Uncovering correlates of decline and critical refuges for a threatened terrestrial mammal Authors: Natalya M Maitz, Michelle S Ward, Jonathan R Rhodes, Ben C Scheele, Hugh P Possingham, Diana O Fisher Published: 2026 ABSTRACT AbstractGlobally, the geographic ranges of numerous species are contracting. Identifying spatiotemporal patterns of threat impact can illuminate why species decline in some parts of their range but persist in others. We developed a correlative approach to identify species’ response thresholds and locate ecological refuge areas associated with persistence to inform conservation planning and management. We integrated changes in realized niche characteristics for a threatened Australian mammal, the brush‐tailed rock‐wallaby (Petrogale penicillata), with the distributions of threatening processes and environmental features to identify predictors of extirpation and persistence. We used occurrence records to construct historical (2004 or before) and contemporary (2005–2022) species distribution models. We then determined differences in niche characteristics between areas of extirpation and persistence. We applied generalized additive mixed models to determine species‐specific thresholds for extirpation, which we used to identify ecological refuges. Across the two periods, brush‐tailed rock‐wallaby habitat declined by 75%, resulting in realized niche shifts. Extirpations occurred where habitat distance to water increased, and red fox (Vulpes vulpes) densities and feral goat (Capra hircus) habitat suitability were relatively high. Relatively high topographic ruggedness and high minimum temperatures were associated with persistence. Species persistence was also positively correlated with increasing dingo (Canis familiaris) habitat suitability, and inappropriate fire regimes (e.g., more frequent fires than expected for a vegetation group), both of which have been suggested as threats to the species. Specific refuge locations were identified in southeastern Queensland and northern New South Wales. Our findings highlight the importance of quantitative species‐specific threat assessments to prevent misguided conclusions from cross‐taxa threat generalizations. As spatial data become increasingly accessible, our novel approach will provide practitioners with evidence on which to base fine‐scale investigations. Our approach can be applied to other species undergoing range contractions.Keywords: anthropogenic threats, applied conservation, conservation actions, Maxent, overgeneralization, range contraction, research–implementation gap, threatened species, Amenazas antropogénicas, brecha entre la investigación y la aplicación, conservación aplicada, especies amenazadas, Maxent, medidas de conservación, reducción del área de distribución, sobregeneralización, 应用型保护, 人为威胁, 保护行动, Maxent, 过度概括, 分布范围缩小, 研究‐实践的空缺, 受威胁物种 INTRODUCTION At least 1 million species are at risk of extinction, and losses are expected to amass in the upcoming decades (IPBES, 2019). In response, the international community has established the Kunming–Montreal Global Biodiversity Framework, which aims to halt and reverse global losses (CBD, 2022). For this effort to be successful, species‐specific information on key threatening processes is needed to guide on‐the‐ground conservation actions because the relative importance of threats varies depending on the species, scale, and metrics considered (Bellard et al., 2022). This is especially important for species undergoing geographic range contractions, a problem affecting many species across the globe (Pacifici et al., 2020). Range contraction not only reduces the number of individuals but also diminishes genetic diversity and connectivity, which contributes to biological annihilation (Ceballos et al., 2017; Rogan et al., 2023). Because the Global Biodiversity Framework is centered on mobilizing action, understanding the main drivers of species’ range contraction is crucial for informing the management needed to curb species declines and prevent future extinctions. Species conservation status is commonly assessed by quantifying range decline, but recent work also highlights the value of examining changes in species’ realized niches in relation to persistence (Breiner et al., 2017; Britnell et al., 2023; Scheele et al., 2023; Sopniewski et al., 2024). To facilitate research on niche shifts in declining species, Scheele et al. (2017) proposed the niche reduction hypothesis. This approach evaluates species’ historic (predecline) and contemporary niche characteristics to identify potential mechanisms that might explain why species decline in certain areas but not in others. The underlying premise is that variable patterns of threat occurrence and magnitude combine with environmentally mediated heterogeneity in species’ threat tolerance. These factors result in nonrandom contractions that disproportionately reduce species’ realized niche breadth. Moreover, by quantifying changes in species’ niche characteristics, key threats, and the environmental factors that support persistence, it is possible to identify ecological refuges (Scheele et al., 2017; von Takach et al., 2020). Ecological refuges are areas that buffer species from anthropogenic impacts (Pavey et al., 2017; Selwood & Zimmer, 2020). Refuge areas within the contemporary distribution can be identified as priorities for management and restoration (“realized refuges”), whereas areas within the extirpated distribution can be targeted for reintroduction efforts (“potential refuges” [Monsarrat et al., 2019]). Comparisons between historical and contemporary ranges can be used to identify changes in preferred environmental features (Rutrough et al., 2019). For example, in northern Australia, the geographic range of the common brushtail possum (Trichosurus vulpecula) has contracted to areas of high rainfall, where increased shrub density likely provides abundant foraging resources and predator protection (Stobo‐Wilson et al., 2019). Recent research has taken historical and contemporary comparisons a step further by directly incorporating threatening processes. However, these comparisons tend to include only a few anthropogenic threats and often focus on landscape dynamics, such as habitat fragmentation (Cheeseman et al., 2024; Wang et al., 2020; Zhang et al., 2023). Australia's mammals contribute disproportionately to global extinctions, and its endemic terrestrial mammal populations continue to decline (Legge et al., 2023; Woinarski et al., 2015). Around 40% of Australian marsupials have experienced substantial range declines (Fisher et al., 2003). Key threats include invasive predators (predominantly red foxes [Vulpes vulpes] and feral cats [Felis catus]), habitat degradation from inappropriate fire regimes (e.g., too frequent, out‐of‐season, or too intense) and overgrazing, invasive vegetation, forest loss and fragmentation, and climate change (Legge et al., 2023; Woinarski et al., 2015). However, the causes of species‐specific declines are often poorly known, and the relative importance of major threats is judged on little evidence (Fisher & Humphries, 2024). For example, a review by Santos et al. (2022) found that inappropriate fire regimes in Australia were cited as an important threat to 88% of threatened land mammals, yet empirical evidence existed for only 40% of these species. The remaining assessments relied on anecdotal evidence or expert reasoning. Because anecdotal evidence and overgeneralizations can mislead conservation management, systematic methods are needed to assess and quantify the relative importance of proposed threats, rather than relying on precedent (Britnell et al., 2021; Caley, 2022; Caughley, 1994; Sutherland et al., 2004). We focused on the brush‐tailed rock‐wallaby (Petrogale penicillata) (hereafterrock‐wallaby), an Australian vulnerable marsupial that has undergone substantial range contraction since European colonization (Short & Milkovits, 1990; Woinarski & Burbidge, 2016). Although threatening processes, including the above mechanisms of Australian mammal loss, have been proposed to drive rock‐wallaby decline, their relative roles have not been quantitatively assessed (Lavery et al., 2021). We applied the niche reduction hypothesis to assess a suite of potential drivers of decline for the rock‐wallaby. Specifically, we aimed to determine whether proposed threats are correlated with extirpation, identify environmental factors associated with persistence, and integrate the resultant information to identify priority ecological refuge areas for management and provide practitioners with a way to assess multiple potential threats to inform fine‐scale investigations for management efforts. Such accessible methods to elucidate decline and guide action are essential if the international community is to reach its goal of reversing biodiversity loss (Williams et al., 2020). METHODS We built on previous approaches used to compare historical and contemporary niche characteristics by identifying environmental features and threatening processes correlated with extirpation or persistence (Alexander et al., 2022; Moore et al., 2019; von Takach et al., 2020). First, we used maximum entropy models to delineate historical and contemporary rock‐wallaby distributions (Maxent 3.4.4; Phillips et al., 2006). Using these models, we then generated an extirpated distribution. To compare ecological predictors between extirpated and contemporary distributions, we used random sampling points and applied a binomial generalized additive mixed model (GAMM) to examine relationships between these predictors and rock‐wallaby extirpation. Data preparation and analyses were conducted in ArcGIS Pro 3.3.3 (Esri Inc., 2023) and R 4.2.3 (R Core Team, 2024). Focal species and study region The rock‐wallaby, a small macropod (5–11 kg), inhabits cliff lines and rugged outcrops in localized colonies from Nanango, Queensland, through New South Wales (NSW), to Victoria's Grampians. This range encompasses three evolutionarily significant units (ESUs) with substantial genetic differentiation that requires unique management strategies (Australian Government, 2021). Historically, the species was widely distributed throughout southeastern Australia, likely inhabiting lower elevations with simple topography rather than being restricted to rugged terrain (Jarman & Bayne, 1997) (Figure 1). Since European colonization, substantial declines have been documented, with the southern and western portions of the historical range particularly affected (Short & Milkovits, 1990). Numerous populations in these regions have been extirpated, including all populations in the Australian Capital Territory (ACT). The last wild sighting in ACT was in 1959 (ACT Government, 2020). Ex situ populations have become crucial, especially in Victoria, where one of the two remaining populations is entirely composed of reintroduced animals after the last individual was brought into captivity in 1999 (Taggart et al., 2016). Brush‐tailed rock‐wallaby (a) occurrence records and α hulls used in historical and contemporary Maxent model construction and (b) binary Maxent model outputs for the historical (2004 or before), contemporary (2005–2022), and extirpated rock‐wallaby distributions. Many of the threats proposed to explain rock‐wallaby decline are associated with the extinction of other Australian mammal species. Losses began with European occupation, and until the 1920s, tens of thousands of rock‐wallabies were hunted for their fur and because of their perceived status as agricultural pests (Short & Milkovits, 1990). Further declines are thought to have been driven by predation from red foxes, feral cats, dingoes (Canis familiaris), habitat loss and degradation, invasion by Lantana camara (a “weed of national significance”), inappropriate fire regimes, competition with introduced herbivores, changes in temperature and precipitation patterns, hydatid disease, and toxoplasmosis (Australian Government, 2021). Although there is considerable confidence in the role of foxes and habitat loss in the decline of rock‐wallabies, uncertainty remains regarding the relative importance and severity of other threats, such as cats, weeds, herbivore competition, and temperature changes (Australian Government, 2021). However, even for high‐confidence threats, evidence remains largely anecdotal because there are few quantitative studies of these threatening processes (Lavery et al., 2021). Species occurrence data We amalgamated rock‐wallaby occurrence records from an array of sources, including national, state, museum, and private expert databases. These included the Atlas of Living Australia (ALA, 2023), Bionet (NSW Government, 2023), Wildnet (Queensland Government, 2023), Victorian Biodiversity Atlas (DEECA, 2023), ACT Government (Ormay, 1996), academic records (J. Murray, M. Eldridge, & H. Hines, personal communication, 2022), and Museums Victoria (personal communication, 2023). We initially screened records by removing those outside the accepted historical range as per government reports (Menkhorst & Hynes, 2011). Records associated with captive breeding efforts, unconfirmed verification status, duplicates, missing species names, or missing time frames were removed. We cross‐checked outlying records against existing literature to validate species occurrence and extirpation, and removed records without documented support to ensure reliability (NSW Government, 2008). We split the refined dataset into two periods: historical records covering 1760–2004 and contemporary records covering 2005–2022. We selected 2004 as the cutoff point because most local extirpations had occurred by then (Australian Government, 2021) (Appendix S1). Furthermore, proposed threats were well established across the species’ entire distribution by 2004 (Appendix S2). Because rock‐wallaby home ranges span 0.06–0.30 km2 (Menkhorst & Hynes, 2011), we spatially thinned each dataset to one record per square kilometer to minimize spatial autocorrelation (Kramer‐Schadt et al., 2013). After thinning, 763 and 499 records remained in the historical and contemporary datasets, respectively. A few historical records, specifically those in western NSW (e.g., the Cobar region), have high location uncertainties (up to 10 km2). We retained these records given their importance in delineating historical extent and the documented evidence that supports the historical occurrence of rock‐wallabies in these regions (Short & Milkovits, 1990). Our historical records are largely limited to post‐European colonization and lack prehistoric archaeological records. Given the early impact of humans on Australian megafauna (van der Kaars et al., 2017), our models may not capture the prehuman range for the species (i.e., >70,000 years ago) (Monsarrat et al., 2019). Environmental and threat data We collated available spatial datasets for key proposed threatening processes and environmental features identified as important in previous rock‐wallaby distribution models (Appendix S3). Of the threatening processes included in the National Recovery Plan for the bush‐tailed rock‐wallaby, we obtained data for most proposed threats, excluding disease and historical hunting impacts (Menkhorst & Hynes, 2011). To avoid model overfitting, we condensed categorical datasets into rock‐wallaby relevant categories; surface geology was sorted into major rock types (e.g., metamorphic, igneous, sedimentary) and vegetation groups into dominate forest types (e.g., rainforest, eucalypt, grasslands, etc.) (Appendices S4 & S5 contain respective details). Rock type was included as an important landscape predictor for the species because it affects the weathering of geological formations and the subsequent development of complex physical refuges (Murray et al., 2009). For dynamic environmental predictors, such as precipitation, temperature, and productivity, we collected data spanning the longest available period (Appendix S3). For datasets that did not extend back to 1760, we assumed that their averages are indicative of the longer‐term trends (von Tackach et al., 2020). Because temporal data for most threats are unavailable, we used a snapshot approach to compare the historical rock‐wallaby distribution with current threat distributions and incorporated threat intensity metrics when they were available. All spatial datasets were converted to raster format and set to the same projection (Geocentric Datum of Australia, 2020), extent, and resolution (approximately 1 km2). Construction of historical, contemporary, and extirpated distribution models We constructed Maxent models to capture the historical ecological niche and contemporary niche of rock‐wallabies. First, we constructed α hulls for historical (1760–2004) and contemporary (2005–2022) niches with the alphahull package (Pateiro‐Lopez & Rodriguez‐Casal, 2022) (Figure 1a). Unlike simple convex polygons, α hulls create concave shapes that better represent species range contours and accurately capture distribution discontinuities rather than overestimating occupied areas. Following International Union for Conservation of Nature (IUCN) recommendations (2024), we set the α parameter to 2 to control boundary detail with smaller values so that shapes would more closely follow point distributions. To ensure we fully captured habitat accessible to the rock‐wallabies, we added a 3‐km buffer to the α hull that reflected the species’ estimated dispersal distance (Hazlitt et al., 2004). Next, we assessed predictor correlations with 1000 randomly sampled points from within the historical and contemporary α hulls. We used a correlation matrix for continuous variables (|Pearson's coefficient| ≥ 0.7) (Appendix S6), chi‐squared tests for categorical predictors (Cramer's V ≥ 0.4) (Appendix S7), and analysis of variance to examine associations between categorical and continuous predictors (η2 ≥ 0.20) (Appendix S8). For Maxent model construction, we omitted predictors with multiple significant correlations (e.g., slope, rainfall, temperature, productivity, horses, land clearing, grazing, and vegetation). We excluded temperature change from the historical model due to its temporal nature. Using the respective α hulls as masks for environmental and threat data, we built Maxent models for both time periods with all remaining predictors. Sampling bias is a significant concern for presence‐only models. Therefore, we included a target bias file that influences how background points are selected to account for areas with disproportionate sampling effort (Barber et al., 2022; Phillips & Dudík, 2008). We used Atlas of Living Australia records for four other common macropod species that have range overlap with the rock‐wallaby: the eastern grey kangaroo (Macropus giganteus), red kangaroo (Osphranter rufus), swamp wallaby (Wallabia bicolor), and red‐necked wallaby (Macropus rufogriseus). These macropod records were combined and split at 2004 to establish historical and contemporary bias files that accounted for temporal differences in survey effort. We then constructed a two‐dimensional Gaussian kernel density grid for each occurrence dataset (Fourcade et al., 2014). Maxent allows parameter fine‐tuning beyond default settings, a practice recommended for single‐species models (Anderson & Gonzalez, 2011; Merow et al., 2013). We explored all feature types, but ultimately selected a combination of linear, quadratic, and product features in favor of model parsimony. Although hinge features marginally improved model performance, they produced overly complex response curves that were difficult to interpret (Syfert et al., 2013). We tested different regularization parameters ranging from 0.5 to 3.0 at 0.5 increments. Using 10‐fold cross‐validation to assess model performance, we proceeded with values that yielded the greatest area under the curve (AUC) (Bang et al., 2022). To transform the resulting average habitat suitability outputs into binary presence–absence distribution maps, we applied the respective maximum sum of sensitivity and specificity values (Liu, Newell, et al., 2016): historical (0.259) and contemporary (0.296). Finally, we created an extirpated distribution by erasing the contemporary distribution from the historical (Figure 1b). Geographic range and niche changes We used area in the α hulls to compare historical and contemporary extents of occurrence. We also calculated the area of suitable rock‐wallaby habitat (i.e., available habitat but not necessarily occupied) using the binary historical and contemporary Maxent models to quantify habitat loss over time (Brooks et al., 2019). To explore differences in niche characteristics between areas where rock‐wallabies still occur and those no longer occupied, we randomly sampled 3000 points within the contemporary and extirpated maps to compare predictor density distributions (Alexander et al., 2022; Rutrough et al., 2019). Sampled points were spaced 3 km apart as a conservative means to reduce spatial autocorrelation. As this step was exploratory, we included all predictors (Appendix S3). Exploring potential drivers of decline and persistence Initially, to explore how predictors were associated with rock‐wallaby extirpation, we used the randomly sampled points (n = 3000 each) from the extirpated and contemporary distributions as response variables in a binomial generalized additive model (GAM). This model was built with the mgcv package (Wood, 2017a), which applies smooth functions and provides a flexible approach to assess nonlinear relationships (Wood, 2017b). To account for spatial autocorrelation, latitude and longitude were included as a two‐dimensional smoothed covariate. Despite incorporating geographic coordinates in the model, significant residual spatial autocorrelation persisted, potentially affecting model inference (Gaspard et al., 2019). Therefore, we employed a GAMM that allowed for incorporation of a spatial autocorrelation structure. We tested exponential, spherical, and Gaussian spatial correlation structures and selected the one with the lowest Akaike information criterion (AIC) (Liu, Wan, et al., 2016). Because GAMMs with spatial structures require more processing resources than GAMs (Wood, 2006), we had to reduce our input data to 400 records from both the contemporary and extirpated distributions to enable convergence within our computational constraints. To construct the multivariate GAMM, we included all predictors except those with high correlation (|r| ≥ 0.7) in our initial Pearson's assessment (e.g., between slope and ruggedness). In such cases, we retained the predictor with the greatest deviance explained in univariate GAMs. To avoid overfitting and aid in model interpretability, we set all basis functions to five. These functions control the complexity of the smoothed curves in the model; fewer basis functions produce simpler, more interpretable relationships. We started with a full model and sequentially removed one predictor at a time, specifically the predictor with the highest p value, until all remaining predictors were significant (p < 0.05) (Wood, 2025). We also assessed predator interactions with ruggedness because rugged environments are understood to provide refuge from predatation (Morris et al., 2020). Although backward selection has limitations, including the tendency to eliminate important variables (Marra & Wood, 2011), it provides a pragmatic compromise given the computational demands of fitting spatial correlation structures. Our considerable sample size (n = 800) likely helped mitigate these concerns by improving the reliability of backward elimination (Heinze et al., 2018). We applied the gam.check and concurvity functions to assess model performance, again removing predictors with high concurvity (worst ≥0.8) (Robinson et al., 2023). Once the above checks were made, we conducted a Moran's I test to ensure no residual spatial autocorrelation remained. Refuge identification Although our Maxent models provided insights into general habitat suitability, we used the GAMM outputs to locate potential refuge areas within the contemporary distribution of the rock‐wallaby. The species’ response curves generated from the GAMM allowed us to determine the key predictors associated with extirpation or persistence, which enabled us to identify features likely to contribute to the long‐term survival of the species (Reside et al., 2019). We defined refuges as areas with an 80% probability of persistence (or 20% probability of extirpation) based on the GAMM's predictions. Our choice of 80% was arbitrary; however, we considered this a sensible value in terms of balancing species survival and ensuring we identified a reasonable amount of refuge area to help guide management. We estimated predictor values at an 80% probability of persistence with the predict function and set all other variables at their median values (mode for categorical). Based on the directional relationships between variables and extirpation probability identified by the GAMM, we conditionally selected pixels from rasters that fell within the contemporary distribution. For example, if higher average monthly rainfall was associated with rock‐wallaby persistence and there was an 80% probability of persistence at a median of 70 mm, we designated pixels within the contemporary distribution ≥70 mm as potential refuge areas. Finally, all rasters to which a threshold was applied were overlaid to locate high‐quality refuge areas, that is, where multiple predictors supporting persistence overlapped. Exploring the role of scale The rock‐wallaby's distribution spans a considerable latitudinal gradient (27°S to 37°S). Therefore, we anticipated that the relative importance of predictors driving extirpation (or persistence) would vary throughout its range. To investigate how this may have potentially affected outputs, we partitioned the same randomly sampled data from the extirpated and contemporary distributions (n = 3000 each) based on their respective states (ACT was included in NSW here). This resulted in 800 records for Queensland (extirpated 395, contemporary 405), 4830 records in NSW (extirpated 2245, contemporary 2585), and 370 records in Victoria (extirpated 360, contemporary 10). Due to the limited contemporary records remaining within Victoria (n = 10), our investigations at the state level were restricted to exploratory comparisons of predictor medians because the data were insufficient for robust GAMM analyses. Although partitioning the range based on the three ESUs may have been more biologically meaningful, the required spatial data delineating these ESUs were not available. Given uncertainty about exact ESU boundaries (Eldridge et al., 2017) and our intent to demonstrate method application across scales, we used state boundaries, which capture some of the geographic variation in threats (Appendix S9) and reflect the administrative units at which management decisions are implemented. RESULTS Changes in geographic distribution and niche characteristics Brush‐tailed rock‐wallabies have undergone a severe range decline since European colonization. Comparison of α hulls revealed that rock‐wallabies have lost 71% (or 356,572 km2) of their historical extent of occurrence (499,578 km2) (Figure 1a). Maxent models constructed using these α hulls indicated excellent performance, with average test AUC values of 0.88 (SD 0.02) for both historical and contemporary distributions. After converting the resulting models to binary presence–absence maps, we estimated that the 101,288 km2 of rock‐wallaby habitat in the historical distribution was reduced to 25,748 km2 (by 75%) in the contemporary distribution. In subtracting the contemporary distribution from the historical, we estimated the area of extirpation to be 76,071 km2 (Figure 1b). Most of the contemporary distribution (25,216 km2) remained within the historical distribution, indicating 98% of the rock‐wallaby habitat change resulted from geographic range contraction and 2% from range shift. Maxent variable response curves showed similar patterns across historical and contemporary periods. Increases in fox and cat density, habitat degradation, and distance to water were associated with a decline in rock‐wallaby habitat suitability (Appendix S10). Meanwhile, increases in dingo habitat suitability, ruggedness, and lantana habitat suitability were associated with increasing rock‐wallaby habitat suitability. Although directional patterns were mostly similar, response values and curve shapes differed for some predictors between time periods. Comparison of the 3000 sampled points from the contemporary and extirpated distributions revealed no significantly different medians for elevation, aspect, or drought across the two distributions (Figure 2; Appendix S11). Median indices for ruggedness and dingo and lantana habitat suitability had substantially higher medians in contemporary distributions. Median ruggedness of the contemporary distribution was considerably (44%) higher than where the species was found historically (from 245 [95% CI 238–253] to 353 [95% CI 345–361], Cliff's δ = −0.35, p < 0.001), dingo habitat suitability was 11% larger (from 0.61 [95% CI 0.61–0.62] to 0.68 [95% CI 0.68–0.69], Cliff's δ = −0.38, p < 0.001), and lantana habitat suitability was 164% larger (from 0.14 [95% CI 0.13–0.15] to 0.37 [95% CI 0.36–0.38], Cliff's δ = −0.34, p < 0.001). Foxes and cats had higher median densities in the extirpated distribution. Median fox density in the contemporary distribution declined by 36% from the extirpated distribution (from 0.45 individuals/km2 [95% CI 0.43–0.47] to 0.29 individuals/km2 [95% CI 0.28–0.30], Cliff's δ = 0.46, p < 0.001), indicating a potential rock‐wallaby shift to areas with low fox densities. Median cat density in the contemporary distribution was only slightly less (9%) than in the extirpated distribution (from 0.22 individuals/km2 [95% CI 0.21–0.22] to 0.20 individuals/km2 [95% CI 0.20–0.20], Cliff's δ = 0.32, p < 0.001). Density distributions and counts of ecological variables from sampled points in the contemporary (n = 3000) and extirpated (n = 3000) Maxent‐derived distributions for the brush‐tailed rock‐wallaby (green, areas of overlap between the two periods; habitat suitability indices: 0, unsuitable habitat; 1, high‐suitability habitat; productivity, measure of vegetation density; ruggedness, index of terrain ruggedness; EUC, eucalyptus; WDL, woodlands; GRSL, grasslands). See Appendix S4 (geology) and Appendix S5 (vegetation) for full category descriptions and groupings. Potential drivers of decline and persistence Response curves from the final multivariate GAMM identified several significant predictors of rock‐wallaby population status. Goat suitability, fox density, and distance to water were positively associated with extirpation, whereas average minimum temperature, dingo suitability, cat density, presence of inappropriate fire regimes, and ruggedness positively correlated with persistence (Figure 3). Although all tested predator terrain interactions (i.e., dingo with ruggedness, cat with ruggedness, and fox with ruggedness) were significant, only the fox–ruggedness interaction reduced model AIC and was thus retained. Persistence was associated with areas of higher topographic ruggedness where fox densities were low, whereas extirpation was associated with areas of lower ruggedness and high fox densities. Generalized additive mixed model response curves for persistence relative to extirpation for the brush‐tailed rock‐wallaby (blue shading, 95% confidence intervals for smoothed variables). Plots are shifted to incorporate the intercept and its uncertainty to represent simulations based on the assumption that other variables are held at their average. For the parametric variables, inappropriate fire regime (i.e., more frequent fires than expected for a relevant vegetation group) and distance to water, partial effects represent the log odds of extirpation. Model was fit with 800 points (400 each from extirpated and contemporary distributions). The final model incorporated an exponential correlation structure to account for spatial autocorrelation and achieved an adjusted R 2 value of 0.57, and we considered this good given the complexity of ecological systems. Using the 5200 randomly sampled points not used in model construction, we validated its ability to differentiate between extirpated and contemporary points. The AUC was 0.92, which indicated strong discriminatory performance. When we iteratively dropped and reinstated individual terms from the model to determine each predictor's contribution to the adjusted R 2, ruggedness and minimum temperature were found to be the most important predictors in determining extirpation (Appendix S12). Refuges identified in northern NSW and Queensland Examination of individual predictors associated with ≥80% probability of rock‐wallaby persistence within the contemporary distribution revealed the following conditions: distance to water ≤321 m (31% of contemporary range), minimum temperature ≥10.21°C (55%), ruggedness ≥345 (49%), goat suitability ≤0.06 (12%), fox density ≤0.20 individuals/km2 (25%), dingo suitability ≥0.73 (19%), cat density ≥0.22 individuals/km2 (13%), and areas with inappropriate fire regimes (27%) (Figure 4c). When considering regions where at least one predictor met its threshold, the total refuge area was 25,376 km2 (99% of the contemporary distribution). However, this area rapidly shrank as we increased the number of co‐occurring thresholded predictors. Where two predictors overlapped, 9836 km2 remained (38% of contemporary distribution); where three predictors overlapped, 7604 km2 remained (30%); where four predictors overlapped, 2602 km2 remained (10%); where five predictors overlapped, 403 km2 remained (1.6%); and where six predictors overlapped, 34 km2 remained (0.13%). There were no regions where seven or eight predictors overlapped. Clusters of high‐quality refuge (four or more overlapping predictors) appeared at D'Aguilar, Main Range, and Mount Barney National Parks in Queensland and in NSW at Oxley Wild Rivers, Guy Fawkes, Nymboida, Chaelundi, and Yengo National Parks (Figure 4b). Overall, northern NSW had the most high‐quality refuges (Figure 4a). When considering land tenure, 9928 km2 (39%) of the contemporary distribution was on freehold land (least restrictive form of Australian property ownership that continues indefinitely), and 11,633 km2 (45%) was on nature conservation reserves. Identified refuge areas for the brush‐tailed rock‐wallaby within the species’ contemporary distribution when each predictor is set at values that predict an 80% probability of persistence based on the genialized additive mixed model, showing the (a) entire contemporary distribution and (b) Nymboida National Park, which illustrates the spatial variation in refuge quality (count of overlapping predictors that reflect an ≥80% probability of rock‐wallaby persistence), and (c) individual predictor values used to locate refuge areas (e.g., all pixels for fox density ≤0.20 individuals/km2). Magnitude of threats and environmental features depend on scale Similar to national patterns, state‐level analyses revealed contemporary distributions were characterized by lower median fox densities and higher ruggedness indices across all three states (Appendix S13). However, the magnitude of these directional shifts varied across states. For instance, median ruggedness in Queensland increased from 122 (95% CI 110–135) in the extirpated distribution to 308 in the contemporary distribution (95% CI 291–326, Cliff's δ = −0.77, p < 0.001), whereas median ruggedness in NSW increased from 244 (95% CI 236–252) to 361 (95% CI 352–370, Cliff's δ = −0.38, p < 0.001). Similarly, the contemporary median for dingo suitability was smaller in Victoria (0.53, 95% CI 0.50–0.56) compared with Queensland (0.64, 95% CI 0.63–0.65) and NSW (0.69, 95% CI 0.68–0.69). Inferences for Victoria were limited by the small contemporary sample size (n = 10). However, Victoria's extirpated data (n = 360) indicated a higher median fox density of 0.63 individuals/km2 (95% CI 0.59–0.68) compared with Queensland (0.44 individuals/km2, 95% CI 0.39–0.48) and NSW (0.42 individuals/km2, 95% CI 0.40–0.44). DISCUSSION Examining environmental and threat correlates of species’ declines across their distributions can provide key insights into threatening processes and identify potential refuge areas. We ranked the importance of potential drivers of decline and estimated species’ response thresholds for contemporary persistence, enabling us to locate environmental refuges to inform conservation management plans. The rock‐wallaby underwent a 71% reduction in geographic range and a 75% loss in habitat area compared with its historical distribution (<2005). This substantial decline falls within previous estimates of 50–90% (Short & Milkovits, 1990). We assessed multiple proposed threats to identify possible drivers of extirpation and found that relatively high fox densities and goat habitat suitability were associated with rock‐wallaby extirpation, providing quantitative support for the IUCN Red List's designation of foxes and feral goats as primary threats (Woinarski & Burbidge, 2016). Some perceived threats, however, showed unexpected positive associations with rock‐wallaby persistence: increasing habitat suitability for dingoes, relatively high cat densities, and presence of inappropriate fire regimes. Some of these unexpected positive associations may be explained by biological mechanisms, such as dingoes suppressing foxes through mesopredator release (Johnson et al., 2007). Similarly, for inappropriate fires, rocky outcrops may buffer rock‐wallabies from direct mortality (Lavery et al., 2021), allowing the species to benefit from improved foraging during subsequent grass regeneration—provided they survive the increased initial postfire predation risk (Doherty et al., 2023). Although our correlative approach identified several significant relationships, we interpret these findings cautiously, particularly when clear biological mechanisms are lacking. For instance, the positive association between cats and rock‐wallaby persistence likely represented an artifact of correlations with rainfall or productivity, rather than suggesting cats directly promote persistence. The lack of negative impact may also reflect that adult rock‐wallabies exceed the critical weight range targeted by feral cats (Burbidge & McKenzie, 1989), although cats could still affect population dynamics through juvenile predation and subsequent reduced recruitment (Australian Government, 2021). Overall, our model suggested that cats are potentially not primary drivers of extirpation and we emphasize that although cats were positively associated with persistence, this does not necessarily mean they were actively promoting persistence per se. We also hypothesized that heterogeneity in environmental features would be associated with persistence. Based on model contributions, the two most important predictors explaining rock‐wallaby persistence were topographic ruggedness and minimum temperature. Other contemporary Australian mammalian niches have shifted toward more complex, rugged terrain, possibly because these landscapes offer greater protection from threats, specifically introduced predators (McDonald et al., 2018; Moore et al., 2019). The positive relationship between minimum temperature and persistence also complements previous research showing rock‐wallaby preference for relatively high minimum temperatures (Ashcroft et al., 2014; Morris et al., 2020). This is possibly driven by their sun basking behavior to avoid cold stress, which places them at increased predation risk (Morris et al., 2020). Areas closer to water sources were associated with persistence, which is consistent with research showing declining rock‐wallaby habitat suitability as the distance from water increases (Thurtell et al., 2021). This pattern may again reflect increased predation risk (including hunting, historically) or vehicle collision when individuals must travel farther from rocky refuges to access water. Rocky environments (alongside montane and riparian) are common landscape‐level refuges for species, primarily in the context of protection from fire (Selwood & Zimmer., 2020). Although research often focuses on refuges in relation to a single stressor, species often face multiple interacting threats, making it important to understand their relative contribution to decline and how species‐specific refuges may mitigate these impacts (Reside et al., 2019). By considering a suite of proposed threats, our method identifies refuges in the context of multiple threats, highlights correlates of decline or persistence, and provides thresholds to directly inform action. For example, we found a focus on fox and goat control, and potentially controlled burns, may be the most effective management actions for rock‐wallabies, whereas lantana and feral cat control may not be as beneficial. Moreover, maintaining fox densities below 0.20 individuals/km2 is advisable if attempting to ensure an 80% probability of persistence. The ability to quantify species’ responses to gradients of threat intensity, while considering the influence of other predictors, provides a valuable conservation tool to help guide management (Ingram et al., 2021) and enable decision triggers to prompt on‐the‐ground interventions (Foster et al., 2019). Conservation strategies commonly focus on threat management in areas of high‐value habitat or on bolstering the viability of remnant populations. Our approach, which focuses on identifying the threat landscape as a unique feature of the habitat, represents an opportunity to simultaneously identify threats and areas of high‐value habitat, which helps elucidate where habitat enhancements may be most beneficial (Selwood & Zimmer, 2020). For example, we found 30% of high‐quality refuge areas (four or more overlapping predictors) occurred on freehold land, locating specific areas where collaboration with landowners on threat mitigation could more effectively secure species within their existing range. The relative emphasis on managing refuges versus restoring historical habitat will depend on species‐specific needs and whether threats can be effectively managed, with both reactive and proactive approaches complementing each other in comprehensive conservation planning (Allan et al., 2019). Furthermore, our approach could be applied to extirpated portions of a species’ distribution to locate potential reintroduction sites that are more feasible to manage and could further incorporate climate change projections to establish true anthropogenic refuges (Monsarrat et al., 2019; Seddon, 2010). However, refuges do not guarantee species persistence and require active management (Kinnear et al., 2002), especially in cases of potential refuges where the original causes of extirpation have not yet been fully mitigated (Britnell et al., 2021). Our framework facilitates refuge identification at multiple spatial scales. First, we identified refuges at the landscape level, a scale often overlooked in refuge studies (Selwood & Zimmer, 2020). The ability to assess declines at the landscape scale is important as threat impact is dependent on surrounding local context (Bellard et al., 2022), and it is at these levels that conservation actions are often implemented. For instance, we identified northern NSW as a general region of high‐quality refuge habitat and highlighted important conservation reserves, including Main Range, Oxley Wild Rivers, and Chaelundi National Parks. Second, we pinpointed specific sites within the national parks, providing practical site‐specific guidance for land managers. The ability to prioritize efforts at broad scales and then at specific locations can maximize efficiency in conservation planning (Gilby et al., 2021). Third, we demonstrated how this approach can be scaled down to extract region‐specific insights. For example, contemporary fox densities in Victoria (0.45 individuals/km2) were nearly identical to those in Queensland's extirpated distribution (0.44 individuals/km2). These comparisons show that densities of foxes in the contemporary distribution in Victoria are similar to those at which extirpation occurred elsewhere, highlighting the utility of regional assessment and potentially providing explanation for the severe rock‐wallaby decline at higher latitudes. The unanticipated associations we found show the importance of quantitatively assessing threats to mitigate potential overgeneralizations and bias (Britnell et al., 2021). Failure to consider all possible threats or misunderstanding their relative importance can lead to ineffective management outcomes. For example, nest parasitism by brown‐headed cowbirds (Molothrus ater) was assumed to be the primary threat to Kirtland's warbler (Setophaga kirtlandii), yet intensive cowbird management provided no obvious benefit. It was only after two consequential wildfires that jack pine regeneration was identified as the critical limiting factor for warbler habitat (Walker et al., 2003). Similarly, rock‐wallaby habitat use had been subject to researcher bias. Early research stated the species used only north‐facing slopes, which became a dogma that pervaded the literature for decades. However, the species is now understood to inhabit all aspects, meaning this north‐facing bias may have misled management efforts by overlooking suitable habitat (Soderquist, 2011). Our findings reinforce the importance of species‐specific threat assessments to help reduce overgeneralizations that can lead to misguided conservation efforts. A correlative approach based on occurrence records has limitations. Presence‐only models are contingent on sampling, occurrence, and detection probability (Yackulic et al., 2013). Although we addressed sampling bias by using a target‐group background approach, current data limitations precluded formally incorporating detection probability—the likelihood of detecting the species when present—into our models. This limitation may affect our interpretations, particularly regarding unexpected relationships. For instance, dingoes may prefer less complex terrain at the site level, which could simultaneously improve observer detectability of rock‐wallabies. Such detection bias could partially explain the unexpected positive association we found between dingo presence and rock‐wallabies. Furthermore, for these models, we assumed that geographical occurrence fully represented the ecological niche (Pagel et al., 2020). Ideally, demographic information would be incorporated to more fully explain population dynamics and further disentangle drivers of decline (Plard et al., 2020). As well, the relative role of threats may change with time (Geary et al., 2023). For example, dingoes have been established for ∼4000 years longer than the other threats we considered; they may have been integrated into ecological systems and may not have been a main driver of decline historically (Dickman et al., 2021). However, since rock‐wallabies now persist in smaller remnant populations because of other threatening processes, the role of dingo predation may become more problematic in this context (Legge et al., 2023). Our methodological choices in model construction could have further influenced our findings. For example, we split the species’ occurrence records at the year 2004, based on the timing of documented extirpations for the species. Using an alternative year, to some extent, would change the characteristics of the historical distribution. This, in turn, could affect our assessment of persistence factors. For rock‐wallabies, occurrence records showed a clear contraction toward the core of the species’ range (Appendix S1). However, the choice of temporal threshold may be more influential in cases where contractions occur more patchily throughout the range over time. Future applications could test multiple temporal cutoffs or implement a sliding window approach to better capture gradual or nonuniform range contractions. Similarly, the thresholds used to convert continuous Maxent distributions to presence–absence maps may affect the extent of the extirpated distribution. The difference between the contemporary and historical habitat suitability values within the extirpated distribution was relatively large (Appendix S14). However, for future assessments, where historical and contemporary suitability values are more similar at a site (e.g., 0.49 and 0.51), we recommend careful consideration of this binary threshold and thorough sensitivity analyses. Interpreting results from correlative studies requires caution because they depend on input data quality and may vary with the incorporation of alternate measures. For example, habitat degradation, land clearing, and grazing intensity were not significant predictors of rock‐wallaby extirpation, yet they are often suggested as primary drivers of decline (Australian Government, 2021). We considered habitat condition only at the scale of a rock‐wallaby occurrence record, which failed to capture surrounding habitat degradation. This is a potential shortcoming that limits our inference about the role of habitat, especially when considering previous models show rock‐wallaby habitat suitability declines rapidly when surrounding habitat loss increases (Ashcroft et al., 2014). Additionally, species response curves and thresholds will likely vary with alternate or additional predictors (Salgueiro et al., 2018). These results should therefore serve only as an initial starting point for field‐based population assessments (Rodríguez, 2002). Ideally, experiments should be conducted within an active adaptive management framework to evaluate action effectiveness (Geary et al., 2023; Grantham et al., 2009; Nicol et al., 2019) and determine optimal management levels to avoid resource expenditure where no additional benefit is perceived (Cattarino et al., 2016). Ecological systems are complex; multiple interacting factors can obscure drivers of species decline. Although field studies and experiments remain the gold standard for testing biological mechanisms (Castle et al., 2021; Cooke et al., 2023; Wilcox et al., 2021), advances in computational models now offer ways to determine mechanistic causes of decline (Tomlinson et al., 2023). However, the extensive resources required for field studies and the preexisting data and expertise needed for complex models limit their widespread application. Our methodology offers a pragmatic approach by leveraging presence‐only occurrence records—data typically available even for data‐poor species. With over half of data‐deficient species at risk of extinction (Borgelt et al., 2022) and limited conservation resources, methods with which existing data can be used are crucial for timely action. By combining readily available occurrence records with increasingly accessible spatial data, our approach provides a straightforward framework to assess multiple potential threats. This method can be applied to other species experiencing geographic range contractions across various scales and thus offers practitioners an evidence‐informed foundation for management decisions. Supporting information Supporting Information ACKNOWLEDGMENTS We thank M. Eldridge, J. Murray, H. Hines, and the ACT Government. We also thank R. Clemens and the EcoCommons platform for their guidance and support. N.M. was supported by a Australian Government Research Training Program (RTP) Scholarship, a Queensland Government Community Sustainability Action Grant to D.O.F (CSAT20077), and a WWF Top‐Up Scholarship. J.R.R. was supported by an ARC Future Fellowship (FT200100096). B.C.S. was supported by an ARC DECRA (DE200100121). Maitz, N. M. , Ward, M. S. , Rhodes, J. R. , Scheele, B. C. , Possingham, H. P. , & Fisher, D. O. (2026). Uncovering correlates of decline and critical refuges for a threatened terrestrial mammal. Conservation Biology, 40, e70314. 10.1111/cobi.70314 Article impact statement: Correlations between threats, environment, and extirpation identify key actions and locations for management of a threatened mammal DATA AVAILABILITY STATEMENT Aggregated data and code used to perform statistical analyses are available on https://github.com/natalyamaitz/2026-correlates-decline-btrw. REFERENCES Associated Data Supplementary Materials Supporting Information Data Availability Statement Aggregated data and code used to perform statistical analyses are available on https://github.com/natalyamaitz/2026-correlates-decline-btrw.