SPECIAL ISSUE Model-based uncertainty in species range prediction Richard G. Pearson1*, Wilfried Thuiller2, Miguel B. Araujo3,4��, �� Enrique Martinez-Meyer5, Llu��s �� Brotons6, Colin McClean7, Lera Miles8, Pedro Segurado9, Terence P. Dawson10 and David C. Lees11 INTRODUCTION Environmental niche models utilize associations between environmental variables and known species distributions to define abiotic conditions within which populations can be maintained. Projection of modelled niches into new regions and under scenarios of future climate change enables the geographical distribution of suitable conditions to be predic- ted. This approach has been widely applied, including in studies investigating the potential impacts of climate change 1Department of Herpetology and Center for Biodiversity and Conservation, American Museum of Natural History, New York, NY, USA, 2Laboratoire d���Ecologie Alpine UMR CNRS 5553 BP53, Grenoble Cedex 9, France, 3Biodiversity Research Group, Oxford University Centre for the Environment, Oxford, 4Biogeography and Conservation Laboratory, The Natural History Museum, London, UK, 5Instituto de Biolog��a,�� Universidad Nacional Autonoma �� de Mexico, �� Ciudad Universitaria, Mexico City, Mexico, 6Centre Tecnolologic ` Forestal de Catalunya, Pujada del Seminari s/n, Solsona, Catalunya, Spain, 7Environment Department, University of York, Heslington, York, 8UNEP World Conservation Monitoring Centre, Cambridge, UK, 9Unidade de Macroecologia e Conservac ��ao, �� Universidade de Evora, �� Estrada dos Leoes, �� Evora, �� Portugal, 10Centre for the Study of Environmental Change and Sustainability, University of Edinburgh, Edinburgh, UK and 11Entomology Department, The Natural History Museum, London, UK *Correspondence: Richard G. Pearson, Department of Herpetology and Center for Biodiversity and Conservation, American Museum of Natural History, Central Park West at 79th Street, New York, NY 10024-5192, USA. E-mail: pearson@amnh.org ��Present address: Department of Biodiversity and Evolutionary Biology, National Museum of Natural Sciences, CSIC, C/Jose �� Gutierrez �� Abascal, 2, 28006 Madrid, Spain. ABSTRACT Aim Many attempts to predict the potential range of species rely on environmental niche (or ���bioclimate envelope���) modelling, yet the effects of using different niche-based methodologies require further investigation. Here we investigate the impact that the choice of model can have on predictions, identify key reasons why model output may differ and discuss the implications that model uncertainty has for policy-guiding applications. Location The Western Cape of South Africa. Methods We applied nine of the most widely used modelling techniques to model potential distributions under current and predicted future climate for four species (including two subspecies) of Proteaceae. Each model was built using an identical set of five input variables and distribution data for 3996 sampled sites. We compare model predictions by testing agreement between observed and simulated distributions for the present day (using the area under the receiver operating characteristic curve (AUC) and kappa statistics) and by assessing consistency in predictions of range size changes under future climate (using cluster analysis). Results Our analyses show significant differences between predictions from different models, with predicted changes in range size by 2030 differing in both magnitude and direction (e.g. from 92% loss to 322% gain). We explain differences with reference to two characteristics of the modelling techniques: data input requirements (presence/absence vs. presence-only approaches) and assumptions made by each algorithm when extrapolating beyond the range of data used to build the model. The effects of these factors should be carefully considered when using this modelling approach to predict species ranges. Main conclusions We highlight an important source of uncertainty in assessments of the impacts of climate change on biodiversity and emphasize that model predictions should be interpreted in policy-guiding applications along with a full appreciation of uncertainty. Keywords Bioclimate envelope modelling, biodiversity, Cape Flora, climate change, conservation biogeography, distribution modelling, environmental niche mod- elling, Proteaceae, South Africa, species biodiversity. Journal of Biogeography (J. Biogeogr.) (2006) �� 2006 The Authors www.blackwellpublishing.com/jbi 1 Journal compilation �� 2006 Blackwell Publishing Ltd doi:10.1111/j.1365-2699.2006.01460.x

on biodiversity (e.g. Peterson et al., 2002 Midgley et al., 2003 Thomas et al., 2004 Hannah et al., 2005 Thuiller et al., 2005 Araujo �� et al., 2006), conservation prioritization (e.g. Araujo �� & Williams, 2000 Ferrier et al., 2002 Raxworthy et al., 2003 Williams et al., 2005), range filling (Svenning & Skov, 2004), niche evolution (Peterson et al., 1999 Martinez-Meyer et al., 2003 Graham et al., 2004 Mart��nez-Meyer �� & Peterson, 2006), factors governing species distributions (Coudun & Gegout, �� 2006 Luoto et al., 2006) and the geographical ecology of invasive species (Higgins et al., 1999), agricultural pests (Baker et al., 2000) and disease vectors (Costa et al., 2002). However, whilst the modelling approach is generic, studies have employed a number of different techniques for defining potential ranges (e.g. Nix, 1986 Stockwell & Peters, 1999 Pearson et al., 2002 Thuiller, 2003 Thuiller et al., 2003 Miles et al., 2004 Segurado & Araujo, �� 2004 McClean et al., 2005 Maggini et al., 2006) and the impact that the specific method has on model predictions is an important consideration in model applications (Thuiller et al., 2004a). Here we assess consistency in predictions from nine of the most widely applied environmental niche modelling approa- ches. Using identical input variables, each model was used to simulate current and potential future distributions for four species of Proteaceae that are endemic to South Africa���s Cape Floristic Kingdom. We compare predictions by testing agree- ment between observed and simulated distributions, and by assessing consistency in predictions of changes in range size under future climates. Previous studies have demonstrated important differences between predictions arising from different data sample sizes (Stockwell & Peterson, 2002) and species range sizes (McPherson et al., 2004 Segurado & Araujo, �� 2004). Our focus here is on differences between predictions from different modelling techniques (see also Loiselle et al., 2003 Segurado & Araujo, �� 2004 Thuiller, 2004 Araujo �� et al., 2005b). We highlight significant differences between models and demonstrate that the magnitude of variation between predictions can be very large. Further analysis of our results enables two key factors causing differences between model predictions to be identified: data input requirements and model extrapolation assumptions. METHODS The modelling approaches we tested were: artificial neural networks, with two alternative parameterizations, ANN1 (Pearson et al., 2002) and ANN2 (Thuiller, 2003) the climate envelope range (CER) (Nix, 1986 similar to BIOCLIM) the constrained Gower metric (CGM) (Miles et al., 2004 similar to DOMAIN) classification tree analysis (CTA) (Thuiller et al., 2003) genetic algorithm (GA) (McClean et al., 2005) the generalized additive model (GAM) (Segurado & Araujo, �� 2004) genetic algorithm for rule-set prediction (GARP) (Stockwell & Peters, 1999) and the generalized linear model (GLM) (Thuiller, 2003). Each modelling technique was implemented with close adherence to published studies (see Appendix S1 in Supplementary Material for details) and using the same five climatically derived input variables. We studied a region of the Western Cape extending from 17��86�����20��79�� E and 31��91�����34��83�� S. Five model input vari- ables considered to be critical to plant physiological function and survival were gridded for this region at a spatial resolution of 1�� �� 1�� (Schulze, 1997 Midgley et al., 2002). The variables used were mean minimum temperature of the coldest month, heat units calculated as the annual sum of daily temperatures (��C) exceeding 18 ��C, annual potential evaporation (calculated as the sum of mean monthly A-pan equivalent potential evaporation figures derived using the Penman���Monteith method), winter soil moisture days and summer soil moisture days. Soil moisture days are calculated by a hydrological model and are defined as those days on which soil moisture is above a critical level for plant growth (Midgley et al., 2002). Input variables under a climate warming scenario (IS92a) for 2030 were calculated using projections from the general circulation model HadCM2 interpolated to 1�� �� 1�� resolution (as detailed in Schulze & Perks, 1999). Two species and two subspecies whose distributions had contrasting spatial characteristics were selected so as to test model performance across a range of distribution types. Distributions were characterized by the number of occupied grid cells (occupancy) and the straight-line distance between the two most distant occupied grid cells (extent of occurrence) (Segurado & Araujo, �� 2004). The species studied were: Diastella divaricata subsp. divaricata (restricted area of occupancy and low extent of occurrence) Leucospermum hypophyllocarpoden- dron subsp. hypophyllocarpodendron (restricted area of occu- pancy and high extent of occurrence) Leucospermum tomentosum (large area of occupancy and low extent of occurrence) Protea longifolia (large area of occupancy and high extent of occurrence). The inclusion of two subspecies was considered appropriate since in each case the subspecies have distributional and functional characteristics that are strong enough to distinguish a taxonomic grouping that is driven by climate. Leucospermum hypophyllocarpodendron subsp. hypo- phyllocarpodendron is distinguished from its sister subspecies by the absence of leaf pubescence and differing leaf shape. These characteristics are associated with the less arid and cooler climate of the southern Cape lowlands, rather than the warmer western lowlands which are occupied by its sister. In the case of D. divaricata subsp. divaricata, this species is strongly distin- guished from its sister by leaf size and shape, and is associated with the much warmer and drier conditions of the lowlands, rather than the cooler montane environment of its sister subspecies. In neither case are the subspecies sympatric with their sisters. Species distributional data were available as presence and absence for 3996 sampled sites (Rebelo, 1992). Each sampled site was located within a different 1�� �� 1�� cell distributed across the gridded study region (total of 23,875 cells). Environmental niches were defined using each technique based on an identical randomly selected 70% of the sampled sites. The remaining 30% of the sampled data were used to test the agreement between modelled and observed distributions (Araujo �� et al., 2005a). The 70:30 ratio of this random split approximately R. G. Pearson et al. 2 Journal of Biogeography �� 2006 The Authors. Journal compilation �� 2006 Blackwell Publishing Ltd