![]() ![]() |
![]() |
|
UFDC Home | myUFDC Home | Help |
![]() ![]() ![]() ![]()
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||
Full Text | ||||||||||||||||||||||||||||||||||||||||||||||||||||||
PAGE 1 Apparentsurvivalestimationfromcontinuousmark recapture/resightingdata AndrewB.Barbour 1 *,Jos eM.Ponciano 2 andKaiLorenzen 1 1 SchoolofForestResourcesandConservation,ProgramofFisheriesandAquaticSciences,UniversityofFlorida,7922NW 71stStreet,Gainesville,FL,32653,USA;and 2 DepartmentofBiology,UniversityofFlorida,Gainesville,FL32611,USA Summary 1. Therecentexpansionofcontinuous-resightingtelemetrymethods(e.g.acousticreceivers,PITtagantennae) hascreatedaclassofecologicaldatanotwellsuitedfortraditionalmark recapturestatistics.Estimatingsurvival whencontinuousrecapturedataisavailableensuesapracticalproblem,becauseclassicalcapture recapture modelswerederivedunderadiscretesamplingschemet hatassumessamplingevent sareinstantaneouswith respecttotheintervalbetweenevents. 2. Toinvestigatetheuseofcontinuousdatainsurvivalan alysis,weconductedamodelstructureadequacysimulationthattestedtheCormack Jolly Seber(CJS)andBarkerjointdatasurvivalestimationmodels,which mainlydierthroughtheBarkersinclusionofsecond aryperiodinformation.Wesimulatedapopulationin whichsurvivalanddetectionoccurredasanearcontinuous (daily)processandcollapseddetectioninformation intomonthlysamplingbinsf orsurvivalestimation. 3. Whilebothmodelsperformedwellwhensurvivalwastimeindependent,theCJSwassubstantiallybiasedfor lowsurvivalvaluesandtime-dependentconditions.Add itionally,unliketheCJS,th eBarkermodelconsistently performedwellovermultiplesamplesizes(numberofmarkedindividuals).However,thehighnumberofparametersintheBarkermodelledtoconvergencedicultie s,resultinginaneedforanalternativeoptimization method(simulatedannealing). 4. WerecommendtheuseoftheBarkermodelwhenusingcontinuousdataforsurvivalanalysis,becauseitoutperformedtheCJSoverabiologicallyreasonablerangeofpotentialparametervalues.However,thepracticaldifcultyofimplementingtheBarkermodelcombinedwithitsshortcomingsduringtwosimulationsleavesroom forthespecicationofnovelsta tisticalmethodstailoredspe cicallyforcontinuousmark resightingdata. Key-words: Barkerjointdata,Cormack Jolly Seber,modelstructureadequacy,telemetry Introduction Reliablebiologicalinferencesabouttheprocessesdriving survivalofindividualsinapopulationdependontheproper formulationofstochasticproce ssmodelsthatareconfronted withcapture recapture/resightingdata.Suchmodelstranslatefundamentalbiologicalque stionsintotestablehypothesesthatfurtherourunderstandingofthesystemofinterest (Cohen2004;Gimenez etal. 2007).Whensuchmodels areinappropriatelyformulated,biascausedbystructural errorscanleadtounreliablestatisticalinferences(Pradel& Sanz-Aguilar2012). Forcapture recapture/resightingdata,formulationofan appropriatestochasticprocessmodelrequiresconsideration ofthestructureofthedatacollected(e.g.discretevs.continuoussamplingevents),thetypeofdatacollected(e.g. recapture,resightingordead recovery)andthebiological characteristicsofthestudysystem(e.g.openvs.closed populations).Forexample,multiplemodelshavebeen developedtoestimatesurvivalfromopenpopulationswhen usingdiscrete-resightingdata(Hightower,Jackson& Pollock2001;McClintock&White2009;Johnson etal. 2010)ordiscrete-recap turedata(Lebreton etal. 1992).The recentexpansionofcontinuous-resightingtelemetrymethods (e.g.acousticreceivers,PITt agantennae;Heupel&Simpfendorfer2002;Barbour&Adams2012)hascreatedaclass ofecologicaldatanotwellsuitedforstandardstatistical methodswhenfatesareunknown(Kie etal. 2010).Without aninvestigationofpropermodelformulation,theinformationcontainedinthisdatawillnotbefullyharnessed,and statisticalinferencesmaybeweakormisleading(Strong etal. 1999). Severalprevioussurvivalstudiesusingcontinuousresightingdatacollapsedcontinuousresightingsintodiscretetimeintervalsandappliedexistingdiscrete-timemodels.For example,Heupel&Simpfendorfer(2002)appliedHightower, Jackson&Pollock(2001)sdisc rete-timemodeltocontinuousresightingdatabycollapsingres ightingsintoweeklysampling bins.Similarly,Adams etal. (2006)collapsedcontinuousresightingdataintoweeklyinte rvalsandestimatedapparent survivalwiththediscreteCormack Jolly Seber(CJS)model. *Correspondenceauthor.E-mail:snook@u.edu 2013TheAuthors.MethodsinEcologyandEvolution 2013BritishEcologicalSociety MethodsinEcologyandEvolution 2013doi:10.1111/2041-210X.12059 PAGE 2 Duringamultiyearstudy,Cameron etal. (1999)collapsed 4months(NovemberthroughFebruary)ofcontinuousresightingsintoasingleencounteroccasionlabelledasJanuary1st eachyearandthenestimatedannualsurvivalwithadiscrete multistatemodel.Hewitt etal. (2010)tookasim ilarapproach, butusedadiscreteCJSmodel. Theuseofcontinuousdataindiscrete-timemodelsviolates theassumptionthatsamplingocc asionsareinstantaneouswith respecttotheintervalbetweenp eriods(e.g.acohortismarked inasingleday,aprolongedperiodoftimeelapses[e.g.a month],thenasubsequentcapture recaptureeventoccurs overasingleday;Pollock etal. 1990).Somestudieshaverecognizedandaccountedforthisissue(Barbour,Boucek& Adams2012a;Bowerman&Budy2012;Ruiz-Guti errez etal. 2012;Mintzer etal. 2013),butitisunknownhowviolatingthis assumptionbiasessurvivalpr obabilitiesinstudiesthathave not.Here,weexplorethisissu ebysimulatingapopulationof markedindividualsthatareresightedonarelativelycontinuous(daily)basisandcollapsingthesecontinuousresightings intodiscrete-timebins.Weth enestimatetheknownsurvival valueswithtwosurvivalestimationmodelstodetermine whetheramodelcurrentlyexiststhatisappropriatefor estimatingsurvivalfromcontinuousresightings.MaterialsandmethodsMODELSTRUCTUREADEQUACYWeusedamodelstructureadequacy (MSA)approach(Taper,Staples &Shepard2008)totestwhethertwosurvivalestimationmodelscould beusedforunbiasedestimationofsur vivalfromcontinuous-resighting data.MSAselectsmodel sbasedontheirabilitytoanswerspecicscienticquestionsgiventhecurrentunderstandingoftherelevantaspects oftherealworld.UndertheMSAapproach,amechanisticsimulation modeliscreatedtorepresenttheunderlyingprocessofinterest,and candidatemodelsareusedtoestimat e/predicttherelevantmetricfrom simulateddata.Thisallowsinvestigationoftwotypesoferrorinthe testedmodels:structural(errors ofapproximation)andestimation (uncertaintyinparameterestimates;Taper,Staples&Shepard2008). Inadditiontotheseerrortypes,theMSAapproachitselfissubjecttoa thirdtypeoferror.Formulatione rroroccursduetodierences betweenthemechanisticsimulat ionmodelandthetrueunderlying processes. Accordingly,weformulatedamec hanisticsimulationmodelofa markedpopulationinwhichindividua lsurvivalanddetectionoccurred asanearcontinuous(daily)process. Wethengenerateddatasetsfrom thesimulationmodelusingarangeofparametervaluesthatfully encompassedbiologicallyplausibleconditions.Foreachofthescenarios,wetestedtheabilityoftwoestimationmodels(CJSandBarker jointdata)torecoverthebasicprope rtiesofthesurvivalparameter.We iteratedthisprocessforeachparameterset100times.Weevaluated structuralerrorbycalculatingrelativebiasandpercentcoverageof survivalestimatesfromeachestimat ionmodelaftersimulatingpopulationsfrommultipleknownparametervalues.Weassessedestimation errorinasecondsimulationbyvaryingthenumberofmarkedindividualsinthesimulatedpopulation.Finally,toevaluatetherobustnessof themodelinferencestounavoidable formulationerrors,weaddedan additionalbiologicalprocess,aseveredisturbanceevent,inathird simulation.SURVIVALESTIMATIONMODELSWeemployedtwosurvivalestimationmodels,theCJS(Lebreton etal. 1992)andtheBarkerjointdata(Barker1997,1999).TheCJSmodel assumessamplingperiodsthatareinstantaneouscomparedtothe intervalbetweensamplingevents(Pollock etal. 1990).Incomparison, theBarkermodeliscomposedofbothinstantaneousprimaryperiods ( i and i + x )andcontinuoussecondaryperiods( i i + x ),withsecondaryperiodsbeingtheinterval( x )betweenprimaryperiods(Fig.1). Duringprimaryperiods,individualsarecapturedandrecapturedinan identicalfashiontotheCJSapproach.However,secondaryperiods occurbetweenmarkingperiodsan dallowmarkedindividualstobe resightedaliveordeadonacontinuousbasis. TheCJSmodelestimatestwoparam eters:(1)survival,estimatedas eitherapparentsurvival( ;survivalconfoundedbyemigration)when emigrationoccursandtruesurvival( s )whenemigrationdoesnotoccur and(2)recaptureprobability( p ).TheBarkermodelestimatesseven parametersduetotheadditionalinformationfromcontinuoussecondaryperiods(Table1).TheBarkermodelestimatestruesurvival( s ) whensecondaryperiodsareconductedovertheentirerangeofa markedpopulationorwhenemigrationdoesnotoccur,and otherwise.Oursimulationmodelsdidnotincludeemigration;therefore,all survivalestimateswillhe reafterbereferredtoas s .SIMULATION1:ST RUCTURALERRORTosimulatetheuseofcontinuousdatafordiscretesurvivalestimation, wesimulatedapopulationthatsurv ived/diedandwasdetected/not detectedonadailybasisandcolla psedthesedailydetectionsinto monthlysamplingbins.Weassumedasystemclosedtoemigrationin whichallindividualsweremarkedduringtherstdaywithnotagging mortality.Therefore,staticpara meters(thoseheldconstantoverall iterations)includedthenumberofmarkedindividuals( n = 1000)and thenumberofdaysforthesimulation( d = 180).Thevariableparameters(thosewealteredbe tweeniterations)ofthesimulationmodelwere limitedtotruemonthlysurvival( sm)andtruemonthlyrecaptureprobability( pm).Tofullyencompassthebiologicallyplausiblerangeof parametervalues,wecreatedsimulationmodelsusing50known sm(a Fig.1. Schematicdiagramofprimaryandsecondarysamplingperiods fortheCormack Jolly SeberandBarkerjointdatamodels.Primary samplingperiodsaredenotedby i i + 1 etc.andendinallsimulations atmonth m = 6.TheBarkermodelincludessecondarysampling information,whichconsistsoftheopenintervalbetweenprimaryperiods.Here,weusedevenintervalsoflength x = 1;however,uneven intervals,andintervalsnotequalto1,couldbeused. 2013TheAuthors.MethodsinEcologyandEvolution 2013BritishEcologicalSociety, MethodsinEcologyandEvolution2 A.B.Barbour,J.M.Ponciano&K.Lorenzen PAGE 3 sequencefrom0 5to1 0)and50known pm(asequencefrom0 02to 1 0)values.Thisresultedin2500variableparametercombinations. Weconstructedthemechanisticsimulationmodel(AppendixS1)in theprogramR(RDevelopmentCoreTeam2011).Foreachindividual, weconductedaBernoullitrial(abinomialcoinip)eachdaytodeterminewhethertheindividualsurvivedordiedwithadailysurvivalprobability( sd)ofeqn1: sd s 1 = 30 m: eqn1 Eachdayanindividualsurvived, asecondBernoullitrialwasconductedtodeterminewhethertheindividualwasdetected.Toconvert monthlyrecaptureprobabilityt odailyrecaptureprobability( pd),we calculatedthedailyprobabilityofn otbeingrecapturedandsubtracted thisvaluefromone,eqn2: pd 1 1 pm 1 = 30 : eqn2 Therecaptureprobabilityneedsto becomputedthiswaysincethere aremanypossiblecombinationsforanindividualtobedetectedatleast onceinagivenmonth,butthereisonlyonepossiblewaytonotbe recaptured.Subtractingtheprobabilityofnondetectionfromone accountedforallpossiblerecapturecombinations. Afterrunningthemechanisticsimulationmodel(AppendixS1)fora givenvariableparameterset,wecollapseddailydetectionsinto monthlybins( m = 6),inwhichindividualswereeitherdetectedornot, tocreatecapturehistoriesforeac hindividual.FortheCJS,these monthlybinsrepresentedprimaryperiods,butwereusedasthesecondaryperiodsintheBarker.FortheBarkermodel,wesetthecapturehistoryvaluesinallprimaryperiods,withtheexceptionofthetagging event,tozero.WecreatedcapturehistoriesfortheCJSbytwomethods.Intherstmethod,whichmirroredAdams etal. (2006),wecollapseddailydetectionsinto6monthlybinsasdescribedabove, meaningthetaggingeventwasincludedintherstmonthofdetections. Weleftalltimeintervalsasthedefaultlengthofone.Inthesecond method,wesetthemarkingeventasanindependentprimaryevent, therebycreatingaseventhbin(sixintervals)inthecapturehistory. Whenusingthesecondmethod,weadjustedforuneventimeintervals withintheRMARKpackage(Laake&Rexstad2008)forR.Usingthe midpointofeachresightingmonthasourreferencepoint,wesetthe rstinterval(betweenmarkingandtherstresightingmonth)toequal alengthof0 5months.Thus,eachsubsequenttimeintervaloccurred betweenthemidpointsoftheresightingmonthsandwasoflength1 0. Sinceshorter-termsurvivalestima tesmaybeofinterestincertain studies,weranaseparatesimulati onusing10-daybins,insteadof30daybins.WecreatedcapturehistoriesfortheCJSbythesecond method,treatingthemarkingeventse paratelyfromresightinginformation. WeestimatedsurvivalwiththeCJSandBarkermodelsusing programMARK(White&Burnham1999)accessedbytheRMARKpackage(Laake&Rexstad2008).FortheBarkermodel,wexed F at1and F at0asnoemigrationoccurred,andwexed p and r to0sincenorecaptureoccurredduringprimaryperiods, andwedidnotsimulatedeadrecoveries(parameterdenitionsin Table1).Thesimulationsdidnotincludetimevariabilityinsurvivalorrecaptureprobability;ther efore,weusedti me-independent estimatesfor s R and R intheBarker,and s and p intheCJS estimationmodel. Weranthemechanisticsimulatio nmodel100timesforeachofthe 2500variableparametercombinations,resultingin250000totaliterations.Foreachiteration,wecomputedtherelativedepartureoftheestimatedsurvivalfromthetruesurvivalas,eqn3: Relativedeparture ^ sm sm s 1 m: eqn3 Therelativebiaswasthenestimatedastheaverageoveralliterationsof theserelativedeparturesforagivenvariableparametercombination. Additionally,wequantiedpercentcoveragebycountingthenumber ofsuccessfuliterationspervariableparametersetinwhichthetrue valuefor smwasincludedinanestimationmodels95%condence intervalof ^ sm.Forthefullsimulationrun(all250000iterations),we rstusedthedefaultNewton RaphsonoptimizationmethodinprogramMARK,andthenreranthefullsimulationwithanalternative optimizationmethod(simulatedan nealing)fortheBarkermodelsince thismodelfailedtoconvergeinmultipleinstances.SIMULATION2:ESTIMATIONERRORWerepeatedsimulationonewithtwoalterationstothemechanistic simulationmodel.First,wexedthevariableparameters( smand pm) toavalueof0 9,asthesevaluesapproximatedinitialestimatesfrom aknownstudysystem(Barbour,Boucek&Adams2012a;Barbour etal. 2012b).Second,wemadethenumberofmarkedindividuals ( n )avariableparameter,withvaluesrangingfrom n = 50to n = 1000byincrementsof50.Wethenreranthesimulationaspreviouslydescribedanditeratedthesimulation1000timesforeach n valuewhileusingsimulatedanne alingforoptimizationwiththe Barkerandthesecondmethodofca pturehistorycreationforthe CJS.Wedeterminedpercentcoverageandestimatedrelativebiasof survivalateach n valueforeachestimationmodelinidenticalfashiontosimulationone.Besidesco mputingtherelativebiasasthe average,relativedeparturefromthetruesurvivalprobability,we kepttrackofthe2 5and97 5percentilesofthedistributionofthese relativedepartures. ToaddresscoverageissueswiththeCJS,weconductedaparallel simulationthatusedparametricbo otstrappingtocreatecondence intervals.For n = 200and500,werepeatedthepriorsimulationforthe CJSfor500iterations,butusedthere sultsofeachiterationtorun1000 bootstrappedsimulations.Weusedthe2 5and97 5percentilesof Table1. BarkerjointdatamodelparameterdenitionsinprogramMARKParameterDenition siTheprobabilitythatananimalaliveat i isaliveat i + 1 piTheprobabilitythatananimalatriskofrecaptureat i isrecapturedat i riTheprobabilitythatananimaldiesin i i + 1isfounddead RiTheprobabilityananima lthatsurvivesfrom i to i + 1isresighted(alive)sometimebetween i and i + 1 R0 iTheprobabilityananimalthatdiesin i i + 1withoutbeingfounddeadisresightedalivein i i + 1beforeitdied FiTheprobabilitythatananimalatriskofrecaptureat i isatriskofrecaptureat i + 1 F0 iTheprobabilitythatananimalnotatriskofrecaptureat i isatriskofrecaptureat i + 1(thisdenitiondiersfromBarker(1997)in ordertoforceprobabilitydriveninternalconstraints;White&Burnham1999) 2013TheAuthors.MethodsinEcologyandEvolution 2013BritishEcologicalSociety, MethodsinEcologyandEvolutionSurvivalestimationfromcontinuousdata 3 PAGE 4 maximumlikelihood(ML)survivalestimatesfromthese1000bootstrappediterationstoconstructcondenceintervalsforeachofthe500 iterationspersamplesize( n ).Weusedtheseparametricbootstrapcondenceintervalstotestcoverageofthetruesurvivalvalue.SIMULATION3:FORMULATIONERRORTodeterminetheabilityoftheestimationmodelstoaccountforadditionalbiologicalcomplexityintheformofadisturbanceevent,we alteredthemechanisticsimulationmodeltoincludeamonthoflowsurvival.First,wexed smto0 90and pmto0 90andmaintained n = 1000.Then,forthethirdmonthofthesimulation(days61 90),we lowered smto0 30torepresentaseveredisturbanceevent. Wecreatedtwomodelstructurestoaccountforthedisturbance event,andweusedthemtoestimatesurvivalusingtheCJSandBarker models.Therstmodelstructureallowedforfulltimedependencewith respecttosurvival, sm( t ).Oursecondmodelstructure, sm( d ),representedthetruth-generatingprocess, themechanisticsimulationmodel. Inthismodel, smforthedisturbancemonthw asestimatedseparately fromtheother,time-independent smperiods.Forallestimationmodels, theotherparameterswerecalculatedinidenticalfashiontosimulation one. Weiteratedthissimulation1000 timesandusedsimulatedannealingforoptimizationwiththeBarkerandthesecondmethodofcapturehistorycreationfortheCJS.Weselectedthemostparsimonious modelstructureforeachestimationmodelaftereachiterationby identifyingthemodelwiththeminimumAkaikesInformationCriterion(AIC;Akaike1973)score.Generally,modelswith D AICvalues < 2havesubstantialsupport,andmodelswith D AIC > 10have nosupport(Burnham&Anderson2004).Wethensummarizedand plottedthesimulateddistributionoftheMLsurvivalestimates. Finally,wedeterminedthepercentcoverageasthenumberof successfuliterationsduringwhichthe95%condenceintervalof ^ smforthegiveniterationincludedthetruevalueof sm.ResultsTheBarkermodelestimatedsurvivalfromcontinuousresightingdatawithminimals tructuralerror,whiletheCJS modelonlyperformedwellundertime-independentconditions withhighsurvival.UnliketheCJS,theBarkermodelperformedwellacrossmultiplesamp lesizesofmarkedindividuals ( n ).Additionally,theBarkermode lreliablyestimatedsurvival whenweaddedbiologicalcompl exitytothemechanisticsimulationmodel.However,theBarke rmodelsoptimizationfailed toconvergeforsomecombinationsofparametervaluesusing Newton Raphsonsmethod,necessitatingtheuseofsimulated annealing.Wesummarizethesi mulationresultsinaseriesof contourplots(Figs2 4)inwhichweplottedtherelativebias (subguresaandb)andpercentcoverage(subguresc andd)ateachofthe2500parametercombinationforagiven simulationrun.SIMULATION1:ST RUCTURALERRORWhenconstructingcapturehistoriesundermethodone,the CJSmodelmoderatelyunderestimated sm(Fig.2a)andrarely demonstratedanacceptablelevel ofcoverage(Fig.2c).Creatingcapturehistoriesundermethodtwo,whichseparated markingfromresightinginformation,resultedinrelatively unbiasedestimates(Fig.2b)withpropercoverageexceptwhen survivalwaslowandespeciallywhencombinedwithhigh recaptureprobability(Fig.2d).Movingfrommonthlyto10daybinswhileusingmethodtwoofcapturehistorycreation didnotsubstantiallyaectresultsfortheCJS(Fig.3a,c). Incomparison,theBarkermodelestimated smwithaconsistent,minorpositivebias,(Fig.4a),butfailedtoconvergemultipletimeswhenusingNewton Raphsonoptimization (Figs4cand5).Thiswaslikelyduetothehighnumberof estimatedparametersleadingtolocalminimaduringnumericaloptimization.However, 100%ofmodelrunsconverged whenusingsimulatedannealingandcoverageestimates consistentlyranged90 98%(Fig.4b,d).Whenusing10day insteadofmonthlybins,theperformanceoftheBarker modelwasreducedatlowrecapt ureprobabilities,withcoverageapproaching0%andrelativebiasexceeding 10 0% (Fig.3b,d).SIMULATION2:ESTIMATIONERRORWhenalteringthenumberofmarkedindividuals( n ),the Barkermodelreliabilitycoveredthetruevalueof smin c .95% oftheiterationsforevery n tested(Fig.6).Relativebiasforthe Barkermodelwasnear0 0%,withvariabilityinthedeparture fromthetruthdecreasingwithincreasing n .TheCJSmodel coveredthetruevalueof smin95%oftheiterationswhen n waslow( n = 50,100),butas n increasedto1000,coveragefell below86%(Fig.6).Thisoccurre dbecausecondenceintervalsbecamenarroweras n increased,whilerelativebiaswas maintainedat0 005%(Fig.6).Thus,theprobabilityofcovering smwiththeCJSdecreasedwithincreasing n .Whenusing parametricbootstrappingtoaddresstheCJSpoorcoverage, coveragedecreasedfrom92 4%to88 4%at n = 200andfrom 88 8%to81 4%at n = 500. 05 06 07 08 09 10 020610 05 06 07 08 09 10 05 06 07 08 09 10 05 06 07 08 09 10 010 008 006 004 002 000 002 004 00 02 04 06 08 10 Monthly survivalMonthly recapture probabilityRelative bias % Coverage020610 020610020610(a) (b) (c) (d) Fig.2. Estimatedrelativebias(a,b)andpercentcoverage(c,d)of monthlysurvivalestimatesfortheCormack Jolly Seber(CJS)estimationmodelwhenusing30-daysamplingbins.Capturehistorieswere createdbytwoseparateapproaches,asdetailedinthemethods.Plots include:(a)methodonerelativebias,(b)methodtworelativebias,(c) methodonepercentcoverageand(d)methodtwopercentcoverage. Forvisualclarity,relativebiasvalueslessthan 5 0%wereplotted as 5 0%:thisisapparentinthelowerright-handcornerof(b). 2013TheAuthors.MethodsinEcologyandEvolution 2013BritishEcologicalSociety, MethodsinEcologyandEvolution4 A.B.Barbour,J.M.Ponciano&K.Lorenzen PAGE 5 SIMULATION3:FORMULATIONERRORWhena30-daydisturbancewasincorporatedintothesimulationmodel,AICsupporteddierentmodelstructuresforthe CJSandBarkerestimationmodels(Table2).FortheCJS, AICselectedthefullytime-dependentmodelforsurvival [ sm( t )],everyiterationwithnosupportgiventothetruth-generatingmodel, sm( d )(Table2).FortheBarkerestimationmodel, thestructuralmodelrepresenti ngthetruth-generatingprocess, time-independentestimatesforallperiodsexceptforthedisturbancemonth, sm( d ),wastheminimumAICmodelin89 9% ofthe1000iterations.Howeve r,thetime-dependentmodel receivedconsiderableAICsupport(Table2). Forthetruth-generatingandtime-dependentmodelstructures,wecomparedtheestimatedrelativebiasinsurvival obtainedbytheCJSandtheBarkermodel.Whenthestructuralmodelwasthe sm( d )model,theCJSmodelcoveredthe truevalueof smforthedisturbancein0 1%oftheiterations withanestimatedrelativebiasof33 3%.Coveragefornondisturbance smwas0 0%,witharelativebiasof 6 3%.Incomparison,theBarkermodelcoveredthetruevalueof smduring thedisturbancein93 8%oftheiterations,witharelativebias of2 5%.Coveragefornondisturbance smwas83 7%,witha relativebiasof1 0%.Whenwecomparedresultsfromthe sm( t )model,theBarkeroutperformedtheCJSmodel(Fig.7). TheCJSmodelprovidedrelativ elyunbiasedestimatesforall monthsexcepttwoandthree,attributingasubstantialproportionofthe smdeclineinmonththreetomonthtwo(Fig.7). Forthedisturbancemonth,theCJSestimationmodelcovered thetruevalueof smin0%oftheiterationswithameanrelative biasof44 2%.TheBarkermodelcoveredthetruevalueof smduringthedisturbancein76 4%oftheiterationswithamean relativebiasof5 6%.DiscussionWepresentedtherstassessmentofthestatisticalpropertiesof survivalestimatorswhencontinuous-timedataareavailable yetadiscrete-timesamplingmodelisusedforestimation. Usingamodelstructuraladequacyapproach(Taper,Staples &Shepard2008),wedemonstratedthatsubstantialbiasexists whencontinuouscapture recaptureinformationisdiscretized forsurvivalestimation.Theextentofthebiasdependsupon theestimationmodelused,withtheBarkerjointdatamodel outperformingtheCJS. TherstmethodofcapturehistorycreationfortheCJS introducedsubstantialbiasbeca usewecodedallindividualsas aliveinmonthonedespitetherebeing29daystosuccumbto mortality.Usingthesecondmethodofcapturehistorycreation,theCJSfailedtoreliablyestimatesurvivalinmostsimulationsandresultedinbiasatlow s withbiasworseningathigh p .Thecomplexitiesofhowthesurvivalanddetectionprocesses 05 06 07 08 09 10 020610 05 06 07 08 09 10 05 06 07 08 09 10 05 06 07 08 09 10 010 005 000 005 010 00 02 04 06 08 10 Monthly survivalMonthly recapture probability(a)(b) (c)(d)Relative bias % Coverage020610 020610 020610 Fig.3. Estimatedrelativebias(a,b)andpercentcoverage(c,d)for survivalestimatesfromtheCormack Jolly Seber(CJS)(a,c)andBarkerjointdata(b,d)estimationmodelswhenusing10-daybins.CJS capturehistorieswerecreatedundermethod2,andBarkeroptimizationBarkerwasconductedbysimulatedannealing.Plotsinclude:(a) CJSrelativebias,(b)Barkerrelativ ebias,(c)CJSpercentcoverageand (d)Barkerpercentcoverage.Forvisualclarity,relativebiasvaluesless than 10 0%wereplottedas 10 0%:thisisapparentinthelower left-handcornerof(b). 05 06 07 08 09 10 020610 020610 020610 020610 05 06 07 08 09 10 05 06 07 08 09 10 05 06 07 08 09 10 010 008 006 004 002 000 002 004 00 02 04 06 08 10 Monthly survivalMonthly recapture probability(a)(b) (c)(d)Relative bias % Coverage Fig.4. Estimatedrelativebias(a,b)andpercentcoverage(c,d)for monthlysurvivalestimatesfromtheBarkerjointdataestimation modelwhenusing30-daybins.Optimizationwasconductedbytwo separatemethods:Newton Raphson(NR)(a,c)andsimulatedannealing(SA)(b,d).Plotsinclude:(a)NRrelativebias,(b)SArelativebias, (c)NRpercentcoverageand(d)SApercentcoverage. 0204060810 05 06 07 08 09 10 000 005 010 015 020 025 030 Monthly survivalMonthly recapture probabilityConvergence failure % Fig.5. Percentageof100iterationsateachknownparametercombinationthatfailedtoconvergewhenusingtheNewton RaphsonoptimizationmethodwiththeBarkerestimationmodeland30-day samplingbins. 2013TheAuthors.MethodsinEcologyandEvolution 2013BritishEcologicalSociety, MethodsinEcologyandEvolutionSurvivalestimationfromcontinuousdata 5 PAGE 6 operatejointlymakeitdiculttounequivocallyascertainwhy biasincreasesatlow s .Onepossibilityisthatthesampleinformationatsuchvaluesislowenoughtogenerateparameter identiablyproblemsassociatedwithproblematicjointprole likelihoods(Ponciano etal. 2012).Findinganapproximation ofthebiasinaverysimplecase forwhichthelikelihoodfunctionallowsananalyticaltreatmentoftheproblemmayshed lightonthisissue. Despiteanexpectationthatincreasingsamplesize( n )would improveCJSmodelperformance,increasing n resultedinan unchangedbiasanddecreasingcoverageduetooverlynarrow condenceintervals.Thus,wecreatedparametricbootstrap condenceintervalssincetheyhavebettercoverageproperties whentheMLestimateisunbiased(Efron&Tibshirani1993). However,ourimplementationofparametriccondenceintervalsexacerbatedthecoverageproblem.Theconstantbiasin parameterestimatesacrosssa mplesizessuggeststhataparametricbootstrapconstantbia scorrection(constantacross dierentvaluesofthetrueparameters)oftheestimatemay improvecoveragepropertiesandthuswarrantsadetailed simulationstudyexploringthisissue. Duringthedisturbanceeventsimulation,weusedanarbitrarymethodofcapturehistorycreationfortheCJS,inwhich wehadtheabilitytoperfectlybracketthedisturbanceevent withinasinglesamplingperiod.Evenwiththisprescient knowledge,theCJSreturnedbiasedestimates,makingitunlikelytoperformwellundereldconditionswheresuchknowledgedoesnotexist.TheCJSspoorcoverageanddicultlyin dealingwithbiologicalcomplexityseemtomakethismodela poorchoiceforusewithcontinuousdatainreal-lifeapplicationsthatrequiretime-depen dentestimates.However,while estimateswerebiasedinthetime-dependentsimulation,the Percent coverage Relative bias0850 0875 0900 0925 0950 0050 0025 0000 0025 0050 02505007501000Number markedValue Model Barker CJS Fig.6. PercentcoverageandrelativebiasfortheBarkerandCormack Jolly Seber(CJS)undervariousnumbersofmarkedindividuals( n ).CJS capturehistorieswerecreatedundermethodtwo,andoptimizationfortheBarkerwasconductedbysimulatedannealing.Relativebiascalculatedas theaverage,relativedeparturefromthetru esurvivalprobabilityover1000iterationsper n andwasplottedwiththe2 5 97 5percentilesofthedistributionoftheserelativedepartures.Onthe x -axis, n wasreducedby10fortheCJSandincreasedby10fortheBarkermodelforvisualclarity. Table2. AkaikesInformationCriterion(AIC)tableresultsfrom simulationthree,whichincludeda1-monthdisturbanceevent Model k MeanAICcMean AICc%AICcselected (a) sm( t )74620 10 0100 sm( d )34812 0191 90 0 (b) sm( d )44618 10 089 9 sm( t )84622 24 110 1 Monthlysurvival( sm)estimateswereeithertime-dependent sm( t )or timeindependentexceptforthedisturbanceperiod sm( d ).Thenumber ofestimatedparameters( k ),themeanAICscoreover1000simulated iterationsandthepercentofitera tionsgivingAICsupporttoamodel aregiven.Thesimulationranforthe:(a)Cormack Jolly Seberand(b) Barkerjointdatamodels. 04 06 08 10 12345 Time periodSurvival estimateModel Barker Truth CJS Fig.7. Time-dependentestimatesofmonthlysurvival, sm( t ),forthe Cormack Jolly Seber(CJS)andBarkerestimationmodelsplotted againstthetruevaluefor sm.CJScapturehistorieswerecreatedunder methodtwo,andoptimizationfortheBarkerwasconductedbysimulatedannealing.Meansurvival estimatesplottedwiththe2 5 97 5percentilesofthedistributionofmaximumlikelihoodestimatesfromthe 1000iterations. 2013TheAuthors.MethodsinEcologyandEvolution 2013BritishEcologicalSociety, MethodsinEcologyandEvolution6 A.B.Barbour,J.M.Ponciano&K.Lorenzen PAGE 7 CJSsuccessfullyapproximatedth eoverallsurvivalduringthe simulation.Thus,ifastudyis designedtomeasureoverall survivaltheCJSmaybeanappropriatechoice. ThesuccessoftheBarkerjoint datamodelisnotsurprising, sincethemodelwasformulatedforasituationinwhichcontinuousresightingsoccurredbetweendiscretesamplingintervals (Barker1997).AlthoughtheBar kermodelperformedwell,it didnotreach95%coverageofthetruesurvivalvalueinall simulations.Additionally,the practicalitiesofimplementing theBarkermodelwithcontinuousdatawerenotwithoutdiculties.Inoursimulation,wedi dnotconductdiscretesampling events,whichallowedustocomputethelikelihoodbyxing thevaluesoffourparameters.Inrealsituations,however,all sevenparametersinthemodelmayneedtobeestimated.Since theparametersmayvarybytime,group,orbeassociatedwith covariates,theBarkermodelrequiressubstantialexperienceto properlyformulatean apriori modelset.Ifthelikelihoodsurfaceswereproblematicwithon lythreeparameters,wewould expectnontrivialmaximizationproblemswhenthefullmodel isimplemented,particular lywhenthenumberofunknown parametersislargerelativetothedatasetathand.Thepracticaldicultyofimplementingt heBarkermodelcombinedwith itsshortcomingsduringtwosimulationsleavesroomforthe specicationofnovelstatistica lmethodstailoredspecically forcontinuousmark resightingdata.Astartingpointto achievesuchagoalcouldbeworkingwithcontinuous-time survivalstochasticprocessmod elswhosetransitionprobability matrixcorrespondexactlytothetransitionmatrixofafamily ofdiscrete-timestochasticprocesses(Allen2010). Oursimulationrepresentst herststeptowardsunderstandinghowtobestusecontinuousdatainsurvivalestimation.Withtheexceptionofthedisturbanceevent,weonly simulatedtime-independents urvivalanddetectionprobabilitywhileexplicitlyignoringem igration,whichisnotlikelyto bereectiveofbiologicalreality.AlthoughtheBarkermodel isdesignedtoaccountforrandomemigration(Barker1997) andhasbeenshowntoeectivelyhandlesuchmovement (Horton&Letcher2008),themodelsrobustnesstoemigrationwhentheparametersdesignedtodealwithemigration (FandF )arexedisunknown.Whileourstudyignored theissue,wearecurrentlyusingempiricaldatafromour eldresearchsitetodeterminerealisticratesofemigration, whichwillbeusedtoextendthiswork(A.B.Barbour, unpublisheddata). Althoughwefocusedsolelyontheuseofcontinuousdatain theestimationofdiscretesurvival,thisproblemmayhaveparallelsinothercontexts.Forinstance,datamaybegrouped alongspaceinsteadofthetimeaxis,despitetheacknowledged importanceofspatialheterogeneity(VanKirk&Lewis1997; Neubert&Caswell2000).Itisu nknown,forexample,ifthe discretizationofmodern,large -scaleGIS(GeographicInformationSystems)dataofspatialabundancedistributionsmay leadtobiasedabundanceestimators(Kleiber&Hampton 1994;Sibert etal. 1999;Adam&Sibert2002).Themathematicalintricaciesofndingthecorrecttime-scalerepresentation formodelling,estimationandtes tingofthebiolo gicalprocess ofinterestineachcasearenot trivial.Inthecontextofmark recapturemodels,itisnecessarytoinvestigatewhentheunderlyingdiscrete-timeMarkovia nstructureintheBarkermodel canbeapproximatedwithacontinuous-timeMarkovprocess (e.g.seeKarlin&Taylor1981,chap.15,section2.F). Reliableunderstandingandpredictionofcomplexecologicaldatahingesontheformulatio nofproperstatisticalmodels toquantifybiologicalprocesse swhileaccountingforthesamplingschemeused.However,theecologicalliteratureislled withexampleswhereo-the-shelfstatisticalmodelshaveproventobeaninsucienttooltogenerateunderstandingofthe biologicalprocessesofinterestsimplybecausetheyarenottailoredtotheapplicationathandandassuch,areunabletoharnesstheinformationinthedataeectively(e.g.Strong etal. 1999).Here,wefocusedoninformingtheoreticiansandpractitionersalikeabouttheinferentialproblemsassociatedwith temporalgroupingpracticesinsurvivalestimation.Thiswork shouldbetakenasapositive rststeptowardsseekinga model-centredsolutiontosuchdiculties.AcknowledgementsWethankJ.Nichols,J.HinesandM.Connerforcommentshelpfulinthedesign ofthesimulationandM.Allen,A.Adams,D.Behringer,DavidKoonsandone anonymousreviewerfortheirvaluableinsights.ABBwassupportedbya NationalScienceFoundationGraduate ResearchFellowshipunderGrantNo. DGE-0802270.K.L.acknowledgesfundingfromtheFloridaFishandWildlife ConservationCommission,ProjectNo.11409.ReferencesAdam,M.S.&Sibert,J.R.(2002)Populationdynamicsandmovementsofskipjacktuna( Katsuwonuspelamis )intheMaldivianshery:analysisoftagging datafromanadvection-diusion-reactionmodel. AquaticLivingResources 15 ,13 23. Adams,A.J.,Wolfe,R.K.,Pine,W.E.&Thornton,B.L.(2006)EcacyofPIT tagsandanautonomousantennasystemtostudythejuvenilelifestageofan estuarine-dependentsh. EstuariesandCoasts 29 ,311 317. Akaike,H.(1973)Informationtheoryasa nextensionofthemaximumlikelihood principle. SecondInternationalSymposiumonInformationTheory (edsB.N. Petrov&F.Csaki),pp.267 281.AkademiaiKiado,Budapest. Allen,L.J.S.(2010) AnIntroductiontoStochasticProcesseswithApplicationto Biology ,2ndedn.ChapmanandHall,UpperSaddleRiver,NewJersey. Barbour,A.B.&Adams,A.J.(2012)Biologgingtoexaminemultiplelifestagesof anestuarine-dependentsh. MarineEcologyProgressSeries 457 ,241 250. Barbour,A.B.,Boucek,R.E.&Adams,A.J.(2012a)Eectofpulsedgastric lavageonapparentsurvivalofajuvenileshinanaturalsystem. Journalof ExperimentalMarineBiologyandEcology 422 423 ,107 113. Barbour,A.B.,Adams,A.J.,Yess,T.,B ehringer,D.C.&Wolfe,R.K.(2012b) Comparisonandcost-benetanalysisofPITtagantennaeresightingand seine-netrecapturetechniquesforsurvivalanalysisofanestuarine-dependent sh. FisheriesResearch 121 122 ,153 160. Barker,R.J.(1997)Jointmodelingoflive-recapture,tag-resight,andtag-recovery data. Biometrics 53 ,666 677. Barker,R.J.(1999)Jointanalysisofmark recapture,resightingandrecovery datawithage-dependenceandmarking-eect. BirdStudy 46 ,82 91. Bowerman,T.&Budy,P.(2012)Incorporatingmovementpatternstoimprove survivalestimatesforjuvenilebulltrout. NorthAmericanJournalofFisheries Management 32 ,1123 1136. Burnham,K.P.&Anderson,D.R.(2004)Multimodelinference:understanding AICandBICinmodelselection. SociologicalMethods&Research 33 ,261 304. Cameron,C.,Barker,R.,Fletcher,D.,Slooten,E.&Dawson,S.(1999)ModellingsurvivalofHectorsdolphinsaroundBanksPeninsula,NewZealand.JournalofAgricultural,Biological,andEnvironmentalStatistics 4 ,126 135. Cohen,J.E.(2004)Mathemati csisbiologysnextmicroscope,onlybetter;biology ismathematicsnextphysics,onlybetter. PLoSBiology 2 ,e439. Efron,B.&Tibshirani,R.(1993) AnIntroductiontotheBootstrap .Chapmanand Hall,NewYork,NY,USA. 2013TheAuthors.MethodsinEcologyandEvolution 2013BritishEcologicalSociety, MethodsinEcologyandEvolutionSurvivalestimationfromcontinuousdata 7 PAGE 8 Gimenez,O.,Rossi,V.,Choquet,R.,Dehais,C.,Doris,B.,Varella,H.,Vila,J.P. &Pradel,R.(2007)State-spacemodell ingofdataonmarkedindividuals. EcologicalModelling 206 ,431 438. Heupel,M.R.&Simpfendorfer,C.A.(2002)Estimationofmortalityofjuvenile blacktipsharks, Carcharhinuslimbatus ,withinanurseryareausingtelemetry data. CanadianJournalofFisheriesandAquaticSciences 59 ,624 632. Hewitt,D.A.,Janney,E.C.,Hayes,B.S.& Shively,R.S.(2010)Improvinginferencesfromsheriescapture recapturestudiesthroughremotedetectionsof PITtags. Fisheries 35 ,217 231. Hightower,J.E.,Jackson,J.R.&Pollock,K.H.(2001)Useoftelemetrymethods toestimatenaturalandshingmortalityofstripedbassinLakeGaston,North Carolina. TransactionsoftheAmericanFisheriesSociety 130 ,557 567. Horton,G.E.&Letcher,B.H.(2008)Movementpatternsandstudyareaboundaries:inuencesonsurvivalestimationincapture mark recapturestudies. Oikos 117 ,1131 1142. Johnson,H.E.,Mills,S.L.,Wehausen ,J.D.&Stephenson,T.R.(2010)Combininggroundcount,telemetry,andmark-resightdatatoinferpopulation dynamicsinanendangeredspecies. JournalofAppliedEcology 47 ,1083 1093. Karlin,S.&Taylor,H.M.(1981) ASecondCourseinStochasticProcesses .AcademicPress,SanDiego,California,USA. Kie,J.G.,Matthiopoulos,J.,Fieberg,J.,Powell,R.A.,Cagnacci,F.,Mitchell, M.S.,Gaillard,J.M.&Moorcroft,P.R.(2010)Thehome-rangeconcept:are traditionalestimatorsstillrelevantwithmoderntelemetrytechnology? PhilosophicalTransactionsoftheRoyalSociety.B,BiologicalSciences 365 ,2221 2231. Kleiber,P.&Hampton,J.(1994)ModelingeectsofFADsandislandsonmovementofskipjacktuna( Katsuwonuspelamis ):estimatingparametersfromtaggingdata. CanadianJournalofFisheriesandAquaticSciences 51 ,2642 2653. Laake,J.&Rexstad,E.(2008)RMark analternativetobuildinglinearmodels inMARK. ProgramMARK:AGentleIntroduction (edsE.Cooch&G. White),pp.C1 C115.http://www.phidot.org/software/mark/docs/book/. Lebreton,J.D.,Burnham,K.P.,Clobert,J.&Anderson,D.R.(1992)Modeling survivalandtestingbiologicalhypothesesusingmarkedanimals:aunied approachwithcasestudies. Ecology 62,67 118. McClintock,B.T.&White,G.C.(2009)Alesseld-intensiverobustdesignforestimatingdemographicparameterswithmark-resightdata. Ecology 90 ,313 320. Mintzer,J.V.,Martin,A.R.,daSilva,V.M.F.,Barbour,A.B.,Lorenzen,K.& Frazer,T.K.(2013)EectofillegalharvestonapparentsurvivalofAmazon Riverdolphins( Iniageorensis ). BiologicalConservation 158 ,280 286. Neubert,M.G.&Caswell,H.(2000)Demogr aphyanddispersal:calculationand sensitivityanalysisofinvasionspeedforstructurespopulations. Ecology 81 1613 1628. Pollock,K.H.,Nichols,J.D.,Brownie,C.&Hines,J.(1990)Statisticalinference forcapture recaptureexperiments. WildlifeMonographs 107 ,3 97. Ponciano,J.M.,Burleigh,J.G.,Bra un,E.L.&Taper,M.L.(2012)Assessing parameteridentiabilityinphylogeneticmodelsusingdatacloning. Systematic Biology 61 ,955 972. Pradel,R.&Sanz-Aguilar,A.(2012)Modelingtrap-awarenessandrelatedphenomenaincapture recapturestudies. PLoSONE 7 ,e32666. RDevelopmentCoreTeam(2011) R:ALanguageandEnvironmentforStatistical Computing .RFoundationforStatisticalComputing,Vienna.http://www. R-project.org Ruiz-Guti errez,V.,Doherty,P.F.Jr,Santana,E.C.,Mart nez,S.C.,Schondube, J.,Mungu a,H.V.&I ~ nigo-Elias,E.(2012)Survivalofresidentneotropical birds:considerationsforsamplinganalysisbasedon20yearsofbird-banding eortsinMexico. TheAuk 129 ,500 509. Sibert,J.R.,Hampton,J.,Fournier,D.A.&Bills,P.J.(1999)Anadvectiondiusion-reactionmodelfortheestimationofshmovementparametersfrom taggingdata,withapplicationtoskipjacktuna( Katsuwonuspelamis ). CanadianJournalofFisheriesandAquaticSciences 56 ,925 938. Strong,D.R.,Whipple,A.V.,Child,A .L.&Dennis,B.(1999)Modelselection forasubterraneantrophiccascade:root-feedingcaterpillarsandentomopathogenicnematodes. Ecology 80 ,2750 2761. Taper,M.L.,Staples,D.F.&Shepard,B.B.(2008)Modelstructureadequacy analysis:selectingmodelsonthebasisoftheirabilitytoanswerscienticquestions. Synthese 163 ,357 370. VanKirk,R.W.&Lewis,M.A.(1997)Integr odierencemodelsforpersistencein fragmentedhabitats. BulletinofMathematicalBiology 59 ,107 137. White,G.C.&Burnham,K.P.(1999)ProgramMARK:survivalestimationfrom populationsofmarkedanimals. BirdStudy 46 ,120 138. Received25January2013;accepted3April2013 HandlingEditor:RobertB.OHaraSupportingInformationAdditionalSupportingInformationmaybefoundintheonlineversion ofthisarticle. AppendixS1. Mechanisticsimulationmodelfunctionasrunin programR. 2013TheAuthors.MethodsinEcologyandEvolution 2013BritishEcologicalSociety, MethodsinEcologyandEvolution8 A.B.Barbour,J.M.Ponciano&K.Lorenzen |