1. Emerging infectious diseases can have serious consequences for wildlife populations, ecosystem structure and biodiversity. Predicting the spatial patterns and potential impacts of diseases in free-ranging wildlife are therefore important for planning, prioritizing and implementing research and management actions.
2. We developed spatial models of environmental suitability (ES) for infection with the pathogen Batrachochytrium dendrobatidis, which causes the most significant disease affecting vertebrate biodiversity on record, amphibian chytridiomycosis. We applied relatively newly developed methods for modelling ES (Maxent) to the first comprehensive, continent-wide data base (comprising >10000 observations) on the occurrence of infection with this pathogen and employed novel methodologies to deal with common but rarely addressed sources of model uncertainty.
3. We used ES to (i) predict the minimum potential geographic distribution of infection with B. dendrobatidis in Australia and (ii) test the hypothesis that ES for B. dendrobatidis should help explain patterns of amphibian decline given its theoretical and empirical link with organism abundance (intensity of infection), a known determinant of disease severity.
4. We show that (i) infection with B. dendrobatidis has probably reached its broad geographic limits in Australia under current climatic conditions but that smaller areas of invasion potential remain, (ii) areas of high predicted ES for B. dendrobatidis accurately reflect areas where population declines due to severe chytridiomycosis have occurred and (iii) that a host-specific metric of ES for B. dendrobatidis (ES for Bdspecies) is the strongest predictor of decline in Australian amphibians at a continental scale yet discovered.
5.Synthesis and applications. Our results provide quantitative information that helps to explain both the spatial distribution and potential effects (risk) of amphibian infection with B. dendrobatidis at the population level. Given scarce conservation resources, our results can be used immediately in Australia and our methods applied elsewhere to prioritize species, regions and actions in the struggle to limit further biodiversity loss.
Emerging infectious diseases can have serious consequences for wildlife populations, ecosystem structure and biodiversity (Crowl et al. 2008). Alarmingly, their incidence appears to be rising as a result of anthropogenic influences that favour the growth, dispersal and transmission of pathogens (Daszak, Cunningham & Hyatt 2000; Jones et al. 2008). Assessing the extent, effects and dynamics of diseases in host populations are therefore important for predicting disease emergence and its consequences, and to plan, prioritize and implement research and management actions.
The potential distribution of Bd is, however, still relatively poorly understood; the native range has not been delineated and Bd may still be expanding its range worldwide (Lips et al. 2008; Rohr et al. 2008; James et al. 2009). In Australia, it is now widely accepted that the invasion and spread of Bd is the probable cause of many frog declines (Skerratt et al. 2007). Despite this, little quantitative data on risk of disease have been available to researchers and managers at broad spatial scales, hampering efforts to pinpoint areas and species warranting immediate management attention. Tools for predicting the spread or establishment of Bd and for identifying areas of high disease risk are therefore critical for policy makers, researchers and managers charged with detecting this pathogen, developing management actions and prioritizing resource expenditure (Gascon et al. 2007; Skerratt et al. 2008).
Predicting the dispersal and potential range of organisms is commonly approached by characterizing environmental suitability (ES) with correlative species distribution models (SDMs) (Guisan & Thuiller 2005; Kearney & Porter 2009). ‘Presence-only’ SDMs are being used increasingly for their application to species occurrence data sets for which no reliable absence records may be available (e.g. museum/herbarium collections, atlases, non-targeted surveys etc.) (Pearce & Boyce 2006). Rarely used in studies of infectious disease, presence-only SDMs appear well suited to investigating the distribution of infection with some pathogens because, analogous to verifying true absence of rare or endangered species (Gibson, Barrett & Burbidge 2007), it is a statistical and sampling challenge to assert ‘freedom from disease’ (Digiacomo & Koepsell 1986; Ziller et al. 2002; Skerratt et al. 2008). This challenge is rarely met for wildlife pathogens because the cost of sufficient sampling (including diagnostics, personnel, logistics, etc.) at broad spatial scales is typically prohibitive. Furthermore, pathogen prevalence may be low in a host population or may fluctuate temporally, and the host itself may be difficult to detect, particularly if the pathogen has resulted in host declines as has been the case with Bd (e.g. Lips et al. 2006).
Correlative SDMs will only be appropriate where the distribution of infection with a pathogen is expected to be regulated by spatially quantifiable predictors that capture ES, such as climate or habitat type. For many pathogens, this may be inappropriate if hosts provide a highly regulated ‘habitat’ in which to grow and no stage of the life cycle is exposed to external environmental conditions (e.g. for internal, directly transmitted pathogens of endotherms). In the case of Bd, however, infections occur on ectothermic amphibian hosts and there is a direct effect of the environment (particularly temperature and moisture) on growth and survival of both free-living and parasitic life stages (Johnson & Speare 2003; Berger et al. 2004; Piotrowski, Annis & Longcore 2004; Woodhams et al. 2008). SDMs should thus be highly suited to characterizing ES for infection with Bd to provide important insights into its potential distribution, shed light on the probability of pathogen establishment following invasion into previously naïve areas [as has been demonstrated for other invasive species (Ficetola, Thuiller & Miaud 2007), and to help improve detection probability while reducing cost and effort of surveying for the pathogen in the future (Guisan et al. 2006)]. In an adaptive management context, such models are ideally suited to tailoring future data collection, which can in turn be used to iteratively improve the model (Wintle, Elith & Potts 2005).
We hypothesized that modelling ES for infection with Bd may also provide useful information about the risk to amphibian populations posed by chytridiomycosis. Recently, VanDerWal et al. (2009b) demonstrated that modelling ES broadly predicts an organism’s abundance. For chytridiomycosis, Bd abundance (infection intensity) on the host is a direct determinant of disease development, severity and population effects (Carey et al. 2006; Voyles et al. 2007; Briggs, Knapp & Vredenburg 2010). Indeed, seasonal and elevational variation in the prevalence, intensity and virulence of Bd infections has long implicated climatic suitability as a major factor governing its effects in the wild (Berger 2001; Berger et al. 2004; Woodhams & Alford 2005; Kriger & Hero 2007), and this has been consistently supported by laboratory infection experiments (Woodhams, Alford & Marantelli 2003; Berger et al. 2004; Carey et al. 2006). We would thus expect that our ES results not only reflect proliferation of Bd on the host at the time scale used in model training (average annual) but also the risk of severe chytridiomycosis to populations as a whole, a link we test herein by examining patterns of disease-induced amphibian population declines.
A published SDM already exists for Bd (Ron 2005), in which a correlative, presence-only SDM (GARP) with relatively few (n = 44) presence records in the New World was used to predict the global potential range of this pathogen. While this was of great use at the time of publication, the model appears to exaggerate suitable area in Australia (Fig. S1, Supporting Information), is at a spatial scale too coarse to be useful for regional management or for predicting population declines, and is likely to suffer from several sources of uncertainty inherent to correlative SDMs (e.g. extrapolation beyond the training region, limited sample size, algorithm nuances, inappropriate pseudo-absence selection; see e.g. Araujo & Guisan 2006; Pearson et al. 2007; Peterson, Papes & Eaton 2007; VanDerWal et al. 2009a). These issues together necessitate the development of independent, regionally specific predictions for planning future research and management actions on Bd at finer spatial scales.
To this end, we applied a relatively novel SDM method (Maxent) (Phillips, Anderson & Schapire 2006) to the most comprehensive, continent-wide data base available to date on the occurrence of infection with Bd (Murray et al. 2010) to model ES for this pathogen. We employed novel methodologies to deal with common but rarely addressed sources of SDM uncertainty to provide maximum robustness in our predictions of ES in Australia given the available data. The predictions were used to estimate the minimum potential geographic distribution of infection with Bd in Australia and to test the hypothesis linking ES to disease risk as indicated by patterns of disease-induced population declines. We used our results to identify where chytridiomycosis may pose the greatest risk to endangered species, allowing prioritization of species, regions and actions when considering research and management options given scarce conservation funds (Wilson et al. 2007).
Materials and methods
The SDM software used was Maxent (ver. 3.3.0), for which the underlying theory and assumptions have been described in detail elsewhere (Phillips, Anderson & Schapire 2006; Dudik, Phillips & Schapire 2007). Briefly, Maxent has been shown to generally outperform other correlative (both presence-only and presence-absence) SDM algorithms (Elith et al. 2006; Peterson, Papes & Eaton 2007; Graham et al. 2008; Wisz et al. 2008). It requires presence records only (but uses random background points to sample available environmental space), accounts for interactions among variables and identifies areas that fall beyond the range of environmental conditions used during training when making projections (identified as ‘clamped’ areas). The output of Maxent corresponds with an index of ES for the organism, where higher values correspond to a prediction of better conditions (Phillips, Anderson & Schapire 2006).
We used Bd occurrence records from Murray et al. (2010). Full details of these data and their collection methods are described in the Metadata provided therein. Briefly, this newly compiled data set represents the first comprehensive, continent-wide data base describing occurrence patterns of Bd on wild amphibian hosts. The data base comprises 821 sites in Australia at which frogs or tadpoles have been tested for Bd and includes 10 183 records from >80 contributors spanning collection dates from 1956 to 2007. Bd was detected on 63 (55%) of the 115 species in the data set (c. 28% of Australia’s 223 species) (Table S1, Supporting Information). Two hundred and eighty-four Bd-positive sites had sufficient geographic accuracy for inclusion in the model (Table 1, Fig. S2, Supporting Information). Few localities in the data base comprise statistically defensible absence records given the difficulty of asserting freedom from chytridiomycosis. The data base represents records of clinical and aclinical infection with Bd, which by definition is considered synonymous with chytridiomycosis (ranging from severe and clinical to benign and aclinical) by disease authorities (sensu Berger et al. 1998 and as per the ‘Definitions’ of the OIE’s Aquatic Animal Health Code; see http://www.oie.int/eng/normes/fcode/en_chapitre_1.1.1.htm) but distinct from including records of the free-living stage which may also be detected off-host (Kirshtein et al. 2007; Walker et al. 2007).
Table 1. Summary of Batrachochytrium dendrobatidis (Bd) data base records. Geo-referenced Bd+ sites are those where the pathogen was detected and an accurate geographic coordinate was obtained for input to the distribution model. Individuals tested is a minimum estimate; many site records in the database did not include total number of individuals tested (see Fig. S2 for map and key to State names)
Sites with records
Georeferenced Bd+ sites
Bd’s current occurrence pattern in Australia is highly consistent with the hypothesis that environmental characteristics, such as climate or habitat type, place direct limits on its distribution. Its extensive distribution nation-wide (Fig. S2; Murray et al. 2010) demonstrates that it has had sufficient opportunity to spread great distances and into new geographic areas from its hypothesized point(s) of introduction (major ports) (Murray et al. 2010). The large number of known hosts and the spectrum of potentially susceptible amphibian hosts nationally (e.g. Litoria spp.) in currently uninfected regions strongly suggest that Bd is not limited in Australia by the unavailability of susceptible amphibian host species. Similarly, its presence in some remote and sparsely populated regions of the country and its absence in some populated regions suggest that it is not simply dependent on humans for its establishment and persistence, although in some cases human aided spread seems likely (Morgan et al. 2007; Skerratt et al. 2007). In contrast, Adams et al. (2010) report that Bd occurrence in Oregon and California, USA, does not correlate with any hypothesized environmental factors, but that Bd detectability increases with human influence on the landscape. We thus also evaluated the predictive power of a human-influence hypothesis for predicting Bd’s current occurrence pattern in Australia and compared it with our ES model (Fig. S8, Supporting Information).
We used 19 bioclimatic variables (all continuous), one geo-physical variable (distance to water; continuous) and one vegetation type variable (categorical) at a resolution of c. 250 m (9 arc-seconds) for our models (Table S2, Supporting Information). We knew a priori that many of the variables were correlated and potentially meaningful contributors to the model; to avoid over-sized models (including variables with no predictive value) or over-fitted models (too many parameters for the data set) (Parolo, Rossi & Ferrarini 2008), we first selected the top ranking variables that together contributed c. 90% of the information to a full model run. We then re-ran a ‘pruned’ model with the most important variables (Table S2). Model accuracy was assessed with the area under the curve (AUC) of the receiver operator characteristic (ROC), which is a single measure of discrimination ability (presence from random background, where a value of 1 = perfect prediction, 0·5 = prediction no better than random) of the models (Fielding & Bell 1997). To incorporate uncertainty into our predictions, we used bootstrapping (N = 100) with unique sets of training and testing data (70 : 30% respectively). Many presence-only SDMs require background points (or pseudo-absences), the selection of which can influence the outcome of the models (Phillips et al. 2009; VanDerWal et al. 2009a). We provide an extended discussion of our background point selection in Fig. S2 which we used in order to limit as far as possible the effects of unquantifiable sampling bias and modelling an organism with considerable invasion potential.
To investigate the hypothesized relationship between ES for Bd and the risk of chytridiomycosis to susceptible amphibian populations, we assessed whether our results were consistent with descriptions of population decline attributed to severe chytridiomycosis (Berger et al. 1998, 2004). We anticipated that decline sites would be strongly skewed towards higher values of ES for Bd. Declines attributed to chytridiomycosis have been best described from uplands in the Australian Alps and from montane rainforest areas in Queensland, where ill and dead frogs have been rigorously diagnosed as dying from chytridiomycosis at the time of declines (Berger et al. 1998, 2004; Hines, Mahony & McDonald 1999; McDonald & Alford 1999; Osborne, Hunter & Hollis 1999) (Fig. S3, Supporting Information).
We next averaged our ES predictions across amphibian occurrence records for each species in the data set described by Slatyer, Rosauer & Lemckert (2007 updated 2009, D. Rosauer unpubl. data) to derive a species-specific metric of ES for Bd that we termed ‘ES for Bdspecies’. Slatyer et al.’s extensive data set comprises 291 942 occurrence records for all of Australia’s amphibian species. We removed duplicate records from the same locality (leaving 140 897 records; mean per species = 640) for calculations. Further details of the metric are provided in Fig. S4, Supporting Information. Species range size has previously been identified as the major risk factor for decline and extinction in Australian amphibians after controlling for other life-history and ecological factors (Murray & Hose 2005). We thus tested for an effect of ES for Bdspecies, controlling for the range size effect, in contributing to whether amphibians have experienced declines or not. Amphibian trend classifications were sourced from the IUCN (2008). Range sizes were calculated from extent of occurrence polygons developed for the Global Amphibian Assessment (GAA) (Stuart et al. 2004).
Finally, we calculated mean ES values across Australia’s biogeographic regions to identify those most suitable for infection with the pathogen (Fig. S5, Supporting Information). We related these results to amphibian species richness and endemism statistics from the study of Slatyer, Rosauer & Lemckert (2007) to indicate where infection with Bd most threatens anuran biodiversity in Australia (Fig. S6, Supporting Information).
Model selection, validation and variable contributions
After the pruning step, mean test AUC was 0·900 (range 0·874–0·925) and the model contained eight variables. The jack-knife procedure, which examines the effect of individual variables, indicated that mean diurnal temperature range and annual precipitation had the most useful information as single variables on training data (highest gain scores in isolation) as well as the highest predictive power (highest AUC in isolation) (Fig. 1). Response curves characterizing the relationships between ES and each of the two most influential predictor variables are shown in Fig. S7, Supporting Information. In the comparative analysis incorporating human population density (HPD), predictive performance of the full model was unchanged (0·903, range = 0·852–0·936) and HPD had inferior predictive power in isolation (AUC = 0·763, range = 0·716–0·811) relative to many of the environmental predictors. For subsequent analyses, we thus used the model incorporating environmental variables only (see Fig. S8, Supporting Information for results and further discussion).
The model suggested that infection with Bd should be largely restricted to the eastern and southern seaboards of Australia, with nearly all of inland and northern Australia unsuitable. Figure 2 represents the average Maxent predictions of ES (available for download in Appendix S2, Supporting Information). Clamping indicated that all of Australia fell within the environmental limits used to train the model (data not shown).
Decline sites (mean ESdecline = 0·758; 95% CI = 0·714–0·802, n = 39) were highly skewed towards higher ES values compared to all sites used for model training and testing (mean ESall = 0·577, 95% CI = 0·550–0·604, n = 284) (Fig. 4a,b,d).
Mean ES for Bdspecies varied between population trend categories (Fig. 3a); three extinct species had the highest value, 42 declining species had an intermediate value and 151 stable species exhibited a comparatively low value. In a logistic model in which species were grouped by whether they had declined or not (unknown trend species omitted), ES for Bdspecies was a highly significant predictor of decline (Δdev = 20·932, d.f. = 1, P <0·001), even after controlling for a significant influence of narrow species range size (Δdev = 22·831, d.f. = 1, P <0·001). The best model in terms of AIC contained ES for Bdspecies as a highly significant term (P <0·001), range size as a marginally significant term (P =0·098) and no interaction term. Table S3, Supporting Information presents a list of priority species for research and management indicating where investigation of Bd as a potential threatening process is critical. Table S4, Supporting Information presents the full list of Australian species.
Mean ES varied considerably across biogeographic regions (Fig. 3b); the Wet Tropics (see also Fig. 4a) was predicted to have the highest mean suitability for Bd, followed by the Central Mackay Coast (Fig. 4b), Tasmania’s southern ranges, northern slopes, north-east (Ben Lomond) and King Island (Fig. 4e) and the NSW north coast (Fig. 4c). South-east Queensland (Fig. 4c), the Australian Alps (Fig. 4d), the Swan Coastal Plain (around Perth) (Fig. 4f) and the Tasmanian south-east also showed high mean ES values. Many regions with low mean ES nevertheless showed limited areas of very high ES as indicated by their maximum values (e.g. Brigalow Belts, Einasleigh Uplands, NSW south-western slopes) (Fig. 3b).
Infection with B. dendrobatidis occurs across a broad range of climates in Australia, in areas that are at times very hot, cold, dry or wet. Those locations range from the hot, humid coastal lowlands of north-eastern Australia to the highest peaks of the Australian Alps, where winter snow occurs. Despite its broad tolerance of conditions, the model suggested that specific environmental conditions will restrict infection with Bd to the generally cooler and wetter areas of Australia (Figs 2 and 4). In this respect, our model was highly consistent with that of Ron (2005) (Fig. S1; Fig. 2); however, our results suggested that Bd should be more restricted, with the majority of central (arid) Australia being broadly unsuitable for Bd persistence (see also Fig. S8).
The model indicated that ES increased with annual precipitation (with a minimum extreme of c. 500 mm) (Fig. S7). This is not surprising since desiccation is known to rapidly kill Bd in vitro (Berger 2001; Johnson et al. 2003) and the presence of permanent water is known to be an important feature for sustaining Bd, probably because the transmission stage for Bd is an aquatic zoospore (Berger et al. 1998). The model also suggested that mean diurnal temperature range was an important variable; the response curve indicated that ES declined rapidly in highly variable temperature regimes, where the difference in daily maxima and minima is greater than c. 11 °C. Variation in temperature of itself has not previously been shown to affect chytridiomycosis (Woodhams, Alford & Marantelli 2003). However, high temperatures are known to be lethal to Bd and the effect of temperature variability may be explained by the observation that areas with higher temperature variability (e.g. the arid/semi-arid interior of the country) also typically exhibit very high maximum temperatures. This suggestion is supported by the response of Bd to maximum temperature of the warmest month, which showed maximum ES in the range of maximum temperatures 18–30 °C beyond which there is a precipitous decrease (data not shown). Our results are thus highly consistent with those of previous studies indicating that high temperatures are detrimental to Bd (Kriger & Hero 2007; Muths, Pilliod & Livo 2008; Puschendorf et al. 2009).
Two key results from this study are that (i) our predictions of ES are strikingly consistent with known associations between Bd and amphibian population declines in Queensland/New South Wales (Fig. S3 and Fig. 4a–c) (Hines, Mahony & McDonald 1999; McDonald & Alford 1999) and in the Australian Alps (Osborne, Hunter & Hollis 1999; Berger et al. 2004) (Fig. 4d) and (ii) the species-specific metric of ES for Bd (termed ES for Bdspecies) was a very strong predictor of amphibian decline at a national level. These findings support the hypothesis that ES for infection with the pathogen as modelled here is broadly predictive of suitability for, and severity of, the disease chytridiomycosis via a theoretical and empirical link with intensity of infection (VanDerWal et al. 2009b). We thus interpret our ES for Bd results as being a highly useful source of quantitative information relevant to explaining the potential effects of infection with B. dendrobatidis (disease risk).
This association should be interpreted cautiously, however, as in order for ES for Bdspecies to translate to risk of decline a number of other conditions relevant to the epidemiology of chytridiomycosis must be fulfilled, most importantly transmission. Species with more aquatic life-histories and an association with permanent water are most susceptible and at greatest risk of severe disease (Berger et al. 1998, 2004). Further, species inhabiting different micro-habitats can vary in their relative risk of infection within a single location (Woodhams & Alford 2005; Skerratt et al. 2008). As such, actual disease risk will be a product of the ES for the pathogen, the susceptibility of the species and the factors that make it susceptible to decline (e.g. see Bielby et al. 2008) given the former.
This is an important consideration as it will be necessary to stratify host life-history traits for prioritization purposes (Table S3). Bielby et al. (2008) found that small range size, altitude and an aquatic life stage are risk factors for rapid decline in Bd-positive species. However, applying these risk factors to all species as they do is a considerable extrapolation because not all species are equally susceptible to infection. For example, very high-risk values in that study were assigned to many species with largely terrestrial life-histories, including many of Australia’s microhylid frogs (e.g. Cophixalus sp). While some of these also exhibit high ES for Bdspecies values as described herein, neither index identifies actual risk from Bd as species in this group appear far less susceptible than stream-dwelling and permanent water-associated species from the same region (N = 557 negative results in areas that are Bd-positive; K. Hauselberger & D. Mendez et al. unpubl. data; Skerratt et al. 2008). A more sophisticated risk analysis can be performed when more information is available about the innate susceptibilities of different amphibian species to chytridiomycosis and to decline. Integrating host-life history and ecological traits with the pathogen’s environmental requirements (as modelled here) to predict infection and decline is the focus of our current research efforts (Murray et al. in press).
We have shown that ES for Bdspecies is a strong predictor of decline at a continental scale. This result was independent of a previously reported, dominant effect of narrow geographic range size. Our study provides a species-specific metric, representing the environmental requirements of the pathogen, with which to begin to assess this risk and calls for targeted vigilance in sampling for this disease and monitoring for its potentially insidious effects (Murray et al. 2009; Pilliod et al. 2010).
Bd records exist from most regions that were deemed suitable by the model, indicating that it has probably reached its broad geographic limits on this continent. There are, however, at least two areas that show marginal suitability for Bd beyond the known range of the pathogen where testing has failed to detect it: Cape York (Skerratt et al. 2008) (Fig. 4a) and Tasmania’s World Heritage south-west (Pauza & Driessen 2008) (Fig. 4e). It is possible that Bd has simply not yet dispersed to these regions, as they are at the extreme limits of the distribution in northern and southern Australia. However, the results of this study may also suggest that establishment or disease risk could be relatively low in these regions. Our results provide a testable hypothesis and surveys should continue in these areas where suitability is predicted to be highest (Fig. 4a,e). Prevention of spread nevertheless remains the best management strategy and these areas should not be regarded as areas in which Bd could not establish and cause mortalities (Fig. 2). Hygiene protocols should therefore be enforced for people entering these areas (Phillott et al. 2010).
Several other marginal to highly suitable regions exist where no sampling has occurred. These regions represent important areas for future sampling to establish the actual geographic limit of Bd in Australia and to establish amphibian population health. Identification of naive populations at high disease risk is a particular priority. Examples include uplands in the far north of Queensland (Cape York; Fig. 4a), upland areas in the Brigalow Belt (Fig. 4c), a large expanse of the western slopes of the Great Dividing Range in NSW (Fig. 4c), the south-west central tablelands of NSW and the regions surrounding Mount Gambier and the Mt Lofty Ranges in South Australia (Fig. 2).
Conversely, our results suggest that several declining species for which chytridiomycosis is a suspected threatening process may have relatively low ES for Bdspecies (e.g. Litoria piperata and Litoria castanea), which provides a way forward when considering management and research activities. Similarly, some declining species may be at high risk of disease only in a subset of populations (e.g. Litoria spenceri, as indicated by maximum vs. mean ES values – Table S3). We are not asserting that Bd will be absent from these species or populations, but that other factors may also be involved in their decline, an example of where our results provide some useful and testable hypotheses that should be pursued.
Regions of high disease risk in association with high host endemism should be the highest priority for population monitoring and/or management activity. Of the 11 centres of exceptional anuran endemism identified by Slatyer, Rosauer & Lemckert (2007), six occur in regions predicted to be highly suitable for Bd, including the Wet Tropics (Fig. 4a), Central Mackay Coast (Mackay/Eungella; Fig. 4b), Gladstone (Kroombit Tops), South-east Queensland (Gympie-Coffs Harbour; Fig. 4c) and south-west Western Australia (Walpole and Bunbury-Augusta; Fig. 4f) (see Fig. S5 for ecoregion names and Fig. S6 for endemism/richness). Records of Bd exist from all of these areas. An additional two areas (Townsville and Cape York) are predicted to have more restricted regions that are marginally or highly suitable for Bd (Fig. 4a). Three endemism hotspots are predicted to be at negligible risk from Bd (Kakadu and the Arnold River region in the NT and the Mitchell Plateau in WA). Establishing and maintaining a disease-free status should be their regional priority.
The methods and results from this study can be used as a tool for establishing cost-sharing arrangements, prioritizing future efforts to detect and manage this pathogen (e.g. disease surveys, preventing further spread to naïve areas), for prioritizing monitoring programmes for Bd and Australia’s anuran fauna (e.g. Skerratt et al. 2008) and for identifying priority species for potential emergency captive-breeding programmes (Gascon et al. 2007) (Table S3). We envisage this to be an iterative process, with models such as ours regularly updated and scrutinized as new systematically collected data accrue (Wintle, Elith & Potts 2005). Critically, our methods can be directly and rapidly applied to other regions of the world experiencing amphibian declines; such results will aid in the task of developing informed management and surveillance decisions for Bd (Skerratt et al. 2008) and will help to make the most of limited conservation funds for prioritizing species, regions and actions for biodiversity conservation outcomes (Wilson et al. 2007).
Limitations and future directions
While our model had high predictive performance and clamping indicated a well sampled environmental space, relatively little sampling has occurred on the western margins of the Great Dividing Range and in inland Australia, and Queensland and Western Australia were over represented compared with other regions. In addition, frogs may take refuge in environments that are not captured by interpolated bioclimatic or vegetation mapping data (see also Fig. S8) and we have limited ability to incorporate microclimatic features into our models given the enormous diversity of amphibian hosts and their habitats in this country. Similarly, beyond considerations of HPD (Fig. S8) we were unable to incorporate models of pathogen dispersal given very limited knowledge regarding how this pathogen is spread. We model the realized niche of this invasive species in an invaded range; there is thus the possibility that Bd’s distribution in Australia has not approached an equilibrium state, potentially resulting in an underprediction of its potential range. We consider this an unlikely source of major bias in our results given Bd’s extensive distribution nation-wide and the spectrum of potentially susceptible amphibian hosts (e.g. Litoria spp.) and hypothesized vectors (e.g. humans) in currently uninfected regions. Nevertheless, our model represents a baseline, minimum potential distribution rather than a finite prediction of this organism’s fundamental niche; we encourage scrutiny and ongoing iteration (e.g. integrated use of new systematically collected data), particularly to increase representation of apparently disease-free areas into future models. Dispersal models should also be a future priority, particularly in areas that are newly invaded. Finally, genetic differentiation has been noted geographically (Morgan et al. 2007; James et al. 2009), and strains may undergo local adaptation (Fisher et al. 2009) and/or show strain specific differences in adaptive plasticity so distribution in Australia with respect to available environmental space may not necessarily correspond exactly to other regions or to the results of other predictive models. Comparison of these and future studies will thus identify important areas and avenues for further research and it is imperative that the predictions of any SDM be independently compared with other SDM methods and data sources (see Elith et al. 2006), other methods (e.g. mechanistic models; K.A.M. unpubl. data) (Morin & Thuiller 2009) and by comprehensive field surveys during sampling periods that maximize detection probability (Skerratt et al. 2008).
We are indebted to the many authors and contributors named in Murray et al. (2010) for the production of the Bd occurrence data base. In particular, we thank K. McDonald, K. Aplin, H. Hines, D. Mendez, A. Felton, P. Kirkpatrick, D. Hunter, R. Campbell, M. Pauza, M. Driessen, S. Richards, M. Mahony, A. Freeman, A. Phillott, J-M. Hero, K. Kriger and D. Driscoll. KAM thanks D. Segan, M. Watts and C. Klein for GIS wisdom and spatial data, R. Wilson and H. Possingham for lab space, B. Sutherst, M. Zalucki and D. Kriticos for fruitful discussions and M. Araujo and R. Pearson for running a timely SDM workshop at the University of Queensland. We also thank Dr Marc Cadotte, Professor Christl Donnelly and three anonymous reviewers for excellent comments and discussion on earlier versions of the manuscript. KAM was supported by an Australian Postgraduate Award, an Australian Biosecurity CRC professional development award and a Wildlife Preservation Society of Australia student research award. Part of this work was conducted when RWRR was supported by the Australian Research Council, the School of Public Health and Tropical Medicine, James Cook University and a National Science Foundation Integrated Research Challenges in Environmental Biology grant awarded to J. Collins at Arizona State University, USA. RWRR thanks J. Collins and C. Carey.