Sunteți pe pagina 1din 10

5/14/2017 EViewsHelp:PanelUnitRootTesting

UsersGuide:AdvancedUnivariateAnalysis:UnivariateTimeSeriesAnalysis:PanelUnitRootTesting

PanelUnitRootTesting

PerformingPanelUnitRootTestsinEViews
PanelUnitRootDetails
TestswithCommonUnitRootProcess
Levin,Lin,andChu
Breitung
Hadri
TestswithIndividualUnitRootProcesses
Im,Pesaran,andShin
FisherADFandFisherPP
SummaryofAvailablePanelUnitRootTests

Recentliteraturesuggeststhatpanelbasedunitroottestshavehigherpowerthanunitroottests
basedonindividualtimeseries.EViewswillcomputeoneofthefollowingfivetypesofpanelunit
roottests:Levin,LinandChu(2002),Breitung(2000),Im,PesaranandShin(2003),Fishertype
testsusingADFandPPtests(MaddalaandWu(1999)andChoi(2001)),andHadri(2000).

Whilethesetestsarecommonlytermedpanelunitroottests,theoretically,theyaresimply
multipleseriesunitrootteststhathavebeenappliedtopaneldatastructures(wherethepresenceof
crosssectionsgeneratesmultipleseriesoutofasingleseries).Accordingly,EViewssupportsthese
testsinsettingsinvolvingmultipleseries:asaseriesview(iftheworkfileispanelstructured),asa
groupview,orasapoolview.

PerformingPanelUnitRootTestsInEViews
Thefollowingdiscussionassumesthatyouarefamiliarwiththebasicsofbothunitroottestsand
panelunitroottests.

Tobegin,selectView/UnitRootTestfromthemenuofanEViewsgrouporpoolobject,orfrom
themenuofanindividualseriesinapanelstructuredworkfile.HereweshowthedialogforaGroup
unitroottesttheotherdialogsdifferslightly(fortestingusingapoolobject,thereisanadditional
fieldintheupperlefthandportionofthedialogwhereyoumustindicatethenameofthepoolseries
onwhichyouwishtoconductyourtestfortheseriesobjectinapanelworkfile,theUsebalanced
sampleoptionisnotpresent).

http://www.eviews.com/help/helpintro.html#page/content/advtimeserPanel_Unit_Root_Testing.html 1/10
5/14/2017 EViewsHelp:PanelUnitRootTesting

Ifyouwishtoacceptthedefaultsettings,simplyclickonOK.EViewswillusethedefaultSummary
setting,andwillcomputeafullsuiteofunitroottestsonthelevelsoftheseries,alongwitha
summaryoftheresults.

Tocustomizetheunitrootcalculations,youwillchoosefromavarietyofoptions.Theoptionsonthe
lefthandsideofthedialogdeterminethebasicstructureofthetestortests,whiletheoptionson
therighthandsideofthedialogcontroladvancedcomputationaldetailssuchasbandwidthorlag
selectionmethods,orkernelmethods.

Thedropdownmenuatthetopofthedialogiswhereyouwillchoosethetypeoftesttoperform.
Therearesixsettings:Summary,CommonrootLevin,Lin,Chu,CommonrootBreitung,
IndividualrootIm,Pesaran,Shin,IndividualrootFisherADF,IndividualrootFisher
PP,andHadri,correspondingtooneormoreofthetestslistedabove.Thedropdownmenu
labelsincludeabriefdescriptionoftheassumptionsunderwhichthetestsarecomputed.Common
rootindicatesthatthetestsareestimatedassumingacommonARstructureforalloftheseries
IndividualrootisusedfortestswhichallowfordifferentARcoefficientsineachseries.

WehavealreadypointedoutthattheSummarydefaultinstructsEViewstoestimatethefirstfiveof
thetests,whereapplicable,andtoprovideabriefsummaryoftheresults.Selectinganindividual
testtypeallowsyoubettercontroloverthecomputationalmethodandprovidesadditionaldetailon
thetestresults.

Thenexttwosetsofradiobuttonsallowyoutocontrolthespecificationofyourtestequation.First,
youmaychoosetoconducttheunitrootontheLevel,1stdifference,or2nddifferenceofyour
series.Next,youmaychoosebetweensetsofexogenousregressorstobeincluded.Youcanselect
Individualinterceptifyouwishtoincludeindividualfixedeffects,Individualinterceptsand
individualtrendstoincludebothfixedeffectsandtrends,orNonefornoregressors.

TheUsebalancedsampleoptionispresentonlyifyouareestimatingaPooloraGroupunitroot
test.Ifyouselectthisoption,EViewswilladjustyoursamplesothatonlyobservationswhereall
seriesvaluesarenotmissingwillbeincludedinthetestequations.

Dependingontheformofthetestorteststobecomputed,youwillbepresentedwithvarious
advancedoptionsontherightsideofthedialog.Forteststhatinvolveregressionsonlagged
differenceterms(Levin,Lin,andChu,Breitung,Im,Pesaran,andShin,FisherADF)theseoptions
relatetothechoiceofthenumberoflagstobeincluded.Forthetestsinvolvingkernelweighting
(Levin,Lin,andChu,FisherPP,Hadri),theoptionsrelatetothechoiceofbandwidthandkernel
type.

Foragrouporpoolunitroottest,theEViewsdefaultistouseautomaticselectionmethods:
informationmatrixcriterionbasedforthenumberoflagdifferenceterms(withautomaticselection
ofthemaximumlagtoevaluate),andtheAndrewsorNeweyWestmethodforbandwidthselection.
Forunitroottestsonaseriesinapanelworkfile,thedefaultbehaviorusesuserspecifiedoptions.

Ifyouwishtooverridethesesettings,simplyentertheappropriateinformation.Youmay,for
example,selectafixed,userspecifiednumberoflagsbyenteringanumberintheUserspecified
field.Alternatively,youmaycustomizethesettingsforautomaticlagselectionmethod.Alternative
http://www.eviews.com/help/helpintro.html#page/content/advtimeserPanel_Unit_Root_Testing.html 2/10
5/14/2017 EViewsHelp:PanelUnitRootTesting

criteriaforevaluatingtheoptimallaglengthmaybeselectedviathedropdownmenu(Akaike,
Schwarz,HannanQuinn,ModifiedAkaike,ModifiedSchwarz,ModifiedHannanQuinn),andyoumay
limitthenumberoflagstotryinautomaticselectionbyenteringanumberintheMaximumlags
box.Forthekernelbasedmethods,youmayselectakerneltypefromthedropdownmenu
(Bartlett,Parzen,Quadraticspectral),andyoumayspecifyeitheranautomaticbandwidth
selectionmethod(Andrews,NeweyWest)oruserspecifiedfixedbandwidth.

Asanillustration,weperformapanelunitroottestsonrealgrossinvestmentdata(I)intheoft
citedGrunfelddatacontainingdataonR&Dexpenditureandothereconomicmeasuresfor10firms
fortheyears1935to1954foundinGrunfeld_Baltagi.WF1.Wecomputethesummarypanelunit
roottest,usingindividualfixedeffectsasregressors,andautomaticlagdifferencetermand
bandwidthselection(usingtheSchwarzcriterionforthelagdifferences,andtheNeweyWestmethod
andtheBartlettkernelforthebandwidth).Theresultsforthepanelunitroottestarepresented
below:

Thetopoftheoutputindicatesthetypeoftest,exogenousvariablesandtestequationoptions.Ifwe
wereinsteadestimatingaPoolorGrouptest,alistoftheseriesusedinthetestwouldalsobe
depicted.Thelowerpartofthesummaryoutputgivesthemaintestresults,organizedbothbynull
hypothesisaswellasthemaintainedhypothesisconcerningthetypeofunitrootprocess.

Alloftheresultsindicatethepresenceofaunitroot,astheLLC,IPS,andbothFishertestsfailto
rejectthenullofaunitroot.

Ifyouonlywishtocomputeasingleunitroottesttype,orifyouwishtoexaminethetestsresultsin
greaterdetail,youmaysimplyrepeattheunitroottestafterselectingthedesiredtestinTesttype
dropdownmenu.Here,weshowthebottomportionoftheLLCtestspecificoutputforthesamedata:

http://www.eviews.com/help/helpintro.html#page/content/advtimeserPanel_Unit_Root_Testing.html 3/10
5/14/2017 EViewsHelp:PanelUnitRootTesting

Foreachcrosssection,theautoregressioncoefficient,varianceoftheregression,HACofthe
dependentvariable,theselectedlagorder,maximumlag,bandwidthtruncationparameter,andthe
numberofobservationsusedaredisplayed.

PanelUnitRootDetails
Panelunitroottestsaresimilar,butnotidentical,tounitroottestscarriedoutonasingleseries.
Here,webrieflydescribethefivepanelunitroottestscurrentlysupportedinEViewsforadditional
detail,weencourageyoutoconsulttheoriginalliterature.Thediscussionassumesthatyouhavea
basicknowledgeofunitroottheory.

Webeginbyclassifyingourunitroottestsonthebasisofwhethertherearerestrictionsonthe
autoregressiveprocessacrosscrosssectionsorseries.ConsiderafollowingAR(1)processforpanel
data:

(37.45)

where crosssectionunitsorseries,thatareobservedoverperiods

The representtheexogenousvariablesinthemodel,includinganyfixedeffectsorindividual

trends, aretheautoregressivecoefficients,andtheerrors areassumedtobemutually

independentidiosyncraticdisturbance.If , issaidtobeweakly(trend)stationary.Onthe

otherhand,if then containsaunitroot.

Forpurposesoftesting,therearetwonaturalassumptionsthatwecanmakeaboutthe .First,one

canassumethatthepersistenceparametersarecommonacrosscrosssectionssothat for

all .TheLevin,Lin,andChu(LLC),Breitung,andHadritestsallemploythisassumption.

Alternatively,onecanallow tovaryfreelyacrosscrosssections.TheIm,Pesaran,andShin
(IPS),andFisherADFandFisherPPtestsareofthisform.

TestsWithCommonUnitRootProcess

Levin,Lin,andChu(LLC),Breitung,andHadritestsallassumethatthereisacommonunitroot
processsothat isidenticalacrosscrosssections.Thefirsttwotestsemployanullhypothesisof
aunitrootwhiletheHadritestusesanullofnounitroot.

LLCandBreitungbothconsiderthefollowingbasicADFspecification:
http://www.eviews.com/help/helpintro.html#page/content/advtimeserPanel_Unit_Root_Testing.html 4/10
5/14/2017 EViewsHelp:PanelUnitRootTesting

(37.46)

whereweassumeacommon ,butallowthelagorderforthedifferenceterms, ,to


varyacrosscrosssections.Thenullandalternativehypothesesforthetestsmaybewrittenas:

(37.47)

(37.48)

Underthenullhypothesis,thereisaunitroot,whileunderthealternative,thereisnounitroot.

Levin,Lin,AndChu

ThemethoddescribedinLLCderivesestimatesof fromproxiesfor and thatare


standardizedandfreeofautocorrelationsanddeterministiccomponents.

Foragivensetoflagorders,webeginbyestimatingtwoadditionalsetsofequations,regressing
both ,and onthelagterms (for )andtheexogenousvariables

.Theestimatedcoefficientsfromthesetworegressionswillbedenoted and ,
respectively.

Wedefine bytaking andremovingtheautocorrelationsanddeterministiccomponents


usingthefirstsetofauxiliaryestimates:

(37.49)

Likewise,wemaydefinetheanalogous usingthesecondsetofcoefficients:

(37.50)

Next,weobtainourproxiesbystandardizingboth and ,dividingbytheregression


standarderror:

(37.51)

where aretheestimatedstandarderrorsfromestimatingeachADFinEquation(37.46).

Lastly,anestimateofthecoefficient maybeobtainedfromthepooledproxyequation:

(37.52)

http://www.eviews.com/help/helpintro.html#page/content/advtimeserPanel_Unit_Root_Testing.html 5/10
5/14/2017 EViewsHelp:PanelUnitRootTesting

LLCshowthatunderthenull,amodifiedtstatisticfortheresulting isasymptoticallynormally
distributed

(37.53)

where isthestandardtstatisticfor , istheestimatedvarianceoftheerrorterm ,

isthestandarderrorof ,and:

(37.54)

Theremainingterms,whichinvolvecomplicatedmomentcalculations,aredescribedingreaterdetail
inLLC.Theaveragestandarddeviationratio, ,isdefinedasthemeanoftheratiosofthelong
runstandarddeviationtotheinnovationstandarddeviationforeachindividual.Itsestimateis
derivedusingkernelbasedtechniques.Theremainingtwoterms, and areadjustment
termsforthemeanandstandarddeviation.

TheLLCmethodrequiresaspecificationofthenumberoflagsusedineachcrosssectionADF
regression, ,aswellaskernelchoicesusedinthecomputationof .Inaddition,youmust
specifytheexogenousvariablesusedinthetestequations.Youmayelecttoincludenoexogenous
regressors,ortoincludeindividualconstantterms(fixedeffects),ortoemployindividualconstants
andtrends.

Breitung

TheBreitungmethoddiffersfromLLCintwodistinctways.First,onlytheautoregressiveportion
(andnottheexogenouscomponents)isremovedwhenconstructingthestandardizedproxies:

(37.55)

where , ,and areasdefinedforLLC.

Second,theproxiesaretransformedanddetrended,

(37.56)

Thepersistenceparameter isestimatedfromthepooledproxyequation:

http://www.eviews.com/help/helpintro.html#page/content/advtimeserPanel_Unit_Root_Testing.html 6/10
5/14/2017 EViewsHelp:PanelUnitRootTesting

(37.57)

Breitungshowsthatunderthenull,theresultingestimator isasymptoticallydistributedasa
standardnormal.

TheBreitungmethodrequiresonlyaspecificationofthenumberoflagsusedineachcrosssection
ADFregression, ,andtheexogenousregressors.NotethatincontrastwithLLC,nokernel
computationsarerequired.

Hadri

TheHadripanelunitroottestissimilartotheKPSSunitroottest,andhasanullhypothesisofno
unitrootinanyoftheseriesinthepanel.LiketheKPSStest,theHadritestisbasedontheresiduals
fromtheindividualOLSregressionsof onaconstant,oronaconstantandatrend.Forexample,
ifweincludeboththeconstantandatrend,wederiveestimatesfrom:

(37.58)

Giventheresiduals fromtheindividualregressions,weformtheLMstatistic:

(37.59)

where arethecumulativesumsoftheresiduals,

(37.60)

and istheaverageoftheindividualestimatorsoftheresidualspectrumatfrequencyzero:

(37.61)

EViewsprovidesseveralmethodsforestimatingthe .SeeUnitRootTestingforadditional
details.

AnalternativeformoftheLMstatisticallowsforheteroskedasticityacross :

(37.62)

Hadrishowsthatundermildassumptions,

(37.63)

http://www.eviews.com/help/helpintro.html#page/content/advtimeserPanel_Unit_Root_Testing.html 7/10
5/14/2017 EViewsHelp:PanelUnitRootTesting

where and ,ifthemodelonlyincludesconstants( issetto0forall ),and


and ,otherwise.

TheHadripanelunitroottestsrequireonlythespecificationoftheformoftheOLSregressions:
whethertoincludeonlyindividualspecificconstantterms,orwhethertoincludebothconstantand
trendterms.EViewsreportstwo statisticvalues,onebasedon withtheassociated

homoskedasticityassumption,andtheotherusing thatisheteroskedasticityconsistent.

Itisworthnotingthatsimulationevidencesuggeststhatinvarioussettings(forexample,small ),
Hadri'spanelunitroottestexperiencessignificantsizedistortioninthepresenceofautocorrelation
whenthereisnounitroot.Inparticular,theHadritestappearstooverrejectthenullofstationarity,
andmayyieldresultsthatdirectlycontradictthoseobtainedusingalternativeteststatistics(see
HlouskovaandWagner(2006)fordiscussionanddetails).

TestsWithIndividualUnitRootProcesses

TheIm,Pesaran,andShin,andtheFisherADFandPPtestsallallowforindividualunitroot
processessothat mayvaryacrosscrosssections.Thetestsareallcharacterizedbythe
combiningofindividualunitrootteststoderiveapanelspecificresult.

Im,Pesaran,AndShin

Im,Pesaran,andShinbeginbyspecifyingaseparateADFregressionforeachcrosssection:

(37.64)

Thenullhypothesismaybewrittenas,

(37.65)

whilethealternativehypothesisisgivenby:

(37.66)

(wherethe maybereorderedasnecessary)whichmaybeinterpretedasanonzerofractionof
theindividualprocessesisstationary.

AfterestimatingtheseparateADFregressions,theaverageofthetstatisticsfor fromthe

individualADFregressions, :

(37.67)

isthenadjustedtoarriveatthedesiredteststatistics.
http://www.eviews.com/help/helpintro.html#page/content/advtimeserPanel_Unit_Root_Testing.html 8/10
5/14/2017 EViewsHelp:PanelUnitRootTesting

Inthecasewherethelagorderisalwayszero( forall ),simulatedcriticalvaluesfor

areprovidedintheIPSpaperfordifferentnumbersofcrosssections ,serieslengths ,andfor


testequationscontainingeitherintercepts,orinterceptsandlineartrends.EViewsusesthesevalues,
orlinearlyinterpolatedvalues,inevaluatingthesignificanceoftheteststatistics.

InthegeneralcasewherethelagorderinEquation(37.64)maybenonzeroforsomecross
sections,IPSshowthataproperlystandardized hasanasymptoticstandardnormaldistribution:

(37.68)

TheexpressionsfortheexpectedmeanandvarianceoftheADFregressiontstatistics,

and ,areprovidedbyIPSforvariousvaluesof and anddifferingtestequation


assumptions,andarenotprovidedhere.

TheIPSteststatisticrequiresspecificationofthenumberoflagsandthespecificationofthe
deterministiccomponentforeachcrosssectionADFequation.Youmaychoosetoincludeindividual
constants,ortoincludeindividualconstantandtrendterms.

FisherADFAndFisherPP

AnalternativeapproachtopanelunitroottestsusesFishers(1932)resultstoderiveteststhat
combinethepvaluesfromindividualunitroottests.ThisideahasbeenproposedbyMaddalaand
Wu,andbyChoi.

Ifwedefine asthepvaluefromanyindividualunitroottestforcrosssection ,thenunderthe

nullofunitrootforall crosssections,wehavetheasymptoticresultthat

(37.69)

Inaddition,Choidemonstratesthat:

(37.70)

where istheinverseofthestandardnormalcumulativedistributionfunction.

EViewsreportsboththeasymptotic andstandardnormalstatisticsusingADFandPhillipsPerron
individualunitroottests.ThenullandalternativehypothesesarethesameasfortheasIPS.

ForbothFishertests,youmustspecifytheexogenousvariablesforthetestequations.Youmayelect
toincludenoexogenousregressors,toincludeindividualconstants(effects),orincludeindividual
http://www.eviews.com/help/helpintro.html#page/content/advtimeserPanel_Unit_Root_Testing.html 9/10
5/14/2017 EViewsHelp:PanelUnitRootTesting

constantandtrendterms.

Additionally,whentheFishertestsarebasedonADFteststatistics,youmustspecifythenumberof
lagsusedineachcrosssectionADFregression.ForthePPformofthetest,youmustinsteadspecify
amethodforestimating .EViewssupportsestimatorsfor basedonkernelbasedsumof
covariances.SeeFrequencyZeroSpectrumEstimationfordetails.

SummaryOfAvailablePanelUnitRootTests
Thefollowingtablesummarizesthebasiccharacteristicsofthepanelunitroottestsavailablein
EViews:

Test Null Alternative Possible Autocorrela


Deterministi tion
c Correction
Component Method
Levin,Lin Unitroot NoUnitRoot None,F,T Lags
andChu
Breitung Unitroot NoUnitRoot None,F,T Lags
IPS UnitRoot Somecross F,T Lags
sections
withoutUR
FisherADF UnitRoot Somecross None,F,T Lags
sections
withoutUR
FisherPP UnitRoot Somecross None,F,T Kernel
sections
withoutUR
Hadri NoUnit UnitRoot F,T Kernel
Root

NonenoexogenousvariablesFfixedeffectandTindividualeffectandindividualtrend.

Lastupdated:Mon,27Feb201720:50:08PST

Backtotop

http://www.eviews.com/help/helpintro.html#page/content/advtimeserPanel_Unit_Root_Testing.html 10/10

S-ar putea să vă placă și