Testing species distribution models across space and time: high ...

A. Eskildsen et al.plot (AUC; Fielding & Bell, 1997) and the true skill statistic(TSS; Allouche et al., 2006). A non-parametric Wilcoxon’smatched pairs test was used to test whether the explanatorypower and predictive accuracy differed significantly between thebaseline climate-only models and each of the three sets of morecomplex models. Furthermore, the relationship between speciescharacteristics and modelling accuracy was investigated usingSpearman rank correlation tests.For evaluation of AUC values we used the approach recommendedby Swets (1988): excellent AUC > 0.90; good 0.80 0.75; good 0.40 < TSS < 0.75; and poorTSS < 0.40.The relative importance of predictor variables was assessedfrom models calibrated on all data from t 1. In GAM models,variable importance was calculated as each variable’s drop contribution(i.e. change in deviance associated with exclusion of agiven variable from a model containing all the other variables).In the BRT models, the relative importance of predictors basedon the frequency of variable selection during model buildingand the reduction in deviance associated with each variable’sinclusion in models was calculated. For both methods variables’contributions were scaled to sum to 100, with higher numbersindicating stronger influence on the response variable.Range margin shifts during climatic warmingTo compare observed and modelled shifts at northern rangemargins, we used the mean latitude of the five northernmostrecords in t 1 and t 2 to estimate species’ northern range limits.Modelled distributions were obtained from our simple climatemodel and the full climate-soil-land-cover models, which werecalibrated with 100% of data from time t 1. Due to missing informationabout species whose range margins in time t 2 may haveexceeded the northernmost border of Finland, only species witha poleward range limit more than 100 km south of Finland’snorthern border in t 1 were used in the analysis. Furthermore,four migratory species (Pieris brassicae, Pieris rapae, Colias hyaleand Vanessa atalanta) were excluded. Thus, the analysis included56 species. We used Spearman’s rank correlations betweenobserved and predicted range margins to quantify the ability ofSDMs to predict distributional changes at expanding rangemargins.RESULTSEvidence of observed climate changeThe two periods, 1991–99 and 2001–09, differed significantly ingrowing degree-days (mean GDD5 SE: t 1 = 910.9 24.3,t 2 = 1032.1 25.6; Wilcoxon’s signed rank test, W < 0.1,P < 0.001) and mean temperature of the coldest month (meanTEMP C SE: t 1 =-8.80 0.31 °C; t 2 = -8.69 0.28 °C;W = 2423, P = 0.002). No significant change was observed formean annual precipitation between the two periods (meanPRE SE: t 1 = 594.9 3.0 mm; t 2 = 588.3 4.7 mm; W =4050, P = 0.071).Validation of SDM performance using independentand non-independent dataOn average, models validated using t 1 occurrence data (i.e. nonindependentvalidation) showed fair to good AUC values(0.77 AUC 0.80) and good TSS values (0.46 TSS 0.51). Mean AUC values derived from models validated using t 2occurrence data (i.e. independent validation) also showed fair togood results (i.e. 0.72 AUC 0.76), while average TSS valuessometimes showed poor results (0.37 TSS 0.43) (Table 1).Overall, models with non-independent validation yielded significantlyhigher predictive accuracies than model validationswith independent t 2 data (Table 1; Wilcoxon’s signed rank test,P < 0.001 for all).Adding land cover and soil informationFor models validated with non-independent t 1 data, the inclusionof soil type variables did not lead to model improvementwhen evaluated by AUC and performed significantly poorerwhen evaluated by TSS. The inclusion of land cover variablesonly resulted in significantly higher AUC values for GAM andBRT models (Table 1). For models validated with independent t 2data, model performance was significantly improved comparedwith the simple climate-only models for all models except forthe GAM models including soil type and land cover, as evaluatedby TSS (Table 1). In the latter case, the models performed significantlypoorer than the climate-only model. In nonindependentmodel tests, the inclusion of land cover yieldedconsistently higher predictive accuracies than the inclusion ofsoil (Table 1). However, in independent model tests with t 2 data,there was no clear difference in model accuracy for either soiltype or land cover independently. For non-independent modeltests, the joint inclusion of land cover and soil type variables ledto significant improvement of GAM and BRT models evaluatedby AUC, compared with the climate-only model, while GAMmodels evaluated by TSS performed significantly poorer(Table 1). In contrast, for independent model tests, significantimprovement of performance was observed for all model types,except for GAM models evaluated by TSS (Table 1).When considering the average effect of environmental variableson species distributions across all species, growing degreedays contributed most strongly, followed by mean temperatureof the coldest month, cover of built-up areas and cover of pastures(Fig. 1). Soil types and other land cover classes generallymade rather small contributions to explain species distributions(Fig. 1), but their integration considerably improved model predictions(Table 1).Overall, the predictive accuracies of climate-only andclimate-soil-land-cover SDMs were fairly strongly correlated(0.54 r 0.78, P < 0.001 for all) (see Figure S1). Interestingly,there was a drop in correlation coefficients between the4Global Ecology and Biogeography, ••, ••–••, © 2013 John Wiley & Sons Ltd

Testing SDMs across space and timeTable 1 The predictive accuracy of generalized additive models (GAM) and boosted regression trees (BRT) models validated using non-independent observations from 1992 to 1999 andindependent observations from 2002 to 2009.Climate only Climate and soil Climate and land cover Climate, soil and land coverNon-independent Independent Non-independent Independent Non-independent Independent Non-independent IndependentAUCGAM 0.78 0.10 0.73 0.12 0.77 0.10 0.75 0.11*** 0.80 0.10** 0.75 0.11*** 0.78 0.09* 0.75 0.11***BRT 0.79 0.10 0.72 0.11 0.78 0.09 0.75 0.11*** 0.80 0.09** 0.76 0.11*** 0.80 0.09** 0.76 0.10***TSSGAM 0.51 0.20 0.40 0.22 0.46 0.20*** 0.41 0.20 0.50 0.20 0.40 0.21* 0.46 0.19* 0.40 0.20BRT 0.50 0.20 0.37 0.21 0.47 0.17* 0.40 0.20*** 0.50 0.18 0.42 0.20*** 0.48 0.17 0.43 0.19***The table shows mean values and standard deviations for area under the curve of a receiver operating characteristic plot (AUC) and true skill statistic (TSS) for GAM and BRT models including climate, soil typeand land cover variables. Asterisks indicate significant differences between models and the simple climate-only model, with bold numbers indicating significant improvement (P-values for Wilcoxon signed ranktests. *P < 0.05, **P < 0.01, ***P < 0.001).climate-only GAM model and the combined climate, soil andland cover GAM model, indicating that some species distributionsare better predicted when using independent validation incombination with more complex models (see Appendix S1).Species characteristics and modelling accuracySpecies characteristics showed low (r < 0.20) to high (r > 0.60)correlations with SDM accuracy for the independent data usingthe full BRT model, including climate, soil and land cover variables(Fig. 2). Because species characteristics were only weakpredictors of model accuracy in combined climate-soil-landcover GAMs, only BRT models were considered for this analysis.Of the four species characteristic variables, highly significantnegative effects on modelling accuracy were found for mobility,length of flight period and prevalence (Spearman’s rank correlationand P-values given in Fig. 2). Removal of the five outlyingspecies with the longest flight periods (Fig. 2c) resulted inslightly stronger correlations (r=-0.50 for AUC and r=-0.53for TSS).Changes at northern range marginsObserved range margin shifts tended to be greater for speciesoccurring at higher latitudes (Fig 3a). On average, the observednorthwards range margin shift across all species between t 1 andt 2 was 54.5 km ( 15.0), with a maximum observed range shiftof 454 km (silver-washed fritillary, Argynnis paphia Linnaeus,1758). Predicted shifts in range margins varied more widely thanobserved range shifts, with a tendency for over-prediction forspecies occurring at lower latitudes (Fig 3b). Predicted rangemargin shifts tended to be larger than the observed shifts,with GAMs predicting values closer to observed shifts(59.8 13.9 km for the simple climate-only model, and73.3 10.5 km for the combined climate-soil-land-covermodel) while BRTs considerably over-predicted range marginshifts (106.5 17.8 km for the simple climate-only model, and165.5 25.4 km for the combined climate-soil-land-covermodel). Correlations between observed and predicted rangeshifts were generally weak (Fig. 3c and d), indicating that rangedynamics at leading edges were not well predicted. The onlysignificant positive correlation was found for the combinedclimate-soil-land-cover GAM model (Spearman’s rank correlationgiven in Fig. 3d, P < 0.05). Predictions of northern rangemargins for three example species, the highly mobile blackveinedwhite (Aporia crataegi Linnaeus, 1758) and peacock(Nymphalis io Linnaeus, 1758) and the less mobile Chestnutheath (Coenonympha glycerion Borkhausen, 1788), showed highaccuracy in time t 1 (Fig. 4a,c,e) . However, when models wereused to predict the species’ distributions and range shifts in timet 2, satisfactory results were only obtained for the Chestnut heath,which showed no range shift, and peacock, which showed anobvious range shift (Fig. 4d,f). In contrast, the large observedrange shift of the black-veined white butterfly at the expandingnorthern range edge was not adequately captured by the SDM(Fig. 4b).Global Ecology and Biogeography, ••, ••–••, © 2013 John Wiley & Sons Ltd 5

A. Eskildsen et al.Figure 1 Box plot of the variable importance of environmental variables in the generalized additive models (GAM) (white boxes) andboosted regression tree (BRT) models (grey boxes) for 77 butterfly species. Boxes span the 25th–75th percentiles, black dots indicate meanvalues, solid lines the median and whiskers indicate the extreme values. The average combined importance of climatic variables, i.e. annualgrowing degree-days above 5 °C (GDD), mean temperature of the coldest month (TEMP C), and precipitation (PRE) was 39.6% for GAMand 45.9% for BRT. The average combined importance of soil type variables, i.e. rock, sand and gravel, clay, and peat was 21.1% for GAMand 19.7% for BRT. The average combined importance of all land cover variables, i.e. built up-areas (built), arable land (arable), pasture,sparse vegetation and wetland was 39.3% for GAM and 34.4% for BRT.Figure 2 Relationships between speciescharacteristics and modelling accuracyfor boosted regression tree (BRT) modelsbased on climate, soil type and landcover variables. Both area under thecurve of a receiver operatingcharacteristic plot (AUC) and true skillstatistic (TSS) values derived fromindependent model validation areshown. Rho values and P-values forSpearman’s rank correlation are given.6Global Ecology and Biogeography, ••, ••–••, © 2013 John Wiley & Sons Ltd

Testing SDMs across space and timeFigure 3 Observed and modelled northern range margins based on the mean latitude of the five northern-most grid cells for 56 species ofFinnish butterflies. (a) The observed northern range margin at time t 1 (1992–99) and time t 2 (2002–09); (b) the observed range margin attime t 2 and the modelled range margin at time t 2 (using the mean latitude of the five northern-most predicted occurrences from generalizedadditive models (GAM) and boosted regression tree (BRT) models including climate, soil type and land cover variables); (c) therelationship between observed and predicted range margin shifts using models including climate only; (d) the relationship betweenobserved and predicted range margin shifts using models including climate, soil type and land cover variables. Dotted lines indicate the 1:1relationship. For (c) and (d) Pearson’s r is shown. A significant correlation was found for the full GAM model, but not for the other models(*P < 0.05).DISCUSSIONOur study provides a real-world test of the capacity of SDMs toforecast range shifts of high latitude species under globalchange. Our results offer new insights into the predictive performanceof SDMs, especially their limited performance in predictingrange dynamics at the expanding edge. Our findingsfurther support earlier suggestions (Araújo et al., 2005) thatmodel validation with non-independent data from the sametime period (t 1) provides overly optimistic assessments of predictiveaccuracy compared with validation based on independentdata from another time period (t 2). However, our resultsbroaden the previous knowledge because data stem from twoperiods characterized by a significant increase in summer andwinter temperatures, allowing us to directly test the performanceof SDMs under observed climate warming (cf. Kharoubaet al., 2009). SDMs obtained strong model fits when using nonindependentvalidations, but model fits became weaker whentested with independent datasets.The importance of climatic and other abiotic variables islikely to be scale dependent (Wiens, 1989; Pearson et al., 2004;Luoto, 2007), with climate thought to play the most importantrole at broad spatial scales and other factors becoming increasinglyrelevant at finer spatial scales (Pearson & Dawson, 2003).For species distribution data recorded at a resolution of 10 or20 km, previous work has shown that the integration of soil typeand land cover variables may significantly improve spatial prediction(Pearson et al., 2004; Coudun et al., 2006; Luoto et al.,2006; Heikkinen et al., 2007; Titeux et al., 2009). Our resultssupport this idea, showing that significant improvements ofmodel performance can be obtained when soil type and landcover variables are included into SDMs. As a consequence, correlationsbetween the accuracy of models (for both GAMs andBRTs) validated with independent versus non-independent datawere high for simple climate-only models, but the discrepancybetween independent and non-independent validation becamelarger when using the full climate-soil-land-cover models. Thisshows that, at least for some species, the predictive potential ofSDMs may be considerably improved when adding land coverand soil, and that non-independent validation might underestimatethis potential increase compared with independent validations.The main drivers of Finnish butterfly distributions at10-km resolution are growing degree-days, average temperatureof the coldest month, cover of built-up areas and cover of pastures– land cover thus appear to have a larger effect on speciesdistributions than soil type. In ecological terms, growing degreedaysis related to the developmental threshold for larvae and themean temperature of the coldest month to over-wintering survival(Hill et al., 2003), while land cover variables can be seen asa surrogate for the occurrence of potential breeding habitatsGlobal Ecology and Biogeography, ••, ••–••, © 2013 John Wiley & Sons Ltd 7

A. Eskildsen et al.Figure 4 Observed and predicted distributions of the highlymobile black-veined white (Aporia crataegi) (a, b) and peacock(Nymphalis io) (c, d) and the less mobile Chestnut heath(Coenonympha glycerion) (e, f). Observed occurrences of the threespecies are indicated by black dots, while predicted occurrencesbased on full-boosted regression tree (BRT) models calibrated on100% of t 1 data using climate, soil type and land cover areindicated by grey grid cells. Model performance (area under thecurve of a receiver operating characteristic plot (AUC) and trueskill statistic (TSS) for the three species was found to be highlysimilar in time t 1 (left column), while model performance in timet 2 (right column) was markedly better for the low-mobilityChestnut heath and peacock. RSO = range shift observed (km)and RSP = range shift predicted (km). Range shift estimates arebased on the mean of the five northernmost cells in t 1 and t 2.Projection: Finnish National Coordinate System (Gauss–Krügerwith central longitude 27°).(Luoto et al., 2006). Overall, these results underline that broadscaledata recorded at a coarse spatial resolution are able tocapture the effects of climate, land cover and soil type on speciesdistributions, all of which contribute to improving model fit andpredictive power.Ecological traits can be important in determining the distributionalresponses of species to climate change (Pöyry et al.,2009; Angert et al., 2011). In our study, the modelling accuracyof butterfly distributions was negatively correlated with threespecies characteristics, namely mobility, flight period and prevalence.This is consistent with previous findings (Luoto et al.,2005; Pöyry et al., 2008) and indicates that species with largedistributions and the ability to fly long distances over a shortperiod of time are inherently difficult to model. The mobilityindex used in our study was significantly correlated with bothprevalence (Pearson’s correlation, r = 0.638, P < 0.001) andflight period (Pearson’s correlation, r = 0.591, P < 0.001), suggestingthat the effects of these three parameters on modellingaccuracy may be to some extent confounded (Pöyry et al., 2009).Overall, these results draw attention to the importance of speciescharacteristics for influencing the capacity of different species totrack rapid climate change, and to the need of identifying keytraits that could confound model predictions.We documented an average northbound range shift of54.5 km for 56 species of Finnish butterflies between the periods1992–99 and 2002–09, indicating a strong and rapid overallresponse to climatic warming among high-latitude butterflyspecies (see also Pöyry et al., 2009). However, some species(19%) were observed to show slight range contractions. A smallnumber of species had very large range shifts, whereas otherspecies only showed minimal responses, with the observed rangemargin shifts tending to be stronger for species occurring athigher latitudes. With respect to modelling success, our resultssuggest that while SDMs may provide credible predictions of thechanges in the overall distribution of some species over shorttime periods (e.g. 10 years), they may fail to provide accuratepredictions for other species, especially at their expandingnorthern range margins. These findings have important implicationsfor SDM modelling under climate change and suggestthat SDMs ability to predict the exact locations of new populationsat expanding range margins is at best weak, possiblybecause of the exclusion of important parameters, such asdemographic rates and/or dispersal (Schurr et al., 2012). Ourillustrative examples of the observed and predicted range shiftsof three species of Finnish butterflies, the highly mobile blackveinedwhite and peacock and the more sedentary Chestnutheath, show that mobility and dispersal behaviour play animportant role in determining species responses to changing8Global Ecology and Biogeography, ••, ••–••, © 2013 John Wiley & Sons Ltd

Testing SDMs across space and timeenvironmental conditions. Mismatches between observed andpredicted range shifts may thus be related to species-specifictraits and among-species variation in demographic (birth anddeath) rates.Novel implementations, such as dynamic range models(DRMs), could potentially overcome such problems for specificspecies (Schurr et al., 2012). Another option to reduce mismatchesbetween observed and predicted range margins inhighly mobile species could be to separate the extreme outlyingand temporally sporadic local occurrences with no permanentpopulations (established for 1 or 2 years under exceptionallyfavourable weather conditions as in the case of A. crataegi innorthern Finland in 2004) from those occurrences at the rangemargins with more permanent populations. However, thiswould require the availability of spatially comprehensive butterflyatlas survey data recorded on an annual basis, which are bothexpensive and difficult to collect.In conclusion, SDMs assessed with non-independent splitsamplevalidations show good model fits, but their potential forpredicting range dynamics in non-equilibrium conditions, suchas climate change, might be limited. The inclusion of land coverand soil types can improve SDM predictions as suggested by ourindependently validated models, and species characteristics mayaccount for why predictions are more reliable for some speciesthan for others. However, even after the inclusion of these additionalfactors, poor predictions of changes at species rangemargins are typical, suggesting that range dynamics cannot befully captured even in more complex SDMs. Thus, our resultssuggest that SDMs have limitations in their ability to predictspecies distributional responses to climate change at expandingrange margins, and we propose that incorporation of dynamicprocesses, such as demographic rates and dispersal, mightstrengthen the predictive power of the models. There is a clearneed for more studies testing the potential of SDMs to predictdistributional changes between two or more time periods andunderstanding what determines the discrepancy in accuracybetween non-independently and independently validatedmodels. Future work should also include collecting and compilingspatially and temporally comprehensive datasets on thechanges in species distributions and making them available forsuch model testing.ACKNOWLEDGEMENTSWe thank editor-in-chief David Currie, editor Jeremy Kerr andtwo anonymous referees for comments. This work acknowledgessupport from the Nordforsk TFI Network ‘Effect Studiesand Adaptation to Climate Change’ (2011–2014) and 15. JuniFonden. WDK acknowledges a starting independent researchergrant (11-106163) from the Danish Council for IndependentResearch | Natural Sciences. MSW received support from theGreenland Climate Research Centre project number 6505. PRand ML acknowledge funding from the Academy of Finland(Project Number 1140873). We are grateful to Kimmo Saarinenfor kindly providing the NAFI Butterfly Atlas for the purposes ofthis study.REFERENCESAlgar, A.C., Kharouba, H.M., Young, E.R. & Kerr, J.T. (2009)Predicting the future of species diversity: macroecologicaltheory, climate change, and direct tests of alternative forecastingmethods. Ecography, 32, 22–33.Allouche, O., Tsoar, A. & Kadmon, R. (2006) Assessing the accuracyof species distribution models: prevalence, kappa and thetrue skill statistic (TSS). Journal of Applied Ecology, 43, 1223–1232.Angert, A.L., Crozier, L.G., Rissler, L.J., Gilman, S.E., Tewksbury,J.J. & Chunco, A.J. (2011) Do species’ traits predictrecent shifts at expanding range edges? Ecology Letters, 14,677–689.Araújo, M.B. & Rahbek, C. (2006) How does climate changeaffect biodiversity? Science, 313, 1396–1397.Araújo, M.B., Pearson, R.G., Thuiller, W. & Erhard, M. (2005)Validation of species–climate impact models under climatechange. Global Change Biology, 11, 1504–1513.Atlas of Finland (1988) Climate, Folio 131. National Board ofSurvey and Geographical Society of Finland, Helsinki.Coudun, C., Gegout, J.-C., Piedallu, C. & Rameau, J.-C. (2006)Soil nutritional factors improve models of plant species distribution:an illustration with Acer campestre (L.) in France.Journal of Biogeography, 33, 1750–1763.Devictor, V., van Swaay, C., Brereton, T. et al. (2012) Differencesin the climatic debts of birds and butterflies at a continentalscale. Nature Climate Change, 2, 121–124.Elith, J., Graham, C.H., Anderson, R.P. et al. (2006) Novelmethods improve prediction of species’ distributions fromoccurrence data. Ecography, 29, 129–151.Elith, J., Leathwick, J.R. & Hastie, T. (2008) A working guide toboosted regression trees. Journal of Animal Ecology, 77, 802–813.Elith, J., Kearney, M. & Phillips, S. (2010) The art of modellingrange-shifting species. Methods in Ecology and Evolution, 1,330–342.European Environment Agency (1994) CORINE land covertechnical guide. Report No. EUR 12585 EN edition. Luxemburg:Commission of the European Communities.Fielding, A.H. & Bell, J.F. (1997) A review of methods for theassessment of prediction errors in conservation presence/absence models. Environmental Conservation, 24, 38–49.Fitzpatrick, M.C. & Hargrove, W.W. (2009) The projection ofspecies distribution models and the problem of non-analogclimate. Biodiversity and Conservation, 18, 2255–2261.Gaston, K.J. (2003) The structure and dynamics of geographicranges, 1st edn. Oxford University Press, Oxford.Heikkinen, R.K., Luoto, M., Araújo, M.B., Virkkala, R., Thuiller,W. & Sykes, M.T. (2006) Methods and uncertainties in bioclimaticenvelope modelling under climate change. Progress inPhysical Geography, 30, 751–777.Heikkinen, R.K., Luoto, M., Virkkala, R., Pearson, R.G. &Korber, J.-H. (2007) Biotic interactions improve prediction ofboreal bird distributions at macro-scales. Global Ecology andBiogeography, 16, 754–763.Global Ecology and Biogeography, ••, ••–••, © 2013 John Wiley & Sons Ltd 9

A. Eskildsen et al.Heikkinen, R.K., Marmion, M. & Luoto, M. (2012) Does theinterpolation accuracy of species distribution models come atthe expense of transferability? Ecography, 35, 276–288.Hewitt, G. (2000) The genetic legacy of the Quaternary ice ages.Nature, 405, 907–913.Hijmans, R.J. & Graham, C.H. (2006) The ability of climateenvelope models to predict the effect of climate change onspecies distributions. Global Change Biology, 12, 2272–2281.Hill, J.K., Thomas, C.D. & Huntley, B. (2003) Modelling presentand potential future ranges of European butterflies usingclimate response surfaces. Butterflies: ecology and evolutiontaking flight (ed. by C.L. Boggs, W.B. Watt and P.R. Ehrlich),pp. 149–167. University of Chicago Press, Chicago, IL.Kattsov, V.M., Källén, E., Cattle, H.P., Christensen, J., Drange,H., Hanssen-Bauer, I., Jóhannesen, T., Karol, I., Räisänen, J.,Svensson, G. & Vavulin, S. (2005) Future climate change:modeling and scenarios for the Arctic. Arctic climate impactassessment, pp. 99–150. Cambridge University Press,Cambridge.Kharouba, H.M., Algar, A.C. & Kerr, J.T. (2009) Historicallycalibrated predictions of butterfly species’ range shift usingglobal change as a pseudo-experiment. Ecology, 90, 2213–2222.Komonen, A., Grapputo, A., Kaitala, V., Kotiaho, J.S. & Päivinen,J. (2004) The role of niche breadth, resource availability andrange position on the life history of butterflies. Oikos, 105,41–54.Landis, J.R. & Koch, G.G. (1977) The measurement of observeragreement for categorical data. Biometrics, 33, 159–174.Luoto, M. (2007) The role of land cover in bioclimatic modelsdepends on spatial resolution. Global Ecology and Biogeography,16, 34–42.Luoto, M., Pöyry, J., Heikkinen, R.K. & Saarinen, K. (2005)Uncertainty of bioclimate envelope models based on the geographicaldistribution of species. Global Ecology and Biogeography,14, 575–584.Luoto, M., Heikkinen, R.K., Pöyry, J. & Saarinen, K. (2006)Determinants of the biogeographical distribution ofbutterflies in boreal regions. Journal of Biogeography, 33,1764–1778.Menéndez, R., Megías, A.G., Hill, J.K., Braschler, B., Willis, S.G.,Collingham, Y., Fox, R., Roy, D.B. & Thomas, C.D. (2006)Species richness changes lag behind climate change. Proceedingsof the Royal Society B: Biological Sciences, 273, 1465–1470.Mitikka, V., Heikkinen, R.K., Luoto, M., Araujo, M.B., Saarinen,K., Poyry, J. & Fronzek, S. (2008) Predicting range expansionof the map butterfly in Northern Europe using bioclimaticmodels. Biodiversity and Conservation, 17, 623–641.Parmesan, C., Ryrholm, N., Stefanescu, C., Hill, J.K., Thomas,C.D., Descimon, H., Huntley, B., Kaila, L., Kullberg, J.,Tammaru, T., Tennent, W.J., Thomas, J.A. & Warren, M.(1999) Poleward shifts in geographical ranges of butterflyspecies associated with regional warming. Nature, 399, 579–583.Pearson, R.G. & Dawson, T.P. (2003) Predicting the impacts ofclimate change on the distribution of species: are bioclimateenvelope models useful? Global Ecology and Biogeography, 12,361–371.Pearson, R.G., Dawson, T.P. & Liu, C. (2004) Modelling speciesdistributions in Britain: a hierarchical integration of climateand land-cover data. Ecography, 27, 285–298.Pöyry, J., Luoto, M., Heikkinen, R.K. & Saarinen, K. (2008)Species traits are associated with the quality of bioclimaticmodels. Global Ecology and Biogeography, 17, 403–414.Pöyry, J., Luoto, M., Heikkinen, R.K., Kuussaari, M. & Saarinen,K. (2009) Species traits explain recent range shifts of Finnishbutterflies. Global Change Biology, 15, 732–743.Prentice, I.C., Bartlein, P.J. & Thompson, W., III (1991) Vegetationand climate change in Eastern North America since thelast glacial maximum. Ecology, 72, 2038–2056.Raynaud, X. (2004) Soil characteristics play a key role in modelingnutrient competition in plant communities. Ecology, 85,2200–2214.Ridgeway, G. (2006) gbm: generalized boosted regression models.R package version 1.6-3. Available at: http://www.ipensieri.com/gregr/gbm.shtml(accessed 16 March 2012).Roberts, D.R. & Hamann, A. (2012) Method selection for speciesdistribution modelling: are temporally or spatially independentevaluations necessary? Ecography, 35, 792–802.Saarinen, K., Lahti, T. & Marttila, O. (2003) Population trends ofFinnish butterflies (Lepidoptera: Hesperioidea, Papilionoidea)in 1991–2000. Biodiversity and Conservation, 12, 2147–2159.Schurr, F.M., Pagel, J., Cabral, J.S., Groeneveld, J., Bykova, O.,O’Hara, R.B., Hartig, F., Kissling, W.D., Linder, H.P., Midgely,G.F., Schröder, B., Singer, A. & Zimmermann, N. (2012) Howto understand species’ niches and range dynamics: a demographicresearch agenda for biogeography. Journal of Biogeography,12, 2146–2162.Sinclair, S.J., White, M.D. & Newell, G.R. (2010) How useful arespecies distribution models for managing biodiversity underfuture climates? Ecology and Society, 15, 8.Swets, J. (1988) Measuring the accuracy of diagnostic systems.Science, 240, 1285–1293.Tietäväinen, H. (2010) Annual and seasonal mean temperaturesin Finland during the last 160 years based on gridded temperaturedata. International Journal of Climatology, 30, 2247–2256.Titeux, N., Maes, D., Marmion, M., Luoto, M. & Heikkinen, R.K.(2009) Inclusion of soil data improves the performance ofbioclimatic envelope models for insect species distributions intemperate Europe. Journal of Biogeography, 36, 1459–1473.Venäläinen, A. & Heikinheimo, M. (2002) Meteorological datafor agricultural applications. Physics and Chemistry of theEarth, Parts A/B/C, 27, 1045–1050.White, P. & Kerr, J.T. (2006) Contrasting spatial and temporalglobal change impacts on butterfly species richness during the20th century. Ecography, 29, 908–918.White, P.J.T. & Kerr, J.T. (2007) Human impacts onenvironment–diversity relationships: evidence for biotichomogenization from butterfly species richness patterns.Global Ecology and Biogeography, 16, 290–299.10Global Ecology and Biogeography, ••, ••–••, © 2013 John Wiley & Sons Ltd

Testing SDMs across space and timeWiens, J.A. (1989) Spatial scaling in ecology. Functional Ecology,3, 385–397.Wood, S.N. (2006) Generalized additive models: an introductionwith R. Chapman & Hall/CRC, Boca Raton, FL.Yates, C.J., Elith, J., Latimer, A.M., Le Maitre, D., Midgley, G.F.,Schurr, F.M. & West, A.G. (2010) Projecting climate changeimpacts on species distributions in megadiverse South AfricanCape and Southwest Australian Floristic Regions: opportunitiesand challenges. Austral Ecology, 35, 374–391.Yee, T.W. & Mitchell, N.D. (1991) Generalized additive modelsin plant ecology. Journal of Vegetation Science, 2, 587–602.SUPPORTING INFORMATIONAdditional supporting information may be found in the onlineversion of this article at the publisher’s web-site.Figure S1 Relationships between the accuracy of generalizedadditive models (GAM) and boosted regression tree (BRT)models validated with independent and non-independentobservations, evaluated with area under the curve of a receiveroperating characteristic plot (AUC) and true skill statistic (TSS).Figures in the left column are based on climate-only models,while figures in the right column are based on combinedclimate-soil-land-cover models. Rho (r) values for Spearman’srank correlations (P < 0.001 for all), as well as linear regressionfits are shown. Dotted lines indicate the 1:1 relationship. A dropin r values between climate-only models (left) and combinedclimate-soil-land-cover models (right) indicates that somespecies distributions are better predicted when using independentvalidation in combination with more complex models.BIOSKETCHAnne Eskildsen is a PhD student at AarhusUniversity’s Department of Bioscience under thesupervision of Toke T. Høye and W. Daniel Kissling.Her main areas of research are the effects of land useand climate change on the persistence of butterflypopulations in Western Europe. The team of co-authorsis interested in macroecology, biogeography, speciesdistributions, population ecology and global change.Editor: Jeremy KerrGlobal Ecology and Biogeography, ••, ••–••, © 2013 John Wiley & Sons Ltd 11