habitats such as beech forests), the Berlese–Tullgren apparatus has generally proved to be very efficient for estimating the composition of taxa such as oribatid mites (see Migliorini, Petrioli & Bernini 2002). Note that the main result of the analysis based on neutral models did not depend on whether we calculated community dissimilarity using presence/absence data or abundance data (for the latter, see Fig S4 in Appendix S1, Supporting Information). Based on these pieces of information, we can state that the most important results of the study (rejection of neutral models, high beta diversity and strong spatial structure from fine to broad scales) are fairly robust to temporal changes and sampling bias; nevertheless, temporal replication of spatial data could provide important information for a more mechanistic interpretation of the main patterns presented in this study. For example, well-replicated temporal data would allow us to test the hypothesis that the various species have demographic peaks in different seasons, which can greatly increase the spatiotemporal region of coexistence by avoiding competition through temporal segregation (Chase & Leibold 2003). Previous data on the study area (Migliorini, Petrioli & Bernini 2002) are fairly consistent with this hypothesis. As for the unmeasured fine-scale variations in environmental variables, note that changes in soil structure have been proposed as crucial determinants for soil animals, which in turn also positively contribute to soil structure (Wardle 2002; Bardgett 2005). We broadly quantified soil structure by the percentage of coarse material (>2 mm) in the soil: this was sufficient to distinguish the soils of the two habitats, but was later found to have a low discriminatory power at finer scales. We also roughly quantified organic matter in a broad sense by measuring its overall amount in the soil by ignition; however, stoichiometric ratios of key elements such as C, N and P may have a much greater explanatory potential (as demonstrated for the most disparate soil organisms), considering that these ratios drive the microbiological community on which most microarthropod species feed (Bardgett 2005). As for the latter point, note that the feeding habits of oribatids and other microarthropods such as springtails have been investigated in detail over the last years using different approaches ranging from food choice experiments to isotopic analysis (e.g. Schneider et al. 2004; Schneider & Maraun 2005). Results indicate that these taxa have different degrees of specialization and highly variable feeding habits. If one of the life traits accounting for niche differences was scale dependent or highly unpredictable, neutral-like and niche processes could interplay over a range of spatial scales, possibly resulting in complex spatial patterns such as the multiple patterns observed in our study. To address this point in the future, we propose that studies such as those conducted by Schneider et al. (2004) and Pollierer et al. (2009), who investigated trophic levels and habits through a careful and standardized use of stable isotope ratios, should be coupled with an analysis of community structure such as the one presented herein. For example, a critical issue is that stable isotopes revealed a continuity of feeding spectra within taxonomic assemblages such as oribatid assemblages, which are classically believed to be relatively homogeneous in terms of feeding ecology. There are a few species that show isotopic signatures close to the litter layer instead; these can thus be considered true decomposers. Within the same assemblage, there are even a few species that show isotopic signatures typical of predators in the ground above. As most species are in an intermediate position, the classical concept of discrete and well-defined trophic levels simply does not apply to soil food webs. Strictly speaking, the neutral theory should be applied to a community of species belonging to the same trophic level and potentially competing for a relatively homogenous set of resources (Hubbell 2001), although this definition is rather broad. Soil animal assemblages such as oribatid mite assemblages exemplify what animal ecologists have recently started to observe in the soil: sets of communities that are intermediate between the classical definitions of the community analysed by neutral models and real systems with several more or less discrete trophic levels. A new generation of theories and observations that take into account multiple spatiotemporal scales and different environments are thus required to unravel the causes of the complex community patterns arising in animal communities because of interacting stochastic and deterministic processes.