Comparison of habitat models for scarcely detected species
Résumé
When performinghabitatmodels,modellershavetochoosebetweenpresence-absenceandpresence-
only modelstoestimatethehabitatpreferencesofaspecies.Primarily,thischoicedependsonthedata
that areavailableandwhethereffortdataarerecordedinparalleltosightingdata.Forspeciesthatare
rare orscarce,themodelshavetoaddressagreatnumberofzeros(i.e., no animalseen)thatweakensthe
ability tomakesoundecologicalinferences.Wetestedtwotypesofhabitatmodels(presence-absence
vs. presence-only)todeterminewhichtypebestdealtwithdatasetscontaininganexcessofzeros,and
we appliedourmodelstoasightingdatasetthatincludedthecommon(Delphinus delphis) and striped
(Stenella coeruleoalba) dolphin (approximately92%zeros).Weusedtwotypesofpresence-absencemod-
els (GeneralisedAdditivemodels–GAMs,GeneralisedLinearModel–GLM)andonepresence-onlymodel,
a MaxEntmodel,andweusedvariouscriteriatocomparethesemodels(i.e., AIC, deviances,rootograms
and distributionpatternspredictedbythemodels).Overall,weobservedthatthepresence-absencemod-
elsmadebetterpredictionsthanthepresence-onlymodel.Amongthepresence-absencemodels,theGAM
with aNegativeBinomialdistributionwasbetteratpredictingsmalldelphinidshabitats,eventhough
the GAMwithaTweediedistributionexhibitedsimilarresults.However,thezero-inflatedPoissondis-
tributions exhibitedlessconvincingresultsandwascontrarytowhatwasexpected.Finally,despite92%
zeros, ourdatasetwasnotzero-inflated.Ourstudydemonstratestheimportanceofselectingappropriate
models tomakereliablepredictionsofhabitatuseforspeciesthatarerareorscarce.