5/6/2017 repec.org/bocode/x/xtabond2.
html
helpforxtabond2
"Difference"and"system"GMMdynamicpanelestimator
xtabond2depvarvarlist[ifexp][inrange][weight][,level(#)
svmattwosteprobustcluster(varname)noconstantsmall
noleveleqorthogonalgmmopt[gmmopt...]ivopt[ivopt...]
pcacomponents(#)artests(#)arlevelsh(#)nodiffsargan
nomata]
wheregmmoptis
gmmstyle(varlist[,laglimits(##)collapseorthogonalequation({diff
|level|both})passthrusplit])
andivoptis
ivstyle(varlist[,equation({diff|level|both})passthrumz])
aweights,pweights,andfweightsareallowed.fweightsmustbeconstant
overtime.Seehelpweights.
xtabond2isforusewithcrosssectiontimeseriesdata.Youmusttsset
yourdatabeforeusingxtabond2;seehelptsset.
Allvarlistsmaycontaintimeseriesoperatorsand,inStataversion11
orlater,factorvariables.Seehelpvarlist.
by...:maybeusedwithxtabond2ifnotimeseriesoperatorsareused
inthecommandline.Thebyclausewillnotrestrictthesamplefrom
whichlagsaredrawninbuildinginstruments.Seehelpby.
xtabond2sharesfeaturesofallestimationcommands;seehelpestcom.
Thesyntaxofpredictfollowingxtabond2is
predict[type]newvarname[ifexp][inrange][,statistic]
[difference]
wherestatisticis
xbbx_it,fittedvalues(thedefault)
residualse_it,theresiduals
Description
http://repec.org/bocode/x/xtabond2.html 1/13
5/6/2017 repec.org/bocode/x/xtabond2.html
xtabond2canfittwocloselyrelateddynamicpaneldatamodels.The
firstistheArellanoBond(1991)estimator,whichisalsoavailablewith
xtabond,thoughwithoutthetwostepstandarderrorcorrectiondescribed
below.Itissometimescalled"differenceGMM."Thesecondisan
augmentedversionoutlinedbyArellanoandBover(1995)andfully
developedbyBlundellandBond(1998).Itisknownas"systemGMM."
Roodman(2009)providesapedagogicintroductiontolinearGMM,these
estimators,andxtabond2.Theestimatorsaredesignedfordynamic
"smallT,largeN"panelsthatmaycontainfixedeffectsandseparate
fromthosefixedeffectsidiosyncraticerrorsthatareheteroskedastic
andcorrelatedwithinbutnotacrossindividuals.Considerthemodel:
y_it=x_it*b_1+w_it*b_2+u_iti=1,...,N;t=1,...,T
u_it=v_i+e_it,
where
v_iareunobservedindividualleveleffects;
e_itaretheobservationspecificerrors;
x_itisavectorofstrictlyexogenouscovariates(onesdependenton
neithercurrentnorpaste_it);
w_itisavectorofpredeterminedcovariates(whichmayincludethelag
ofy)andendogenouscovariates,allofwhichmaybecorrelated
withthev_i(Predeterminedvariablesarepotentiallycorrelated
withpasterrors.Endogenousonesarepotentiallycorrelated
withpastandpresenterrors.);
b_1andb_2arevectorsofparameterstobeestimated;
andE[v_i]=E[e_it]=E[v_i*e_it]=0,andE[e_it*e_js]=0foreachi,j,t,s,
i<>j.
Firstdifferencingtheequationremovesthev_i,thuseliminatinga
potentialsourceofomittedvariablebiasinestimation.However,
differencingvariablesthatarepredeterminedbutnotstrictlyexogenous
makesthemendogenoussincethew_itinsomeD.w_it=w_itw_i,t1is
correlatedwiththee_i,t1inD.e_it.FollowingHoltEakin,Newey,and
Rosen(1988),ArellanoandBond(1991)developaGeneralizedMethodof
Momentsestimatorthatinstrumentsthedifferencedvariablesthatarenot
strictlyexogenouswithalltheiravailablelagsinlevels.(Strictly
exogenousvariablesareuncorrelatedwithcurrentandpasterrors.)
ArellanoandBondalsodevelopanappropriatetestforautocorrelation,
which,ifpresent,canrendersomelagsinvalidasinstruments.
AproblemwiththeoriginalArellanoBondestimatoristhatlaggedlevels
arepoorinstrumentsforfirstdifferencesifthevariablesarecloseto
arandomwalk.ArellanoandBover(1995)describehow,iftheoriginal
equationinlevelsisaddedtothesystem,additionalinstrumentscanbe
broughttobeartoincreaseefficiency.Inthisequation,variablesin
levelsareinstrumentedwithsuitablelagsoftheirownfirst
differences.Theassumptionneededisthatthesedifferencesare
uncorrelatedwiththeunobservedcountryeffects.BlundellandBondshow
http://repec.org/bocode/x/xtabond2.html 2/13
5/6/2017 repec.org/bocode/x/xtabond2.html
thatthisassumptioninturndependsonamorepreciseoneaboutinitial
conditions.
xtabond2implementsbothestimatorstwice.TheversioninStatasado
programminglanguageisslowbutcompatiblewithStata7and8.TheMata
versionisusuallyfaster,andrunsinStata10.0orlater.Thextabond2
optionnomatapreventstheuseofMataevenwhenitisavailable.
TheMataversionalsoincludestheoptiontousetheforwardorthogonal
deviationstransforminsteadoffirstdifferencing.ProposedbyArellano
andBover(1995)theorthogonaldeviationstransform,ratherthan
subtractingthepreviousobservation,subtractstheaverageofall
availablefutureobservations.Theresultisthenmultipliedbyascale
factorchosentoyieldthenicebutrelativelyunimportantpropertythat
iftheoriginale_itarei.i.d.,thensoarethetransformedones(see
ArellanoandBover(1995)andRoodman(2009)).Likedifferencing,taking
orthogonaldeviationsremovesfixedeffects.Becauselaggedobservations
ofavariabledonotentertheformulaforthetransformation,they
remainorthogonaltothetransformederrors(assumingnoserial
correlation),andavailableasinstruments.Infact,forconsistency,
thesoftwarestorestheorthogonaldeviationofanobservationoneperiod
late,sothat,aswithdifferencing,observationsforperiod1are
missingand,foraninstrumentingvariablew,w_i,t1enterstheformula
forthetransformedobservationstoredati,t.Withthismove,exactly
thesamelagsofvariablesarevalidasinstrumentsunderthetwo
transformations.
Onbalancedpanels,GMMestimatorsbasedonthetwotransformsreturn
numericallyidenticalcoefficientestimates,holdingtheinstrumentset
fixed(ArellanoandBover1995).Butorthogonaldeviationshasthe
virtueofpreservingsamplesizeinpanelswithgaps.Ifsomee_itis
missing,forexample,neitherD.e_itnorD.e_i,t+1canbecomputed.But
theorthogonaldeviationcanbecomputedforeverycompleteobservation
exceptthelastforeachindividual.(Firstdifferencingcandono
bettersinceitmustdropthefirstobservationforeachindividual.)
Notethat"differenceGMM"isstillcalledthatevenwhenorthogonal
deviationsareused.Wewillrefertotheequationindifferencesor
orthogonaldeviationsasthetransformedequation.InsystemGMMwith
orthogonaldeviations,thelevelsoruntransformedequationisstill
instrumentedwithdifferencesasdescribedabove.
xtabond2reportstheArellanoBondtestforautocorrelation,whichis
appliedtothedifferencedresidualsinordertopurgetheunobservedand
perfectlyautocorrelatedv_i.AR(1)isexpectedinfirstdifferences,
becauseD.e_i,t=e_i,te_i,t1shouldcorrelatewithD.e_i,t1=
e_i,t1e_i,t2sincetheysharethee_i,t1term.Sotocheckfor
AR(1)inlevels,lookforAR(2)indifferences,ontheideathatthis
willdetecttherelationshipbetweenthee_i,t1inD.e_i,tandthe
e_i,t2inD.e_i,t2.Thisreasoningdoesnotworkfororthogonal
deviations,inwhichtheresidualsforanindividualareall
mathematicallyinterrelated,thuscontaminatedfromthepointofviewof
detectingARinthee_it.Sothetestisrunondifferencedresiduals
evenafterestimationindeviations.Autocorrelationindicatesthatlags
ofthedependentvariable(andanyothervariablesusedasinstruments
thatarenotstrictlyexogenous),areinfactendogenous,thusbad
instruments.Forexample,ifthereisAR(s),theny_i,tswouldbe
correlatedwithe_i,ts,whichwouldbecorrelatedwithD.e_i,ts,which
wouldbecorrelatedwithD.e_i,t.
xtabond2alsoreportstestsofoveridentifyingrestrictionsofwhether
http://repec.org/bocode/x/xtabond2.html 3/13
5/6/2017 repec.org/bocode/x/xtabond2.html
theinstruments,asagroup,appearexogenous.Foronestep,nonrobust
estimation,itreportstheSarganstatistic,whichistheminimizedvalue
oftheonestepGMMcriterionfunction.TheSarganstatisticisnot
robusttoheteroskedasticityorautocorellation.Soforonestep,robust
estimation(andforalltwostepestimation),xtabond2alsoreportsthe
HansenJstatistic,whichistheminimizedvalueofthetwostepGMM
criterionfunction,andisrobust.xtabond2stillreportstheSargan
statisticinthesecasesbecausetheJtesthasitsownproblem:itcan
begreatlyweakenedbyinstrumentproliferation.TheMataversiongoes
further,reportingdifferenceinSarganstatistics(really,
differenceinHansenstatistics,exceptinonesteprobustestimation),
whichtestforwhethersubsetsofinstrumentsarevalid.Tobeprecise,
itreportsonetestforeachgroupofinstrumentsdefinedbyanivstyle()
orgmmstyle()option(explainedbelow).Soreplacinggmmstyle(xy)ina
commandlinewithgmmstyle(x)gmmstyle(y)willyieldthesameestimate
butdistinctdifferenceinSargan/Hansentests.Inaddition,including
thesplitsuboptioninagmmstyle()optioninsystemGMMsplitsan
instrumentgroupintwofordifferenceinSargan/Hansenpurposes,one
eachforthetransformedequationandlevelsequations.Thisis
especiallyusefulfortestingtheinstrumentsforthelevelsequation
basedonlaggeddifferencesofthedependentvariable,whicharethemost
suspectinsystemGMMandthesubjectofthe"initialconditions"inthe
titleofBlundellandBond(1998).Inthesamevein,insystemGMM,
xtabond2alsotestsalltheGMMtypeinstrumentsforthelevelsequation
asagroup.Allofthesetests,however,areweakwhentheinstrument
countishigh.DifferenceinSargan/Hansentestsarearecomputationally
intensivesincetheyinvolvereestimatingthemodelforeachtest;the
nodiffsarganoptionisavailabletopreventthem.
AslinearGMMestimators,theArellanoBondandBlundellBondestimators
haveoneandtwostepvariants.Butthoughtwostepisasymptotically
moreefficient,thereportedtwostepstandarderrorstendtobeseverely
downwardbiased(ArellanoandBond1991;BlundellandBond1998).To
compensate,xtabond2makesavailableafinitesamplecorrectiontothe
twostepcovariancematrixderivedbyWindmeijer(2005).Thiscanmake
twosteprobustestimationsmoreefficientthanonesteprobust,
especiallyforsystemGMM.
Standarderrorscanalsobe"bootstrapped"butnotwiththebootstrap
command.Thatcommandbuildstemporarydatasetsbysamplingtherealone
withreplacement.Andhavingmultipleobservationsforagiven
observationalunitandtimeperiodviolatespanelstructure.Instead,use
jacknife,perhapswiththecluster()option,clusteringonthepanel
identifiervariable,inordertodropeachobservationalunitinturn.
Thesyntaxofxtabond2differssubstantiallyfromthatofxtabondand
xtdpdsys.xtabond2almostcompletelydecouplesspecificationof
regressorsfromspecificationofinstruments.Asaresult,most
variablesusedwillappeartwiceinanxtabond2commandline.xtabond2
requirestheinitialvarlistofthecommandlinetoincludeall
regressorsexceptfortheoptionalconstantterm,betheystrictly
exogenous,predetermined,orendogenous.Variablesusedtoform
instrumentsthenappearingmmstyle()orivstyle()optionsafterthe
comma.Theresultisalossofparsimony,butfullercontroloverthe
instrumentmatrix.Variablescanbeusedasthebasisfor"GMMstyle"
instrumentsetswithoutbeingincludedasregressors,orviceversa.
Thegmmstyle()andivstyle()optionsalsohavesuboptionsthatallow
furthercustomizationoftheinstrumentmatrix.
http://repec.org/bocode/x/xtabond2.html 4/13
5/6/2017 repec.org/bocode/x/xtabond2.html
Citation
xtabond2isnotanofficialStatacommand.Itisafreecontributionto
theresearchcommunity.Pleaseciteitassuch:
Roodman,D.2009.Howtodoxtabond2:Anintroductiontodifference
andsystemGMMinStata.StataJournal9(1):86136.
Options
level(#)specifiestheconfidencelevel,inpercent,forconfidence
intervalsofthecoefficients;seehelplevel.Thedefaultis95.
svmattellsxtabond2tosavetheX,Y,Z,H,andweightmatricesase()
returnmacros.Thesearenotincludedbydefaultbecausethematrices
canbelargerthanthedatasetitself.Ifthepcaoptionisused,
svmatwillalsosavetheeigenvectorsmatrixas
xtabond2_eigenvectors.Thisoptionisavailableonlywhenusingusing
theMataimplementationinMata'sspeedfavoringmode.Dataare
storedinbalancedmatricesandsortedbyindividual,equation(for
SystemGMM),thentime.Rowsandcolumnsarelabelledforclarity.
Theinstrumentmatrixtypicallycontainsallzerocolumns,whichdo
notaffectestimation.ForcompatibilitywithStatacolumnlabeling
conventions,instrumentssubjecttothebackwardorthogonal
deviationstransform(seebelow)arestilldenotedwitha"D."
operator.
twostepspecifiesthatthetwostepestimatoristobecalculatedinstead
oftheonestep.
robust:Foronestepestimation,robustspecifiesthattherobust
estimatorofthecovariancematrixoftheparameterestimatesbe
calculated.Theresultingstandarderrorestimatesareconsistentin
thepresenceofanypatternofheteroskedasticityandautocorrelation
withinpanels.Intwostepestimation,thestandardcovariance
matrixisalreadyrobustintheorybuttypicallyyieldsstandard
errorsthataredownwardbiased.twosteprobustrequests
Windmeijersfinitesamplecorrectionforthetwostepcovariance
matrix.
cluster(varname)overridesthedefaultuseofthepanelidentifier(as
setbytsset)asthebasisfordefininggroups.cluster(varname)
impliesrobustinthesensesjustdescribed.Forexample,intwostep
estimation,itrequeststheWindmeijercorrection.Changingthegroup
identifierwiththisoptionaffectsonestep"robust"standard
errors,alltwostepresults,theHansenanddifferenceinHansen
tests,andtheArellanoBondserialcorrelationtests.
noconstantsuppressestheconstantterminthelevelsequation.By
default,thetermisincludedasaregressorandIVstyleinstrument.
UnlikextabondandDPD(theoriginalimplementationofthese
estimators),xtabond2doesnotincludetheconstantterminthe
transformedequationindifferenceGMM.Rather,theconstantis
transformedout.
smallrequestststatisticsinsteadofzstatisticsandanFtestinstead
ofaWaldchisquaredtestofoverallmodelfit.
http://repec.org/bocode/x/xtabond2.html 5/13
5/6/2017 repec.org/bocode/x/xtabond2.html
noleveleqspecifiesthatlevelequationshouldbeexcludedfromthe
estimation,yieldingdifferenceratherthansystemGMM.
nodiffsarganpreventsdifferenceinSargan/Hansentests,whichareare
computationallyintensivesincetheyinvolvereestimatingthemodel
foreachtest.Theoptionhasnoeffectontheadoversionof
xtabond2,whichdoesnotperformdifferenceinSargan/Hansentesting
anyway.
nomatapreventstheuseofMatacodeevenwhenthelanguageisavailable
(inStata10.0orlater).ItisnotnecessaryinStata79.
Ordinarilythisswitchdoesnotaffectresults.However,ifsome
variablesarecollinearornearlyso,thetwoversionsoftheprogram
maydroppeddifferentones,whichcanaffecttheresults.Theycan
evendifferinhowmanytheydrop,sincetheversionsusedifferent
routinesandtolerancesfordeterminingcollinearity.Inaddition,
theMataversiondoesnotperfectlyhandlestrangeandunusual
expressionslikegmm(L.x,lag(11)).(Documentationforthe
gmmstyle()optionisbelow.)Thisexpressionisthesameasgmm(x,
lag(00))inprinciple.ButtheMatacodewouldinterpretitby
laggingx,thuslosingtheobservationsofxfort=T,thenunlagging
theremaininginformation.Theslow,adoversionwouldnotlosedata
inthisway.
orthogonalrequeststheforwardorthogonaldeviationstransforminstead
ofdifferencing.
ivstyle()specifiesasetofvariablestoserveasstandardinstruments,
withonecolumnintheinstrumentmatrixpervariable.Normally,
strictlyexogenousregressorsareincludedinivstyleoptions,in
ordertoentertheinstrumentmatrix,aswellasbeinglistedbefore
themaincommaofthecommandline.Theequation()suboption
specifieswhichequation(s)shouldusetheinstruments:
firstdifferenceonly(equation(diff)),levelsonly
(equation(level)),orboth(equation(both)),thedefault.Alsoby
default,theinstrumentsaretransformed(intodifferencesor
orthogonaldeviations)foruseinthetransformedequationand
entereduntransformedforthelevelsequation.Thesuboption
passthrumaybeusedafterequation(diff),orwhentheoption
noleveleqisinvoked,topreventthistransformation.equation()is
usefulforproperhandlingofpredeterminedvariablesusedas
IVstyleinstrumentsinsystemGMM.Forexample,ifxis
predetermined,itisavalidinstrumentforthelevelsequationsince
itisassumedtobeuncorrelatedwiththecontemporaneouserrorterm.
However,xbecomesendogenousinfirstdifferences,soD.xisnota
validinstrumentforthetransformedequation.ivstyle(x)would
thereforebeinappropriate.TheuseofxasanIVstyleinstrument
inlevelsonlycouldbespecifiedbyiv(x,eq(level)).
Ifthesuboptionmzisincludedinanivstyleoption,missingvalues
intheinstrumentsareconvertedtozeroes.mzdoesnotchangethe
precisemomentconditionsgeneratedbyivstyletheystillapplyonly
totheerrortermsofobservationswhichhavedataforthe
instruments.Rather,mzallowsobservationsthataremissingdata
fortheinstrumentsinquestiontononethelessstayintheregression
iftheinstrumentsarenotalsoregressors.(Observationsmissing
valuesforregressorsmuststillbedropped.)
gmmstyle()specifiesasetofvariablestobeusedasbasesfor
http://repec.org/bocode/x/xtabond2.html 6/13
5/6/2017 repec.org/bocode/x/xtabond2.html
"GMMstyle"instrumentsetsdescribedinHoltzEakin,Newey,and
Rosen(1988)andArellanoandBond(1991).Bydefaultxtabond2uses,
foreachtimeperiod,allavailablelagsofthespecifiedvariables
inlevelsdatedt1orearlierasinstrumentsforthetransformed
equation;andusesthecontemporaneousfirstdifferencesas
instrumentsinthelevelsequation.Thesedefaultsareappropriate
forpredeterminedvariablesthatarenotstrictlyexogenous(Bond
2000).Missingvaluesarealwaysreplacedbyzeros.Theoptional
laglimits(ab)suboptioncanoverridethesedefaults:forthe
transformedequation,laggedlevelsdatedtatotbareusedas
instruments,whileforthelevelsequation,thefirstdifference
datedta+1isnormallyused.aandbcaneachbemissing(".");a
defaultsto1andbtoinfinity.Theycanevenbenegative,implying
"forward"lags.Ifa>bthenxtabond2swapstheirvalues.(Notethat
ifa<=b<0thenthefirstdifferencedatedtb+1isnormallyusedas
aninstrumentinthelevelsequationinsteadofthatdatedta+1,
becauseitismorefrequentlyintherange[1,T]ofvalidtime
indexes.Or,forthesamereasons,ifa<=0<=borb<=0<=a,the
firstdifferencedatedtisused.)Sincethegmmstyle()varlist
allowstimeseriesoperators,therearemanyroutestothesame
specification.E.g.,gmm(w,lag(2.)),thestandardtreatmentforan
endogenousvariable,isequivalenttogmm(L.w,lag(1.)),thus
gmm(L.w).
Theequation()suboptionofgmmstyle()worksmuchlikethatof
ivstyle()(seeabove),withoneimportantexception.Inresponseto
equation(level),xtabond2generatesthefullsetofavailable
instrumentsforthelevelsequationsinceitisnolongerthecase
thatmostaremademathematicallyredundantbythepresenceofthe
fullsetofmomentconditionsforthetransformedequation.Tobe
precise,ifthelaglimitsareaandb,thenlagsofthespecified
variablesindifferencesdatedtbtotaareused.equation(diff)
hasnoeffectindifferenceGMM.
Thepassthrusuboptionofgmmstyle()ismeaningfulonlyinsystem
GMM,andonlyforvariablesforwhichequation(level)hasalsobeen
specified.Itdirectsxtabond2tocreateinstrumentsforthelevels
equationthatusenotthefirstdifferencesofthespecified
variablesbuttheoriginallevelsofthesamedates.Forexample,
equation(level)passthrulaglimits(1.)requeststhatalllagged
levelsbeusedasinstruments.Underthestandardassumptions,these
instrumentsarenotvalid.
Theorthogonalsuboptiontellsxtabond2toapplythebackward
orthogonaldeviationstransformtotheinstrumentsforthe
transformedequation.Essentially,instrumentsarereplacedwith
theirdeviationsfrompastmeans.Sincetheresultinginstruments
dependonallpastvaluesoftheunderlyingvariables,theregressors
inthetransformedequationshouldnotbesimilarlytransformed.
Otherwisetheinstrumentsmaybecorrelatedwiththeerror.Thatis,
ifthissuboptionisusedtheorthogonaloptionshouldalsobe
included(outsideagmmstyle()option).Insimulations,Hayakawa
(2009)findsthat"DifferenceGMM"withthiscombinationbackword
orthogonaldeviationsfortheinsturmentsandforwardforthe
regressorsislessbiasedandmorestablethantraditional
DifferenceGMMforastandardAR(1)modelwhenT>=10.(ForanAR(p)
model,heusesonlythemostrecentpinstrumentlags,equivalentto
gmm(L.y,orthoglag(1p)).)Thisoptiondoesnotaffecthe
instrumentsforthelevelsequation.
Thesplitsuboptionofgmmstyle()isalsomeaningfulonlyinsystem
http://repec.org/bocode/x/xtabond2.html 7/13
5/6/2017 repec.org/bocode/x/xtabond2.html
GMM,andthenonlywhenneithereq(diff)noreq(level)isspecified.
Itssoleeffectistosplitthespecifiedinstrumentgroupintwofor
purposesofdifferenceinSargan/Hansentestingoneinstrumentset
forthetransformedequationandoneforthelevelsequation.
Thecollapsesuboptionofgmmstyle()specifiesthatxtabond2should
createoneinstrumentforeachvariableandlagdistance,ratherthan
oneforeachtimeperiod,variable,andlagdistance.Inlarge
samples,collapsereducesstatisticalefficiency.Butinsmall
samplesitcanavoidthebiasthatarisesasthenumberof
instrumentsclimbstowardthenumberofobservations.(When
instrumentsaremany,theytendtooverfittheinstrumentedvariables
andbiastheresultstowardthoseofOLS/GLS.)collapsealsogreatly
curtailscomputationaldemandsbyreducingthewidthofthe
instrumentmatrix,and(relevantfortheadoversionoftheprogram)
helpskeepthematrixwithinStata'ssizelimit.
Forexample,ifamodelassumesthatE[w_is*D.e_it]=0foralls<t,
thisisexpressedinstandardArellanoBondestimationas:
sum_i(w_is*D.e_it)=0foreachsandt,s<t.
Thistranslatesintocolumnsintheinstrumentmatrixoftheform:
w_i100000...
0w_i1w_i2000...
000w_i1w_i2w_i3...
.........
.........
collapsedividesthe"GMMstyle"momentconditionsintogroupsand
sumstheconditionsineachgrouptoformasmallersetofconditions
oftheform:
sum_i,t(w_i,tj*D.e_it)=0foreachj>0.
Thisisequivalenttocombiningcolumnsoftheinstrumentmatrixby
addition,yielding:
w_i100...
w_i2w_i10...
w_i3w_i2w_i1...
......
......
Similarly,thestandardinstrumentsforthelevelsequation(in
systemGMM)collapsefrom:
D.w_i200...
0D.w_i30...
00D.w_i4...
......
Tothesinglecolumn:
http://repec.org/bocode/x/xtabond2.html 8/13
5/6/2017 repec.org/bocode/x/xtabond2.html
D.w_i2
D.w_i3
D.w_i4
.
.
pcatellsxtabond2toreplacethe"GMMstyle"instrumentswiththeir
principalcomponentsinordertoreducetheinstrumentcountina
minimallyarbitraryway(KapetaniosandMarcellino2010;BaiandNg
2010;Mehrhoff2009).Principalcomponentsanalysisisrunonthe
correlation,notcovariance,matrixofthe"GMMstyle"instruments.
Bydefaultxtabond2willselectallcomponentswitheigenvaluesat
least1,andwillselectmoreifnecessarytoguaranteethat
instrumentsareatleastasnumerousasregressors,favoringthose
withlargesteigenvalues.
components(#)allowstheusertooverridethedefaultnumberof
componentsdescribedjustabove.
artests(#)specifiesthemaximumorderoftheautocorrelationteststobe
reported.Thedefaultis2.
arlevelsspecifiesthattheautocorrelationtestsshouldbeappliedto
theresidualsfromthelevels,notfirstdifference,equation.It
cannotbespecifiedalongwithnoleveleq.Iftherearefixed
effects,thenautocorrelationinlevelsisexpectedandwouldnot
callthespecificationintothequestion.
h(#)controlstheformofH,theaprioriestimateofthecovariance
matrixoftheidiosyncraticerrors.InonesteplinearGMM,the
inverseofZ'HZ,whereZistheinstrumentmatrix,proxiesforthe
covariancematrixofthemoments,andisusedtoweightthesample
momentswhosemagnitudesarejointlyminimized.SinceHmerely
controlstheweightsoninstrumentsbelievedexogenous,forany
nondegeneratechoiceofH,onestepestimateswillbeconsistent.
Andtwostepestimateswillbeasymptoticallyefficient(Baum,
Schaffer,andStillman2003).SothepriorityindesigningHis
minimizingarbitrariness.Halwayshasblockdiagonalform,withall
blocksthesame.Let*indicatevariablestransformedbyorthogonal
deviationsordifferencingandMbethe(T1)xTmatrixthatperforms
thechosentransform.WeassumeforthepurposesofdesigningHthat
var[e]=I,theidentitymatrix.Then,fordifferenceGMM,the
(T1)x(T1)blocksofHbydefaultareMM',whichisvar[u*]
(=var[e*])whenvar[e]=I(seeRoodman2009).Fororthogonal
deviations,MM'=I.Fordifferencing,itis:
210...
121...
012...
......
ToperformsystemGMM,xtabond2treatsthetransformeddataasbeing
forperiods2toTandlevelsdataasbeingforperiodsT+1to2T.
TheblocksofHarethen(2T1)x(2T1)aprioriestimatesofthe
covarianceofthecompoundvector[u*'u']'.Ifweassume,in
additiontovar[e]=I,thatvar[v]=0(nofixedeffects),thenthe
blocksofHare
http://repec.org/bocode/x/xtabond2.html 9/13
5/6/2017 repec.org/bocode/x/xtabond2.html
MM'M'
MI
However,morethanonechoiceforHispresentintheliterature.In
xtabond2,h(3),thedefault,specifiesthematricesdescribedabove.
h(2)differsinthatforsystemGMMtheupperrightandlowerleft
quadrantsofthedepictedHarezeroedout.Thiscopiescurrent
versionsofDPDforGaussandOx(ArellanoandBond1998;Doornik,
Arellano,andBond2002).h(1)specifiesthatH=Iforbothdifference
andsystemGMM.Htookthisvalueintheoriginalimplementationof
thesystemGMMestimator,inBlundellandBond(1998).Inonestep
GMM,settingH=Iessentiallygives2SLS.
TheMatasystemparametermatafavorinfluencesthebehavioroftheMata
versionofxtabond2.Typemata:matasetmatafavorspeedormata:
matasetmatafavorspacebeforerunningxtabond2toinfluencethe
tradeoffitmakesbetweenspeedandmemoryuse.Addthe,perm
optiontothesecommandstomakethechangepermanent.Note:
IncreasingtheamountofmemoryavailableforStatadatasetsusing
thesetmemorycommandreducesthatavailabletoMata.SoifMata
xtabond2isrunningoutofmemory,usuallyindicatedbyanunableto
allocaterealmessage,alsotryreducingStatamemorywithset
memory.
Optionsforpredict
xb,thedefault,calculatesthelinearprediction.
residualscalculatestheresidualerrorofthedependentvariablefrom
thelinearprediction.
differencerequeststhatthefirstdifferencesofthedependentvariable,
ratherthanthelevels,bepredicted.
Returnvalues
Scalars
e(N)Numberofcompleteobservationsinuntransformeddata
>(systemGMM)ortransformeddata(differenceGMM)
e(sargan)Sarganstatistic
e(sar_df)DegreesoffreedomforSarganstatistic
e(sarganp)pvalueofSarganstatistic
e(hansen)HansenJstatistic
e(hansen_df)DegreesoffreedomforHansenstatistic
e(hansenp)pvalueofHansenstatistic
e(artests)NumberofARtestsrequested
e(ari)AR(i)teststatistic
e(arip)pvalueofAR(i)statistic
e(df_m)Modeldegreesoffreedom
e(df_r)Residualdegreesoffreedom(ifsmallspecified)
e(chi2)Waldchisquaredstatistic(ifsmallnotspecified)
e(chi2p)pvalueofWaldstatistic(ifsmallnotspecified)
e(sig2)Estimatedvarianceofthee_it
http://repec.org/bocode/x/xtabond2.html 10/13
5/6/2017 repec.org/bocode/x/xtabond2.html
e(sigma)Squarerootthereof
e(F)Fstatistic(ifsmallspecified)
e(F_p)pvalueofFstatistic(ifsmallspecified)
e(g_min)Lowestnumberofobservationsinanincludedindividu
>al
e(g_max)Highestnumberofobservationsinanincludedindivid
>ual
e(g_avg)Averagenumberofobservationsperincludedindividua
>l
e(h)Valueofh()option(defaultis3)
e(j)Numberofinstruments
e(j0)Numberofinstruments,includingcollinearones
e(N_g)Numberofincludedindividuals
e(N_clust)Numberofclusters
e(components)Numberofcomponentsextractedifpcaoptioninvoked
e(kmo)KaiserMeyerOlkinmeasureofsamplingadequacyifpc
>aoptioninvoked
e(pcaR2)Sumofeigenvaluesofincludedcomponentsdividedby
>sumofall
Macros
e(predict)"xtab2_p"
e(artype)"firstdifferences"or"levels"
e(vcetype)"Robust"foronesteprobust,"Corrected"fortwostep
>robust,emptyotherwise
e(twostep)"twostep"fortwostep
e(small)"small"forsmall
e(esttype)"system"or"difference"
e(pca)"pca"ifpcaoptioninvoked
e(gmminstsi)Variableslistedingmmstylegroupi
e(ivinstsi)Variableslistedinivstylegroupi
e(transform)"firstdifferences"or"orthogonaldeviations"
e(depvar)Dependentvariable
e(clustvar)Clusteringgroupidentifier
e(tvar)Timevariable
e(ivar)Individual(panel)variable
e(cmd)"xtabond2"
e(cmdline)Fullcommandline
e(diffgroupi)variablesinithgroupsubjecttodifferenceSargan/H
>ansentesting
Matrices
e(b)Coefficientvector
e(V)Variancecovariancematrix
e(A1)FirststepGMMweightingmatrix
e(A2)SecondstepGMMweightingmatrix(iftwostepspecifie
>d)
e(Ze)Z'EwhereE=2ndstepresiduals,usedincomputingHan
>senstatistic
e(eigenvalues)EigenvaluesofprincipalcomponentsofGMMstyleinst
>ruments(ifpcaspecified)
e(diffsargan)TableofdifferenceinSargan/Hansentests
e(ivequation)Valueofequation()suboptionforeachivstyle()opti
>on,inorder
(0=level,1=diff,2=both)
e(ivpassthru)Valueofpassthruoptionforeachivstyle()option.
e(ivmz)Valueofmzsuboptionforeachivstyle()option
e(gmmequation)Valueofequation()suboptionforeachgmmstyle()opt
>ion
(0=level,1=diff,2=both)
e(gmmpassthru)Valueofpassthruoptionforeachgmmstyle()option
e(gmmpasscollapse)Valueofcollapseoptionforeachgmmstyle()option
e(gmmlaglimits)Laglimitsforeachgmmstyle()option
http://repec.org/bocode/x/xtabond2.html 11/13
5/6/2017 repec.org/bocode/x/xtabond2.html
e(gmmorthogonal)Valueoforthogonaloptionforeachgmmstyle()option
e(X)Matrixofrightsidevariablesusedinestimation,if
>svmatinvoked
e(Y)Columnofdependentvariableusedinestimation,ifs
>vmatinvoked
e(Z)Instrumentmatrixusedinestimation,ifsvmatinvoke
>d
e(H)Hmatrixusedinestimation,ifsvmatinvoked
e(wt)Weightvectorusedinestimation,ifsvmatinvokedan
>dweightsused
e(eigenvectors)Principalcomponentscores,ifsvmatandpcainvoked
Functions
e(sample)Marksestimationsample
Examples
usehttp://www.statapress.com/data/r7/abdata.dta
xtabond2nl.nl(0/1).(wk)yr1980yr1984,gmm(l.nwk)iv(yr1980yr1984,
passthru)noleveleqsmall
xtabond2nl.nl(0/1).(wk)yr1980yr1984,gmm(l.nwk)iv(yr1980yr1984,
mz)robusttwostepsmallh(2)
xtabond2nl(1/2).nl(0/1).wl(0/2).(kys)yr1980yr1984,gmm(l.nwk)
iv(yr1980yr1984)robusttwostepsmall
*Nexttwoareequivalent,assumingidisthepanelidentifier
ivreg2ncap(w=kysrec)[pw=_n],cluster(ind)orthog(rec)
xtabond2nwcap[pw=_n],iv(capkys,eq(level))iv(rec,eq(level))
cluster(ind)h(1)
*Samefornexttwo
regressnwk
xtabond2nwk,iv(wk,eq(level))smallh(1)
*Andnexttwo,assumingxtabondupdatedsinceMay2004withupdate
command.
xtabondnyr*,lags(1)pre(w,lags(1,.))pre(k,endog)robustsmall
noconstant
xtabond2nL.nwL.wkyr*,gmm(L.(wnk))iv(yr*)noleveleqrobustsmall
*Andnexttwo
xtdpdnL.nL(0/1).(wk)yr1978yr1984,dgmm(wkn)lgmm(wkn)
liv(yr1978yr1984)vce(robust)twohascons
xtabond2nL.nL(0/1).(wk)yr1978yr1984,gmm(L.(wkn))
iv(yr1978yr1984,eq(level))h(2)robusttwostep
*Threewaystoreducetheinstrumentcount
xtabond2nL.nL(0/1).(wk)yr1978yr1984,gmm(L.(wkn))
iv(yr1978yr1984,eq(level))h(2)robusttwosteppca
xtabond2nL.nL(0/1).(wk)yr1978yr1984,gmm(L.(wkn),collapse)
iv(yr1978yr1984,eq(level))h(2)robusttwostep
xtabond2nL.nL(0/1).(wk)yr1978yr1984,gmm(L.(wkn),lag(11))
iv(yr1978yr1984,eq(level))h(2)robusttwostep
*EstimationalaHayakawa2009
xtabond2nL.nL(0/1).(wk)yr1979yr1984,gmm(L.(wkn),lag(11)
orthog)iv(yr1979yr1984,eq(level))h(2)robusttwosteporthog
noleveleq
Threesamplefilesareincludedwiththepackagedownloadedwiththis
command.abest.doreproducestwosamplefilethatcomeswithDPDforOx,
whichinturngeneratemostoftheGMMresultsinArellanoandBond
(1991).bbest.doreproducesanothersamplefilethatcomeswithDPDfor
Ox,basedonBlundellandBond(1998).Todownloadthem,typethe
followingcommandorclickonit:sscinstallxtabond2,allreplace.
Thiswillsavethefilestoyourcurrentdirectory,assetbythecd
command.greene.doreproducesanexampleinGreene(2002).
http://repec.org/bocode/x/xtabond2.html 12/13
5/6/2017 repec.org/bocode/x/xtabond2.html
References
Arellano,M.andS.Bond.1991.Sometestsofspecificationforpanel
data:MonteCarloevidenceandanapplicationtoemployment
equations.TheReviewofEconomicStudies58:27797.
Arellano,M.andS.Bond.1998.DynamicPaneldataestimationusing
DPD98forGauss:Aguideforusers.
Arellano,M.andO.Bover.1995.Anotherlookattheinstrumental
variableestimationoferrorcomponentsmodels.Journalof
Econometrics68:2951.
Bai,J.,andS.Ng.2010.InstrumentalVariablesEstimationinaData
RichEnvironment.EconometricTheory26(6):15771606.
Baum,C.F.,M.E.Schaffer,andS.Stillman.2003.Instrumentalvariables
andGMM:Estimationandtesting.StataJournal3:131.
Blundell,R.,andS.Bond.1998.Initialconditionsandmoment
restrictionsindynamicpaneldatamodels.JournalofEconometrics
87:11543.
Bond,S.2002.Dynamicpaneldatamodels:Aguidetomicrodatamethods
andpractice.WorkingPaper09/02.InstituteforFiscalStudies,
London.
Doornik,J.A.,M.Arellano,andS.Bond.2002.Paneldataestimation
usingDPDforOx.http://www.nuff.ox.ac.uk/Users/Doornik.
Greene,W.H.2002EconometricAnalysis,5thed.PrenticeHall.
Hayakawa,K.2009.Asimpleefficientinstrumentalvariableestimatorfor
panelAR(p)modelswhenbothNandTarelarge.EconometricTheory
25:87390.
HoltzEakin,D.,W.Newey,andH.S.Rosen.1988.Estimatingvector
autoregressionswithpaneldata.Econometrica56:137195.
Kapetanios,G.,M.Marcellino.2010.FactorGMMestimationwithlarge
setsofpossiblyweakinstruments.ComputationalStatistics&Data
Analysis54(11):265575.
Mehrhoff,J.2009.Asolutiontotheproblemoftoomanyinstrumentsin
dynamicpaneldataGMM.DiscussionPaperSeries1.No31/2009.
Roodman,D.2009.HowtoDoxtabond2:AnIntroductionto"Difference"and
"System"GMMinStata.StataJournal9(1):86136.
Windmeijer,F.2005.Afinitesamplecorrectionforthevarianceof
linearefficienttwostepGMMestimators.JournalofEconometrics
126:2551.
Author
DavidRoodman
SeniorFellow
CenterforGlobalDevelopment
Washington,DC
[email protected]
Alsosee
Manual:[U]23Estimationandpostestimationcommands,
[U]29OverviewofStataestimationcommands,
[XT]xtabond
Online:helpforxtabond,ivreg,ivreg2,estcom,postest;xtgee,
http://repec.org/bocode/x/xtabond2.html 13/13
