Our study provides a first integrated analysis of structural patterns among networks of antagonistic interactions quantified in different parts of the world. Drawing on these diverse data, our explicit control of sampling effects allows us to critically evaluate hypotheses emerging from current theory on geographical patterns in species diversity and proposed trends in ecological specialisation. Overall, we found strong support for the prediction that network specialisation appears higher in studies focusing on smaller interaction matrices. There was less support for the prediction that taxonomically diverse networks will exhibit high levels of network specialisation: although significant changes in network structure were observed as taxonomic diversity of networks increased, the patterns detected were not consistently indicative of higher specialisation. Most surprisingly, we found no support for the hypotheses that network specialisation increases towards the tropics, or that – when controlling for latitude – specialisation within networks varies depending on the ecological guild of the component host species. In this regard, these antagonistic host-parasitoid networks seem to differ from networks of mutualistic interactions (Schleuning et al. 2012), suggesting that different processes may be moulding the architecture of antagonistic and mutualistic networks. Below, we discuss each of these findings in turn.
Sampling issues in network studies
Both within and across the large number of antagonistic networks examined, we found a strong association between network size and measures of network structure. These patterns concerned network size in the sense of both the number of interactions examined and the taxonomic coverage of hosts. While previous network studies have addressed the importance of sampling intensity (sensu mean number of interactions per possible link in a network) and network dimensions (sensu number of species), they have not addressed the impact of the taxonomic diversity (Blüthgen et al. 2007, 2008; Dormann et al. 2009). Our approach of correcting for both matrix size and host taxonomic diversity allowed us to separate statistical patterns from underlying biology, revealing more evidence for the former.
Our findings come with important implications for network studies across habitats and regions. In recent years, there has been a marked increase in the number of studies involving the collection of replicated interaction networks between experimental treatments, or across ecological or land-use gradients (e.g. Schleuning et al. 2011; Devoto et al. 2012; and references in Supplementary Information). While the magnitude of variation in matrix size between replicates within individual studies is likely to be much smaller than among studies in our global data set, there is nevertheless a risk that methodological artefacts will affect the results. Some authors have controlled for differences in host abundance or host species richness among networks in their analyses (e.g. Tylianakis et al. 2007), but others have made uncontrolled comparisons. In future work, we propose that ecologists should control for potential effects of matrix size before conducting comparisons – just as ecologists comparing species richness between sites or treatments routinely standardise sampling effort using approaches such as rarefaction (Gotelli & Colwell 2001).
Standardising for effects of taxonomic diversity among networks may prove more problematic. For this reason, we suggest that any patterns emerging from studies integrating webs of highly differing taxonomic diversity should be supported by more in-depth comparisons between webs of similar taxonomic breadth. Alternatively, the effects of taxonomic breadth may be isolated by using descriptors such as Δ (the taxonomic diversity index; Clarke & Warwick 1998) as a covariate, or by simply subsampling taxa from larger networks before comparison to metrics from smaller networks. The latter approach will be similar to that proposed above for compensating for matrix size.
Latitudinal patterns in network structure and specialisation
While communities vary tremendously in species richness and taxonomic complexity, our analysis of antagonistic interaction networks offers no evidence for consistent variation in network structure with latitude. This contrasts with the findings of Baiser et al. (2012) for aquatic networks inhabiting pitcher plants. Focusing on qualitative descriptors of network structure, these authors found increasing linkage density with increasing latitudes across North America – a pattern which they attributed to increasing species richness. However, for terrestrial insects, an often-assumed decrease in niche breadth (i.e. increasing diet specialisation) with decreasing latitude (Dyer et al. 2007) is expected to result in more specialised networks in the tropics. For this hypothesis, we found no support. Rather, the fact that network structure does not vary consistently with latitude suggests that there may be basic rules for how antagonists interact with each other across the globe, irrespective of the size and diversity of the networks that they comprise.
The lack of variation in the structure of the networks with latitude observed in our study across a global scale is consistent with the patterns observed by Kaartinen & Roslin (2011, 2012) at a landscape scale. These authors found pronounced spatiotemporal consistency in the structure of tens of host-parasitoid networks, despite major variation in both species richness and species identity in space and time. Thus, antagonistic networks may be similarly structured at both global and landscape scales.
Our findings from antagonistic networks contrast with reports of latitudinal patterns in the structure of mutualistic networks. Studying hummingbird-plant networks across the Americas, Dalsgaard et al. (2011) found greater biotic specialisation at lower latitudes. In contrast, Schleuning et al. (2012), using a global data set of quantitative mutualistic networks, found that biotic specialisation of mutualistic networks was significantly lower at tropical than at temperate latitudes. These authors attributed the pattern detected to a response of pollinators and seed dispersers at high latitudes to low plant diversity, since specialisation decreased with increased plant diversity. In our study, the same argument could be used to predict increasing specialisation of parasitoids towards the host-poor Arctic, but no such patterns were evident after correcting for matrix size and taxonomic diversity. We therefore suggest that host-parasitoid networks are structured by other forces – perhaps by basic constraints on the ability of parasitoids to locate diverse sets of hosts or to cope with their behavioural or immune responses.
Since Schleuning et al. (2012) used a metric designed to be insensitive to matrix size (Blüthgen et al. 2006), the discrepancy in latitudinal patterns among different types of interaction networks appears to be real rather than methodological. As such, it suggests that different forces structure networks of different interaction types. The specialisation of our host-parasitoid networks also appears much higher overall (mean H2' = 0.65, with a high s.d. of 0.31) than the equivalent values for mutualistic webs (Fig. 1c in Schleuning et al. 2012). Consequently, even if network specialisation does not change consistently with latitude for antagonistic networks, it may still be higher across antagonistic (specifically host-parasitoid) than mutualistic networks – regardless of latitude. A study by Poisot et al. (2011) found the opposite pattern, with mutualistic networks more specialised than antagonistic networks; although here, too, there was larger variation among the latter. However, the networks included in the study of Poisot et al. included antagonistic and mutualistic interaction types well beyond those included in the present study, and well beyond those studied by Schleuning et al. (2012). This lack of standardisation in interaction types may have contributed to the observed pattern.
Again, a general discrepancy between plant-pollinator or seed-disperser vs. host-parasitoid systems may be attributable to different constraints on resource selection. For parasitoids, host selection is likely to be affected by traits involved in host location and host defences, thus potentially restricting host use to fewer and more similar taxa than plants used by pollinators or seed dispersers.
At a general level, the idea that antagonistic and mutualistic networks may be structured differently is far from new. The stability of mutualistic networks is thought to be promoted by a highly connected and nested architecture, whereas stability in antagonistic networks is promoted by a compartmented and weakly connected structure (Thébault & Fontaine 2010; but see also James et al. 2012, 2013; Saavedra & Stouffer 2013). Highly connected networks will – by their nature – be characterised by lower specialisation than weakly connected antagonistic networks; as a consequence, networks of these two different types are likely to respond differently along a latitudinal gradient of species richness. If structure is related to stability, then the patterns unravelled here and by Schleuning et al. (2012) suggest that mutualistic networks may be more stable towards the tropics, whereas antagonistic networks (or at least those involving insect hosts and parasitoids) may be equally stable across latitudes.
Network specialisation and insect host guilds
Across our data set, we found no evidence for network structure differing consistently among host guilds. This observation runs contrary to the expectation that the morphological and chemical defences of the host would affect specialisation at the level of parasitoids (Gauld et al. 1992; Hawkins 1994) and thus networks. Indeed, our analysis is among the first to address specialisation of parasitoids at the network level, but provides no evidence that specialisation in antagonistic networks is affected by the ecological guild of the host species, after accounting for variation in the size and diversity of the networks. In this context, we stress that the conclusions reached are conditional on the data examined: while our data set encompassed representatives of five major insect host guilds, analysing host-parasitoid networks across a wider range of host guilds might still reveal differences in network specialisation.