全球变化知识资源中心

To estimate future overlap of currently isolated congeneric species, we used previously published bioclimatic models built for 2,954 New World birds (n = 1,818), mammals (n = 723) and amphibians (n = 413; ref. 24). Bioclimatic models relate species’ current distributions to historical climate. These models are then used to project the distribution of suitable climates into the future on the basis of output from global climate models. The models were built in R using random forest classifiers^{25, 26}. Although other modelling approaches could potentially provide different projections of future climatic suitability, we used an approach that proved to be more accurate at projecting current ranges when compared with five other approaches applied to a subset of the species used in the present study²⁷. For each species, presences and absences were modelled as a function of current climate. Species distribution data were taken from digital range maps for birds²⁸, mammals²⁹ and amphibians (data available on-line, http://www.globalamphibians.org). Species ranges were mapped to a 50-km grid. Both current and future climatic conditions were represented by 37 bioclimatic variables (Supplementary Table 1). These included both annual and seasonal variables, and basic climate variables (for example, temperature and precipitation) as well as derived variables (for example, a moisture index and growing degree days). To represent current climatic conditions, historical climate data were downscaled from the University of East Anglia Climatic Research Unit’s (CRU’s) CL 1.0 (ref. 30), CL 2.0 (ref. 31) and TS 2.1 (ref. 32) climate data sets to the 50-km by 50-km grid using locally weighted, lapse-rate-adjusted interpolation. The current climate was represented by averaged climatic variables over a 30-year period (1961–1990). Future climate data were taken from 10 general circulation model (GCM) simulations archived by the World Climate Research Programme’s (WCRP’s) Coupled Model Intercomparison Project phase 3 (CMIP3). These climate projections were averaged over a 30-year period from 2071 to 2100 and represent climates simulated for a mid-to-high (SRES A2) greenhouse gas emissions scenario³³. Models included in this analysis had a mean accuracy rate of 99% for absences and 92% for presences in a subset of locations not used during model construction. For a more detailed description of the modelling approach, see ref. 24.

We measured the prevalence of current range overlap for 9,577 congeneric species pairs (note that many species have multiple congeners), including 3,858 avian, 1,661 mammalian and 4,058 amphibian species pairs. Ranges included only breeding distributions for migratory species. Congener status was determined on the basis of shared genus names in 2009. We estimated end-of-century (2071–2100) areas of suitable climate for each species on the basis of projected climate from 10 GCMs using the A2 greenhouse gas emissions scenario, which represents the mid-to-high range of the scenarios described in the Intergovernmental Panel on Climate Change (IPCC) Special Report on Emissions Scenarios³³ (SRES). We then calculated the number of non-overlapping species pairs whose projected future distributions overlap. We used model agreement among 6 or more of the 10 GCMs as a cutoff for determining which species pairs are projected to come into contact by the end of the century.

We used a generalized linear mixed model (GLIMMIX procedure, SAS University Edition; SAS Institute) to test whether current geographic range size, taxonomic class (birds, mammals and amphibians), or the number of species classified as tropical in a species pair (neither, one, or both species with >50% geographic range between latitudes −25° and 25°), as well as interactions between these factors, significantly influenced the occurrence of future overlap of species pairs. The GLIMMIX procedure models normal and non-normal data with correlated responses. We used a binomial distribution and logit link function for the response variable (overlap versus non-overlap). As most species had multiple congeners and therefore occurred in multiple species pairs, we included species as an R-side (residual) effect, and we modelled the covariance structure using variance components. An R-side effect is equivalent to a repeated measures effect, but the GLIMMIX procedure does not provide type III (analysis of variance) estimates for variance components. Thus, we accounted for within-subject correlations in our analysis, but we did not test for the significance of species. We estimated degrees of freedom for F-tests using the Kenward–Roger method to suppress inflation of type 1 error³⁴. Additionally, we used analyses of variance to evaluate differences in current range size among taxa and to test whether the degree of future range overlap and asymmetry increases with the number of tropical species in a pair. Finally, we used linear regressions to examine relationships between current geographic range size and proportion of future range in overlap, and between current and future geographic range size. We used a natural log transformation of all range size and overlap data, as distributions of these variables were right-skewed²³.