Island biogeography of mutualistic interaction networks

Authors


Abstract

Aim

The seminal theory of island biogeography, based on changing rates of immigration and extinction, should be seen in a geological context, as an island's maturity influences the richness of its biota. Here, we develop an island biogeography of biotic interactions, recognizing that, besides species richness, biodiversity also encompasses the multitude of interactions among species. By sampling interactions between plants and pollinators across the Canarian archipelago, we illustrate how the local richness, specialization and endemism of biotic interactions vary with island age and area.

Location

Canary Islands (27.62° N–29.42° N and 13.33° W–18.17° W).

Methods

On five islands, covering the full age range of the archipelago, plant–pollinator interactions were catalogued and their strength estimated. Network parameters (e.g. interaction richness and specialization) and the number of single-island interactions (equivalent to single-island endemics) were estimated from interaction matrices and related to island area and age.

Results

Plant species richness, interaction richness and average degree of specialization of pollinator species showed hump-shaped relationships with island age. Pollinator richness varied with island area and plant richness. Plant specialization increased with island age, and the proportion of single-island interactions (pSII) exhibited a U-shaped relationship with age.

Main conclusions

The previously reported hump-shaped relationship between species richness and island age, both on the scale of islands and of habitats, was confirmed for plant species in local networks. Both plants and pollinators were more generalized on the youngest island, which may be due to a predominance of generalist colonists. Pollinator specialization peaked on mid-aged islands, whereas plants showed the highest specialization on old islands, potentially reflecting their different life histories. The U-shaped relationship between the proportion of single-island interactions and island age might be explained by (1) young islands having a high proportion of unique interactions, due to interactions between generalists, and (2) old islands having unique interactions due to an accumulation of unique pairwise interactions that have evolved through time. Thus, island age – which not only captures time per se, but also the geomorphological changes of islands – may act as a regional driver of local network structure, and so the contemporary networks we observed across the Canarian archipelago illustrate the development of a network through geological time.

Introduction

Islands are fascinating because they offer outstanding opportunities in the study of community assembly and evolution (Brown & Lomolino, 2000; Emerson, 2002; Losos & Ricklefs, 2009). Consequently, important ideas about the structure and dynamics of island biotas have been developed; of these, the theory of island biogeography (MacArthur & Wilson, 1967) has remained the most influential. MacArthur & Wilson's (1967) theory was recently expanded to incorporate island ontogeny, as island age was envisaged to have a strong impact upon the biota of islands (Gruner, 2007; Stuessy, 2007; Whittaker et al., 2008). Whittaker et al. (2008) developed a general dynamic model of oceanic island biogeography (GDM) which describes how rates of immigration, extinction and speciation – and consequently species richness – might change with island age and area. Thus, their model combines MacArthur & Wilson's (1967) theory with information about island ontogeny. As new volcanic islands emerge, they increase in area and elevation, but eventually their growth is surpassed by erosion and sometimes substantial subsidence, resulting in shrinkage and final submergence (Menard, 1983; Stuessy, 2007; Whittaker et al., 2008). One of the strongest predictions of the GDM is a hump-shaped relationship between species diversity and island age due to changing rates of immigration, speciation and extinction during island development. However, because change in species richness has not been catalogued for any island over longer time-spans, we have to rely on indirect information. This may be provided by archipelagos created by a mantle hotspot, as their individual islands differ in age and therefore offer temporal snapshots of island development. The GDM predicts that:

display math

where diversity is either species richness or the number (nSIE) or proportion (pSIE) of single-island endemics (SIE: species endemic to one island). These variables have been used as metrics offering insight into the evolutionary dynamics of species on islands and as a conservation measure (e.g. Emerson & Kolm, 2005; Fattorini et al., 2012). The GDM has been tested empirically and generally shows good explanatory power across systems (Borges & Hortal, 2009; Fattorini, 2009; Bunnefeld & Phillimore, 2012; Steinbauer et al., 2012; Cameron et al., 2013). However, the model has shortcomings and divergent patterns may be seen because of, for example, taxa with specific habitat preferences (e.g. cave-dwelling animals, whose available habitats disappear rapidly with island age), varying dispersal abilities of the studied species, or archipelago age (e.g. a decrease in species richness with age might not be observed in young archipelagos because old eroded islands are not yet present) (Borges & Hortal, 2009; Cameron et al., 2013). Steinbauer et al. (2012) focused upon vegetation zones in the Canaries and showed that, especially within ‘coastal scrub’ and ‘thermophilous forest’, pSIE followed a hump-shaped pattern with island age. As noted by Whittaker et al. (2008, 2010), the GDM is a simplification of the plethora of processes influencing island biotas, but it serves in the present study as an important platform in our exploration of network topology in an island biogeographical context.

There is strong interest in comprehending the factors that shape species richness on islands, and island area, isolation, age, habitat heterogeneity and species diversity per se receive the most attention (Borges & Brown, 1999; Emerson & Kolm, 2005; Borges & Hortal, 2009; Santos et al., 2010). However, only Sugiura (2010) has placed entire ecological networks of interacting species into an island biogeographical context. For the Japanese Ogasawara Islands, he showed that the number of interactions in ant–plant networks increased with island area, whereas nestedness (a pattern in which interactions of specialized species are subsets of the interactions of more generalist species) and connectance (the proportion of realized interactions in a network) decreased with area. Therefore, island or regional characteristics influenced the structure of local ecological networks. Here, we combine the approaches developed by Whittaker et al. (2008) and Sugiura (2010) and analyse the influence of island age and area on the topology of local pollination networks studied across the Canary Islands. By doing so, we move beyond mere species numbers and look into the detailed network structure, by focusing upon interaction richness, species specialization level, and the number and proportion of single-island interactions (nSII and pSII, i.e. interactions only observed on a single island, equivalent to the species richness parameters nSIE and pSIE, see above).

We put forward the following hypotheses.

  1. That network species richness of plants and pollinators shows a hump-shaped pattern with island age. If so, patterns on the scale of islands (Whittaker et al., 2008) and of habitat (Steinbauer et al., 2012) are discernible at a finer scale of resolution, i.e. in local communities of interacting species [equivalent to Holt's (1992) ‘within-sample species richness’].
  2. Because nSIE has a hump-shaped distribution with island age and such species are, by definition, involved in interactions not observable on other islands, we also expect a hump-shaped relationship between pSII and island age.
  3. If colonization of young islands is favoured by the possession of generalist characteristics (Richardson et al., 2000; Piechnik et al., 2008), and if those species that are most inclined to survive on old eroded islands are the most generalist ones (Aizen et al., 2012; Hagen et al., 2012), then a hump-shaped relationship between specialization and island age can be anticipated.

Materials and methods

The Canary Islands

Seven large islands in the Atlantic Ocean form the Canarian archipelago (27.62° N–29.42° N and 13.33° W–18.17° W). In order to cover the full age range of the archipelago, fieldwork was conducted on five islands, with one locality on each of El Hierro (27.8047° N, 17.8959° W), La Gomera (28.0398° N, 17.2267° W), Gran Canaria (27.9041° N, 15.4331° W) and Fuerteventura (28.5643° N, 13.8919° W), two on Tenerife [Teno Bajo (28.3531° N, 16.9123° W) and Fasnia (28.2222° N, 16.4173° W)], and finally one locality on the African mainland (Western Sahara; 26.1610° N, 14.4222° W). Table 1 summarizes island characteristics, i.e. maximum age of subaerial material and island area.

Table 1. Characteristics of the Canary Islands included in the study
 El HierroLa GomeraTenerifeGran CanariaFuerteventura
  1. a

    Guillou et al. (1996);

  2. b

    Ancochea et al. (2006);

  3. c

    age of Anaga, Teno and the Central Shield, respectively; Guillou et al. (2004);

  4. d

    van den Bogaard & Schmincke (1998);

  5. e

    Carracedo et al. (1998), Stillman (1999).

Max. age of subaerial material (Ma)1.12a10.8b4.9/6.2/11.9c14.5d20.6e
Area (km2)280380205815321677

The island of Tenerife is the result of a coalescence of three shield volcanoes, namely Anaga, Teno and the Central Shield, with estimated maximum ages of 4.9, 6.2 and 11.9 Ma, respectively (Guillou et al., 2004). Anaga and Teno constitute the north-eastern and north-western part of Tenerife, respectively, and the Central Shield forms a possibly larger core part of Tenerife, with Roque del Conde in the south-west as its most striking outcrop (Guillou et al., 2004; Longpre et al., 2009). These three pre-islands merged during the development of the Las Cañadas edifice between 1.9 and 0.2 Ma (Longpre et al., 2009). The present disjunct distribution of some species groups reflects this dramatic geological history (beetles: Emerson et al., 1999; skinks: Brown et al., 2000; geckos: Gübitz et al., 2000; grasshoppers: Hochkirch & Görzig, 2009). The phylogeny of Tarentola delalandii (Gekkonidae) populations on Tenerife, for example, shows three major clades, each located on or near each of the three pre-islands (Gübitz et al., 2000). Thus, populations of species living on these pre-islands have to some extent remained isolated. Our study sites on Tenerife were located at Teno Bajo and Fasnia, i.e. sites located on the pre-islands Teno and Central Shield. Because these two islands vary considerably in age, and ‘island age’ is one of our main predictor variables, we used the ages of these pre-islands (Guillou et al., 2004). Thus, our study islands are El Hierro, La Gomera, Teno, Central Shield, Gran Canaria and Fuerteventura. Although Gran Canaria experienced violent geological events during the Roque Nublo eruptions 5.3–3.5 Ma, research suggests that it is unlikely that Gran Canaria was completely sterilized during these events (Anderson et al., 2009). Thus, we used the maximum age of the subaerial material of Gran Canaria, i.e. 14.5 Ma.

In general, the use of single age estimates is a rough simplification, because volcanic islands are built up over extended periods within which there are episodes of inactivity and occasional landslides (Anderson et al., 2009; Longpre et al., 2009). At present, however, these are the best estimates.

Study habitat and design

Our main criterion of choice of study site was the presence of a large population of the shrub Euphorbia balsamifera Aiton, which occurs in Western Sahara and on all the islands in the Canaries. Besides E. balsamifera, most localities shared several other perennial plants, e.g. Launaea arborescens, Lycium intricatum, Kleinia neriifolia, Fagonia albiflora, Periploca laevigata and Rubia fruticosa. Thus, the localities were taxonomically and structurally similar in their plant communities. Euphorbia balsamifera is dioecious, and the two sexes attract somewhat different pollinator species (K.T., pers. obs.), for which reason, they were treated separately in the interaction matrices.

At each locality, two replicated areas 50–200 m apart were surveyed, adding up to a total of seven localities and 14 replicates. The centre of each replicate had a large stand of E. balsamifera, and the shape and size of the replicates was selected to include as many as possible of the perennial plant species in the area (average size ± SD was 5686 ± 1930 m2). However, even between replicates at the same locality the plant species composition varied. Each replicate was visited twice between 16 January and 28 March 2010 and every flowering perennial plant species was surveyed in four intervals of 15 minutes each at each census. The surveyed individuals were selected randomly, although if flower number per plant varied strongly we chose only from the most flower-rich individuals. If a plant only flowered during one of the two visits, it was surveyed for 1 hour whereas if it flowered at both visits it was surveyed for a total of 2 hours. Likewise, plant species might be present in one or both replicates at a locality. In total, 42 observation days, totalling 295 hours of observation were spent collecting flower visitation data.

Every flower visitor landing on or sitting in the flowers was operationally defined as a pollinator. As identification in the field was not possible for most species, specimens (c. 1300) were collected for later identification.

Network parameters

All visitation data were organized into interaction matrices, with one matrix per replicate, having pollinator species in rows and plant species in columns. Richness of pollinator and plant species in a given network are abbreviated A and PSR, respectively. Matrix cells had a non-zero element (aij) whenever pollinator species i visited plant species j. Here, aij represents interaction strength, defined as number of visits per flower per 15 minutes. Interaction richness (I) is the number of observed pairwise interactions in the whole network, and network connectance (C) is the proportion of potential interactions actually observed: I/(A×PSR).

Specialization at the network level was measured with a standardized two-dimensional Shannon entropy index (H2′), whereas specialization at species level was measured by a standardized Kullback–Leibler distance (d′) (Blüthgen et al., 2006). They are:

display math

where m is total sum of interaction strengths of the quantitative matrix, di is the degree of specialization of species i, Ai is the total of row i, and Aj is the total of column j. The advantages of these indices are that they are based on quantitative data, and that the specialization level of a species includes information about the specialization level of its partners (see Appendix S1 in Supporting Information). Both H2′ and d′ are standardized to range from 0 (highly generalized) to 1 (highly specialized). The average standardized degree of specialization of pollinators and plants are abbreviated as dpol and dpla, respectively. Specialization indices were calculated with the package bipartite in R 2.15.1 (R Development Core Team, 2012).

Single-island interactions (SII) were defined as interactions only observed in the network(s) (one or both of the two replicates) on a single island and, as such, were not observed on any other island or in the networks from Western Sahara (we used the pre-island approach and thus regarded Teno Bajo and Fasnia as being part of two distinct islands). SII is the interaction analogue to single-island endemic species (SIE) used in studies of island biogeography (e.g. Emerson & Kolm, 2005; Borges & Hortal, 2009; Steinbauer et al., 2012). As species and their interactions were matched across the archipelago, and SIIs were strongly influenced by the occurrence of uniquely observed species, i.e. species only observed in the network(s) on a single island, we applied the following procedure. First, in all analyses involving SIIs, interaction matrices were restricted to species for which the taxonomic affiliation (either species or morphospecies) was certain, reducing the problem of less well-resolved species influencing the overall results. Second, we investigated to what extent the changes in pSII across the archipelago were driven by uniquely observed species. In order to do this, we used a variable (pSIIres) consisting of the residuals from a generalized linear model with a binomial family between pSII and the number of unique species on a given island.

Additionally, we intended to investigate how the nestedness (Bascompte et al., 2003; Almeida-Neto et al., 2008) and modularity (Olesen et al., 2007) of interaction networks varied with island age and area, but none of the networks were significantly nested and only a few were significantly modular. NODF was used as nestedness value and empirical values were assessed against 1000 matrix randomizations using null model Ce in the program aninhado (Guimarães & Guimarães, 2006; Almeida-Neto et al., 2008). We used NetCarto with simulated annealing as the optimization algorithm to test for modularity, and the empirical values were assessed against 100 matrix randomizations (Guimerà & Amaral, 2005).

Statistical analysis

Analytical approach

As the biota on each island has experienced distinct historical processes, the two replicates on each island are not independent. Thus, in order to reduce any effects of pseudoreplication, we used mixed effect models with ‘islands’ (El Hierro, La Gomera, Teno, Central Shield, Gran Canaria and Fuerteventura) as a random factor [Western Sahara was excluded from these analyses]. This analytical approach has gained a foothold in island biogeography as a way of accommodating the specific statistical challenges in island ecology (Bunnefeld & Phillimore, 2012; Steinbauer et al., 2012; Cameron et al., 2013). Area and time are important variables influencing species diversity and network properties (Whittaker et al., 2008; Sabatino et al., 2010; Sugiura, 2010; Fattorini, 2011; Triantis et al., 2012) and we therefore expected these variables to have a strong influence on our mutualistic networks across the Canary Islands. We emphasize that our focus is within-sample species richness (sensu Holt, 1992), i.e. we do not analyse species richness of entire islands but species richness in local communities.

We applied an information-theoretical approach based on the Akaike information criterion (AIC) to evaluate our models (Burnham et al., 2011). Models with the lowest AIC value had the highest explanatory power, although models within ΔAIC < 2 (compared to the best model) were considered to be among the most parsimonious ones. Furthermore, model probabilities (Akaike weights, wAIC) and variable importance (sum of the Akaike weights for models in which the fixed parameter occurs; Cameron et al., 2013) were used to validate the relative likelihood of each model and variables, given the current data set (Burnham et al., 2011). Mixed effect modelling was performed using the lme4 package (v. 0.999999-0) in R 2.15.1 (R Development Core Team, 2012), and we used a binomial family when dealing with parameters expressed as percentages (pSII). The dredge function in the package MuMIn (v. 1.9.0) was used to run models covering all possible combinations of the fixed parameters.

Correction of network variables

Generally, interaction richness (I) and connectance (C) correlate strongly with number of species in networks (Jordano, 1987), and the number of pollinator species (A) is a robust predictor of I, whereas the number of plant species (PSR) is a robust predictor of C (Trøjelsgaard & Olesen, 2013). This was also the case in this study, as I correlated strongly with A [adjusted R2 (R2adj) = 0.81, = 46.9; < 0.001] and C correlated strongly with PSR (R2adj = 0.84, = 59.8, < 0.001). Thus, in order to correct these variables for any confounding effects of species richness, we used Icorrected and Ccorrected, which were the residuals from linear regressions of I against A, and C against PSR, respectively.

As network asymmetry [i.e. species asymmetry given as: (A – PSR)/(PSR)] influences the specialization indices dpol and dpla (Blüthgen et al., 2007), we examined how this might affect our results by using corrected metrics (abbreviated dpol,corrected and dpla,corrected), derived from the residuals from a linear regression of dpol and dpla against network asymmetry.

Sampling intensity

The comparison of ecological networks is complicated by variation in sampling effort (Nielsen & Bascompte, 2007; Hegland et al., 2010; Rivera-Hutinel et al., 2012; Trøjelsgaard & Olesen, 2013), and the ‘true’ total numbers of species and interactions in a given community might not be reached even after several years of sampling (Petanidou et al., 2008; Chacoff et al., 2012). It is important to note, however, that a cross-community comparison, like the current study, is more dependent on equal sampling intensity than on the full species count (Olesen et al., 2011).

Although accumulation curves for pollinator species suggest that the asymptotic richness was not reached for any of our networks, they display quite similar shapes, suggesting that we reached the same level of species accumulation in all networks during our census (see Appendix S2). Moreover, sampling intensity at each replicate, measured as the sampling incompleteness (expected number of unobserved pollinators, calculated with the Chao2 estimate; Chao et al., 2009; Chacoff et al., 2012), did not change significantly with any island characteristic or network parameter (Appendix S2). Thus, sampling intensity was regarded as being approximately equal across the archipelago and assumed not to confound our results.

Results

Species and interaction richness

For all networks (including those from Western Sahara), we list the estimated network parameter values in Appendix S3. The Western Saharan networks were excluded from the mixed effect analyses, because we did not find any meaningful way of incorporating the age and area of this locality. Furthermore, we investigated the nested and modular structure of all our networks, but none were nested and only a few were significantly modular when compared to randomized networks. One likely explanation for these findings is the relatively small size of the networks (number of pollinator and plant species varied over the ranges 38–68 and 7–18, respectively). Therefore, we did not explore these network metrics further in relation to island ontogeny.

Plant species richness (PSR) in each network correlated significantly and positively with pollinator species richness (A) (R2adj = 0.35, = 7.0, = 0.024), and plant species had more interactions than pollinators (mean ± SD of network averages; plants: 7.4 ± 1.2 links and pollinators: 1.7 ± 0.2 links). The most parsimonious model describing changes in network plant species richness included both Age, Age2 and log(Area), and the support for this model in terms of wAIC was about twice as high as for the second-best model (Table 2). This suggested that networks on younger and older islands had lower plant species richness than mid-aged islands, but also that island area played an important role (Fig. 1a). Network pollinator richness, on the other hand, was more susceptible to variation in area (Fig. 1b), because log(Area) occurred in all selected models and was the only fixed parameter in the highest-ranked model and thereby also received the largest relative importance. Interaction richness (Icorrected) was especially dependent on island age, because both Age and Age2 occurred in the two most parsimonious models, and both had a relative importance almost three times that of log(Area). This suggested that Icorrected changed in a hump-shaped way across the archipelago, with mid-aged islands having the largest Icorrected (Fig. 1c). Icorrected did not correlate significantly with PSR (R2adj = 0.21, = 4.0, = 0.07), despite both metrics displaying qualitatively similar relationships with island age. Network connectance (Ccorrected) was poorly explained by both Age and Area, as the most parsimonious model did not include any fixed parameters. The no-parameter model had a wAIC (0.387) that was more than twice as large as the second highest-ranked model, which included log(Area) (wAIC = 0.144). Thus, Ccorrected was not influenced by any of the studied island characteristics.

Table 2. Model performance of island characteristics in explaining variation in network parameters across the Canary Islands. Listed are the most parsimonious (ΔAIC < 2) linear mixed effect models, all including islands as a random effect. Also provided are estimates of the fixed parameters, AIC, ΔAIC, AIC weights (wAIC) and variable importance (the sum of the wAIC for the models in which the fixed parameter occurs, as defined in Cameron et al., 2013)
ResponseExplanatory n InterceptAgeAge2Log(Area)AICΔAICwAIC
  1. Icorrected, number of interactions corrected for number of pollinator species; dpol and dpla, average specialization of pollinator and plant species, respectively; dpol,corrected and dpla,corrected, specialization indices corrected for network asymmetry [(A – PSR)/(A + PSR)]; nSII and pSII, number and proportion of single-island interactions, respectively; pSIIres, proportion of single-island interactions corrected for the number of uniquely observed species. *Notice that when using pSII as response variable, we employed a generalized linear mixed-effect model with a binomial family and our islands as random effect, and that we used binomial proportions as our variable (nSII, number of interactions – nSII). Furthermore, H2′ and Ccorrected were not included in the table because the highest-ranked model among the most parsimonious models did not include any parameters, suggesting a minimal effect of island characteristics on these network variables.

Plants (PSR)Age + Age2 + log(Area)12−1.3700.605−0.0424.19457.49600.535
Plants (PSR)Age2 + log(Area)12−3.211 −0.0185.75559.2411.7450.223
 Variable importance:0.5350.7580.758   
Pollinators (A)log(Area)126.197  14.27386.68800.280
Pollinators (A)Age2 + log(Area)12−0.889 −0.02017.63786.8430.1550.259
Pollinators (A)Age + Age2 + log(Area)122.5291.124−0.06414.73887.6470.9590.173
Pollinators (A)Age + log(Area)120.413−0.334 17.39887.7951.1070.161
 Variable importance:0.3340.4320.873   
I corrected Age + Age212−2.9841.939−0.116 83.86300.435
I corrected Age + Age+ log(Area)127.9402.285−0.126−4.37785.1651.3030.227
 Variable importance:0.6620.6620.227   
d pol Age + Age2120.3720.017−9 × 10−4 −46.16300.615
d pol Age + Age2 + log(Area)120.3190.015−9 × 10−40.021−44.9771.1860.340
 Variable importance:0.9550.9550.340   
d pla Age2120.473 6 × 10−4 −26.20300.333
d pla Age120.4240.013  −25.5920.6110.245
d pla Age + Age2120.4530.0054 × 10−4 −24.4551.7480.139
d pla Age2 + log(Area)120.469 6 × 10−40.001−24.2041.9990.123
 Variable importance:0.3840.5950.123   
d pol,corrected Age + Age212−0.0420.009−3 × 10−4 −55.82000.488
d pol,corrected Age + Age2 + log(Area)12−0.0750.008−3 × 10−40.013−54.5581.2620.260
 Variable importance:0.7480.7480.260   
d pla,corrected Age12−0.0810.007  −23.16000.303
d pla,corrected Age + Age212−0.1120.016−4 × 10−4 −21.9551.2050.166
d pla,corrected Age212−0.043 3 × 10−4 −21.6331.5270.141
d pla,corrected Age + log(Area)12−0.1210.007 0.015−21.2151.9450.115
 Variable importance:0.5840.3070.115   
nSIIlog(Area)123.996  12.29190.84100.295
nSIIAge2 + log(Area)12−1.862 −0.01615.07391.9841.1430.166
nSII 1241   92.2581.4170.145
nSIIAge + log(Area)12−0.588−0.265 14.76892.4521.6110.132
 Variable importance:0.1320.1660.593   
pSII*Age + Age2120.417−0.0630.004 14.78500.314
pSII*Age2120.137 0.001 15.1090.3240.267
pSII*Age + Age2 + log(Area)12−0.019−0.0760.0040.17116.2441.4590.151
 Variable importance:0.4650.7320.151   
pSIIresAge212−0.095 0.001 −0.33800.225
pSIIresAge2 + log(Area)120.455 0.001−0.194−0.0340.3040.193
pSIIresAge + Age2120.022−0.0270.002 0.5810.9190.142
pSIIresAge12−0.1250.013  1.0221.3600.114
pSIIres 120.017   1.1861.5240.105
pSIIresAge + log(Area)120.4600.020 −0.2191.3301.6690.098
 Variable importance:0.3540.5600.291   
Figure 1.

Influence of island characteristics on species and interaction richness in local networks on the Canary Islands. (a) Dependence of plant species richness of local networks on both island age and area, with a hump-shaped relationship with island age and an increase with log(Area). (b) Change in pollinator species richness of local networks as a function of island area. (c) Icorrected (interaction richness, corrected for number of insect pollinator species) displays a hump-shaped relationship with island age. All curves are based on coefficients from the highest-ranked model (Table 2). Minimum and maximum untransformed island areas are 280 and 2058 km2, respectively, and island ages are in millions of years (Ma).

Specialization

Neither Age nor Area caused any changes in network specialization (H2′) across the archipelago, because the highest-ranked model had no parameters included. However, the average degree of specialization of the pollinators (dpol) changed in a hump-shaped way with Age, and the relative importance of both Age and Age2 was almost three times that of log(Area) (Table 2, Fig. 2b). The average degree of specialization of plant species (dpla) also depended on Age, because the three highest ranked models all included combinations of Age and Age2. Although Age2 had the highest relative importance, the overall tendency of dpla was to increase with Age (Fig. 2a).The corrected metrics of specialization (dpol,corrected and dpla,corrected) displayed qualitatively similar patterns. In fact, almost the same models were selected as the most parsimonious ones for the asymmetry-corrected variables as for the original variables. For dpla,corrected, the ranking changed slightly, as a linear relationship with Age now had the highest wAIC, and Age had the highest relative importance (Table 2, Fig. 2c,d). Additionally, because dpol,corrected displayed a similar relationship with Age as PSR and Icorrected, we investigated how these variables covaried, but dpol,corrected did not correlate with either PSR (= 0.31) or Icorrected (= 0.87).

Figure 2.

Variation in specialization indices with island characteristics on the Canary Islands. (a) Average degree of specialization of plant (dpla) and (b) pollinator species (dpol) in local interaction networks against island age. A hump-shaped relationship explains most of the variation in dpol, whereas dpla increases quadratically with island age. Specialization indices corrected for network asymmetry (ratio between number of pollinator and plant species) display qualitatively similar responses because dpol,corrected maintains a hump-shaped relationship with island age (d) and dpla,corrected increases linearly (c). All curves are based on coefficients from the highest-ranked model (Table 2). Island ages are in millions of years (Ma).

Single-island interactions

The number of single-island interactions (nSII) was most responsive to changes in log(Area), because this was the only variable included in the highest-ranked model. The relative importance of log(Area) was 3.5–4.5 times as high as the relative importance of any Age variables (Table 2, Fig. 3). The proportion of single-island interactions (pSII), on the other hand, was more dependent on island age and changed in a U-shaped way (Fig. 3b). In particular, Age2 had a high relative importance among variables in the most parsimonious models. To investigate whether the number of uniquely observed species, i.e. species only observed on one island, was driving this U-shaped pattern, we removed the effect of uniquely observed species from pSII, by analysing the residuals (pSIIres) produced from a generalized linear model between the two. When using pSIIres instead of pSII, the relationships got weaker, but Age still remained an important variable. This suggested that pSII, to some extent, was controlled by number of uniquely observed species, but also that pSII in a pollination network might increase with Age in particular, irrespective of the number of uniquely observed species (Table 2, Fig. 3).

Figure 3.

Relationship between island characteristics and single-island interactions in local interaction networks on the Canary Islands. The number of single-island interactions (nSII) increases with island area (a), whereas the proportion of single-island interactions (pSII) displays a negative unimodal relationship (b). (c) pSIIres (residuals from a correlation between pSII and number of uniquely observed species) increases quadratically with island age. All curves are based on coefficients from the highest-ranked model (Table 2). Minimum and maximum untransformed island areas are 280 and 2058 km2, respectively, and island ages are in millions of years (Ma).

Discussion

Species and interaction richness

Central to this paper is to what extent regional drivers influence local processes and patterns, i.e. to what extent do island age and area influence local biotic complexity, here depicted as the structure of local pollination networks? We emphasize that both age and area are indicators of environmental properties and drivers, and that age is not only time per se, but also incorporates the specific geomorphological conditions that islands experience during their ontogeny (Menard, 1983; Stuessy, 2007). Furthermore, by ‘local’, we mean on a spatial scale where individuals have the potential to interact physically (sensu Holt, 1992, 2010). This fundamental issue is discussed at length in the literature (e.g. Ricklefs & Schluter, 1994; Whittaker & Fernández-Palacios, 2007), and Holt (2010), in particular, analysed what he termed the regional–local interface, i.e. island biogeography–food web ecology. The present study shows that the hump-shaped pattern relating plant species richness and island age at the island scale (Whittaker et al., 2008; Bunnefeld & Phillimore, 2012; Cameron et al., 2013) and confirmed for individual vegetation zones (Steinbauer et al., 2012) can be detected at the scale of local interaction networks. Thus, regional processes, i.e. island maturation and its accompanying biodiversity effects, influence species composition in local communities, as argued by Ricklefs (2008). Gruner (2007) found that local arthropod biodiversity on the tree Metrosideros polymorpha also displayed broad-scale patterns, whereby trees on mid-aged Hawaiian islands harboured the highest biodiversity. He argued that local fertility and nutrient availability might be important factors, as these factors peak on mid-aged islands. On our study islands, we found that total numbers of insects and plants (Arechavaleta et al., 2010) correlated with local network numbers of pollinator (A) and plant species (PSR) (insect species per island versus network pollinator species: R2adj = 0.50, = 0.006; plant species per island versus network plant species: R2adj = 0.48, = 0.007), confirming the presence of a link between regional and local processes. Island plant diversity may be a function of habitat diversity, which peaks on mid-aged islands due to their higher elevation and their relief being strongly structured by erosion gorges and diverse lowland sediment beds (Stuessy, 2007; Whittaker & Fernández-Palacios, 2007). This complexity, together with island area per se, may drive the regional (i.e. island) diversity of plant species. However, it is possible that insects, being motile organisms of a higher trophic status, have lower habitat specificity and are instead more dependent upon island area. Thus, although the numbers of plant and pollinator species in a network were positively correlated, they responded differently to island area and age (Table 2, Fig. 1).

In the Canary Island networks, both plant species richness (PSR) and Icorrected showed a hump-shaped relationship with age. Although PSR and Icorrected were only marginally and non-significantly correlated, interaction richness is likely to be related to the richness pattern of plants across the archipelago. On average, plant species had more links than pollinators, and the fewer number of plant species on younger and older islands might cause the observed decrease in Icorrected. Thus, on the scale of the local network, Icorrected may be driven by changes in pollinators and plants; that is, Icorrected depends strongly on species dynamics and turnover. The observed link between regional and local biodiversity means that interactions might be governed at the island scale by the same processes that influence species richness – immigration, speciation and extinction (MacArthur & Wilson, 1967; Whittaker et al., 2008).

Contrary to Sugiura (2010), we found no effect of island area on connectance (Ccorrected) and only a weak effect on Icorrected (Table 2). When using mixed effect models on the uncorrected variables (I and C), log(Area) occurred in all the parsimonious models for these two network variables (K. Trøjelsgaard, unpublished), suggesting that uncorrected metrics are more sensitive to island area. Thus, this discrepancy between the results of Sugiura (2010) and our study is probably partly due to our use of corrected network metrics and partly due to the different nature of the two kinds of network (plant–pollinator versus ant–plant networks).

Specialization

Albrecht et al. (2010) showed that on a shorter time-scale (130 years) along a chronosequence of a glacier foreland, network specialization (H2′) decreased significantly with the age of the site, because pollinators exhibited more generalist behaviour at older sites. Operating on a scale of millions of years, this study found no effect of island age on network specialization, probably as a consequence of the two functional communities responding quite differently. Whereas the average specialization of pollinators (dpol) was related to island age in a hump-shaped way, plant specialization (dpla) increased quadratically with age (Fig. 2a,b). Similar responses were observed for the asymmetry-corrected specialization indices (Fig. 2c,d), suggesting that network asymmetry only had a minor effect on these relationships.

For both plants and pollinators, a more generalist interaction behaviour was observed, especially on the youngest island (Fig. 4), which might be a result of young islands being occupied predominantly by generalist colonists (Richardson et al., 2000; Piechnik et al., 2008). Pollinator specialization peaked on mid-aged islands and then became more generalized on older islands. Older islands have reduced habitat diversity and may thus favour generalist animals, in accordance with findings showing that specialists are more vulnerable to habitat reduction and fragmentation than generalists (Aizen et al., 2012; Hagen et al., 2012). Plant species, on the other hand, became more specialized with island age and this trend might be driven by other forces. All the plants we considered were perennial species, and their life history is therefore fundamentally different from the mobile and annual pollinators. This difference might allow plant species to evolve more intimate interactions during the ontogeny of the island. If this truly is the case, it does lend support to the suggestion by Whittaker et al. (2008, 2010) that intimate interactions might be more prominent on older islands. Alternatively, the observed increase in average plant specialization with island age might only tell us the first half of the story, i.e. show us the first half of a hump-shaped relationship. If Canarian islands older than the present ones had existed they would have been more eroded and contained more homogenous habitats. This would be likely to favour a generalist behaviour, leading to a decreased plant specialization level. If so, the relationship between specialization and island age may be the same for plants and pollinators, with the distinction that the former display a strong time delay in the Canaries (Fig. 4).

Figure 4.

Summary of trends revealed in the present study based on data from the Canary Islands, with curves being based on the coefficients from the highest-ranked models in Table 2 (i.e. the ones depicted in Figs 1-3). Values on the y-axis are omitted as the intention is only to visualize how these metrics change with island age. Note that the curves are normalized to run between zero and one by dividing by the maximum value and adding a positive value if negatives occur. Furthermore, the curves are sorted vertically to minimize overlap and increase visual clarity. The metrics displayed are: (1) interaction richness, corrected for number of pollinator species (Icorrected); (2) pollinator specialization, corrected for network asymmetry (dpol,corrected); (3) plant specialization, corrected for network asymmetry (dpla,corrected); and (4) proportion of single-island interactions (pSII).

Single-island interactions

The number and proportion of single-island endemic species (nSIE and pSIE) have been used in island biogeography as metrics to disentangle evolutionary dynamics on islands (e.g. Emerson & Kolm, 2005; Triantis et al., 2008; Borges & Hortal, 2009). Here, we introduced the concept of single-island interactions (SII) and showed that the proportion of SII (pSII) displayed a U-shaped relationship with age (Table 2, Fig. 3b). SII are tightly connected to the occurrence of uniquely observed species (species only observed in the network(s) on a single island), for which reason we also regressed pSIIres (residuals from a correlation analysis between pSII and number of uniquely observed species) against island characteristics. Although the relationship changed slightly, pSIIres still increased quadratically with island age (Fig. 3c). Thus, the variation in pSII and pSIIres suggests that both young islands and, more noticeably, old islands, have a higher proportion of unique interactions than mid-aged islands.

The U-shaped relationship between pSII and island age contradicts our expectations, as this relationship was expected to be driven by endemic species, and these vary in a positively hump-shaped way with island age (Whittaker et al., 2008; Steinbauer et al., 2012). The increase in pSII on old islands might be the legacy of the presence of intimate interactions that evolved over millions of years, offering support to the idea that tight, and in this case also unique, mutualisms might be an attribute of old islands (Whittaker et al., 2008, 2010). In line with the time-for-speciation hypothesis of Stephens & Wiens (2003), time may also be an important factor in the evolution of unique pairwise interactions. Alternatively, the diminishing ecological space on older islands (Stuessy, 2007) might select for the establishment of novel and less efficient interactions that are not observed on mid-aged islands, which also will result in a higher pSII. The increase in pSII on the youngest island, El Hierro, might be explained by pollinators and plants exhibiting a pronounced generalist behaviour (Fig. 4), i.e. the networks on this young island are still dominated by the effect of colonization. A core of interacting generalists contains many unique interactions, leading to an elevated pSII. It could also be an effect of El Hierro harbouring species that have gone extinct on the other islands. However, the pSIIres results indicate that El Hierro maintains a relatively high proportion of single-island interactions compared to mid-aged islands, even when uniquely observed species are taken into account (Fig. 3c).

In summary, pSII seems to follow a distinct trajectory compared to pSIE. Sufficient time to evolve these unique interactions may overrule the influence of island degradation, and the establishment of such unique links may assure the persistence both of the interactions and of the species involved in those interactions on older islands.

Conclusions

In this paper, we extended studies of island biodiversity from merely looking at species to also including their interactions. Our study is constrained by the modest size of the data set but we regard it as a timely contribution in the efforts to merge island ecology and species interaction biology, which we here term the ‘island biogeography of biotic interactions’. A potential next step would be to investigate how general our results are, i.e. to what extent they apply to other kinds of biotic networks and archipelagos.

We observed a coupling between regional (island) and local (network) biodiversity. Local pollinator richness was mostly influenced by island area, whereas local plant species richness followed a hump-shaped relationship with island age, suggesting that habitat diversity might be an important driver. This is in accordance with earlier studies operating at coarser spatial scales – at the scale of islands (Whittaker et al., 2008) and within vegetation zones across islands (Steinbauer et al., 2012).

We also focused on the way interaction properties vary with island ontogeny. Like plant species richness, interaction richness followed a hump-shaped relationship with island age, suggesting that interactions possess dynamics similar to island species, i.e. driven by immigration and extinction, either directly or indirectly through plant species immigration–extinction dynamics. Using single-island endemic species (SIE) as an evolutionary metric has been a profitable research line, and the study of single-island interactions (SII) may turn out to be the same. We found that the proportion of single-island interactions was higher on older islands and a sufficient time span for the evolution of these unique pairwise interactions might be an explanation. Specialization of pollinators and plants either behaved differently with respect to island ontogeny, suggesting different underlying drivers, or they behaved similarly but with plant specialization displaying a strong time delay, in which case, the observed decrease for pollinators might appear for plants in a few million years.

Whittaker et al. (2008, 2010) proposed that island ontogeny and the accompanying species responses might produce tight mutualisms on older islands: this was partly supported by our study as both the proportion of single-island interactions and plant specialization level were higher on older islands.

Acknowledgements

We are grateful to Maria Christensen, Dennis B. Boll and Nanna M.E. Pedersen for valuable field assistance, to Claus Rasmussen and Antoni Ribes Escolà for insect identification, and to David D. Padilla for his help with the logistics on Tenerife. Pedro Jordano, Alfredo Valido, Kostas Triantis, Robert Whittaker and three anonymous referees greatly assisted in improving the manuscript. The study was financed by Aarhus University Research Foundation (K.T.) and the Danish Research Council FNU (J.M.O., K.T.).

Biosketch

Kristian Trøjelsgaard is interested in large-scale patterns of interaction networks and how they vary on both island biogeographical and macroecological scales. Besides the spatial components, he is also interested in the temporal dynamics that affect and structure communities of interacting species.

Author contributions: K.T. and J.M.O. conceived the ideas and collected the data; K.T. analysed the data; K.T. and J.M.O. led the writing; M.B., X.E., P.O. and F.R. identified the pollinators; and all authors contributed to the writing of the manuscript.

Ancillary