Experimental species removals impact the architecture of pollination networks

1. Introduction

Network analysis is a powerful tool for understanding system-level ecological dynamics, such as the effects of perturbations on the stability of ecosystems [1,2]. Despite this potential, our ability to apply network predictions to natural systems remains limited. Modelling has progressed much more quickly than experimentation, particularly in mutualistic networks (e.g. seed dispersal and pollination networks [3,4]), where few experiments have been conducted. A common assumption in mutualistic network models is that following species removals, network linkage patterns remain static except for node losses (e.g. [3–5]), though there are exceptions for both food webs (e.g. [6]) and mutualistic networks (e.g. [7]). But organisms are known to shift their niches in response to changes in inter- and intra-specific competition, as predicted theoretically [8,9] and demonstrated empirically from a very wide range of natural systems [9].

A small but growing body of network modelling work exploring mutualistic interactions has incorporated structural dynamics in response to perturbations, including topological shifts [10,11] and changes in link intensity [12]. But those efforts are disconnected from experiments, which include plant community structuring [13,14], nutrient additions [15] and plant genetic diversity crossed with nutrient additions [16]. To our knowledge, only two experiments have focused on species removals [17,18] and neither removed pollinators.

To address this gap, we experimentally removed single bumblebee species from field plots and assessed architectural changes in flower visitation networks. Previous pollinator removal experiments have shown changes in foraging patterns of remaining pollinators [19,20], but have exclusively focused on bumblebee foraging, not entire pollination network responses. Pollinator removals could have cascading effects—even on pollinators from highly divergent clades and with substantially different dietary niches—but if such cascades occur, and how they may impact network structure, remains unknown. Our species removals focused on bumblebees, as they are abundant and consume a wide range of floral resources in the study area. We removed the most common species in each plot as we expected these to have the largest potential to trigger changes in floral resource use of other pollinator species in the community.

We hypothesized that post-species-removal, the remaining pollinators would expand their foraging niches to use newly available floral resources, with resulting increases in niche overlap between species (e.g. [8,9,19,20]). We thus predicted (in species-removal manipulations relative to controls):

• — decreased pollinator specialization.

• — increased network connectance (proportion of realized links).

• —  increased resource overlap/decreased niche partitioning and complementarity between pollinators.

• —  increased nestedness, given greater connectance and diet overlap.

• — no change in the total number of plant species visited by non-manipulated pollinator taxa, with hypothesized niche increases compensated by higher niche overlap.

2. Material and methods

3. Results

We observed 7637 flower visits to 43 plant species across 15 control-manipulation pairs [23]. Multiple structural metrics were altered in pollination networks following single pollinator species removals (tables 1 and 2; figure 1). Network metrics are reported moving from control to the manipulated state; for all results N = 15 sites. As predicted, we found marginally increased connectance (p = 0.064), increased resource overlap (niche.overlap.HL: p = 0.012) and decreased niche partitioning/complementarity (C.score.HL: p = 0.0020). In contrast with our predictions, we found no effect of specialization (d′, p = 0.8381) or nestedness (p = 0.33), and species richness of plants visited in the manipulated state marginally decreased relative to controls (p = 0.051).

• Download figure
• Open in new tab
• Download PowerPoint

4. Discussion

Following experimental single pollinator species removals, several elements of network structure changed. Consistent with our hypotheses, we found increases in pollinator niche overlap and connectance (marginal), and a significant decrease in niche partitioning/complementarity. Inconsistent with our predictions, we found no effect on nestedness or specialization and marginally fewer plant species were visited in the manipulated state relative to the control state (again, not including the removed pollinator species), i.e. a guild-level niche contraction. This result is particularly surprising given that we simultaneously documented increased connectance. Still, such decoupled niche responses have been documented previously in other taxa in response to removal of an interspecific competitor [24,25].

These network structural changes are likely to have negative effects on pollination functioning. Fewer visited plant species following species removals suggests the potential for negative effects on specialist and/or rare plants. In addition, we found reduced complementarity/niche partitioning, which is known to play a key functional role in the relationship between biodiversity and ecosystem functioning [26].

Conducting manipulative experiments on entire networks is logistically complex and challenging, and our design involved several trade-offs. We held flowering plant resources constant between control and manipulation states, which along with the difficulty of insect removals in the field necessitated characterization of only short-term network responses. Second, we identified pollinators observationally, since destructive sampling of pollinators in one experimental state would have likely impacted the pollinator community sampled in the other experimental state. Third, we removed the most abundant bumblebee in each network, and removal effects are likely correlated to rank abundance; thus, removing less-abundant pollinators may have a weaker effect on network structure. Finally, network metrics measure different, but often subtly related, facets of structure and are often correlated. Of the 15 pairwise correlations between our six metrics (electronic supplementary material, table S1), two were significant (niche.overlap.HL with both C.score.HL and d′) and one was marginal (C.score.HL with d′) with a Bonferroni correction for multiple comparisons.

Our findings highlight several future research directions. First, there is a need to integrate mechanistic ecological processes, going beyond experiments with just two interacting species; investigating indirect effects [27]; and assessing interactions other than competition. Second, more manipulative experiments are needed, particularly spatially replicated studies that manipulate one variable of interest while holding other system variables constant. Third, to enhance the integration of modelling and empirical studies, we suggest models that incorporate systems-level ecological mechanisms as simply as possible, such as the shifts in resource use in response to competition we document here. This can also lead to explicitly field-testable predictions in models [12,27]. By conducting thoughtful, manipulative experiments on ecological networks and by enhancing integration between theory and empirical data, we will improve our ability to understand and predict how ecological networks will respond to the diverse pressures generated by anthropogenic environmental change.

Data accessibility

Network data: http://dx.doi.org/10.5061/dryad.b5h65 [23].

Authors' contributions

B.J.B. conceived of the study; B.J.B., H.M.B. and K.N. designed the study; K.N., H.M.B. and B.J.B. collected field data; K.N. and B.J.B. analysed data; B.J.B. and H.M.B. drafted the manuscript. All authors gave final approval for publication and agree to be accountable for all aspects of the work.

Competing interests

We declare no competing interests.

Funding

U.S. National Science Foundation (DEB-1120572 and DEB-1406262); Emory University (University Research Council; Institute for Quantitative Theory and Methods); UC-Santa Cruz; the Rocky Mountain Biological Laboratory.