Rolf Tschernig holds the chair of Econometrics at the Department of Economics and Econometrics at the University of Regensburg 93040 Regensburg Germany Phone 499419432737 Email rolftschern ID: 91940
Download Pdf The PPT/PDF document "Regensburger zur Wirtschaftswissenschaf..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Regensburger zur Wirtschaftswissenschaft University of Regensburg Working Papers in Business, Economics and Management Information Systems Fractionally Integrated VAR Models with a Fractional Lag Operator and Deterministic Trends: and Two-step Estimation , Roland Weigandng memory, maximum likelihood estimation, fractional Rolf Tschernig holds the chair of Econometrics at the Department of Economics and Econometrics at the University of Regensburg, 93040 Regensburg, Germany. Phone: +49-941-943-2737, E-mail: rolf.tschernig[at]wiwi.uni-regensburg.de Enzo Weber holds the chair of Empirical Economics, especially Macroeconometrics and Labour Markets at the Department of Economics and Econometrics at the University of Regensburg, 93040 Regensburg, Germany, and is head of the research department Forecasts and Structural Analysis of the Institute for Employment Research (IAB), 90478 Nuremberg, Germany. Phone: +49 -911-179-7643, E-mail: enzo.weber[at]iab.de Roland Weigand is a researcher at the research department Forecasts and Structural Analysis of the Institute for Employment Research (IAB), 90478 Nuremberg, Germany. Phone: +49 -911-179-3291, E-mail: roland.weigand[at]iab.de FractionallyIntegratedVARModelswithaFractionalLagOperatorandDeterministicTrends:FiniteSampleIdenticationandTwo-stepEstimationRolfTscherniga,*,EnzoWebera,b,andRolandWeigandbaUniversityofRegensburg,DepartmentofEconomics,D-93040RegensburgbInstituteforEmploymentResearch(IAB),D-90478NurembergJanuary2013AbstractFractionallyintegratedvectorautoregressivemodelsallowtocapturepersistenceintimeseriesdatainavery exibleway.Additional exibilityfortheshortmemoryproper-tiesofthemodelcanbeattainedbyusingthefractionallagoperatorofJohansen(2008)inthevectorautoregressivepolynomial.However,italsomakesmaximumlikelihoodestimationmoredicult.Inthispaperwerstidentifyparametersettingsforunivariateandbivariatemodelsthatsuerfrompooridenticationinnitesamplesandmaythereforeleadtoesti-mationproblems.Second,weproposetoinvestigatetheextentofpooridenticationbyusingexpectedlog-likelihoodsandvariationsthereofwhicharefastertosimulatethanmultivari-atenitesampledistributionsofparameterestimates.Third,weprovidealineofreasoningthatexplainsthendingfromseveralunivariateandbivariatesimulationexamplesthatthetwo-stepestimatorsuggestedbyTschernigetal.(2010)canbemorerobustwithrespecttoestimatingthedeterministiccomponentsthanthemaximumlikelihoodestimator.Keywordsfractionalintegration,longmemory,maximumlikelihoodestimation,fractionallagoperator. Correspondingauthor.Email:Rolf.Tschernig@wiwi.uni-regensburg.de.Phone:(+49)941943-2737Aformerversionwastitled:Conditionalversusapproximateconditionalmaximumlikelihoodestimationinfractionallyintegratedvectorautoregressivemodelswithafractionallagoperatoranddeterministictrends.1 1IntroductionFractionallyintegratedvectorautoregressive(VAR)modelshavebecomeavaluableextensionofVARmodelswithintegerordersofintegration.Recently,Tschernigetal.(2010)introducedthefractionallagoperator(seeJohansen,2008)intostandardfractionallyintegratedVARmodelsinordertoavoidcertainshortcomingsinimpulseresponseanalysisunderlong-runidenticationrestrictions.Theadditionalmodeling exibilityduetothefractionallagoperator,however,makesmaximumlikelihoodestimationmoredicult,inparticularifdeterministiccomponentsareincluded.Inthispaperwerstidentifyparametersettingsforunivariateandbivariatemodelver-sionsthatsuerfrompooridenticationinnitesamplesandmaythereforeleadtoestimationproblems.Second,weproposetoinvestigatetheextentofpooridenticationbyuseofexpectedlog-likelihoodsandvariationsthereofwhicharefastertosimulatethanmultivariatenitesam-pledistributionsofparameterestimates.Third,weprovidealineofreasoningthatexplainsthendingfromseveralunivariateandbivariatesimulationexamplesthatthetwo-stepes-timatorsuggestedbyTschernigetal.(2010)canbemorerobustwithrespecttoestimatingthedeterministiccomponentsthanthemaximumlikelihoodestimator.Withinthemaximumlikelihoodapproachtheestimatorofthedeterministiccomponentsanditspropertiesdependonthesimultaneouslyestimatedfractionalparameters.Ifthelatteraresubjecttopoornitesampleidentication,thedeterministiccomponentsmaybepoorlyestimatedwhichaddstothedicultiesofestimatingthefractionalparameters.Wethereforesuggesttoapplythetwo-stepestimatorinpractice.Section2brie ypresentsfractionallyintegratedVARmodelswithafractionallagoperatorwhileSection3treatsthemaximumlikelihoodestimatorforthesemodels.InSection4wediscussprominentcasesofunivariatedatageneratingprocessesthatmaybepronetopooridentication.Section5explainshowtheexpectedlog-likelihoodandvariationsthereofallowtovisualizethemagnitudeofpooridenticationforagivendatageneratingprocess.InSection6weshowhowtheproblemofpooridenticationworsensoncedeterministiccomponentshavetobeestimatedaswell.Finally,Section7extendstheanalysistobivariatefractionallyintegratedVARprocesseswithafractionallagoperator.2 2FractionallyintegratedVARbmodelsInthissectionweconsiderfractional(vector)autoregressiveprocesseswithafractionallagoperatoranddeterministicterms.Usingthefractionaldierenceoperator(1L)d=1Xj=0(jd) (d)(j+1)Lj;where()denotesthegammafunction,thefractionallagoperatorLbisdened(seeJohansen,2008)asLb=1(1L)b=c1L+c2L2+withb0:Thedegreeoffractionalintegrationbisrequiredtobepositiveinordertoguaranteethatapplyingthefractionallagoperatordoesnotchangethedegreeofintegration.Forb=1oneobtainsthestandardlagoperatorL.Afractionallyintegratedk-variatevectorautoregressiveprocesswithfractionallagoperator(FIVARb)processforxtisgivenbyA(Lb)(L;d)xt=ut;utWN(0;);t=1;2;:::(1a)(L;d):=diag(1L)d1;(1L)d2;:::;(1L)dk:(1b)Heretheerrorsutaremultivariatewhitenoise(WN)withmeanzeroandhomoscedasticcovariancematrix.Forthepth-ordervectorautoregressivelagpolynomialA(z)=IA1zApzpwerequirethestabilityconditionofJohansen(2008,Corollary6)tohold.ItprovidesaconditionsuchthateachelementinthevectorprocesstgivenbyA(Lb)t=ut;t=:::;2;1;0;1;2;:::;(2)isI(0).TherootsofjA(z)j=0havetobeoutsideCb,whichistheimageoftheunitcircleunderthemappingf:z7!1(1z)b.ThisconditiondependsbothonA(.)andonbandcaneasilybecheckedonceparametervaluesaregiven.Sincez=1liesonCbregardlessofthevalueofb,stabilityoftheA(Lb)polynomialexcludestheunitrootcaseandalsoimpliesnonsingularityofA(1).Underthestabilityconditiontheparameterbaddssome exibilitytotheshort-runprop-ertiesoftheprocessratherthanhavingin uenceontheintegrationorders.Forb=1oneobtainsastandardfractionallyintegratedVARprocess(e.g.Nielsen,2004a).Inthesequeldenotetheparametervectorwithallparametersofthemodelby2SwhereSistheparameterspacethatcontainsallparametersthatfullthestabilitycondition.3 2.1TreatmentofpresamplevaluesInordertoobtainasolutiontotheprocess(1)fornonstationaryxtwemakeuseofthetrun-catedoperatornotation(Johansen,2008,AppendicesA.4,A.5).Let(L)denoteaninnitematrixpolynomialandI()theindicatorfunction.Then+(L)xt=I(t1)Pt1i=0ixtiand(L)xt=(L)xt+(L)xt.1Notealsothat+(L)canalwaysbeinvertedbyexpanding(z)1aroundzeroandtakingthersttterms.Thepresamplevaluesxt,t=0;1;:::,arenotmodeledalthoughtheyareallowedtobestochastic.Undermildconditionsonthepresamplevalues,seee.g.JohansenandNielsen(2012a),anequivalentrepresentationoftheFIVARbprocess(1)isgivenbyA+(Lb)+(L;d)xt=ut+mt;t=1;2;:::;(3)mt=[A+(Lb)(L;d)+A(Lb)+(L;d)+A(Lb)(L;d)]xtwithitssolutiongivenbyxt=A+(Lb)1+(L;d)1ut+t;t=A+(Lb)1+(L;d)1mt;(4)wheretcapturestheimpactofthepresamplevalues.Notethattheforecasterrorimpulseresponses hforhorizonhcanbecomputedfromthetruncatedlagpolynomial +(L)=A+(Lb)1+(L;d)1,replacingin(4)tbyt+h.2.2DeterministiclineartrendsThemodelingoflineardeterministictimetrendsisrestrictedtothesampleofmodeleddatabyassumingthatyt=8]TJ ; -1; .63; Td; [00;:0+1t+xt;ift1;xtift0:(5)Therefore,theFIVARbmodelwithdeterministictrendsisgivenbyA(Lb)(L;d)(yt01t)=ut;t=1;2;::::(6) 1Notethatforaproduct(L)=(L)(L)ofVARpolynomials(L),(L)onehas+(L)=+(L)+(L);(L)=+(L)(L)+(L)+(L)+(L)(L):4 ByexplicitlystatingthedependenceonpresamplevaluesoneobtainstherepresentationsA+(Lb)+(L;d)yt=ut+A+(Lb)+(L;d)(0+1t)+mt;t=1;2;:::;(7)yt=A+(Lb)1+(L;d)1ut+(0+1t)+t;t=1;2;::::(8)Ifitisassumedthatallpresamplevaluesarezero,xt=0,t0,thenmt=t=0,t=1;2;:::.3MaximumlikelihoodestimationInthefollowingwestatetheconditionalmaximumlikelihoodestimatorforgivenpresamplevaluesxt,t=0;1;:::;Tp.Forbrevityitiscalledmaximumlikelihoodestimatorthroughoutthepaper.Let=vec(A1;:::;Ap)denotethevectorofallVARcoecientsandY=yTp;:::;y1;y0;y1;:::;yTthevectorofobservablepresampleandsamplevalues.Foranobservedtimeseries,themaximumlikelihoodestimatorsforthegeneralmodel(6)allowingfordeterministictrendsisgivenby^d;^b;^0;^1;^;^=argmaxd;b;0;1;;L(d;b;0;1;;;Y);(9)wherethemaximizationiscarriedoutoveranappropriateparameterspace.Further,weassumenormallydistributederrorsinordertoderivethelog-likelihoodfunctionexplicitly:L(d;b;0;1;;;Y)=Tk 2log2T 2logjj1 2TXt=1ut(d;b;0;1;)01ut(d;b;0;1;);(10)whereut(d;b;0;1;)isobtainedbyrearranging(6)asut(d;b;0;1;)=(L;d)(yt(0+1t))| {z }zt(d;0;1)A1Lb(L;d)(yt(0+1t))| {z }zt1(d;b;0;1)ApLpb(L;d)(yt(0+1t))| {z }ztp(d;b;0;1);t=1;2;::::(11)Sincezt(d;0;1)doesnotdependonitfollowsfrom(11)thatforgivend,b,0,1theVARcoecientmatricesA1,...,Ap,ifunrestricted,areobtainedbyleastsquares,regressingzt(d;0;1)onzt1(d;b;0;1);:::;ztp(d;b;0;1).Togetherwithconcentratingoutthissimpliesthemaximizationconsiderablyandleadstotheconcentratedlog-likelihoodLc(d;b;0;1;Y)=Tk 2(log2+1)1 2logTXt=1ut(d;b;0;1)ut(d;b;0;1)0;(12)5 thatcanbemaximizedinsteadof(10).Forexample,inthebivariatecasethemaximizationofthisconcentratedlog-likelihoodfunctiononlyrequiresanonlinearoptimizationonsevenparametersindependentlyoftheVARorderp.AvailablerelatedasymptoticresultsarederivedbyJohansenandNielsen(2012a)inaframeworkoffractionalcointegrationundertheassumptionofnodeterministictrends.Whileintheirsettingidenticaldisareassumed,thisdoesnotimplythattheindividualfractionalordersofintegrationareidenticalduetothepossibilityoftrivialcointegrationwith=(1;0)0.However,weexcludethepossibilityoffractionalcointegrationbythestabilityrestrictionontheVARpolynomialA(Lb).ForthestandardFIVARmodelunderthestabilityconditionwithb=1asymptoticresultsarederivedbyNielsen(2004a)andimpliedbyHualdeandRobinson(2011).ItremainstobecheckedifHualdeandRobinson(2011)coverstheFIVARbmodelforb6=1andzeropresamplevalues.Withrespecttounivariateprocessestheasymptoticbehaviorofthemaximumlikelihoodestimator(9)isinvestigatedbyNielsen(2004b),whorequiresb=1,astableARpolynomial,andzeropresamplevaluesbutallowsforadeterministictrend,andbyJohansenandNielsen(2010),whoallowfornonzeropresamplevalues,b6=1aswellasforunitrootsintheARpolynomial.Theyexcludedeterministictrends,however.JohansenandNielsen(2012b)derivetheasymptoticsecond-orderbiasduetopresamplevaluesforapureunivariatefractionalprocessdxt=ut.WiththeexceptionofJohansenandNielsen(2010)andJohansenandNielsen(2012a)allmentionedresultsrequiretheA(Lb)polynomialtobestable.4Poornitesampleidenticationincaseofunivariatepro-cessesItiswellknownthatparameterestimationmaybemoredicultiftheparametervaluesofthedatageneratingprocessareclosetotheboundaryoftheparameterspacewhereallorsomeparametersarenotidentied.Togiveanexample,letthedatageneratingprocessbeasmoothtransitionautoregressiveprocessthatisveryclosetoalinearautoregressiveprocess.Sincetheparametersofasmoothtransitionautoregressivemodelarenotidentiedifthedatageneratingprocessisinfactlinear,aweaknonlinearstructuremaynotbedetectableinmanysamplessuchthatestimationtakesplaceasifparametersarenotidentied.AnotherwellknownexamplearecommonrootsinARMA(p;q)models.Ifthetrueorders6 arep0andq0,thenestimatinganARMA(p0+1;q0+1)modelsuersfromcommonrootswhichcausestheparametersofthemodeltobenotidentied.Incontrasttothepreviousexamplewhereaweaknonlinearstructureinthedatageneratingprocesscausesestimationproblems,itistoolargeamodelorderinthelatterexample.WhenestimatingFIVARbprocessessimilarscenariosmayoccurwheretheorderoftheautoregressivepolynomialistoolarge,pp0:a)p0=0andp=1:AssumethatthedatageneratingprocessisaunivariatewhitenoisebutaunivariateFIVARbmodeloforderoneistted,(1a1Lb)dxt=ut;utWN(0;2);t=1;2;::::(13)ThismodelwillhenceforthbecalledaFARbmodeloforderone.i)Insertingthetrueparametervaluesa1;0=0andd0intothelagpolynomialsdelivers(10Lb)d0=d0sothatbcantakeanypositivevalue.Thus,bisnotidentiedbutdis.ii)However,onemayalsoinserta1;0=1.Then(1Lb)d=b+dwhichhastobeequaltod0.Thenthereisacontinuumofcombinationsforbanddforwhichb+d=d0holds.b)p0=1andp=2:Since(1a1Lba2L2b)d=(11Lb)(12Lb)d,insertingthetrueparameters0a1;01,b0]TJ/;༕ ;.9; ;Tf 1;.51; 0 ;Td [;0,andd0]TJ/;༕ ;.9; ;Tf 1;.51; 0 ;Td [;0intoaFARbmodelwithorder2deliverstworepresentationsusing1=a1;0:i)2=0andthusa2=0,beingequivalenttoaFARbmodeloforderone.ii)2=1.Then(1a1;0Lb0)d0=(1a1;0Lb0)b0d0b0=(1a1;0Lb0)(11Lb0)d0b0=(1a1Lb0a2L2b0)d0b0wherea1=a1;0+1,a2=a1;0.Notethatinii)thestabilityconditionisviolatedsince2=1.IncontrasttoCasea),theparametersforeachscenarioarelocallyidentiedbutnotglobally.Thus,onemayexpectthelog-likelihoodtobebimodal.7 Eveniftheparametersareidentied,theymaybeclosetotheboundaryof(partial)non-identication.Assumingp0=1and0a1;01suchthatthestabilityconditionholds,itcanstillhappenthatthesampleinformationisnotsucienttokeeptheestimateofa1reasonablyfarawayfromzerosothatanidenticationproblemmayresultinnitesamples,resemblingcasea)i)above.Similarly,ifa1;0issmallerthanunitybutnotdistinguishablefromoneinanitesample,thend=d0+b0b(14)givestheapproximatelocationsofestimateddandb.Onlyd+bisappropriatelyidentied.Wecallthesescenariospoornitesampleidentication.ForhigherorderFARbmodels,p1,poornitesampleidenticationmayalsoresultfromthepossibilitythatanestimateofbisclosetozero.ThenonehasajLjb0=aj1(1L)b00(15)independentlyofthevalueofaj.Therefore,ifaDGPthatiswhitenoiseismodeledbyaFARbprocesswithorderp,theARparametersaj,j=1;2;:::;parepoorlyidentiedforvaluesofbclosetozero.Hencealsoforthisreasonitisimportanttoaimatusingcorrectlagorders.Dependingonthepurposeofthemodel,poornitesampleidenticationmayormaynotbeharmful.Ifoneisinterestedinthelong-rundynamicsimpliedbythedegreeofintegration,thenitcanbeveryproblematic.Asanexample,comparethedegreeofintegrationimpliedbyaprocesswitha1closetoonetothecasea1=1foridenticald0andb0:whiletheformerisanI(d0)process,thelatterisanI(d0+b0)process.Whetherpoornitesampleidenticationisanissueforagivensamplemaybecheckedbyinvestigatingthelog-likelihoodfunctionontherelevantrangeoftheparametersofinterest.FortheFARbmodel(13)oforderone,onemayplottheconcentratedlog-likelihoodonagridfordandbandvisuallycheckwhethertherearetwopeaksormountainridgesthatindicatepoornitesampleidentication.Incaseonewantstocheckthepotentialofpoornitesampleidenticationpriortosampling,onemaysimply\average"thelog-likelihoodfunction(10)overpossiblesamplesbytakingexpectationsofthelog-likelihoodfunctionatthetrueparametervector.Thisdeliverstheexpectedlog-likelihoodfunction(16)furtherdescribedinthenextsection.Finally,ifthetrueparametersarefarenoughawayfromtheboundaryof(partial)non-identication,poornitesampleidenticationshouldnotbeamajorissue.Thus,incaseofthe8 FARbmodel(13),onemayexpectreasonablenitesampleidenticationincaseofa1;0=0:6.5Visualizingtheexpectedlog-likelihoodIntheprevioussectionitwasarguedthatpoornitesampleidenticationmaybecheckedwithoutreferringtoanysamplebyinvestigatingtheexpectedlog-likelihoodfunction.Thissuggestiondiersfromthecommonlyusedmethodforinvestigatingnitesampleestimationpropertiesbysimulatingthenitesampledistributionof^.Theexpectedlikelihoodcomputa-tionstakeintoaccountadditionalinformationabouttheshapeofpossiblelikelihoodfunctionsawayfromtheirmaxima.Additionally,whenjointlyconsideringmorethantwoparameters,theexpectedlikelihoodcanbesimulatedfasterthanthejointdensityoftheparameterestimator.5.1Theexpectedlog-likelihoodLetE0[]indicatethattheexpectationistakenwithrespecttothedatageneratingprocess.Thentheexpectedlog-likelihoodisgivenbyE0[L(;Y)]=Zlogf(Y;)f(Y;0)dY:(16)Notethat0maynotbeuniquewithouttherestriction02S.Plottingcontourlinesorsurfacesoftheexpectedlog-likelihood(16)isonlypossibleifisoflengthl=2.Onewaytodealwiththecasel2istosplitthe(l1)vectorintoa(21)vectorIthatcontainsthetwoparametersofinterestanda((l2)1)vectorIIofallotherparametersandthenmaximize(16)withrespecttoII:E0[L(I;mII(I);Y)];wheremII(I)=argmaxIIE0[L(I;II;Y)]:(17)FortheunivariateFARbprocess(13)onemaydeneI=(d1;b)0andII=d2;a1;20.Ifaconcentratedlog-likelihoodisavailable,itmaybepreferabletoconsidertheexpectedconcentratedlog-likelihood.IfIIcanbeconcentratedoutcompletely,theexpectedconcen-tratedlog-likelihoodisdenedbyE0hLI;^II(I;Y);Yi(18)throughconcentrating^II(I;Y)=argmaxIIL(I;II;Y):(19)9 Consideringtheexpectedconcentratedlog-likelihoodallowstostudytheindirecteectsthatestimatingIIhasontheestimationofI,whilesuchindirecteectsareignoredinthemaximizationapproach(17).Further,onemayuse(18)tocomparevariousestimatorsfor^II(Y)withrespecttotheirin uenceontheestimationproblemofI.Ifanalternativeestimatorto(19)isused,thenoneobtainsdierentobjectivefunctionsfortheestimationofIwhichcanbecompared.FinallyonemaycombinebothapproachesbyconcentratingoutsomeparametersofIIwhiletakingthemaximumwithrespecttotheremainingones.Asanexamplewithrespectto(10)onemaydeneI=(d1;b)0,II=d2,andIII=(vec(A)0;00;01;vech()0)0.Thentheexpectedconcentratedlog-likelihooddependingonIisgivenbyE0hLI;mII(I);^III(I;mII(I;Y);Yi(20)throughconcentrating^III(I;II;Y)=argmaxIIIL(I;II;III;Y)and'optimizingout'mII(I)=argmaxIIE0hLI;II;^III(I;II;Y);Yi:5.2Expectedconcentratedlog-likelihoodsforFARbprocessesoforderoneInthissectionweusesimulationstocomputetheexpectedconcentratedlog-likelihoodsforvariousFARbprocesses(13)oforderone.Wechoosed0=b0=0:8andconsidera1;0=0:1;0:6;0:9.ForgivenI=(d;b)0andII=(a1;2)0theconcentratedlog-likelihoodiseasilycomputedasdescribedinSection3.Inordertoapproximatetheexpectedconcentratedlog-likelihood(18)wedraw100realizationsforgivendandb.Forobtainingcontourplotswevarytheparametersd2[1;1:5]andb=[0:02;1:5]usinggridswithstepsizeof0.02.Themagnitudeofpoornitesampleidenticationofdandbisvisualizedbytheshapeandsizeoftheareawiththelargestvaluesoftheexpectedconcentratedlog-likelihood,whichcanwellbeseenfromaplotwithcontourlines.ForsamplesizeT=250anda1;0=0:9thecontourlinesoftheexpectedconcentratedlog-likelihoodareshowninthetoppanelofFigure1.Ifa1;0wereexactly1,onewouldexpectfrom(14)thatthelocationofthelargestvaluesof10 theexpectedconcentratedlog-likelihoodisdescribedby^dd0+b0^b=1:6^b.FromthetoppanelofFigure1itisseenthatbylettinga1;0deviateslightlyfromone,thislocationisshiftedsomewhatto^d1:8^b.AsarguedinSection4,pooridenticationinsmallsamplesislessofaproblemifa1;0isneitherclosetozeronortoone,saya1;0=0:6,ascanbeseenfromthemiddlepanelofFigure1.ThelowerpanelofFigure1showsthatthepoornitesampleidenticationissueisagainprominentifa1;0=0:1andthusclosetozero.FromCasea)ii)inSection4itfollowsthatifa1wereexactlyzero,^bcan oatarbitrarilywhile^d0:8.Thisexplainstheupperridgeifa1isestimatedclosetozero.Thelowerridgeisexplainedbyestimatesa11leadingtoanegativetrade-obetweendandb.6DealingwithdeterministictrendsInthissectionweinvestigatewhytheproblemofpoornitesampleidenticationworsensifdeterministictrendsareallowedinthemodel.InordertoestimatetheparametersoftheFIVARbmodel(6)withlineartrendsonemayusethemaximumlikelihoodestimatorbasedon(10)inSection3.Inthefollowingweprovideargumentsandafewsimulationresultsthatatwo-stepestimationismorerobust.6.1PitfallsinthemaximumlikelihoodestimationOnemayrewrite(6)sothat0and1canbeestimatedbyleastsquaresifalltheotherparametersd,b,andaregiven:A(Lb)(L;d)yt=A+(Lb)+(L;d)10+A+(Lb)+(L;d)t1+ut;t=1;2;:::;T:(21)Thereforetheestimatesofthedeterministiccomponentsarein uencedbyboththelongmem-oryparametersdandtheparametersandbdeterminingtheI(0)dynamics.Asaconse-quence,ifoneofthecasesinvestigatedinSections4and5.2occurswhereforagivensamplesizeandDGPonlyd+biswellidentiedwhiledandbarenot,bmaybeestimatedtoolargeanddtoosmall.Toseethepossibleimplications,consideraunivariateDGPwithd0=1,0;06=0,1;0=0,anda1;0closetoone.Suppose^dtakesthetruevalued0=1,thentheregressionforestimating0and1correspondingto(21)is(1a1Lb)yt=(1a1Lb)I(t=1)0+(1a1Lb)I(t1)1+ut;t=1;2;:::;T:(22)11 Then0isestimatedfromonlyoneobservation,t=1,andVar(^0)1forT!1.Thisisnotproblematicastheimpactoftheestimated0vanisheswithgrowingsamplesize.Duetopoornitesampleinformationa1maybeestimatedclosetoone.Sincefora1;0=1,onlyd+bisidentiedby(14),^dmaybeclosetozeroand^bcloseto1+b0.Then,settingd=0,no(fractional)dierencesaretakenandtheerrorsin(1a1Lb)yt=(1a1Lb)I(t1)0+(1a1Lb)tI(t1)1+ut;t=1;2;:::;T(23)exhibitaunitroot.InthiscaseitcanbeshownthatVar(^0)increaseswithsamplesizeT.Then,theestimateof0canbeexpectedtoremainin uential.Thus,theimplicitestimationpropertiesfor0anditsimpactontheotherestimatescruciallydependonthedestimate.Itcanbeshownthatthisalsoholdsforthe1estimate.Suchdicultiesduetoagrosslywrongdestimatescanbeavoidedifthepoornitesampleidenticationproblemiscircumventedwhenestimating0and1.Toachievethis,weinvertA(Lb)in(21),whichisalwayspossibleifthestabilityconditionholds,andconsidertheregression(L;d)yt=+(L;d)10++(L;d)t1+"t;t=1;2;:::;T;(24)wheretheerrors"t=A(Lb)1utareautocorrelatedbutI(0).Inthisregressionbdoesneitherentertheregressandnortheregressor.Usingtheregression(24)workssincetheregressorsaredeterministicandthereforetheautocorrelatederrorsdonotmattermuchforestimating0and1.Therefore,estimatingtheparametersofthedeterministictermsonbasisof(24)onlyrequiresknowledgeofdwhichcanbeestimatedbysomesemiparametricestimatorthatdoesnotsuerfromthenitesampleidenticationproblems.Thisleadsdirectlytothetwo-stepestimatordescribednext.6.2Two-stepestimationWithinthetwo-stepestimationprocedure,thedeterministiccomponentsareestimatedintherststep.Inthesecondstepthelog-likelihoodfunctionismaximizedafterreplacingthedeterministiccomponentsbytheirestimatesfromtherststep.Therststepinvolvesrunningtheregression(24).Thisamountstocomputingtheleast12 squaresestimatorfrom(1L)dsys;t=(1L)ds+10;s+(1L)ds+t1;s+"s;t;s=1;2;:::;k;t=1;2;:::;T:(25)Weobtainthefollowingtwo-stepestimatorsuggestedbyTschernigetal.(2010):Firststep:Foreachseriess=1;2;:::;kestimatethememoryparameterdswiththesemipara-metricexactlocalWhittleestimatorofShimotsu(2010)thatallowsfordeterministictrends.Inordertoobtain~0;sand~1;s,runtheregression(25)aftertakingfractionaldif-ferencesbasedonthe~dsestimate.Dothisfors=1;2;:::;k.Secondstep:Maximizethelog-likelihoodfunction(10)with0=~0and1=~1.Notethatinthebivariatecasethenonlinearoptimizationinthesecondsteponlyincludesthethreeparametersd1;d2;b.6.3Expectedconcentratedlog-likelihoodsforFARbprocessesoforderonewithdeterministictrendsWeconsiderthesamethreeDGPsasinSection5.2butestimateaFARbmodelallowingforadeterministictrend.Figures2,4,and6displaythecontourlinesoftheexpectedconcentratedlog-likelihoodwherethemaximumlikelihoodestimatorisusedforconcentratingout0and1.Figures3,5,and7displaythecontourlinesoftheexpectedconcentratedlog-likelihoodwherethetwo-stepestimatorofSection6.2isusedforestimatingthedeterministicterms.Comparingthecontourlinesofbothestimatorsfora1;0=0:9inFigures2and3showsthattheregionofhighestexpectedconcentratedlog-likelihoodisclosertothetruevaluesforthetwo-stepestimatorthanforthemaximumlikelihoodestimatorasconjecturedintheprevioussubsection.FortheotherFARbwitha1;0=0:1whoseestimationmayalsosuerfrompoornitesampleidentication(closetoCasea)inSection4),themaximumlikelihoodestimatorinFigure6isevenmoreoincomparisontothetwo-stepestimatorinFigure7.Evenforthecaseofa1;0=0:6whichshowstheleastnitesampleidenticationproblemsincaseofnodeterministiccomponentthetwo-stepprocedureseemssuperiorascanbeseenfromcomparingFigures4and5.Whencomparingthecontourlinesfromthetwo-stepprocedure13 withtheresultsincaseofknowndeterministictermsoneobservesthatestimatingdeterministiccomponentsiscostlysincetheregionsofhighestexpectedconcentratedlog-likelihoodnolongerincludethetrueparameters.Inordertocheckwhethertheidenticationproblemreallydependsonthesamplesizewealsocomputedtheexpectedconcentratedlog-likelihoodsfortheFARbwitha1;0=0:9forsamplesizeT=1000.NowbothestimationproceduresworkmuchbetterascanbeseenfromFigures8and9althoughthetwo-stepconcentratedlikelihoodstillshowsahighercurvature.Insum,thetwo-stepestimatorworksbetteraswasexpectedfromthereasoninginSections6.1and6.2.ItremainstobeinvestigatedhowthesendingscarryovertothebivariateFIVARbmodel(6)whichisdonenext.7PoornitesampleidenticationinbivariateFIVARbpro-cessesInthissectionweconsidertheFIVARbprocesses(1)and(6).7.1ProcesseswithpoornitesampleidenticationIfthebivariateDGPhasadiagonalVARcoecientmatrix0@I0@10021ALb1A0@(1L)d100(1L)d21Axt=ut;t=1;2;:::;(26)thenmaximizingthelikelihoodcanbeexpectedtobeclosealthoughnotidenticaltomaximizingthelikelihoodofeachunivariateseries.Thelattergenerallydeviatesfromthejointestimationsincebisestimatedforeachseriesseparately.Ifdandbareknown,bothestimatorscoincide.However,sincebisidenticalacrosstheindividualseries,oneeigenvalueneitherbeingclosetoonenorzeroshouldhelptoestimatebsucientlywellsothattheothereigenvaluedoesnolongercauseproblemsevenifitisclosetooneorzero.Forthisreason,weexpectestimationproblemsfrompoornitesampleidenticationifneitherindividualprocesshelpstodeterminebandweexpectthatsuchaFIVARbprocessinheritsthepoornitesampleidenticationproblemsfromtheindividualFARbprocesses.Inthefollowingwewillinvestigateprocesses(26)with=1=2.Sinceistheeigenvalue(withmultiplicity2)oftheautoregressiveparametermatrixA,itisinterestingtoinvestigatewhetherthepoornitesampleidenticationproblemsdiminishonce14 dependencebetweenthetwoprocessesthroughtheautoregressivepolynomialisintroducedwhilekeepingtheeigenvalueconstant.InordertoobtainamatrixAthatexhibitsthesameeigenvaluebutnonzeroo-diagonalelementsweapplytheJordandecompositionforrealmatrices(Lutkepohl,1996,Section6.2.1(2)).UsinganonsingularJ=(acbd),a(22)matrixAwithoneeigenvaluecanbewrittenasA=J0@101AJ1=0@ac adbca2 adbcc2 adbc+ac adbc1A:(27)Weconjecturethatthepoornitesampleidenticationproblemsdiminishforacommoneigenvalueoncethedependencebetweenthetwoprocessesisincreasedthrougho-diagonalelementsinA.Wealsoconjecturethattheimpactoftheseproblemsmattersmoreforthemaximumlikelihoodestimatorthanforthetwo-stepestimatorlikeintheunivariatecase.Inordertochecktheseclaimsweinvestigatetheexpectedconcentratedlog-likelihoodfunctionsfortwoDGPsnext.7.2Expectedconcentratedlog-likelihoodsforFIVARbprocessesoforderoneAllDGPsconsideredhavenormallydistributederrorswith0=(10:50:51);nodeterministictrend,0;0=1;0=0,andzeropresamplevalues.Tochecktheclaimthatidenticaleigenvaluesclosetoonewithoutautoregressivedependenceposeproblemsweconsider:DGPdiagthediagonalprocess(26)withd1;0=b0=0:8andd2;0=1:8.ThefractionalparametervaluesresemblevaluesthatwereestimatedbyTschernigetal.(2010)forUSrealGDPandpricedata.Inordertoinvestigatetheeectofautoregressivedependence,wechooseJin(27)suchthatthecorrelationbetweeny1tandy2tisabout-0.5when=I.ThisisthecaseforJ0=1111.ThisdeliversthefollowingDGP:DGPdep0@I0@+1 2+1 21 21 21AL0:81A0@(1L)0:800(1L)1:81Axt=ut;t=1;2;:::(28)Supposethattherearenowthreeparametersofinterest:d1,d2andb.Forcomputingtheexpectedlog-likelihoodswerstconcentrateoutallparametersexceptthosethreeandthen15 plotcontourlineswherewetakethemaximumoftheexpectedconcentratedlog-likelihoodwithrespecttothethirdparameter.Asanexamplechoosein(20)I=(d1;b)0,II=d2,andIII=(0;00;01;vech()0)0.WerstdiscusstheresultsofthediagonalDGPdiag.Ifnodeterministiccomponentsareestimated,Figure10showsthatthetrueparametersareclosetothepointofhighestexpectedconcentratedlog-likelihood.Oncedeterministiccomponentsareallowedforintheestimationmodel,thisisnolongerthecaseforthemaximumlikelihoodestimatorascanbeseenfromtheleftcolumninFigure11.Incontrast,thetwo-stepestimatorstilldeliversreasonableresults.ThisndingsupportsthereasoningofSection7.1.Ifdynamicspill-oversbetweentheseriesarepresentasintheDGPdep,thenitturnsoutfrominspectingFigures12and13(forestimationwithoutandwithdeterministiccomponents,respectively)thattheweaknitesampleidenticationproblemislesspronouncedthaninthediagonalcasealthoughbseemstobeestimablelesspreciselythand1andd2.NotablydierenteectsoftheestimatorsforthedeterministictermsarenotpresentinFigure13.8ConclusionWediscussednitesampleestimationpropertiesoffractionallyintegratedVARmodelswherehigh exibilityisintroducedthroughthefractionallagoperatoranddeterministictrends.Weidentifysituationswhereidenticationmaybepoorinnitesamplesandverifytheseclaimsbyplottingexpected(concentrated)likelihoods.Deterministictrendsaggravatetheproblems.Atwo-stepestimatorhelpstocircumventatleastpartofthe awswhicharefacedifthemaximumlikelihoodestimatorisused.Subsequentworkmaybeconcernedwiththeasymptoticpropertiesoftheestimators.Asageneralrecommendation,futureempiricalresultsusingfractionallyintegratedtimeseriestechniquesshouldbecheckedwithrespecttonitesampleidenticationissuesbothtoassessrobustnessoftheresultsandtosupportanappropriateestimatorchoice.16 ReferencesHualde,J.andRobinson,P.M.(2011),\GaussianPseudo-MaximumLikelihoodEstimationofFractionalTimeSeriesModels,"TheAnnalsofStatistics,39,3152{3181.Johansen,S.(2008),\ARepresentationTheoryforaClassofVectorAutoregressiveModelsforFractionalProcesses,"EconometricTheory,24,651{676.Johansen,S.andNielsen,M..(2010),\LikelihoodInferenceforaNonstationaryFractionalAutoregressiveModel,"JounalofEconometrics,158,51{66.|(2012a),\LikelihoodInferenceforaFractionallyCointegratedVectorAutoregressiveModel,"Econometrica,80,2667{2732.|(2012b),\Theroleofinitialvaluesinnonstationaryfractionaltimeseriesmodels,"WorkingPaper1300,Queen'sEconomicsDepartment.Lutkepohl,H.(1996),HandbookofMatrices,Wiley&Sons.Nielsen,M..(2004a),\EcientInferenceinMultivariateFractionallyIntegratedTimeSeriesModels,"EconometricsJournal,7,63{97.|(2004b),\EcientLikelihoodInferenceinNonstationaryUnivariateModels,"EconometricTheory,20,116{146.Shimotsu,K.(2010),\ExactLocalWhittleEstimationofFractionalIntegrationwithUnknownMeanandTimeTrend,"EconometricTheory,26,501{540.Tschernig,R.,Weber,E.,andWeigand,R.(2010),\Long-runIdenticationinaFraction-allyIntegratedSystem,"RegensburgerDiskussionsbeitragezurWirtschaftswissenschaft447,UniversityofRegensburg.17 Figure1:Theexpectedconcentratedlog-likelihoodfunctionforaFAR0:8process(13)with250observations,d=0:8anda1=0:9(above),a1=0:6(middle)anda1=0:1(below)18 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d -3.1 -3.1 -3.1 -3.1 -3.1 -3.1 -3 -3 -3 -3 -3 -3 -3 -3 -3 -2.9 -2.9 -2.9 -2.9 -2.9 -2.9 -2.9 -2.8 -2.8 -2.8 -2.8 -2.8 -2.8 -2.8 -2.8 -2.7 -2.7 -2.7 -2.7 -2.6 -2.6 -2.5 -2.5 -2.4 -2.4 -2.3 -2.2 -2.1 -2 -1.9 -1.8 -1.7 -1.6 -1.5 -1.4 -1.3 -1.2 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d -3.6 -3.6 -3.6 -3.6 -3.6 -3.5 -3.5 -3.5 -3.5 -3.5 -3.5 -3.5 -3.5 -3.5 -3.5 -3.5 -3.5 -3.4 -3.4 -3.4 -3.4 -3.4 -3.4 -3.4 -3.4 -3.4 -3.3 -3.3 -3.3 -3.3 -3.2 -3.1 -3 -2.9 -2.8 -2.7 -2.6 -2.5 -2.4 -2.3 -2.2 -2.1 -2 -1.9 -1.8 -1.7 -1.6 -1.5 -1.4 -1.3 -1.2 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d -2.25 -2.25 -2.2 -2.2 -2.2 -2.2 -2.2 -2.2 -2.2 -2.2 -2.2 -2.15 -2.15 -2.15 -2.15 -2.15 -2.15 -2.15 -2.15 -2.15 -2.15 -2.15 -2.15 -2.15 -2.15 -2.15 -2.15 -2.15 -2.1 -2.1 -2.1 -2.1 -2.1 -2.1 -2.1 -2.1 -2.1 -2.1 -2.1 -2.1 -2.1 -2.1 -2.05 -2.05 -2.05 -2.05 -2.05 -2.05 -2.05 -2.05 -2.05 -2 -2 -2 -2 -2 -2 -2 -1.95 -1.95 -1.95 -1.95 -1.95 -1.9 -1.9 -1.9 -1.9 -1.9 -1.9 -1.85 -1.85 -1.85 -1.85 -1.85 -1.8 -1.8 -1.8 -1.8 -1.75 -1.75 -1.75 -1.7 -1.7 -1.7 -1.65 -1.65 -1.65 -1.6 -1.6 -1.6 -1.55 -1.55 -1.55 -1.5 -1.5 -1.5 -1.45 -1.45 -1.45 -1.4 -1.4 -1.4 -1.35 -1.35 -1.35 -1.3 -1.3 -1.25 -1.25 -1.2 -1.2 -1.15 -1.1 Figure2:Theexpectedconcentratedlog-likelihoodfunctionwithestimateddeterministictrendforaFAR0:8process(13)withd=0:8anda1=0:9with250observations. Figure3:Theexpectedconcentratedtwo-stepapproximatelog-likelihoodfunctionwithestimateddeterministictrendforaFAR0:8process(13)withd=0:8anda1=0:9with250observations.19 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d -2 -2 -2 -2 -2 -2 -2 -1.8 -1.8 -1.8 -1.8 -1.8 -1.8 -1.8 -1.8 -1.8 -1.6 -1.6 -1.6 -1.6 -1.6 -1.4 -1.4 -1.4 -1.4 -1.2 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d -2.9 -2.9 -2.9 -2.9 -2.9 -2.8 -2.8 -2.8 -2.8 -2.8 -2.8 -2.7 -2.7 -2.7 -2.7 -2.7 -2.7 -2.7 -2.6 -2.6 -2.6 -2.6 -2.6 -2.6 -2.5 -2.5 -2.5 -2.5 -2.5 -2.4 -2.4 -2.4 -2.4 -2.3 -2.3 -2.2 -2.1 -2 -1.9 -1.8 -1.7 -1.6 -1.5 -1.4 -1.3 -1.2 -1.1 -1 -0.9 -0.8 -0.7 Figure4:Theexpectedconcentratedlog-likelihoodfunctionwithestimateddeterministictrendforaFAR0:8process(13)withd=0:8anda1=0:6with250observations. Figure5:Theexpectedconcentratedtwo-stepapproximatelog-likelihoodfunctionwithestimateddeterministictrendforaFAR0:8process(13)withd=0:8anda1=0:6with250observations.20 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d -2.2 -2.2 -2.2 -2.2 -2.2 -2.2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -2 -1.8 -1.8 -1.8 -1.8 -1.8 -1.6 -1.6 -1.6 -1.6 -1.4 -1.2 -1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d -3.2 -3.2 -3.2 -3.2 -3.2 -3 -3 -3 -3 -3 -3 -3 -3 -2.8 -2.8 -2.8 -2.8 -2.8 -2.8 -2.8 -2.8 -2.8 -2.6 -2.6 -2.4 -2.4 -2.2 -2 -1.8 -1.6 -1.4 -1.2 -1 -0.8 -0.6 Figure6:Theexpectedconcentratedlog-likelihoodfunctionwithestimateddeterministictrendforaFAR0:8process(13)withd=0:8anda1=0:1with250observations. Figure7:Theexpectedconcentratedtwo-stepapproximatelog-likelihoodfunctionwithestimateddeterministictrendforaFAR0:8process(13)withd=0:8anda1=0:1with250observations.21 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d -0.4 -0.4 -0.4 -0.4 -0.4 -0.4 -0.4 -0.4 -0.4 -0.2 -0.2 -0.2 -0.2 -0.2 -0.2 -0.2 -0.2 -0.2 -0.2 0 0 0 0 0 0 0 0 0 0.2 0.2 0.2 0.2 0.2 0.2 0.4 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 3.2 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d -1.3 -1.2 -1.2 -1.2 -1.2 -1.2 -1.2 -1.2 -1.2 -1.1 -1.1 -1.1 -1.1 -1.1 -1.1 -1.1 -1.1 -1.1 -1.1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -0.9 -0.9 -0.9 -0.9 -0.9 -0.9 -0.9 -0.8 -0.8 -0.8 -0.8 -0.8 -0.7 -0.7 -0.7 -0.6 -0.6 -0.5 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 Figure8:Theexpectedconcentratedlog-likelihoodfunctionwithestimateddeterministictrendforaFAR0:8process(13)withd=0:8anda1=0:9with1000observations. Figure9:Theexpectedconcentratedtwo-stepapproximatelog-likelihoodfunctionwithestimateddeterministictrendforaFAR0:8process(13)withd=0:8anda1=0:9with1000observations.22 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d -9 -9 -9 -9 -9 -9 -9 -9 -9 -9 -9 -8.5 -8.5 -8.5 -8.5 -8.5 -8.5 -8.5 -8 -8 -8 -8 -8 -8 -8 -8 -7.5 -7.5 -7.5 -7.5 -7.5 -7 -7 -7 -6.5 -6 -5.5 -5 -4.5 -4 -3.5 -3 -2.5 -2 -1.5 -1 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d -10 -10 -10 -10 -10 -9.5 -9.5 -9.5 -9.5 -9.5 -9.5 -9 -9 -9 -9 -9 -9 -9 -9 -9 -8.5 -8.5 -8.5 -8.5 -8.5 -8.5 -8 -8 -7.5 -7.5 -7 -6.5 -6 -5.5 -5 -4.5 -4 -3.5 -3 -2.5 -2 Figure10:Theexpectedconcentratedlog-likelihoodfunctionfortheFIVAR0:8processDG-Pdiag(26)withd1=0:8,d2=1:8anda1=0:8with250observations.23 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d1 24.2 24.8 25.3 25.3 25.3 25.8 25.8 25.8 26.4 26.4 26.4 26.4 26.4 26.9 26.9 26.9 26.9 27.4 27.4 27.4 27.4 27.9 27.9 27.9 27.9 28.5 28.5 29 29 29 29.5 29.5 30.1 30.1 30.6 30.6 30.6 31.1 31.1 31.6 31.6 32.2 32.2 32.7 32.7 33.2 33.2 33.8 34.3 34.8 35.3 35.9 36.4 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 0.0 0.5 1.0 1.5 2.0 2.5 b d2 24.3 24.3 27.9 27.9 27.9 27.9 31.5 31.5 31.5 35.1 -1.0 -0.5 0.0 0.5 1.0 1.5 0.0 0.5 1.0 1.5 2.0 2.5 d1 d2 31.4 32 32 32 32 32 32.7 32.7 32.7 32.7 32.7 32.7 32.7 33.3 33.3 33.3 33.3 33.3 33.3 33.9 34.6 35.2 35.8 36.5 Figure11:Leftcolumn:Theexpectedconcentratedlog-likelihoodfunctionwithestimateddeterministictrends.Rightcolumn:expectedconcentratedtwo-stepapproximatelog-likelihoodfunctionwithestimateddeterministictrends.EachfortheFIVAR0:8processDG-Pdiag(26)withd1=0:8,d2=1:8anda1=0:8with250observations.24 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d1 27.9 27.9 28.5 28.5 29 29.5 29.5 30.1 30.1 30.6 30.6 31.1 31.1 31.1 31.6 31.6 32.2 32.2 32.2 32.7 32.7 32.7 33.2 33.2 33.8 33.8 33.8 34.3 34.8 34.8 35.3 35.9 36.4 36.4 36.9 36.9 37.5 37.5 38 38.5 39.1 39.6 40.1 40.6 41.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d1 25.3 25.8 25.8 26.4 26.4 26.9 26.9 27.4 27.4 27.9 27.9 27.9 28.5 28.5 28.5 29 29 29.5 29.5 29.5 30.1 30.1 30.6 30.6 30.6 31.1 31.1 31.6 31.6 31.6 32.2 32.2 32.7 33.2 33.2 33.8 33.8 34.3 34.3 34.8 34.8 35.3 35.9 36.4 36.9 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 0.0 0.5 1.0 1.5 2.0 2.5 b d2 31.5 31.5 31.5 31.5 35.1 35.1 38.7 38.7 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 0.0 0.5 1.0 1.5 2.0 2.5 b d2 27.9 27.9 27.9 27.9 27.9 27.9 31.5 31.5 31.5 35.1 -1.0 -0.5 0.0 0.5 1.0 1.5 0.0 0.5 1.0 1.5 2.0 2.5 d1 d2 36.5 37.1 37.1 37.1 37.1 37.8 37.8 37.8 37.8 38.4 38.4 38.4 39 39.7 40.3 41 -1.0 -0.5 0.0 0.5 1.0 1.5 0.0 0.5 1.0 1.5 2.0 2.5 d1 d2 32.7 32.7 32.7 33.3 33.3 33.3 33.3 33.9 33.9 33.9 33.9 33.9 34.6 34.6 34.6 35.2 35.8 36.5 37.1 Figure12:Theexpectedconcentratedlog-likelihoodfunctionfortheFIVAR0:8processDG-Pdep(28)withd1=0:8,d2=1:8and=0:8butautoregressivedependencewith250observations.25 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d1 -46.9 -43.7 -43.7 -40.4 -40.4 -40.4 -40.4 -37.1 -37.1 -33.8 -33.8 -33.8 -33.8 -30.5 -30.5 -30.5 -27.2 -27.2 -27.2 -23.9 -20.6 -17.3 -14 -10.7 -7.4 -4.1 -0.8 -0.8 2.5 2.5 5.8 5.8 9.1 9.1 12.4 12.4 15.7 15.7 19 19 22.3 22.3 25.6 28.9 32.1 35.4 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 0.0 0.5 1.0 1.5 2.0 2.5 b d2 5.3 5.3 5.3 8.4 8.4 8.4 8.4 8.4 8.4 11.4 11.4 11.4 11.4 11.4 14.4 14.4 17.5 17.5 20.5 20.5 23.6 26.6 29.6 32.7 35.7 35.7 -1.0 -0.5 0.0 0.5 1.0 1.5 0.0 0.5 1.0 1.5 2.0 2.5 d1 d2 1.6 1.6 1.6 1.6 1.6 1.6 1.6 3.4 3.4 3.4 3.4 5.3 5.3 5.3 5.3 5.3 7.1 7.1 7.1 7.1 7.1 9 9 9 9 10.9 10.9 10.9 12.7 12.7 14.6 16.4 18.3 20.2 22 23.9 25.7 27.6 29.4 31.3 33.2 35 Figure13:Leftcolumn:Theexpectedconcentratedlog-likelihoodfunctionwithestimateddeterministictrends.Rightcolumn:expectedconcentratedtwo-stepapproximatelog-likelihoodfunctionwithestimateddeterministictrends.EachfortheFIVAR0:8processDG-Pdep(28)withd1=0:8,d2=1:8and=0:8butautoregressivedependencewith250observations.26 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d1 -17.3 -14 -14 -10.7 -10.7 -10.7 -7.4 -7.4 -4.1 -4.1 -0.8 -0.8 2.5 2.5 5.8 5.8 9.1 9.1 12.4 12.4 15.7 15.7 19 19 22.3 22.3 25.6 25.6 28.9 32.1 35.4 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 -1.0 -0.5 0.0 0.5 1.0 1.5 b d1 -56.8 -46.9 -46.9 -43.7 -43.7 -37.1 -37.1 -33.8 -30.5 -27.2 -23.9 -20.6 -17.3 -14 -10.7 -7.4 -7.4 -4.1 -4.1 -0.8 -0.8 2.5 2.5 5.8 5.8 9.1 9.1 12.4 12.4 15.7 15.7 19 19 22.3 22.3 25.6 28.9 32.1 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 0.0 0.5 1.0 1.5 2.0 2.5 b d2 11.4 11.4 14.4 14.4 14.4 14.4 14.4 14.4 14.4 14.4 17.5 17.5 17.5 20.5 20.5 23.6 23.6 26.6 26.6 29.6 32.7 35.7 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 0.0 0.5 1.0 1.5 2.0 2.5 b d2 2.3 5.3 5.3 5.3 5.3 8.4 8.4 8.4 8.4 8.4 11.4 11.4 11.4 14.4 14.4 14.4 17.5 17.5 20.5 20.5 23.6 23.6 26.6 29.6 32.7 -1.0 -0.5 0.0 0.5 1.0 1.5 0.0 0.5 1.0 1.5 2.0 2.5 d1 d2 7.1 7.1 7.1 7.1 7.1 9 9 9 9 9 9 10.9 10.9 10.9 10.9 12.7 12.7 14.6 14.6 14.6 14.6 16.4 16.4 18.3 20.2 22 23.9 25.7 27.6 29.4 31.3 33.2 35 36.9 -1.0 -0.5 0.0 0.5 1.0 1.5 0.0 0.5 1.0 1.5 2.0 2.5 d1 d2 -0.3 1.6 1.6 1.6 3.4 3.4 3.4 3.4 3.4 5.3 5.3 5.3 5.3 5.3 7.1 7.1 7.1 7.1 9 9 10.9 10.9 12.7 14.6 16.4 18.3 20.2 22 23.9 25.7 27.6 29.4 31.3 33.2