Data from 56 islands from eight oceanic archipelagos were selected based on the availability of reliable faunal lists, and an estimated age of origin (maximum age) of each of the islands (see Appendix S1 in Supporting Information). Oceanic islands are generally considered to be those that have formed over oceanic crust and that have never been connected to continental landmasses (Whittaker & Fernández-Palacios, 2007): they are typically relatively short-lived landmasses and relatively few last longer than a few million years before subsiding and/or eroding back into the sea. Despite the relative simplicity of the geological history of some oceanic island groups, e.g. Hawaii, the dynamics of most archipelagos are more complex than assumed within the GDM (e.g. Courtillot et al., 2003; Neall & Trewick, 2008). Even for hotspot archipelagos with a more or less linear arrangement of islands, at least three distinct types have been identified, based largely on the origin of the plumes (Fig. 4 in Courtillot et al., 2003). Nevertheless, for the majority of the islands considered here, the age of origin (maximum age) is more or less agreed upon (below), and thus we use maximum age for all the analyses here (Table 1).
Table 1. Properties of the oceanic island systems included in the analyses on land snail diversity and endemism. For a full list of data sources, see text
|Island group||No. of islands/unitsa||Area (km2)||Total area (km2)||Elevation (m)||Geological age (Ma)|
|Hawaii||10 (4)||0.2–10,433||16397||4205|| |
|Tristan da Cunha||4 (4)||4–96||179||2060||0.20–18|
Island age data were derived as follows: (1) Hawaiian Islands: Clague (1996). (2) Galápagos Islands: Geist (1996 and unpublished data). (3) Azores: although Johnson et al. (1998) suggest an age of just 0.8 Ma for São Miguel, we use 4.01 Ma and the other ages adopted by Borges et al. (2009) for the rest of the archipelago in formal analyses. (4) Madeiran group: Geldmacher et al. (2005). (5) Canary Islands: Carracedo et al. (2002). In the case of Gran Canaria, an age of c. 3.5 Ma has been used in some previous analyses (see Whittaker et al., 2008 and discussion therein) based on the hypothesis of near-complete sterilization in the catastrophic Roque Nublo ash flow (Marrero & Francisco-Ortega, 2001). However, Anderson et al. (2009) demonstrate that this hypothesis is implausible, hence we use the maximum subaerial age of Gran Canaria, i.e. c. 14.5 Ma (Carracedo et al., 2002; for discussion see Fernández-Palacios et al., 2011). (6) Samoan Islands: Workman et al. (2004), and Neall & Trewick (2008). (7) Society Islands: Clouard & Bonneville (2005). (8) Tristan da Cunha: Ryan (2009).
Sources and treatment of snail data
Our sources of data on snail faunas were as follows:
(1) Hawaiian Islands: Cowie et al. (1995) and Cowie (1995), updated by reference to Pokryszko (1997) for Lyropupa. (2) Galápagos Islands: Dall & Ochsner (1928), Smith (1966), Coppois (1985), Parent & Crespi (2006) and references therein. (3) Azores: Cunha et al. (2010), updated with unpublished data of A.M.F. Martins, R.A.D. Cameron and B.M. Pokryszko. (4) Madeira group: Seddon (2008) with some corrections and minor modifications (following Goodfriend et al., 1996; Cameron et al., 2007). (5) Canary Islands: Núñez & Núñez (2010), updated by Vega-Luz & Vega-Luz (2008), Holyoak & Holyoak (2009), Neiber et al. (2011), and unpublished data of M.R. Alonso and M. Ibáñez. (6) Samoan Islands: Cowie (1998) with additional records from Cowie (2001) and Cowie et al. (2002). (7) Society Islands: Peake (1981) provided overall species numbers, but did not discriminate introduced species. Thus, we use total number of species for this group of islands in our analyses. We have updated this list adding the recently described species from Gargominy (2008). Species lists are available only for the native family Partulidae (Coote & Loève, 2003). (8) Tristan da Cunha: Holdgate (1965) and Preece & Gittenberger (2003).
Many oceanic island snail faunas have suffered extensive extinctions due to human activity (Solem, 1990; Cowie, 1995, 2001; Coote & Loève, 2003). We have included described species extinguished by such activity, but we cannot know about species that became extinct before they had been described, or about segregations that might have resulted had modern techniques been available at the time. Nevertheless, detailed studies of island mollusc faunas started earlier than for most invertebrate groups, and early inventories, including species now extinct, are remarkably complete (Seddon, 2008). Furthermore, in each case we have excluded from the analysis species thought to have been introduced. Fossil evidence for early occurrence of non-endemic snails is available in some cases; in others the judgement of the local workers has been used. All slugs have been excluded because nearly all are introduced (e.g. for Hawaii, see Cowie et al., 1995).
We have generally followed the taxonomic status as given in the source publication, considering only full species. For each dataset, we compiled and recorded three diversity metrics (D), i.e. number of native species/species richness (SR), number of archipelagic endemic species (nEnd), and the number of single-island endemic species (nSIE). The nSIE is a simple metric indicative of evolutionary dynamics that reflects the outcome of in situ speciation, extinction and migration within the islands of an archipelago. Additionally, for Hawaii and the Canaries, we subdivided faunas into major taxonomic groups.
We recognize that our data are incomplete (e.g. Neiber et al., 2011). However, insofar as the data are incomplete, or subject to excessive taxonomic splitting or lumping, we assume that there is unlikely to be significant bias across the islands within a particular archipelago as the same taxonomists usually work on material from all the islands within an archipelago. Small numbers of additional or deleted species make little difference to the trends shown here. A possible exception is shown by Cowie's (1995) path analysis of variation in species richness of Hawaiian land snails, in which the densely populated island of Oahu appears oversampled in comparison to the much larger but sparsely populated island of Hawaii.
The configurations of islands vary through time, not just because of the ontogeny of the islands, but also due to the influence of other factors, e.g. tectonics and eustatic change. In particular, sea-level minima during the Pleistocene produced connections between some adjacent islands, turning them into single islands, while volcanism can both join and sometimes subdivide island territories (general review in Whittaker & Fernández-Palacios, 2007). For some analyses (detailed below), we have treated such groups as single islands so as to test the possible effects of varying configuration of the archipelagos through time (see Appendix S1, Table S2 & Appendix S2).
In studies such as this, the small number of islands per archipelago can lead to low power in detecting trends, instability in parameter estimation and model over-fitting (e.g. Burnham & Anderson, 2002). Bunnefeld & Phillimore (2012) recently suggested the use of linear mixed effect models (LMMs) to overcome such limitations. LMMs are designed to detect general patterns where data come from grouped sources (Bolker et al., 2009; Zuur et al., 2009). We follow their lead in applying LMMs to allow the simultaneous consideration of data from eight archipelagos comprising 56 islands.
LMM predictors are classified into (1) fixed effects: those for which we aim to estimate regression parameters, i.e. slope and intercept; and (2) random effects: those that identify groups conceptually drawn from a larger population (e.g. archipelagos or taxa) within the data and for which we examine variation in a parameter (i.e. slope and intercept) across levels (Bunnefeld & Phillimore, 2012; Hortal, 2012). When considering a factor as fixed we are interested in estimating and comparing regression parameters for the different levels of the factor (e.g. different archipelagos). Considering a factor as a fixed effect leads to the estimation of different regression parameters for continuous variables, such as area, time or elevation for each level of the factor. When studying the fixed effect of a factor, the implicit assumption is made that the levels of the factor considered in the analysis are exhaustive (e.g. contain all the possible oceanic archipelagos), or alternatively that one is not interested in generalizing the results to other levels of the factor not included in the study (e.g. other oceanic archipelagos not included in the study). When a factor is studied as a random effect, instead of fitting regression parameters for each level of a factor, one is interested in estimating the variation in the regression parameters induced by the different levels of the factor (e.g. the variation around the general slope considering all the islands belonging to the same virtual global archipelago). The random effects can be seen as grouping factors drawn as a random sample from a larger (conceptual) population, such as the eight archipelagos considered here (adapted from glossary in Bunnefeld & Phillimore, 2012). In this context, one is not primarily interested in estimating and comparing the relationships under study for the different levels of the factor. The random effect is seen as a source of pseudoreplication [non-independence of data points (here islands) belonging to the same level of the factor (here archipelagos)] that needs to be taken into account.
Bunnefeld & Phillimore (2012) demonstrated the advantages of LMMs when applied to the data that were originally used for testing the ATT2 model (i.e. D = b1 + b2Area + b3Time + b4Time2). Here we follow the same methodological steps: our response variables were the three diversity metrics (SR, nEnd and nSIE). The fixed effects were island area (Area, in km2; log-transformed as generally supported by previous work, e.g. Whittaker et al., 2008; Triantis et al., 2012a), time elapsed since island formation (Time, i.e. date of emergence of each island, in million years ago, Ma), a quadratic term for the time elapsed (Time2), and we also considered elevation (Elevation, in metres, m) as a proxy of environmental heterogeneity. The grouping factor considered as a random effect was the archipelago that each island belongs to, as the values of the intercept and the slopes of the relationships between the diversity metrics, area, time and elevation may vary across archipelagos. To select the best models for describing the diversity metrics we followed a two-step procedure: first, the most parsimonious random effects structures (with all fixed effects included) were selected using model selection based on the small-sample corrected Akaike's information criterion (AICc) (Burnham & Anderson, 2002). The model with the lowest AICc value is considered to fit the data best. However, all models with a ∆AICc value < 2 (the difference between each model's AICc and the lowest AICc) must be considered as having relatively similar levels of support and thus belong to the group of ‘best models’ (i.e. equally parsimonious; Burnham & Anderson, 2002). Accordingly, when several random structures provided indistinguishable AICc values (∆AICc value < 2), they were considered as part of the ‘best random structure group’, and subsequent fixed effect structures were compared. To find the most parsimonious random effect structures we compared models with and without a varying intercept among archipelagos and all possible combinations of varying slopes across archipelagos for the different variables considered (i.e. log(Area), Time, Time² and Elevation). We used the ‘lmer’ function in the ‘lme4’ library (version 0.999375-39) in R 2.14.1 (R Development Core Team, 2011) using restricted maximum likelihood (REML).
After determining the best random effect structures, the most parsimonious combinations of fixed effects were found using model selection based on AICc (models fitted using maximum likelihood). We used the ‘dredge’ function in the ‘MuMIn’ library in R (version 0.13.17) to run a complete set of models with all possible combinations of the fixed effects and determined the subset of ‘best models’ as the ones with ∆AICc value < 2 (as above). We additionally used Akaike weights derived from the AICc (wAICc) to evaluate the relative likelihood of each model, given the dataset and the set of models considered, and to estimate the relative importance of each variable by summing these wAICc across the models in which they were included. Akaike weights are directly interpreted in terms of each model's probability of being the best at explaining the data (Burnham & Anderson, 2002).
To facilitate comparison with the previous LMM applications of the ATT2 (Bunnefeld & Phillimore, 2012), we have applied the above methodological steps for the log-transformed values of the diversity metrics considered (log n + 1 was used for datasets that contain at least one zero value for the diversity metric considered) and also for the untransformed values. As we employ log-area in the analysis, this particular implementation of the ATT² assumes a power law species–area relationship, the most general and widely applied species–area relationship model (Rosenzweig, 1995; Triantis et al., 2012a).
We have undertaken additional analyses for separate snail families for the two most species-rich archipelagos, i.e. Hawaii (10 islands) and the Canaries (7 islands), following the above steps, but with the random effect being ‘Family’ instead of ‘Archipelago’, in order to compare the diversity patterns within the same archipelago but for different taxonomic groupings. For the Hawaiian Islands the taxonomic groupings considered were: Succineidae, Pupilloidea, Helicarionidae, Helicinidae, Endodontoidea, Amastridae and Achatinellidae. For the Canary Islands they were: Vitrinidae, Helicoidea, Ferussaciidae and Enidae.