TESTING ECOLOGICAL EXPLANATIONS FOR BIOGEOGRAPHIC BOUNDARIES

Authors


Abstract

Barriers to dispersal and resulting biogeographic boundaries are responsible for much of life's diversity. Distinguishing the contribution of ecological, historical, and stochastic processes to the origin and maintenance of biogeographic boundaries, however, is a longstanding challenge. Taking advantage of newly available data and methods—including environmental niche models and associated comparative metrics—we develop a framework to test two possible ecological explanations for biogeographic boundaries: (1) sharp environmental gradients and (2) ribbons of unsuitable habitat dividing two highly suitable regions. We test each of these hypotheses against the null expectation that environmental variation across a given boundary is no greater than expected by chance. We apply this framework to a pair of Hispaniolan Anolis lizards (A. chlorocyanus and A. coelestinus) distributed on the either side of this island's most important biogeographic boundary. Integrating our results with historical biogeographic analysis, we find that a ribbon of particularly unsuitable habitat is acting to maintain a boundary between species that initially diverged on distinct paleo-islands, which merged to form present-day Hispaniola in the Miocene.

Barriers to dispersal are responsible for much of life's diversity. Without the oceanic barriers that separate islands and continents, for example, there would likely be no lemurs in Madagascar or kangaroos in Australia. Although the importance of such physical barriers to dispersal is undeniable, many biogeographic boundaries have somewhat less-obvious explanations. Wallace's Line, for example, does not correspond with a particularly profound barrier to contemporary dispersal, but rather with a deep ocean channel marking the point of contact between two historically and tectonically distinct regions. Other biogeographic boundaries, meanwhile, result from ecological processes such as local adaptation and niche conservatism, or perhaps most frequently, some combination of historical and ecological processes (Schneider et al. 1999; Ogden and Thorpe 2002; Jouventin et al. 2006; Wiens et al. 2006; Thorpe et al. 2008; Pearson and Raxworthy 2009; van Rensburg et al. 2009). Although recent molecular phylogenetic and phylogeographic studies identify a growing number of biogeographic boundaries that do not correspond with obvious physical barriers to contemporary dispersal, methods for objectively discriminating ecological, historical, and stochastic explanations for biogeographic boundaries have been slow to develop (but see Pearson and Raxworthy 2009).

Availability of relevant ecological data has long been a major obstacle to identifying ecological barriers to dispersal, but this obstacle is being overcome by rapidly expanding data in the form of geographic information system (GIS) data layers and known sampling localities catalogued in biodiversity collections (reviewed in Kozak et al. 2008). When used to generate environmental niche models (ENMs), these data permit characterization of species’ and populations’ environmental requirements at an unprecedented scale. Comparative studies of ENMs are increasingly used to ask whether environmental requirements have diverged between the species or populations located on the either side of biogeographic boundaries, (e.g., Graham et al. 2004; Kozak and Wiens 2006; Bond and Stockman 2008; Jezkova et al. 2009). One problem with most of these comparative studies, however, is that they tend to find that ENMs generated from related species are more similar than expected by chance, but rarely identical (Warren et al. 2008). For this reason, most available methods are unable to determine if environmental divergence observed among allopatrically or parapatrically distributed populations is sufficient to explain the origin and/or maintenance of a biogeographic boundary (see also Wiens and Graham 2005; Wiens 2008).

Here, we propose a framework that may be used to test for the presence of significant environmental variation across biogeographic boundaries. Using this framework, we can test whether ENM divergence across a biogeographic boundary is greater than expected by chance, with chance expressed as a null distribution of ENM divergence scores between two species or populations when the interface between them is moved to random positions in the landscape. These methods represent a biogeographically informed extension of randomization analyses that we introduced previously (Warren et al. 2008). Our framework may be used to identify two types of environmental variation that potentially contribute to the maintenance of biogeographic boundaries by directly reducing dispersal across the boundary: (1) abrupt environmental changes or steep environmental gradients (Fig. 1B) and (2) ribbons of relatively unsuitable habitat dividing two more highly suitable regions (Fig. 1C). The first type of barrier predicts that ENMs generated from populations on the either side of a given boundary will be more divergent than ENMs generated from pairs of populations generated by random geographic subdivisions of a pooled sample. The second, meanwhile, predicts that ENMs generated from populations on the either side of the boundary will be no more different from one another than expected by chance whereas an ENM generated from occurrence points in the less-suitable region is more divergent from ENMs for the flanking regions than expected by chance (see also Kozak and Wiens 2006).

Figure 1.

Three alternative scenarios investigated by our framework: (A) A biogeographic boundary not associated with significant environmental variation, (B) a biogeographic boundary associated with an abrupt environmental change or a steep environmental gradient, (C) a biogeographic boundary associated with a region of particularly unsuitable habitat that divides two more suitable regions.

Although our framework is focused on identifying contemporary environmental barriers, it may be combined with other lines of evidence to further disentangle the ecological, historical, and stochastic causes for a given biogeographic boundary. Integration with historical biogeographic data, for example, can be used to test whether environmental variation was directly involved in a boundary's origin—via speciation due to divergent natural selection (Schneider et al. 1999; Schilthuizen 2000; Rundle and Nosil 2005) or niche conservatism (Wiens 2004; Kozak and Wiens 2006)—or is merely acting to maintain a boundary that arose through some other mechanism (i.e., vicariant speciation). Integration of information about the fitness of hybrids and population densities along a barrier may also permit an assessment of whether tension zone dynamics are reinforcing the position of an environmental barrier; tension zones result when interspecific crosses produce unfit hybrids and are expected to settle along environmental barriers associated with a reduced population density (Barton and Hewitt 1985).

CASE STUDY: ANOLIS LIZARDS ON HISPANIOLA

We apply our framework to a pair of Hispaniolan Anolis lizards (A. chlorocyanus and A. coelestinus) distributed on the either side of a biogeographic boundary first identified by Mertens (reviewed in Schwartz 1980; Powell et al. 1999) (Fig. 2). Although Mertens’ Line has long been recognized as a boundary between distinct northern and southern Hispaniolan faunas, it does not represent an obvious physical barrier to dispersal and the factors responsible for its origin and maintenance have never been well understood. Previous authors have speculated that endemicity on the either side of Mertens’ Line results from Pleistocene vicariance that occurred when the low-lying plain corresponding with Mertens’ Line was inundated by higher sea levels (Williams 1965) (we refer to this plain, which extends across the boundary between two countries using and acronym - CSVN - that encompasses both the Haitian [Plan du Cul de Sac] and Dominican [Valle de Neiba] names). To explain why A. chlorocyanus and A. coelestinus have not moved across the CSVN plain in the years since, Williams (1965) hypothesized ecological incumbency by suggesting that the two species were “too [ecologically] similar to permit anything but a stand-off, with a very narrow, perhaps fluctuating zone of sympatry.”

Figure 2.

Sampled localities for A. chlorocyanus (black circles) and A. coelestinus (white circles). Shading indicates values for environmental PCA 1, which loads most heavily for temperature and altitude. The pale yellow region corresponding with Mertens’ Line is the xeric, low-lying valley known CSVN plain (blue areas, meanwhile, correspond with relatively cool, montane regions).

Although such an interface may exist for some period of time, one species is generally expected to overtake the range of the other or, if reproductive isolation is incomplete, the two species may simply merge into one. For this reason, the hypothesis that the boundary between A. chlorocyanus and A. coelestinus is the result of Pleistocene vicariance and subsequent niche incumbency is challenged by the recent discovery that these species likely diverged in northern and southern Hispaniola well before the Pleistocene: specimens fossilized in Miocene amber that are indistinguishable from extant A. chlorocyanus and A. coelestinus (Rieppel 1980; de Queiroz et al. 1998; Polcyn et al. 2002) and molecular clock-based analyses (R. E. Glor and A. Larson, unpubl ms.) suggest that the ancestors of A. chlorocyanus and A. coelestinus likely diverged allopatrically prior to the formation of present-day Hispaniola when two paleo-islands merged in the Miocene (Iturralde-Vinent and MacPhee 1999; Powell et al. 1999; Graham 2003). Although subsequent inundation may have reinforced this boundary, the hypothesis that the ranges of these two species have been in sporadic contact for as long as 15 million years suggests that some persistent mechanism(s) may be required to explain the long-term maintenance of the boundary's position.

We test the hypothesis that the CSVN plain represents a significant environmental barrier to present-day terrestrial dispersal. Although both species can be found in the CSVN plain, they are rare in this hot, xeric region and tend to be restricted to localities around human habitat or rivers with broadleaf trees and ample water (Fig. 3). It seems most likely that the CSVN plain is a ribbon of unsuitable habitat that divides two more highly suitable regions, but we also test the hypothesis that it corresponds with a sharp environmental gradient between distinct environmental conditions in northern and southern Hispaniola. In addition to testing whether the CSVN represents a direct barrier to dispersal between northern and southern populations, we also consider the possibility that it is a likely settling spot for a tension zone between hybridizing forms.

Figure 3.

Localities sampled and projected suitability scores from ENMs generated for A. chlorocyanus (A) and A. coelestinus (B) by Maxent. Warmer colors correspond with higher suitability.

Methods

SAMPLING AND ENVIRONMENTAL NICHE MODELING

Environmental niche modeling requires two types of data: (1) presence localities with accurate latitude/longitude coordinates (i.e., georeferenced localities), and (2) environmental data available in the form of GIS data layers. Georeferenced localities for A. chlorocyanus and A. coelestinus are obtained primarily from the HerpNet database (http://www.herpnet.org). HerpNet collates specimen records from 62 natural history collections, including the two museums with the largest holdings of A. chlorocyanus and A. coelestinus (Kansas University's Natural History Collection and Harvard's Museum of Comparative Zoology). Hispaniolan localities in the HerpNet database were georeferenced primarily by staff at the Kansas University Natural History Collection and HerpNet following MaNIS/HerpNet/ORNIS georeferencing guidelines (http://manisnet.org/GeorefGuide.html). To ensure quality of georeferenced localities from HerpNet we (1) correct obvious errors (most frequently involving positive [eastern hemisphere] longitudes), (2) confirm that plotted georeferenced localities correspond with original locality descriptions, (3) remove localities with oceanic coordinates in our environmental data layers, and (4) confirm identification of specimens collected where the two species’ ranges overlap or abut. Records obtained from HerpNet are supplemented with additional samples from private collections held by REG and S. Blair Hedges, for which coordinates were obtained directly from handheld global positioning system (GPS) units. Our final dataset comprises 261 and 160 unique localities for A. chlorocyanus and A. coelestinus, respectively. The number of localities required for accurate niche modeling and associated analyses is not well understood, and highly dependent on both the distribution of available localities and the underlying environmental variation. Although more work is needed on this subject, the number of localities available for each species included in this study exceeds the number considered sufficient for accurate niche modeling in other recent studies (Papes and Gaubert 2007; Pearson 2007; Peterson et al. 2007)

Environmental data layers are all resolved to 1 km2 and comprise the 19 bioclimatically significant variables extracted from the WorldClim database (Hijmans et al. 2005) plus altitude data from CGIAR-CSI GeoPortal digital elevation maps (http://srtm.csi.cgiar.org/). Because many of these 20 layers are strongly correlated, we use principal component analyses (PCA) implemented by the multivariate tools in ArcMAP's (ESRI, Redlands, CA) spatial analyst toolbox to reduce data dimensionality and generate a set of new independent variables from the original unstandardized WorldClim data. The first four PCA axes, which account for more than 99% of the underlying environmental variation, are used for subsequent niche modeling.

We reconstruct niche models using the maximum entropy method implemented by the program MaxEnt 3.3.0 beta (http://www.cs.princeton.edu/~schapire/maxent/) with default parameter settings (Phillips et al. 2006; Phillips and Dudik 2008). We assess the accuracy of niche modeling via the receiver operating characteristic (ROC) and associated area under the ROC curve (AUC) statistic calculated by MaxEnt (see also Peterson et al. 2008). A core assumption of maximum entropy niche modeling is that the species being modeled is not excluded from areas included in the modeling for nonenvironmental reasons. Soberón and Peterson (2005) discussed this issue using what they term the “BAM” approach, an acronym representing the biotic (B), abiotic (A), and mobility (M) factors that may limit a species’ ability to occupy a particular patch of habitat. If areas included in the modeling process are environmentally suitable but inaccessible to the species for other reasons (i.e., barriers to dispersal [M], interspecific competition [B]), Maxent will assume these regions are environmentally inappropriate and will attempt to exclude them from the model, potentially yielding inappropriately low suitability scores when the model is projected into unoccupied areas. Because A. chlorocyanus and A. coelestinus may be excluded from one another's ranges by interspecific competition rather than environmental differences, we examine the impact of the region used for modeling on ENM reconstruction and projection. To do this, we conduct two analyses with each species. In the first analysis, background sampling for MaxEnt is drawn from across the entire island of Hispaniola, which is equivalent to assuming that only the abiotic component (A) is unimportant in determining the ranges of the two species in Hispaniola. In the second analysis, background samples used for niche modeling are drawn only from the region of the island occupied by the species being analyzed with the resulting model then being projected onto the entire island. This latter analysis is relevant if the two species ranges are disjunct due to limited ability to disperse into each others’ habitat for reasons unrelated to available abiotic habitat (A).

IDENTIFYING SIGNIFICANT ECOLOGICAL BARRIERS

Our framework for identifying ecologically significant barriers is a three-step process that tests whether two populations or species separated by a given biogeographic boundary are: (1) characterized by ENMs that are identical, or no more similar than expected if localities are drawn at random from the environmental background, (2) characterized by ENM differentiation that is greater than expected by a random geographic subdivision of their pooled range, (3) separated by a ribbon of habitat that is more divergent from each of its two adjacent regions than expected by chance. We conduct all analyses using the program ENMTools (Warren et al. 2010).

Testing whether the niches of species separated by a biogeographic boundary are more similar than expected by chance, but not identical

We first use two tests introduced by Warren et al. (2008) to ask whether ENMs generated from A. chlorocyanus and A. coelestinus are identical (niche identity test) or no more similar than expected if localities are drawn at random from the environmental background (background similarity test). These tests are based on one of two similar metrics—I or Schoener's D—that summarize similarity of projected suitability scores for each 1-km2 grid cell of a shared landscape and range from 0 (ENMs highly divergent) to 1 (ENMs identical). Because recovery of identical projections (I or Schoener's D = 1) is unlikely even when samples are drawn from a shared underlying distribution, we use the test for niche identity to ask whether the ENMs generated for A. chlorocyanus and A. coelestinus are any more different than ENMs calculated from pairs of samples drawn at random from the a pooled dataset. If ENMs from A. chlorocyanus and A. coelestinus are no more different than pairs of randomly drawn samples, they are diagnosed as effectively identical. We conduct this and all subsequent analyses with 500 replicates. The test for background similarity, meanwhile, is used to test whether ENMs for A. chlorocyanus and A. coelestinus are more or less similar than expected by chance. This test compares the I or Schoener's D values obtained in the comparison of ENM projections for A. chlorocyanus and A. coelestinus to I or Schoener's D values obtained by comparing the ENM generated from the actual localities from one species to ENMs generated from samples drawn randomly from the range occupied by the second species. Although the definition of the range of a species (and hence the appropriate definition of environmental background) can in some cases be problematic, it is unambiguous in the present study. Together, the two species’ occurrence points are spread across the landmass they occupy, and the zone of contact between the two is fairly well defined. As this is the only edge of the two species’ distributions that is not defined by the coast, there is little chance that the ranges of the two species have been misidentified to an extent that would affect our analyses.

Testing for the presence of an abrupt environmental transition or steep environmental gradient

The second test implemented in our framework asks whether the biogeographic boundary in question is associated with an abrupt environmental transition or a steep environmental gradient. To test this hypothesis, we introduce a new, geographically informed randomization analysis called the linear range-breaking test. For this test, we generate a null distribution of I and Schoener's D values through a five-step process involving: (1) pooling samples from both species, (2) randomly drawing a line through the shared geographic range, (3) calculating ENMs for the two sets of localities on the either side of this line, (4) calculating I and Schoener's D value between these ENMs, and (5) repeating this process N times to generate the null distribution of I and Schoener's D values. The null hypothesis that Mertens’ Line is not associated with a significant environmental transition is rejected if the I or Schoener's D values obtained by comparing the ENMs of A. chlorocyanus and A. coelestinus are lower than 95% of the values in the null distribution.

The line dividing the ranges of populations in each pseudoreplicate is drawn by selecting an angle at random and subsequently shifting this line until samples are partitioned into populations containing the same numbers of samples as the empirical samples from A. chlorocyanus and A. coelestinus. If a line with a given starting angle cannot appropriately partition populations, the line is discarded and another line is attempted. This procedure ensures that the values for I and Schoener's D are not biased by differences in sample sizes between the psuedoreplicate datasets and the empirical dataset. Duplicate partitions generated by this procedure are excluded from subsequent analyses. Although the linear partitioning procedure here is easily applied to any geographic region, it may perform poorly due to reduced independence of pseudoreplicates when the spatial extent of the union of the two ranges is very elongate, when sample sizes for the two species are highly asymmetric, or when the ranges of the two species are separated by a very broad region that is unoccupied by either species. Problems may also arise when many duplicate occurrences are included for the same point (either within or among species), particularly if those duplicates constitute a substantial portion of the available occurrences for either of the species.

We also develop a variant of the linear range-breaking test—referred to here as the blob range-breaking test—that geographically partitions pseudo-replicates by starting with a single point and expanding outward from this point until the desired number of samples (for our purposes, the sample size of the less numerous of the two species) are obtained. Although spatial distributions produced by the blob test are less comparable to the empirical data for A. chlorocyanus and A. coelestinus than the linear range-breaking analysis, this method may be more appropriate for other empirical situations. We are not presently aware of any objective criterion that would allow users to determine whether the blob or linear range-breaking test is more applicable to any given system, and recommend that both tests be performed when there is any doubt as to which is most appropriate.

By conducting geographically informed subdivisions and by using only localities from which one of the two species has been sampled, both the linear and blob range-breaking tests provide a more biologically meaningful test for niche divergence across a biogeographic boundary than previous methods, and the niche identity and background similarity tests introduced by Warren et al. (2008) in particular. Rejection of niche identity (see above) is necessary, but not sufficient for the identification of environmental differentiation between the two populations on the either side of a biogeographic boundary; indeed, niche identity is frequently rejected when species with identical niche requirements are distributed across a heterogeneous landscape (Godsoe 2010; A. T. Peterson, unpubl ms.). Similarly, rejection of the background similarity test does not necessarily indicate biological meaningful similarity between the two species being investigated. One reason for this is that randomly drawn background samples are not necessarily localities from which either species—or any related species, for that matter—were sampled. Another reason is that niche similarity may not be rejected by the background similarity test even in cases in which the two species are strongly environmentally differentiated: if the ranges from which these species are sampled are strongly differentiated environmentally, empirical comparison and comparisons of pseudoreplicates generated via random background sampling may yield similarly high divergence metrics.

Testing for the presence of region of unsuitable habitat between two suitable regions

Our final test addresses the hypothesis that populations in highly suitable habitat on the either side of a biogeographic boundary are divided by a ribbon of particularly unsuitable habitat. We are particularly interested in asking whether the hot, xeric valley (i.e., the CSVN plain) corresponding with Mertens’ Line represents such a ribbon of unsuitable habitat. To address this question, we begin by generating ENMs for three populations: (1) samples of A. chlorocyanus from north of the valley, (2) samples of A. coelestinus from south of the valley, and (3) samples from both species from within the valley. We then calculate I and Schoener's D between each of these three models, with the expectation that the values obtained in comparisons of the northern and southern populations will be more similar than either is to the intervening valley samples. To ask whether this pattern is significant, we obtain pseudo-replicates by drawing a ribbon through the combined range of the two species. This procedure is automated by first drawing a line through the combined range in the same manner discussed previously (so that sample sizes on either side of the line are equal to those that were observed in the samples from A. chlorocyanus and A. coelestinus). In this case, however, the line is expanded into a ribbon of predefined width and the samples from within the ribbon are analyzed separately from the two allopatric regions on the either side of it. In our case, the width of the ribbon is defined as the width of the CSVN plain (20 km or 0.18 decimal degrees). By calculating I and Schoener's D values from these pseudoreplicates, we are able to ask whether the ranges on the either side are more similar to one another than either is to the intervening region, and whether this pattern is significantly different from the null expectation. Although the ribbons drawn for different pseudoreplicates may differ in the number of occurrence points they contain, we chose to constrain the width of the ribbon rather than sample size to ensure that the environmental heterogeneity encompassed by each ribbon was more comparable to that in the empirical data. Duplicate pseudoreplicates and pseudoreplicates with ribbons containing fewer than 10 localities were excluded from consideration. We note that this analysis requires that at least one of the species of interest be present in the region that is hypothesized to be unsuitable, and so is not applicable to all biogeographic boundaries.

Results

ENMs from both A. chlorocyanus and A. coelestinus are characterized by high AUC statistics, 0.722 and 0.914, respectively. Anolis chlorocyanus's ENM predicts high suitability across lowland habitats on both the north and the south paleo-islands (Fig. 3). Anolis coelestinus's ENM, meanwhile, exhibits notably higher suitability scores for the south paleo-island than for most regions of the north paleo-island (Fig. 3). The ENM generated with background sampling only from the south paleo-island followed by subsequent projection onto the entire island, however, suggests that the low suitability scores for A. coelestinus in northern Hispaniola likely reflect our choice of background sampling region rather than specialization to environmental conditions on the south paleo-island; using background samples from only the south paleo-island, we recover suitability scores for A. coelestinus on the north paleo-island comparable to those recovered for the south paleo-island (Fig. 4). Because projection of models onto a novel environmental landscape carries its own set of problems, we focus on interpreting the results of our randomization analyses. Although these are always conducted while modeling across the entire island of Hispaniola and are therefore likely under-predict in unoccupied regions, there is no reason to suspect that this will bias the conclusions of our analyses in one direction or the other.

Figure 4.

Suitability projections when ENMs are generated from the region occupied by the species of interest (e.g., north paleo-island for A. chlorocyanus and south-paleo-island for A. coelestinus) and subsequently projected onto the entire island. Warmer colors correspond with higher suitability.

Because we obtain qualitatively indistinguishable results from I and Schoener's D in all of our comparative analyses, we focus the subsequent discussion exclusively on I, which tends to express more variation than D. The identity test and the background similarity test suggest that ENM projections from A. chlorocyanus and A. coelestinus are not identical (identity test, P < 0.002); indeed, they appear to be no more similar than expected by chance (P < 0.574 [A. chlorocyanus vs. A. coelestinus background], P < 0.224 [A. coelestinus vs. A. chlorocyanus background]). Linear range-breaking reinforces the latter conclusion by suggesting that ecological divergence between A. chlorocyanus and A. coelestinus is no greater or less than that observed between pseudoreplicate pairs generated via random geographic fragmentation (P < 0.463, Fig. 5). We therefore reject the hypothesis that a particularly abrupt environmental transition occurs across Mertens’ Line and the CSVN plain. A similar result is obtained when psuedoreplicates are obtained using the blob range-breaking method (P < 0.574, Fig. S1).

Figure 5.

Results from linear range-breaking analyses. Density plot indicates distribution of I values from 242 unique pseudo-replicates obtained after eliminating those involving identical range fragmentation. Vertical black line indicates Schoener's D value between ENM projects for A. chlorocyanus and A. coelestinus.

ENM projections from both species support the hypothesis that Mertens’ Line and the CSVN plain mark a ribbon (or ribbons) of particularly unsuitable habitat between more highly suitable regions on the north and south paleo-island populations. First, low suitability scores are observed for both species in the CSVN plain and the Sierra de Bahoruco (Fig. 3). Even these scores are likely to be overestimates of habitat suitability in the CSVN plain because nearly all of the samples from this region were obtained from locally mesic environments (R. Glor, pers. obs., Williams 1965, HerpNet locality descriptions) that are not distinguished from the prevailing xeric conditions of the CSVN plain by the 1-km2 resolution Worldclim data (recall that Worldclim data are obtained by extrapolation from weather stations) (Fig. 2). Second, our random ribboning analyses confirm that the environmental conditions experienced by populations in the CSVN plain are more different from those experienced by populations in flanking regions than expected by chance (P < 0.008 [A. chlorocyanus sampling vs. valley], P < 0.015 [A. coelestinus sampling vs. valley], Fig. 6), whereas populations in the flanking regions in northern and southern Hispaniola are no more different from one another than expected by chance (P < 0.471, Fig. 6).

Figure 6.

Results of ribboning analyses. (bottom left) Density of I values calculated between the two flanking regions and the ribbon in each pseudoreplicate, dashed line for a comparison of ENMs from populations with sampling comparable to A. chlorocyanus, solid line for ENMs from populations with sampling comparable to A. coelestinus. Overlapping solid and dashed bars indicate the I values from a comparison of an ENM for A. chlorocyanus north of the CSVN plain to an ENM from samples of A. chlorocyanus and A. coelestinus from within the CSVN plain (P < 0.008) and from comparison of an ENM for A. coelestinus from south of the CSVN plain to an ENM from populations of A. chlorocyanus and A. coelestinus sampled from within the valley (P < 0.015). (top left) Density of 333 I values obtained from randomly generated ribbons and a comparison of flanking habitat on the either side of this ribbon. Vertical line indicates I value calculated between populations of A. chlorocyanus and A. coelestinus located on the either side of the CSVN plain (P < 0.471).

Discussion

We develop a simple framework for testing the ecological basis for biogeographic boundaries. Applying this framework to two species of Hispaniolan Anolis lizards, we find evidence that an important biogeographic boundary corresponds with a significant environmental barrier to dispersal. This barrier comes in the form of the hot, xeric conditions prevailing across the CSVN plain, which are particularly unsuitable for both species while dividing two regions in northern and southern Hispaniola that are more highly suitable for both the species (Figs. 2 and 6). Coincidence between a biogeographic boundary and a ribbon of environmental conditions as extreme as that existing across the CSVN plain is unlikely to result from chance (Fig. 6).

In spite of its significance, the environmental barrier identified by our analyses does not tell the whole story of Mertens’ Line. Indeed, historical biogeographic analyses suggest that the initial divergence between the ancestors of A. chlorocyanus and A. coelestinus occurred allopatrically prior to the merger of Hispaniola's north and south paleo-islands in the Miocene (R. E. Glor and A. Larson, unpubl ms.). Thus, the environmental barrier we identify is likely serving primarily to reinforce a barrier with a historical biogeographic origin, rather than as the location of in situ speciation. Even in this role, the simple reduction in dispersal between northern and southern Hispaniola that results from the environmental conditions in the CSVN may be reinforced by additional mechanisms.

One candidate is Williams’ (1965) niche incumbency hypothesis. Our analyses support the possibility of incumbency by suggesting that ENMs from A. chlorocyanus and A. coelestinus are effectively indistinguishable (Figs. 5 and 6). Although habitat suitability scores estimated from ENMs for A. chlorocyanus and A. coelestinus can be dramatically different (Figs. 3 and 4), and the niche identity test is rejected, our range-breaking and ribboning analyses suggest that the differences in environmental niche requirements between these species are no greater than that expected by chance (Figs. 5 and 6).

In addition to direct ecological mechanisms such as reduced dispersal across unsuitable habitat and incumbency when populations come into contact, the position of the contact zone between A. chlorocyanus and A. coelestinus is also consistent with two key predictions of the tension zone hypothesis. First, hybrids appear to be at a selective disadvantage; although hybridization between A. chlorocyanus and A. coelestinus appears possible, putative hybrids are rare in nature and the two parental forms remain distinct where they come into contact (Garcia et al. 1994). Second, the location of the contact zone between the two species is clearly associated with an area of reduced population density. Although the life span of tension zones can be difficult to predict, numerous examples in the literature suggest that they can persist for millions of years (Barton and Hewitt 1985).

Conclusions

Many ancient biogeographic boundaries likely require some combination of historical and ecological explanations. Our framework for statistically testing the ecological factors contributing to the maintenance of Mertens’ Line can easily be applied to other biogeographic boundaries. Combined with historical biogeographic analyses, our methods will be particularly useful for identifying ancient historical boundaries that are associated with significant, although superficially cryptic, environmental variation. The presence of such environmental variation, perhaps in a combination with tension zone dynamics, may go a long way toward explaining why some biogeographic boundaries persist as long as they do.

Simple extensions of our framework would permit application to entire faunas or floras separated by biogeographic boundaries. One might, for example, draw backgrounds samples from the regions separated by a biogeographic boundary and develop ENMs from these samples to ask whether these areas are more different than expected by chance. Our framework may also be applied to other, related questions. For example, it might be used to more rigorously test whether putative cryptic species differ significantly in their environmental niche requirements (sensu Bond and Stockman 2008). In addition, a simple modification of our methods allows for spatially partitioning occurrence points within a single species, which may be helpful in determining the modeling approaches that cope best with strong spatial biases in sample selection.


Associate Editor: M. Hellberg

ACKNOWLEDGMENTS

We thank M. Turelli, A. T. Peterson, M. Hellberg, members of the Glor Lab and two anonymous reviewers for constructive comments on this manuscript. Funding for this work was provided by the Center for Population Biology at UC Davis, the University of Rochester, and NSF grants DEB-0920892 to REG and DBI-0905701 to DW.

Ancillary