Bioclimate envelope models: what they detect and what they hide

In a recent issue of Global Ecology and Biogeography, Pearson Clark et al., 2001; Travis, 2003). These differ greatly across species’ ranges from their expanding to their eroding margins, and so also does the character of the respective populations (Lesica Davis see also Davis see Hampe Santamaría et al., 2003; and references therein for case studies). BEM that simply project the climatic tolerance of current rear edge populations to construct future low-latitude range limits assume inherently that these will be formed by descendants of populations from the current range periphery. This would require that peripheral populations migrate fast enough to match climate change while outcompeting previously existing conspecifics on the migration route and conserving their genetic makeup sufficiently unchanged that they maintain their previous climate tolerance. This is obviously far from realistic. It appears more likely instead that most rear edge populations will go extinct, the overall climatic tolerance of species will therefore decline and their ranges will shrink more towards higher latitudes or altitudes than predicted by BEM. BEM ignore completely that a limited dispersal capacity may constrain future migrations of species, whereas ample ecological and phylogeographical research has underlined its key role for population and range dynamics (Cain et al., 2000; Clark et al., 2001; Jansson Hampe et al., 2003). Again, a clear difference exists between populations at expanding and at retreating range margins. In the first case, changing climate increases both population fecundity and the availability of suitable establishment sites across the landscape. Both factors contribute significantly to reduce dispersal limitation. In contrast, populations at rear edges undergo successive fecundity declines, and they become highly isolated within a matrix of unsuitable habitats, so colonization events and any latitudinal migration are extremely unlikely. Phylogeographical studies show that many glacial refugia separated by only a few hundred kilometres have experienced no tangible gene flow over multiple glacial cycles, although the climatic conditions were often more favourable for population expansions and interchanges than today (Hampe et al., 2003; Petit et al., 2003). Moreover, many extant temperate European plant species have recovered their current distribution ranges from only some of their glacial refugia, while others did not contribute to the recolonization due to geographical dispersal barriers (Hewitt, 2000). Numerous species have expanded much more slowly than their climatic requirements would have allowed them or did not arrive to expand at all into formerly glaciated areas during the Holocene, yet they had been present there during previous interglacial periods (Huntley, 1990; Hewitt, 2000). In conclusion, past range shifts were less deterministic and predictable than current distribution ranges — formed after a period of relative climate stability — may suggest. Pearson Hannah et al., 2002; Travis, 2003). It is doubtful to what degree past range shifts can serve to allow inference of future shifts: high model fits with current ranges do not guarantee high model realism. Moreover, the statistical methods used to date for model validation systematically overestimate the fit achieved by BEM because spatial autocorrelation exists when a given value of a variable at a given point can be predicted from its values at other points of known position. It is an inherent feature of species’ distributions across spatial scales (Koenig, 1999; Legendre et al., 2002), and climate is likewise autocorrelated over distances of up to thousands of km (Koenig, 2002). Spatially autocorrelated data are not mutually independent, and statistics that ignore this fact produce inflated type I errors as a result of pseudoreplication (Legendre et al., 2002). This means, in the present case, that model fits increase systematically with increasing size and continuity of species’ ranges: they are statistically biased. Various statistical treatments have been developed to deal with spatially autocorrelated data (Koenig, 1999; Keitt et al., 2002; Perry et al., 2002), and these need to be incorporated into BEM validations. Anthropogenic climate change is a major threat for the maintenance of biological diversity during the coming decades, and modelling is a crucial tool for evaluating its impact. Pearson & Dawson (2003) suggest using BEM as a first approximation, but the multiple drawbacks of this approach are likely to produce more ‘wildly incorrect’ predictions than Pearson & Dawson (2003) assume. Midgley et al. (2002) provide an illustrative example of how missing criticial detail within BEM may lead to overly optimistic predictions about the maintenance of species and within-species genetic diversity in a biodiversity hotspot. The study demonstrates that future conservation strategies require models that incorporate more detail and attain greater biological realism than BEM can (so far) provide (Hannah et al., 2002). Comments from Rémy Petit improved an earlier version of this manuscript.