Plant-herbivore networks

We produced a comprehensive compilation of sets of local interaction records between host plants and herbivorous insects (S1 and S2 Tables). We used Scopus and Google Scholar to search for studies reporting local plant-herbivore interaction networks using the following keywords: (plant*) and (herbivor*) and (network* or interaction* or web*) and (survey* or list*). In addition, we inspected the Interaction Web Database of the National Center for Ecological Analysis and Synthesis (www.nceas.ucsb.edu/interactionweb) and the cited literature in some key review studies on plant-insect interactions [21–24]. We also included available data from MSc and PhD theses, research reports, and reports from government agencies.

The following criteria were applied to include a plant-herbivore network in our analyses: (1) The original study should provide at least a basic description of the study area to allow categorization of land use intensity; (2) herbivores were recorded on their host plants; (3) all plants should be potential hosts for any herbivorous insect; (4) the network should contain at least five plant and five insect species, totaling at least 10 species; and (5) at least 70% of the herbivores should be identified at the species or genus level.

Land use intensity

In order to assess the degree of habitat alteration to which each plant-herbivore network was subjected, we developed an index of land use intensity. Because most of the compiled studies did not include a detailed description of land use or human-induced modifications that might allow a fine classification of land use intensity, we used a simple classification scheme based on the descriptions of the sampling sites in the original studies. We defined the following four categories of land use intensity: (1) pristine natural habitats (e.g., primary forests); (2) natural habitats moderately impacted by human activities (e.g., secondary forests); (3) agricultural or ruderal habitats with high diversity of plant species (e.g., agroforestry systems and abandoned pastures); and (4) agricultural or urban habitats with low diversity of plant species (e.g., monocultures and urban orchards). This categorization of land use was used as an ordinal variable ranging from 1 to 4.

Exotic plants

For each network, we classified the plant species as native or exotic according to plant databases available for the country where the network was studied (S3 Table). This binary classification of plants according to their origin can be interpreted as a distinction between plants with some (native) or little (exotic) shared evolutionary history with the native host species of herbivorous insects. Plants that were not identified to the species level in the original studies could not be checked for their origin, so we assumed, conservatively, that they were natives (less than 14% of the plant species). No distinction between native and exotic herbivores was made; we assumed that in most networks the large majority of insect species are native to the sampled region. After classifying each plant species in each network as native or exotic, we calculated the proportion of exotic plant species in each plant-herbivore network. To meet normality requirements, we applied the arcsine transformation to the proportion of exotic plants when necessary.

Taxonomic variables

When describing plant-herbivore networks, authors often choose to limit the sampling procedure within a taxon. For example, some networks include only plants from a given family [19], or only insects from a given order [25]. We used the taxonomic span of the sampled plant set and of the herbivorous insects as explanatory variables for network specialization measures. For plant taxonomic levels we used the following categories: Family = 1, Suborder = 2, Order = 3, Superorder = 4, Subclass = 5, Class = 6, Subdivision = 7, Division = 8. For herbivorous insects, the taxonomic levels were Subfamily = 1, Family = 2, Suborder = 3, Order = 4, Subclass = 5, Class = 6.

The taxonomic spans of plants and insects are necessarily related to the phylogenetic diversity of the plant and insect assemblages, respectively. Because of the difficulty to obtain well-resolved phylogenies for insects and plants, we used taxonomic diversity as surrogate of phylogenetic diversity of the plant-insect networks. For this, we used the index Average Taxonomic Distinctness (AvTD), which is based on topological distances between species in the taxonomic tree [26]. This index can be interpreted as the expected number of nodes between any two randomly selected species in the considered community [26]. Schweiger et al. [27] compared different phylogenetic diversity indices and recommended the use of AvTD because it is unbiased by species richness and reflects phylogeny per se. To calculate the AvTD, we used six taxonomic levels for plants (genus, subfamily, family, order, superorder, and subclass) and five taxonomic levels for insects (genus, subfamily, family, superfamily, and order). We used PRIMER 6 software to calculate the AvTD [28].

Estimating network specialization

We used network connectance and the proportion of monophagous herbivores as proxies for network specialization. Connectance is, in fact, an inverse measure of overall interaction specialization because, in networks with higher connectance, species tend to have more connections (i.e., to be less specialized); thus, the higher the connectance, the lower the specialization of plant-herbivore networks. The connectance was calculated as the residuals from a linear regression between the number of realized interactions and the number of potential interactions (both log-transformed) across plant-herbivore networks (hereafter “residual connectance”). By using the residual connectance, we controlled for the negative relationship between network size (i.e., total number of plant and insect species) and connectance, a well-known issue when comparing distinct networks [29]. Residual connectance therefore allows the comparison of different-sized networks in terms of higher or lower connectance than expected based on their size (positive and negative residuals, respectively) [24].

The proportion of monophages (i.e., insects consuming a single plant species) in each network was used as a further measure of network specialization. Note that we measured specialization at the local scale, so that a monophagous species in a given plant-herbivore network may be a generalist herbivore if all its populations are considered [30]. In addition, the proportion of monophages is a good measure of network fragility at the local scale because the coextinction of each monophagous herbivore is caused by the extinction of a single host species.

Data analysis

We used a path analysis approach [31] to quantify the effects of land use intensity and the proportion of exotic host species on the residual connectance and the proportion of monophages in plant-herbivore networks, while controlling for the confounding effects of plant and herbivore taxonomic span and the average taxonomic distinctness of both groups. The path analysis for residual connectance was based on the following rationale (Fig. 1): (1) the index of land use intensity is the explanatory variable of interest and was included in all models as an exogenous variable; (2) the proportion of exotic plants is expected to be positively influenced by land use intensity and to have a positive effect on the residual connectance, being included in all models as both an exogenous and endogenous variable; (3) plant taxonomic span and herbivore taxonomic span are explanatory variables for the residual connectance, for the proportion of exotic plants, and for the AvTD of plants and insects, and both were included as explanatory variables to control for their possible confounding effects on the relationships between the variables of interest; (4) the AvTD of plants is expected to be influenced by plant taxonomic span, the proportion of exotic plants, and land use intensity, and to affect residual connectance, being thus included in the models as both exogenous and endogenous variables; and (5) the AvTD of insects is expected to be influenced by herbivore taxonomic span and by the AvTD of plants and all variables that also potentially influence the AvTD of plants. We used similar models to determine the effects of explanatory variables on the proportion of monophages, and to construct models of the residual connectance and proportion of monophages separately for endophagous and exophagous herbivores.

The model fits were assessed through chi-squared tests and by examination of the Tucker-Lewis Fit Index (TLI), the Comparative Fit Index (CFI), and the Root Mean Square Error of Approximation (RMSEA). Statistical power was calculated using the approach introduced by MacCallum et al. [32] and the R code developed by Preacher and Coffman [33]. Normality was checked using the Mardia test, and outliers were inspected using the Mahalanobis distance. Path analyses were built and fitted in AMOS version 5.0 [34].

Do land use intensity and exotic host plants affect the degree of specialization in plant-herbivore interactions?

Our results show that land use intensification and an increasing proportion of exotic host plant species have opposite effects on the overall interaction specialization of plant-herbivore networks. Contrary to our expectations, land use intensity has a direct negative effect on network connectance and a positive effect on the proportion of monophagous herbivores. The term “direct effect” here means an effect that is not mediated by other variables in the path models, but that can be mediated by non-measured variables. Conversely, connectance increases and the proportion of monophagous herbivores decreases as the proportion of exotic host species increases. Since the proportion of exotic host species tends to increase at higher levels of land use intensity [7–8], the positive direct effect of land use on network specialization is somewhat counterbalanced by the negative effect of exotic host species. Consequently, the overall net effect of land use intensity on the specialization of plant-herbivore interactions is only moderately positive (Tables 2 and 3).

It is possible that network specialization increases with habitat changes, because more specialized herbivorous insects tend to be associated with more resilient plant species [35]. Thus, although specialists would be especially vulnerable to losing their hosts, the resilience of their hosts to human-induced changes might explain their persistence even in strongly altered habitats where many generalists have been lost. In addition, monophages can be more metabolically and behaviorally efficient than generalists in exploiting their host plants [30].

Another explanation for the increase in network specialization in highly impacted habitats could be that, at the local scale, generalist insects tend to become specialized in a few plant species after habitat conversion by human activities. In fact, spatial and temporal variation in the use of host plants might cause herbivores to have few local host species [15, 30]. Although we did not include total richness of native plants (including the non-host plant species) in the models (because that information was not available in the compiled studies), we have strong expectations that it should decrease in more intensively modified habitats. Thus, it is possible that specialists in habitats with high land use intensity are not true specialist insects, but generalists constrained to be local specialists due to the lack of alternative food resources [15]. We therefore suggest that the modifications promoted by land use, through the reduction in the number of potential native host plant species, might restrict the diet of the herbivores and enhance the occurrence of local specialists.

From the standpoint of herbivorous insects, an increase in network specialization can be interpreted as a decline in the number of alternative host-plants, which increases the probability of secondary extinctions of herbivores following the loss of host species [12, 25]. This might be one of the mechanisms by which extreme levels of habitat conversion (as in croplands and pastures) greatly increase the vulnerability of host-parasitoid networks [36]. In plant-herbivore networks, previous results show a positive relationship between high network connectance (i.e., decreased specialization) and increased herbivore persistence [37, 38].

Land use intensity also has a negative indirect effect on specialization mediated by an increase in the proportion of exotic plants. An explanation for our results is that the herbivore fauna associated with exotic plants is restricted, composed of few species [19] that are usually highly polyphagous [39]. An expected effect of exotic plants on the structure of plant-herbivore networks is the replacement of specialist herbivores with generalists [40]. Herbivorous species that have a restricted host range are more sensitive to exotic plant invasion [8]. As a result of the increasing richness and abundance of exotic plants, the richness of more specialized herbivores diminishes rapidly, but the number of interactions diminishes more slowly, because the loss of each specialist removes few interactions. Therefore, the number of realized interactions diminishes much more slowly than the number of potential interactions. This process of disproportional loss of herbivore specialists leads to a net reduction of specialization in plant-herbivore networks.

Discrepant effects on plant-herbivore networks composed by endophages vs. exophages

Our findings provide evidence that external and internal plant feeders have different responses to land use intensity. In the present study, we found that land use intensity has direct and indirect effects on the community-wide specialization of endophagous insects, but little or no effect on the community-wide specialization of exophagous insects.

Endophages tend to be more specialized than exophages [20, 41], and thus, they have a higher degree of dependence on a smaller set of host species [21, 42]. Furthermore, because they develop inside the plant during their immature stages, endophages are better protected than exophages against external microclimatic changes caused by human action. For those reasons, it is expected that the effects of land use intensity on endophages occur mainly via changes in host plant composition. The replacement of native host species with exotic ones has a greater impact on endophagous insects, as some exophages can colonize the novel hosts more rapidly [21].

We recently showed that an increase in the proportion of exotic plants has a greater impact on the diversity of endophagous than on the diversity of exophagous herbivore insects [40]. We also observed this pattern in the current study, as land use intensity had significant effects only on those networks composed of endophagous insects. These insects are most affected because exotic plant species often lead to a decline in the populations of native plants, including species that were dominant before the invasion [9]. For generalist herbivores, this effect might be buffered by relatively broad diets, allowing them to tolerate the loss or reduction of important host species by compensating or switching to alternative hosts.

How does the sampled taxonomic range affect network specialization in plant-herbivore networks?

Several studies on plant-herbivore interaction networks have shown that interactions between species are phylogenetically compartmentalized [2, 43]. Phylogenetically related plants tend to be consumed by phylogenetically related insect species, such as folivorous Lepidoptera in the Brazilian Cerrado [2]. Researchers usually set the taxonomic range of the studied network arbitrarily, and the diversity of recorded interactions is likely to depend on the taxonomic inclusiveness of both plants and insects. However, we are not aware of previous studies that investigated whether the choice of range of taxonomic groups of plants or insects influences estimates of specialization in plant-insect networks.

Our results show that estimates of network specialization are negatively affected by herbivore taxonomic range. Thus, networks studied over a broader taxonomic range (i.e., class or order) are likely to have a smaller proportion of realized interactions than networks that span a more restricted taxonomic unit (i.e., family or genus). This can be explained by the fact that interaction connectivity within a given network compartment is greater than interaction connectivity between the different compartments [44]. The sampling of more inclusive taxonomic levels leads to an increase in the number of compartments in the networks, each of them only weakly linked to the others.

Phylogenetic diversity (as measured by AvTD) of herbivores also influenced network residual connectance; however, the effect was positive. This might occur because phylogenetically diverse assemblages are more likely to contain super-generalist herbivores. Additionally, when sampling a wider taxonomic span of herbivores (e.g., orders) in phylogenetically diverse assemblages, researchers are likely to sample a smaller fraction of the diversity of these taxa, with the overrepresentation of generalist species leading to higher perceived network specialization.