Modelling the probability of meeting IUCN Red List criteria to support reassessments

Glob Change Biol. 2024;30:e17119.   | 1 of 13 https://doi.org/10.1111/gcb.17119 wileyonlinelibrary.com/journal/gcb Received:26June2023 | Accepted:2December2023 DOI:10.1111/gcb.17119 RESEARCH ARTICLE Modelling the probability of meeting IUCN Red List criteria to support reassessments Etienne G. Henry1,2 | Luca Santini3 | Stuart H. M. Butchart4,5 | Manuela González- Suárez6 | Pablo M. Lucas3,7 | Ana Benítez- López8 | Giordano Mancini3 | Martin Jung9 | Pedro Cardoso10,16 | Alexander Zizka11 | Carsten Meyer1,12,13 | H. Reşit Akçakaya14,15 | Alex J. Berryman4 | Victor Cazalis1,17 | Moreno Di Marco3 This is an open access article under the terms of the CreativeCommonsAttribution-NonCommercial License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited and is not used for commercial purposes. ©2024TheAuthors.Global Change Biology published by John Wiley & Sons Ltd. VictorCazalisandMorenoDiMarcocontributedequally. ForAffiliationreferpageon10 Correspondence EtienneG.Henry,ÉcoleNormale Supérieure, 45 rue d'Ulm, Paris 75005, France. Email: etienne.henr[email protected] Victor Cazalis, Leipzig University, Ritterstraße 26, 04109 Leipzig, Germany. Email: [email protected] Funding information DeutscheForschungsgemeinschaft, Grant/AwardNumber:FZT118- 202548816 Abstract Comparative extinction risk analysis—which predicts species extinction risk from correlation with traits or geographical characteristics—has gained research attention asapromisingtooltosupportextinctionriskassessmentintheIUCNRedListof Threatened Species. However, its uptake has been very limited so far, possibly because existing models only predict a species' Red List category, without indicating which Red List criteria may be triggered. This prevents such approaches to be integrated into Red List assessments. We overcome this implementation gap by developing models that predict the probability of species meeting individual Red List criteria. Using dataontheworld'sbirds,weevaluatedthepredictiveperformanceofourcriterion- specificmodelsandcompareditwiththetypicalcriterion-blindmodellingapproach. We compiled data on biological traits (e.g. range size, clutch size) and external drivers (e.g. change in canopy cover) often associated with extinction risk. For each specific criterion, we modelled the relationship between extinction risk predictors and species' Red List category under that criterion using ordinal regression models. We found criterion-specificmodelswerebetteratidentifyingthreatenedspeciescomparedto acriterion-blindmodel(highersensitivity),butlessgoodatidentifyingnotthreatenedspecies(lowerspecificity).Asexpected,differentcovariateswereimportantfor predicting extinction risk under different criteria. Change in annual temperature was important for criteria related to population trends, while high forest dependency was important for criteria related to restricted area of occupancy or small population size. Ourcriteria-specificmethodcansupportRedListassessorsbyproducingoutputs that identify species likely to meet specific criteria, and which are the most important predictors.Thesespeciescanthenbeprioritisedforre-evaluation.Weexpectthis 2 of 13 | HENRY et al. 1 | INTRODUCTION Over recent decades, the IUCN Red List of Threatened species (henceforth ‘Red List’) has become the global standard for species' extinction risk assessments (Betts et al., 2020; Mace et al., 2008; Rodrigues et al., 2006).ARedListassessmentisbasedonatleast oneoffivecomplementarycriteriawithquantitativethresholdsrelating to population and distribution size, structure and trends to assignspeciestocategoriesofextinctionrisk(IUCN,2012; Mace et al., 2008). The Red List now includes assessments for over 150,000speciesofanimals,fungiandplants(IUCN,2022), but, despite its great importance for conservation action and policy (Betts et al., 2020; Hoffmann et al., 2010; Williams et al., 2021), it is insufficientlyfunded(Juffe-Bignolietal.,2016).Asaconsequence,the Red List faces important challenges in keeping assessments up to date (i.e. <10 yearsold)andreducingtheproportionofdatadeficient species (Cazalis et al., 2022; Rondinini et al., 2014). Different methods have been developed to support Red List assessments and address the above challenges (e.g. Bachman et al., 2019; Buchanan et al., 2008; Cazalis et al., 2023; Santini et al., 2019; see Cazalis et al., 2022foranoverview).Amongthem, comparative extinction risk models link extinction risk (i.e. Red List categories) with species' biological traits (e.g. body mass, habitat specialisation, range size) and external drivers of risk (e.g. human density, land-use change, climate change; Chichorro et al., 2019, 2022; Purvis et al., 2000). The models are then used to predict the Red List categories of species that have not yet been assessed for theRedList(Darrahetal.,2017;Zizkaetal.,2021),thatareData Deficient(Bland&Böhm,2016; Luiz et al., 2016) or that need updating(DiMarcoetal.,2014; Lucas et al., 2023), with the objective of accelerating and helping to prioritise the work of Red List assessors by providing additional information on species' extinction risk. These comparative extinction risk models predict that Red List categories met under any of the five possible criteria (based on e.g. distribution, abundance, population trends), thus ignoring potential differencesintheirdrivingforces.This‘criterion-blind’approachassumes all criteria can be predicted from the same covariates, while in reality, the criteria address different sources of risk that are spatially structured (Figure S1), and some are likely easier to predict than others. For instance, predicting species threatened under criterion B1 (restricted extent of occurrence combined with some subcriteria) is relatively easy using range size, while it may be harder to find relevant covariates to predict species threatened under criterion A3(futurepopulationdecline).Thissuggeststhatspeciesclassified under different criteria might have different risk correlates and face different prediction uncertainties, which might have contributed to thelowperformanceofsomecriterion-blindmodelswhentestedon independentsamplesofspecies(DiMarco,2022).Insomecases,a covariate could even have opposite effects on different criteria. For instance, species with large body mass tend to have low population density (Santini et al., 2022;Silva&Downing,1995) and might be more likely to trigger criterion C1 (small population size and decline), butsuchspeciestendtohavelargeranges(Newsomeetal.,2020) hencearelesslikelytotriggercriterionB.Importantly,ignoringthe diversity of criteria limits the uptake of comparative extinction risk modelsbyassessors,whoneedcriteria-specificinformation(Cardillo & Meijaard, 2012; Cazalis et al., 2022). Here,weaimtoovercomethisresearch-implementationgapby developingcriterion-specificextinctionriskmodelsandcomparing theirperformanceandapplicabilitytoaclassiccriterion-blindmodel. While the latter estimates the probability of a species to be threatenedunderanycriterion,ourcriterion-specificmodelestimatessuch aprobabilityindependentlyforeachindividualcriterion(A1,A2,A3, etc).Whilebenefitingfromthepowerofthemulti-speciescomparisons, this approach better encompasses the diversity of reasons that mayqualifyaspeciesasthreatenedontheRedListandprovides assessors with an output that is directly related to the information needed for assessments. We use birds as a study group to test our approach as they are the most consistently assessed group across Red List criteria (Cazalis et al., 2022), with very few data deficient species (0.4%), they have been used in many previous comparative extinctionriskanalyses(e.g.Olahetal.,2018; Richards et al., 2021; White & Bennett, 2015), and they present great variation in their response to human pressure (Cazalis, 2022; Lees et al., 2022). 2 | METHODS We compiled data on avian biological traits associated with extinction risk (e.g. range size, clutch size) as well as external risk drivers operating within species ranges (e.g. change in forest canopy cover, distance to cities). We modelled the relationship between extinction risk predictors and each species' Red List category met under each specific criterion, using ordinal regression models that best match the ordinal structure of the Red List categories (Lucas et al., 2019, 2023; Luiz et al., 2016), and combined these models into a single final prediction. We then compared the performance of these criterion-specificmodelswithacriterion-blindapproach,aswellas new approach to increase the uptake of extinction risk models in Red List assessments,bridgingalong-standingresearch-implementationgap. KEYWORDS assessment,Aves,biodiversityconservation,birds,comparativeanalysis,extinctionrisk, functional traits 13652486, 2024, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/gcb.17119 by Universidad De Sevilla, Wiley Online Library on [31/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License | 3 of 13 HENRY et al. the importance of different predictors in each approach. Finally, we evaluatedthepotentialconservationapplicationsofthecriterion- specificapproach.AllanalyseswereconductedinRversion4.0.2(R Core Team, 2020). 2.1 | Extinction risk predictors For each of the 11,162 bird species included in the Red List version 2021-3(BirdLifeInternational,2022), we gathered information on species biological traits that are associated with extinction risk (e.g. Olahetal.,2018; Ripple et al., 2017; Tobias & Pigot, 2019), considering five types (see details and rationale in Table S1): morphological (bodymass,beaklength,hand-wingindex),behavioural(nocturnality, migratory status), life history traits (clutch size, generation length), ecological (trophic niche, forest dependency, habitat breadth) and geographical (insularity, range size). However, species' intrinsic traits alone cannot predict extinction risk (Chichorro et al., 2022; González-Suárezetal.,2013), and it is key to include measurements oftheimpactofhumanactivitieswithinthespecies'range(DiMarco et al., 2014; Murray et al., 2014). We thus also included proxies for habitat alteration and degradation (extent and change of cropland and forest cover), human encroachment (human density and trends, proportion of rural population, travel time to the nearest city) and past and contemporary climate change (difference in precipitation and temperature) within each species range (see details in Table S1). 2.1.1 | Speciestraitsandcharacteristics WeusedthedistributionmapspublishedinBirdLifeInternational and Handbook of the Birds of the World (2021), filtering polygons with a high probability of presence (‘extant’) and of ‘native’ origin during the breeding season (‘resident’ and ‘breeding season’), while removing polygons coded with other presence (e.g. extinct), origin (e.g. introduced, vagrant) and season (e.g. non-breeding, passage) attributes. As most of the available data on predictors of extinction risk are terrestrial, we removed marine species from our study (N = 365,asdefinedbyBirdLifeInternational). We calculated range size as the area of the filtered distribution maptransformedintoaMollweideequal-areaprojection.Inaddition, fourpredictorswereextractedfromBirdLifeInternational(2022): generation length, migratory status, forest dependency and habitat breadth (calculated as the number of major habitat types that are coded as suitable for each species). Morphological traits relating to beak length (from tip to nares), body mass and hand-wing index were extracted from AVONET (Tobias et al., 2022), alongside ecological information on trophic niches (aggregated into four classes: herbivore, omnivore, invertivore and carnivore; Table S1).Insularityandclutchsizewereobtained from Tobias and Pigot (2019), and information on the diurnal/ nocturnal activity of birds was obtained from Wilman et al. (2014). To address taxonomic mismatches, we matched all taxa to the taxonomyusedbyBirdLifeInternational,usingthesynonymtable from Tobias et al. (2022). The remaining 200 taxonomic mismatches were then corrected manually using the synonyms documented in BirdLifeInternational(2022).Atableofthesematchesisincludedin theExtendedDataS1 provided with this article. 2.1.2 | Extrinsicfactors Similarly to the range size calculation, we only considered breeding range when measuring extrinsic variables for consistency among migratory and resident species. We used a raster layer of percentagetreecanopycoverin2018andchangesincoverduring2000– 2018at300-mresolutionfromRemelgadoandMeyer(2023; using Landsat data to correct some biases in global forest cover maps). We extracted the median value of these predictors within the range ofeachspecies.Similarly,wecalculatedthemedianvalueofrange- wide cropland coverage (in 2019) and cropland changes from 2003 to 2019 obtained from Potapov et al. (2022). We also calculated the median human population density within each species' range, and the difference between median density in 2015 versus 2000 (Global Human Settlement [GHS]; Schiavina et al., 2019), used as a proxy for human direct pressure, although some species are well adapted to it (Cazalis et al., 2021). To account for the direct pressures that species can face in the rural environment (as defined by the GHS product), we also calculated the proportion of the human population living in rural areas within each species' range (Schiavina et al., 2019).Ashumanaccessibilitycanalsodetermine the level of disturbance to which species are exposed, we extracted data from the global map of travel time to cities (Weiss et al., 2018) and calculated the median value of pixels contained within each species' distribution. Finally, we extracted the countries of occurrence of each species from BirdLife International (2022) andcalculatedtheresultingmediangrossdomesticproduct(GDP) per capita from WorldBank data (Worldbank, 2021) as an index of human population development. Inordertoaccountforclimaticcorrelatesofrisk,weextracted the current value and difference between past and current valuesfortwovariablesfromtheCHELSAcrutsdatabaseversion1.0 (Karger et al., 2017;Karger&Zimmermann,2018), choosing mean annual temperature and annual precipitation for their relevance in influencing species' distributions and their ease of interpretation (Supplementary Methods S1). Using data from another study predicting Red List categories (Lucas et al., 2023), we calculated the average value of both bioclimatic variables over two periods based onCHELSAdata(Kargeretal.,2017):1965–1995,torepresentthe pastclimate,and2005–2014,torepresentthecurrentclimate.We then extracted the median value of each variable within the species' range at both periods. We finally calculated the difference between both time periods as a proxy of recent climate change experienced by the species within their geographic range. We extracted raster values within species' distribution polygons usingtheRpackage‘exactextractr’(Baston&ISciences,LLC,2022). 13652486, 2024, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/gcb.17119 by Universidad De Sevilla, Wiley Online Library on [31/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 4 of 13 | HENRY et al. Polygons were reprojected according to each raster's original coordinate system before extraction in order to minimise raster reprojection. Variables that followed a skewed distribution were log-transformed,andallnumericvariableswerescaled(Table S1). Weremoved1864speciesforwhichwecouldnotextractand calculate all predictors (mainly due to gaps for clutch size, insularity[1400speciesmissingdata]anddiurnality[1183speciesmissing data]),leavingafinaldatasetcomposedof8695species.Weensured that the absence of 20% of the data did not introduce any bias into the distribution of species within the Red List categories (Figure S3). 2.2 | IUCN Red List framework WeextractedfromBirdLifeInternational(2022) the Red List category assigned to each species under each criterion (see below), as well as the final listed category (Figure 1; see Figure S1 for spatial distribution of these criteria). For birds, generally all Red List criteria have been evaluated for all species (with the exception of criterion E, which is excludedfromourstudy).Forthreatenedbirdspecies(i.e.thosequalifying as Critically Endangered, Endangered and Vulnerable), all criteria thatqualifyaspeciesasthreatenedshouldbereportedinBirdLife International(2022), not just the one resulting in the highest risk category, although this may be omitted in some specific cases. However, thisisnotthecaseforspeciesqualifyingasNearThreatened(e.g.data are not available on whether a species classified as Endangered under B1qualifiesasNearThreatenedunderB2).Toaccountforthis,wefollowedtwoapproaches.Inthefirst,weclassifiedacriterionas‘missing data’ if not explicitly listed; the results presented in the main text, Figure 1 and Table S2correspondtothatassumption.Inthesecond, we assumed that the species was Least Concern under a criterion unlessitwasexplicitlylisted.Asaconsequence,allmodelswererunon thesamesetofspecies.Theseresultswerequalitativelyequivalent with those obtained from our original model (the performances of individualcriterion-specificmodelswereevenslightlypoorer;Table S4). This sensitivity analysis demonstrates the limited impact of sample size differences on our model outcomes. OnlythreespeciesqualifiedasthreatenedunderA1,hencethe criterion has been excluded from the analysis. 2.3 | Extinction risk modelling Wedevelopeda‘criterion-specific’modellingapproachinwhichwe fitted a separate model for each Red List criterion. Each criterion is thus considered independently, and the Red List category met under that criterion is contrasted with the same set of extinction predictors.Forcomparison,wealsofitteda‘criterion-blind’model,astypically done in comparative extinction risk analyses, using the single listed species' Red List category as the response variable. To investigate the relationship between species traits, extrinsic factors and extinction risk, we used cumulative link models (CLM, also known as ordinal regression models) from the R package ‘ordinal’ (Christensen, 2019), which allow preservation of the ordinal structure of the Red List categories (Lucas et al., 2019, 2023; Luiz et al., 2016). Moreover, CLMs have been demonstrated to be the best algorithms to deal with the ordinal structure of Red List categories when compared to other algorithms traditionally used in comparative models of extinction risk, such as Random Forest, Neural networks or Phylogenetic Generalised Linear Models (PGLS) (Lucas et al., 2023). CLM does not allow to include phylogenetic information in a way as PGLS does to account for potentialnon-independenceinspeciesextinctionrisk,butoverallCLM has demonstrated to predict significantly better than PGLS, possibly due to the necessity of PGLS to transform Red List categories into a binary or discrete variable, losing important information from the original ordinal variable (Lucas et al., 2023). We therefore considered the Red List category as an ordered categorical factor (LC < NT < VU < EN < CR;excludingallspecieswithcategoriesEX, EWandDD),useditastheresponsevariableandcontrastedthis with the predictors. Models varied from n = 7060forcriterionB2ton = 7702for criterionC2,withatotalof8695speciesincludedintheanalysis. We checked that predictors were not highly correlated (Pearson FIGURE 1 Proportionofthe8695bird species included in our analysis currently qualifyingineachRedListcategoryfor each criterion. CR, Critically Endangered; EN,Endangered;VU,Vulnerable;NT,Near Threatened; LC, Least Concern. ‘Missing’ appliestospeciesqualifyingasthreatened for which the given criterion is not explicitly listed. The number of species per criterion is given in Table S2. 13652486, 2024, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/gcb.17119 by Universidad De Sevilla, Wiley Online Library on [31/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License | 5 of 13 HENRY et al. correlationorKruskal-gammacoefficients > |0.70|;Figure S2). To adjust for unbalanced data (Figure 1), we calculated the proportion of threatened and not threatened species under each criterion and weighted our models by the proportion of species in theoppositecategory(i.e.specieswithcategoriesVU,ENorCR were weighted by the proportion of not threatened species, and specieswithcategoriesLCorNTwereweightedbytheproportionofthreatenedspecies).Abackwardstepwisemodelselection was performed using the step function from the R package ‘stats’ (R Core Team, 2020) in order to find the subset of variables that minimisetheAkaikeInformationCriterionforeachcriterion.The proportional odds assumption of a linear relationship was not always met, but this should not impact our results substantially (see Supplementary Methods S2). AsthenumberofspecieslistedinENandCRcategorieswasvery small for some criteria (Figure 1), we anticipated predicting specific categories could be challenging (Table S5); thus, for validation, we focused on a simplified prediction: whether a species was classified as threatened or not. We used a method of taxonomic block validation in which we iteratively excluded one taxonomic family from the data used to train the model and then used the model to predict the Red Listbinarythreatlevelofthespeciesintheleft-outfamily.Aspecies was predicted as threatened under a given criterion if the sum of the probabilitiestobeCR,ENandVUwas>0.5 and predicted as not threatened otherwise. We then compared the predictions with the actual Red List categorisation under that specific criterion (assuming that the current Red List category of each species is correct for each criterion). Specificity and sensitivity were defined, respectively, as thepercentageof notthreatenedspecies(LC, NT)correctly classifiedassuchandthepercentageofthreatenedspecies(VU,EN, CR)correctlyclassifiedassuch.FollowingRedListguidelines(IUCN Standards and Petitions Committee, 2022), we assigned a ‘combined’ categorytoeachspeciesbasedontheninecriterion-specificmodels asthehighestcategorypredictedamongmodels;consequently,a species was classified as not threatened only because it was not predictedasthreatenedinanyofthecriterion-specificmodels.Wealso report the models' performance in predicting the specific Red List category for each species (assigning to each species the category with the highest probability; Table S5). To measure and compare the overall performance of both modelling approaches, we used the True Skill Statistic (TSS), defined as: sensitivity + specificity − 1.TSStakesintoaccountbothomissionand commission errorsandrangesfrom −1to+1, where +1 indicates perfect agreement, values greater than 0.5 indicate a good performance and values of 0 (or less) indicate a performance no better (or worse)thanrandom(Alloucheetal.,2006). Finally, our models were employed to predict the most probable Red List category for all species, which was then transformed into a binary prediction (threatened/not threatened) and compared with the actual Red List category. The mismatches thus identified serve as flags for prioritising reassessments of a specific list of species, as classicallydoneincomparativeextinctionriskanalyses(DiMarco et al., 2014; Lucas et al., 2023). 3 | RESULTS 3.1 | Model performance Models' performance in predicting extinction risk greatly varies among criteria (Figure 2).ThemodelpredictingcriterionD2,related to population structure, had the highest TSS score (0.91), followed bycriteriaB1,D1,B2,C2andC1(0.88–0.69).Allthesemodelshad TSSscoreshigherthanthecriterion-blindapproach(0.61),meaning that models are better at predicting extinction risk for single criteria than for the overall categories. For most criteria, these high TSS scores were the result of both higher specificity and higher sensitivity (Figure 2).Conversely,modelspredictingcriteriaA2–A4,related to population declines, showed the lowest TSS scores among all criteria(0.52–0.59),andtheywereslightlylowerthanthecriterion- blind model. Following Red List guidelines, we assigned a ‘combined’ category to each species as the highest extinction risk category from any of thenineindividualcriterion-specificmodelsandfoundthatthissubstantially reduced the TSS (0.52) compared with applying the models individuallyforeachcriterion(averagemodelTSS = 0.72;Figure 2a). Thisislargelyduetolowerspecificitycomparedto thecriterion- blindapproach(0.68vs.0.83probabilityofcorrectlyclassifyinga not threatened species. Figure 2c; Table S2), this is partly explained because a species had to be classified as not threatened under each ofthenineappliedcriterion-explicitmodelsinordertofallinthis group overall. In contrast, using a criterion-specific approach resultedina0.84probabilityofcorrectlyclassifyingthreatenedspecies,comparedwith0.79forthecriterion-blindapproach(Figure 2b; Table S2). Considering ‘missing’ criterion-specific categories as LC (Table S4) or predicting at the category level rather than binarily contrasting threatened versus not threatened (Table S5) resulted in respectivelyequivalentorlowerperformances. 3.2 | Drivers of extinction risk Thecriterion-blindapproachshowedpositiverelationshipsbetween extinction risk and body mass, carnivore trophic niche, high forest dependency and cropland cover, while showing negative relationships with insularity, range size or percentage canopy cover (Figure 3). Someoftheserelationshipswerealsodetectedbycriteria-specific models. For instance, carnivorous species were generally associated with higher levels of extinction risk, while species with a larger range size or with distributions that had a high tree canopy coverage were less at risk on average. Incontrast,theimportanceandsignificanceofotherpredictors were idiosyncratic between criteria. For instance, body mass generally correlated positively with extinction risk for criteria related to rapidpopulationdeclinesorsmallpopulationsize(A2,A3,A4,C1, C2,D1),whileithadnoinfluenceoncriteriarelatedtorestricted geographic range (criteria B1, B2, D2). Conversely, high forest 13652486, 2024, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/gcb.17119 by Universidad De Sevilla, Wiley Online Library on [31/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 6 of 13 | HENRY et al. dependency was associated with increased extinction risk for criteria B1, B2 and C2, all relating to small population size, restricted area of occupancy and subpopulation structure, but did not, or slightly negatively, influence criteria based on rates of decline alone. Extrinsic factors were mainly important to predict criteria related topopulationreductions(A2,A3andA4);forinstance,changein annual temperature correlating positively with extinction risk for criteriaA2–A4andC2,andGDPpercapitacorrelatingpositivelywith criteriaA2–A4,B2andC1. 3.3 | Criterionspecific approach to prioritise reassessments The criterion-blind model predicts that 15% of species currently classified as not threatened (N = 1337)maybethreatened(Table 1), butthis percentage almostdoubled(to 28%,N = 2467) under the criterion-specificmodel.Conversely,wepredicted104threatened species as not threatened (199 with the criterion-blind model). Predictions for all models and all species are provided in Extended DataS2. 4 | DISCUSSION Inthisstudy,wedevelopedamodellingapproachthatpartitionsextinction risk according to individual Red List criteria and compared itwithacriterion-blindapproach.Onaverage,modellingindividual criteria performed better than the criterion-blind approach, with higherperformanceforsixcriteria(especiallycriteriaB1andD2), while three provided similar or marginally lower performance (criteria A2–A4). This result highlights that predicting extinction risk under some criteria may be intrinsically difficult, at least using the predictorsconsideredhere.Inparticular,criteriarelatedtopopulationtrends(especiallyA3relatedtofuturetrends)aremoredifficult to predict. With these models, we can enhance our understanding of the mechanisms underlying observed correlations and, ultimately, point to distinct drivers of risk. FIGURE 2 Comparisonofmodelperformances.Theleft-handsideofeachplotcomparestheperformanceofthecombinedcriterion- specificmodels(referredtoas‘CS’)withthatofthecriterion-blindapproachtocomparativeextinctionriskmodelling(referredtoas ‘CB’),whiletheright-handsidepresentstheperformanceofeachcriterion-specificmodel.(a)TrueSkillStatistic[−1,1];(b)sensitivity[0,1], proportionofthreatenedspeciescorrectlyclassified;(c)specificity[0,1],proportionofnotthreatenedspeciescorrectlyclassified.Dotted linesrepresentthemeanvalueobtainedfromthenineindependentcriteria-specificmodels. FIGURE 3 Heatmapofpredictorimportanceassociatedwithextinctionriskundereachcriterion.Rowsindicatepredictorsofextinction risk,andcolumnsrelatetocriterion-specificmodels.The‘CB’modeldescribesthecriterion-blindapproachtoextinctionriskmodelling. Colour indicates the sign (blue for negative and orange for positive), with darker tones indicating stronger relationships. Both positive and negativevaluesweredividedintofiveequalgroupsaccordingtotheintervals:−3.71,−0.97,−0.46,−0.27,−0.15and0fornegativeestimates and0,0.16,0.24,0.42,0.89and1.76forpositiveestimates(blanksindicatethatthepredictorhasnotbeenretainedintheoptimalmodel after predictor selection; see Section 2.3; estimates are detailed in Table S3).CLM,cumulativelinkmodels;GDP,grossdomesticproduct. 13652486, 2024, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/gcb.17119 by Universidad De Sevilla, Wiley Online Library on [31/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License | 7 of 13 HENRY et al. Combining the nine criteria-specific models to obtain a single prediction per species led to substantially greater sensitivity (i.e. more likely identification of threatened species as threatened, Figure 2b)butlowerspecificitythantheclassicalcriterion-blindapproach(see,forexample,Orange-frontedParakeetEupsittula canicularis in Figure 4a). Because one of the primary goals of automated 13652486, 2024, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/gcb.17119 by Universidad De Sevilla, Wiley Online Library on [31/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License 8 of 13 | HENRY et al. extinction risk methods is to identify species likely to be at risk of extinction (but not currently recognised as such) to prioritise their reassessment (Cardillo & Meijaard, 2012), a model with high sensitivity will be more valuable than a model with similar TSS and higher specificity (Cazalis et al., 2022). Previous extinction risk models have typically predicted threatened species less accurately than not threatenedspecies(DiMarco,2022; Murray et al., 2014).Ourresults show that a reason behind this observation may be the omission of the diversity of reasons why a species is considered threatened on the Red List, which is represented by the multiple Red List criteria. However, combiningninecriterion-specificmodelsdecreasedthe overall specificity of the prediction (see, for example, Gray's Lark Ammomanopsis grayi in Figure 4b), resulting in an overestimation of the proportion of threatened species and a slightly lower TSS in comparisonwiththecriterion-blindapproach(Figure 2a). This result is explained by the fact that, following Red List guidelines, a species was classified as not threatened only if predicted as such by all ninecriterion-specificmodels.ThisiswellalignedwiththeRedList process and a precautionary approach, but it makes our approach sensitive to misclassification. Hence, increasing the specificity of criterion-specificmodelsisapriorityforthefuture.Possiblewaysof achieving this include improving the performance of the individual models with additional covariates (for example, relating to hunting/ trappingpressureforcriteriaA2–A4),accountingforsharedevolutionary history using phylogenetic models (Purvis, 2008) or developing an approach to combine individual models that accumulates fewer errors from individual models that misclassify a species as threatened under a given criterion. Inaccordancewithpreviousstudies,ourcriterion-blindmodel identified extinction risk as correlating positively with for example, body mass, carnivorous niche, high forest dependency and lower tree canopy cover, and negatively with for example, clutch size, range size and distance to cities (Gaston & Blackburn, 1995; Olahetal.,2018; Richards et al., 2021; Ripple et al., 2017; Tobias & Pigot, 2019; White & Bennett, 2015).Ourfindingshighlightthe importance of considering separately the multiple processes underlying patterns of extinction (Figure 3; Figure S1). They reveal that increases in annual temperature across species' ranges are of particular importance for criteria related to population decline (A2–A4,andtosomeextentC2),consistentlywiththepredicted role of climate change in driving declines in abundance and increasesinspecies'extinctionrisk(Şekercioğluetal.,2012; Mancini et al., 2023). Additionally,body massisof particularimportance for criteria related to population trends and population sizes (A1–A4,C1–C2,D1),suggestingthatspecieswithhighbodymass are more likely to be declining and to have small population sizes (Carvajal-Quinteroetal.,2023). The diversity of relationships between predictors and criteria, along with their ecological meaning, emphasises the importance of accounting for heterogeneity in the predictability of Red List criteria, rather than considering them as equallypredictableasassumedincriterion-blindcomparativeextinctionriskanalyses.Theyalsohighlightthatourcriterion-specific approach can help better understand the diverse mechanisms associated with extinction risk. Further, a better approximation of the causal relationships underlying species classification under different Red List categories can improve our ability to forecast status change based on changes in the drivers. By modelling criteria separately, we increase the applicability of comparative extinction risk models (Cazalis et al., 2022; Owens&Bennett,2000; Ripple et al., 2017). Red List assessors arerequiredtoassesseachspeciesagainstallcriteriaforwhich there is sufficient information (IUCN Standards and Petitions Committee, 2022). Therefore, while our models' outputs do not fundamentallychangethered-listingprocess,theyallowforvarious sources of information to be considered by assessors. As these models are contingent on the availability of life history and distributional data, their use for helping reassess data deficient species—for which these are often unavailable—is likely to be limited. We contend them to be most suited to help prioritise efforts to reassess threatened and non-threatened species for which underlying data are available and to identify knowledge gaps and opportunities for future research. For instance, the Least ConcernRufous-belliedChachalacaOrtalis wagleri (Figure 4c) is predictedtobethreatened(VU)byboththecriterion-blindand thecriterion-specificapproaches.However,whilethecriterion- blind model offers no additional insight, our criterion-specific approach provides assessors with relevant information about why this species might be VU—namely, that it might meet criteria A2,A3,A4,C1andC2.Assessorscouldusethisinformationto focus efforts on investigating past and future population trends to assess criteria A2–A4 and population size and structure to assess criteria C1 and C2, which could be complemented with the specific contributions of different covariates that led to this prediction.Ifthesedataarenotavailable,determiningvaluesfor these parameters may be considered a priority for fieldwork and research. Comparative extinction risk models have often been promoted as useful tools to provide a first prediction of extinction risk for IUCN Red List Prediction criterionblind Prediction criterionspecific Not threatened Threatened Not threatened Threatened Notthreatened 6380(73%) 1337 (15%) 5250 (60%) 2467(28%) Threatened 199 (2%) 779 (9%) 104 (1%) 874(10%) Note:Thepredictionforthecriterion-specificmethodcorrespondstothepredictionafter combining results from the nine individual criterion models. TABLE 1 RedListcategoryprediction ofcriterion-blindandcriterion-specific models compared with the current binarised category (Threatened for Vulnerable, Endangered and Critically Endangered;NotthreatenedforLeast Concern;andNearThreatened). 13652486, 2024, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/gcb.17119 by Universidad De Sevilla, Wiley Online Library on [31/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License | 9 of 13 HENRY et al. speciesnotyetincludedintheRedList(Darrahetal.,2017;Zizka, Andermann,etal.,2022;Zizka,Barratt,etal.,2022), for data deficientspecies(Bland&Böhm,2016; Borgelt et al., 2022; Cazalis et al., 2023; He et al., 2021; Luiz et al., 2016) or to prioritise reassessments(DiMarcoetal.,2014; Lucas et al., 2023). But so far, these have remained largely unmet promises, with hardly any uptake of these modelling approaches in the Red List (Cazalis et al., 2022). By focusing on reassessments and informing the assessors about the specificcriteriaunderwhichaspeciesislikelytoqualify,criterion- specific models could help accelerate the rate of Red List assessments, guide data collection efforts and facilitate the growth and update of the Red List so that it can best inform conservation FIGURE 4 Comparisonofoutputs forselectedspeciesfromacriterion- blindapproachandacriterion-specific approach to comparative extinction risk analysis. The three panels show different species, with their current categories in the Red List, the categories predicted by the models and the criteria triggered for thecriterion-specificapproach.Panel(a) shows a threatened species predicted as suchbythecriterion-specificapproach only;panel(b)showsanon-threatened speciespredictedassuchbythecriterion- blind approach only; and panel (c) shows anon-threatenedspeciespredicted as threatened by both approaches. IllustrationbyÀ.Jutglar,T.Worfolk. Source: © 2022 Cornell University. 13652486, 2024, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/gcb.17119 by Universidad De Sevilla, Wiley Online Library on [31/05/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License