Study area and species selected
We studied the distribution of T. marmoratus and T. pygmaeus in mainland Spain. The two species overlap slightly in the Tagus River basin (Pleguezuelos et al., 2004; Espregueira Themudo et al., 2012), where hybridization events are believed to have occurred (García-París et al., 2001; Espregueira Themudo et al., 2012). Distribution records of the Portugal atlas of amphibians and reptiles (Loureiro et al., 2008) were excluded because they do not distinguish between T. marmoratus and T. pygmaeus. Also, Portugal and France were not included as part of our study area because of incompatibilities between the high-quality climatic predictions available for the different countries (see ‘Distribution modelling’ below). Thus, our distribution models locate the favourable environmental conditions for Triturus populations in mainland Spain.
Spain has a complex climate because of its orography and geographical situation between two continents (Europe and Africa) and two water masses – the Atlantic Ocean and Mediterranean Sea. The annual distribution of precipitation is highly heterogeneous (Font, 2000; Ninyerola et al., 2007). The Mediterranean part of the Iberian Peninsula is part of a biodiversity hotspot (Myers et al., 2000); 41% of the amphibian species are endemic and about 35% are threatened according to the IUCN (Pleguezuelos et al., 2004).
The distribution data for both species were taken from Pleguezuelos et al. (2004). Species distributions are represented on a 10 km × 10 km UTM grid (Fig. 1). The total number of cells studied was 5161; T. pygmaeus occurs in 468 cells and T. marmoratus in 1046 cells. The distribution of both species in Spain occupies 1505 cells, about 30% of the study area. The absence data set corresponded to all grid cells where the modelling taxa were not observed.
Figure 1. The study area in the European context. Current distributions of Triturus pygmaeus and T. marmoratus represented in 10 km × 10 km UTM cells within the study area of mainland Spain (data from Pleguezuelos et al., 2004).
Download figure to PowerPoint
Models forecasting future species distributions should include the impact of climate change and other factors that are not expected to change or whose change is not actually foreseeable (Márquez et al., 2011). We modelled the distribution of taxa using variables related to four explanatory factors: climate, spatial distribution, topography and human activity (see Appendix S1 in Supporting Information). We modelled the distribution of taxa considering two taxonomic options: one considering T. pygmaeus (Tp) and T. marmoratus (Tm) as parapatric species, and another considering them together as two subspecies of the former T. marmoratus species (T).
We took into account uncertainty in the climate predictions, and so climatic variables were used according to different socioeconomic scenarios and atmosphere–ocean general circulation models. We considered two IPCC SRES emission scenarios (Nakićenović et al., 2000) – A2 and B2 – which represent different socioeconomic scenarios but avoid extremes; and two atmosphere–ocean general circulation models – CGCM2 (Canadian Centre for Climate Modelling and Analysis) and ECHAM4 (Max Planck Institut für Meteorologie), which were regionalized to mainland Spain by the Spanish Meteorological Agency (AEMET) (Brunet et al., 2007). This improves quality in the climatic predictions with respect to the forecasts available from WorldClim (Hijmans et al., 2005), which used only 134 stations in Spain, whereas 373 stations for temperature and 2326 stations for rainfall were used for mainland Spain by the AEMET.
For each period 1961–1990, 2011–2040, 2041–2070 and 2071–2100, 12 climate variables (Appendix S1) were obtained. The variables regarding climate, spatial distribution, topography and human activity were obtained following Márquez et al. (2011). We included the square of elevation and slope in the topographical factor to evaluate a possible unimodal response of the species distribution to these variables. In addition, variables related to a polynomial trend-surface analysis (Legendre & Legendre, 1998) were considered in order to include the effect of factors that may involve purely spatial trends, such as history or spatial ecological dynamics (Legendre, 1993).
Present distribution models for each taxon were constructed using forward–backward stepwise logistic regression. These models were performed using separately each set of variables related to one of the following explanatory factors: spatial distribution, topography and human activity. For the climate factor, three models were constructed according to the climate variable values estimated for the period 1961–1990 using all possible combinations of the general circulation models and emission scenarios. ECHAM4 applied to the 1961–1990 period proposes the same precipitation and temperature values for both A2 and B2 scenarios, but these values differ in predictions for the 21st century (Márquez et al., 2011). Thus, we obtained three climate models: CGCM2-A2, CGCM2-B2 and ECHAM4-A2/B2. To avoid type I errors due to multiple tests (Benjamini & Hochberg, 1995), we controlled the false discovery rate (FDR) in every model using the procedure proposed by Benjamini & Hochberg (1995), entering in the model only variables that were significant under an FDR of q < 0.05.
Other models that combined the four explanatory factors (hereinafter combined models) were built by performing a stepwise regression using the set of variables entered in the models based on a single factor. Thus, three combined models were obtained for each taxon (Tp, Tm and T) for the period 1961–1990, each one considering a different climate model. Finally, the favourability function (Real et al., 2006) was applied to the combined models. This function allows direct comparison of favourability values for species differing in their prevalence (Acevedo & Real, 2012).
Five criteria were used to assess the fit of the models: sensitivity; specificity; correct classification rate (CCR), based on the 0.5-favourability threshold – which, in the favourability function, makes probability be equal to overall prevalence; the area under the curve of the receiver operating characteristic (AUC), which is independent of any favourability threshold (Hosmer & Lemeshow, 2000); and the parsimony test based on the Akaike information criterion (AIC; Akaike, 1973). Finally, a model calibration test was applied (Landis & Koch, 1977) to test the extent to which the observed presences fitted those predicted by each model.
The combined models were projected into the future to obtain forecasts about variations in favourability for each taxonomic option. With this aim, the values of the climate variables in the models corrected by the favourability function were replaced by their corresponding future values.
For every favourability forecast (Ff) considering the nine combinations of general circulation models, emission scenarios and time periods for each of the three taxonomic groups, some fuzzy logic parameters were used to assess to what extent the initial favourability (F0, period 1961–1990) was modified because of climate change (Kuncheva, 2001):
where I represents increment and M represents maintenance. In these equations, c(FX) is the cardinality of X favourability – where favourability is treated as a fuzzy set (Estrada et al., 2008) – that is, the sum of all the cells' favourability values (these are treated as degrees of membership in the fuzzy set); and the intersection between future and present favourability values is defined as follows:
Positive values of increment (I) indicate a net increase in favourability for the taxon, that is, a gain in favourable areas, whereas negative values of I mean a net loss of favourable areas. Maintenance values (M) indicate the degree to which the favourable areas in F0 overlap with the favourable forecasted areas.
Assessment of uncertainty derived from taxonomic ambiguity
The uncertainty in forecasts regarding the taxonomic options was analysed by comparing the predicted distribution areas and the fit of the models based on Triturus spp. with the union of the predicted areas of occurrence for T. marmoratus and T. pygmaeus as separate species. The differences between the favourability forecasts for Triturus spp. (FT) and for the combination of both species (FTp ∪ FTm) were visualized using ArcGIS 9.3 (ESRI, Redlands, CA, USA) by mapping the degree of discrepancy (D) attributable to the taxonomic criterion:
where ; D = 0 indicates no uncertainty, D = +1 indicates maximum discrepancy when the predicted favourability for Triturus spp. is the highest, and D = −1 indicates maximum discrepancy when the predicted favourability for the united model for separate species is the highest. A total discrepancy (TD) value was quantified by adding all absolute discrepancies throughout the study area:
where n = 5161, that is, the total number of cells in which the presence/absence of taxa has been recorded in the study area.
The uncertainty analysis regarding climate models and scenarios was performed to improve assessment of the importance of the uncertainty derived from taxonomic ambiguity. The uncertainty of models that forecast species distributions according to different emission scenarios should be reasonably low to be useful for policy planning. The consistency of forecasts derived from applying different methodological options to the same emission scenario should be higher than the coincidence between forecasts derived from applying the same methodological options to different scenarios (Real et al., 2010). Consistency, when using different circulation models, was formulated as follows:
where c(FX) is the cardinality of X favourability (see above in the previous subsection) and FC and FE are the predicted favourability for taxa according to the circulation models CGCM2 or ECHAM4, respectively.
When using different taxonomic criteria, consistency was formulated as:
Coincidence was mathematically calculated as:
where, for a given taxonomic criterion and circulation model, FA2 and FB2 are the favourability for taxa predicted using emission scenarios A2 or B2, respectively. Consistencies and coincidences were compared using the Student's t-test.