Professional Documents
Culture Documents
1
SteveGibbons(SERCandLSE)2
HenryOverman(SERCandLSE)3
Abstract:Wearguethatidentificationproblemsbedevilappliedspatialeconomicresearch.Spatial
econometricsusuallysolvestheseproblemsbyderivingestimatorsassumingthatfunctionalforms
areknownandbyusingmodelcomparisontechniquestoletthedatachoosebetweencompeting
specifications.Wearguethatinmanysituationsofinterestthisachieves,atbest,onlyveryweak
identification.Worse,inmanycases,suchanapproachwillbeuninformativeaboutthecausal
economicprocessesatwork,renderingmuchappliedspatialeconometricresearchpointless,
unlessthemainaimisdescriptionofthedata.Weadvocateanalternativeapproachbasedonthe
experimentalistparadigmwhichputsissuesofidentificationandcausalityatcentrestage.
ThetitleisareferencetoAngristandPischkes(2009)MostlyHarmlessEconometricswhichoutlinesthe
experimentalistparadigmandarguesthatfanciereconometrictechniquesareunnecessaryandpotentially
dangerous.
2
SteveGibbons,SpatialEconomicsResearchCentreandDepartmentofGeographyandEnvironment,London
SchoolofEconomics,HoughtonStreet,London,WC2A2AE,UK.Email:s.gibbons@lse.ac.uk.
3
HenryG.Overman,SpatialEconomicsResearchCentreandDepartmentofGeographyandEnvironment,
LondonSchoolofEconomics,HoughtonStreet,London,WC2A2AE,UK.Email:h.g.overman@lse.ac.uk.
1.INTRODUCTION
Thelasttwodecadeshaveseeneconomistsbecomeincreasinglyinterestedingeographicalissues
(Martin,1999,BehrensandRobertNicoud,2009).Thishasbeenvariouslyattributedtotheoretical
developments,agrowinginterestincitiesorsimplythegreateravailabilityofgeoreferenceddata.
Theresulthasbeengreaterinteractionbetweeneconomicgeographers,regionalscientistsand
economistsinterestedinspatialaspectsoftheeconomy.Morerecently,asimilarprocesshasseen
mainstreameconometrictheoristsbecomingincreasinglyinterestedinspatialprocesses,
traditionallythepreserveofagroupofspatialeconometricians.4Onemightthinkthatthenextstep
wouldbeconvergencebetweenthetoolsdevelopedbyspatialeconometriciansandthemethods
usedbyappliedeconomiststoassesstheempiricalvalidityofmodelsofspatialeconomics.We
arguethatthisisunlikelybecause,whiletheremayhavebeenconvergencebetweenmainstream
andspatialeconometrictheory,mostappliedeconomicresearchistakingadifferentpath.
Inmany(micro)economicfieldsparticularlydevelopment,education,environment,labor,health,
andpublicfinanceempiricalworkisincreasinglyconcernedwithquestionsaboutcausality(Angrist
&Pischke,2010).Ifweincreaseanindividualsyearsofeducation,whathappenstotheirwages?If
wedecreaseclasssizes,whathappenstostudentgrades?Thesequestionsarefundamentallyofthe
typeifwechangex,whatdoweexpecttohappentoy.Justaswitheconomicsmoregenerally,
suchquestionsarefundamentaltoourunderstandingofspatialeconomics.Whenmoreskilled
peopleliveinanarea,whathappenstoindividualwages?Ifajurisdictionincreasestaxes,what
happenstotaxesinneighbouringjurisdictions?
Inanexperimentalsetting,agents(individuals,firms,governments)wouldberandomlyassigned
differentxandtheoutcomesyobserved.Measuringwhetherdifferentxareassociatedwith
differentoutcomeswouldthengivethecausaleffectofxony.Thefundamentalchallengeto
answeringthesequestionsfor(most)economicdataisthatxisnotrandomlyassigned.Instead,we
jointlyobservexandysowelackthecounterfactual,thatis,whatwouldhavehappenedifxhadnot
beenchanged.Fortunately,appliedeconomicshascomealongwayinitseffortstofindcredibleand
creativewaystoanswersuchquestionsbyconstructingcounterfactualsfromobservationaldata.
Agoodstartingpointforthinkingaboutwhetheraquestionaboutcausalitycanbeansweredand
howtoanswerit,istoconsideranidealexperiment.Theexperimentmaynotbefeasible,butwith
thedesigninminditiseasiertothinkofwaystofindsourcesofvariationinthedatathatmimicor
approximatetheidealexperiment.Theexperimentalistparadigm(AngristandKrueger1999,
AngristandPischke,2009,2010)doesthisbyusingsimplelinearestimationmethods,takingcareto
pinpointandisolatesourcesofvariationinxthatcanplausiblybeconsideredexogenous.Theaimof
thesemethodsistomimic,asfaraspossible,theconditionsofanexperimentinwhichagentsare
randomlyassigneddifferentxandoutcomesyobserved.Thecentralideaistofindotherwise
comparableagents(e.g.twins,siblings,neighbours,regions)whoforsomereasonhavebeen
exposedtodifferentx.Thisapproachisstill'econometric'itdrawsontheorytoguidethe
questionsaskedandthinkingaboutthecausalprocessesatwork.However,thefundamental
SeeAnselin(2010).Manyofthespecialisedeconometricianswhodevelopedthefieldarerecognisedby
FellowshipoftheSpatialEconometricAssociation.Seehttp://spatialeconometr.altervista.org/
attractionisthattheassumptionsrequiredforidentificationofcausaleffectsareusuallyclearly
specifiedandunderstandablewithoutreferencetospecific(anduntested)economictheories.Put
anotherway,theaimistoobtainplausibleestimatesofcausaleffectswithoutrelyingonadhoc
functionalformsandexclusionrestrictionsimposedarbitrarily,orderivedfromuntestedtheories
aboutwhichthereisnoconsensus.5Thisapproachisparticularlyattractiveinareas,likemuchof
spatialeconomics,whereavailablestructuralmodelsdonotcloselycapturethecomplexitiesofthe
processesforwhichwehavedata.6Unfortunately,althoughtheseissuesmaybewellunderstoodby
moreexperiencedpractitioners,theyarenotwidelydiscussedinmanyofthestandardspatial
econometricsreferences.7Forthisreasonthereisadangerthatpeopleenteringtheworldof
appliedeconomicresearchusingspatialeconometricswillignoretheseinsightsintoframing
questionsandachievingcredibleresearchdesigns.
Whyisitthecasethatthespatialeconometricsliteratureoftenignorestheseissues?Wesuspect
thisispartlybecausetheunderlyingtheorydevelopedfromtimeseriesfoundations,sothat
questionsaboutcausalityhavenotbeencentrestage.Thestandardapproachtospatial
econometricshasbeentowritedownaspatialmodel(e.g.thespatialautoregressivemodel),to
assumeitaccuratelydescribesthedatageneratingprocessesandthentoestimatetheparameters
bynonlinearmethodssuchas(quasi)maximumlikelihood(ML).Becauseestimationisnotalways
simple,muchefforthasgoneintodevelopingtechniquesthatallowestimationofarangeofmodels
forlargedatasets.Questionsofidentification(i.e.doesanestimatedcorrelationimplythatxcauses
y?)havebeenaddressedbyaskingwhichspatialprocessesbestfitthedata.Whilethissounds
straightforward,inpractice,aswediscussbelow,itishardtodistinguishbetweenalternative
specificationsthathaveverydifferentimplicationsforwhichcausalrelationshipsareatwork.
Inthisarticleweexplainwhythestandardspatialeconometrictoolboxisunlikelytoofferasolution
totheproblemoftheidentificationofcausaleffectsinmanyspatialeconomicsettings.Ofcourse,
muchstandard(i.e.nonspatial)empiricaleconomicanalysisfallssomewayshortoftheloftyideals
ofidentifyingcausaleffectsfromrandomvariationinthevariableofinterest(x).Findingsourcesof
trulyexogenousorrandomvariationinxisdifficult,butgoodappliedworkaimedatcausalanalysis
mustsurelymakesomecredibleattempttodoso.Thisisnottosaythatnoncausalassociationsare
neverwithoutmerit,becausedescriptionandcorrelationcanprovideessentialinsights.However,
identificationofcausaleffectsremainsthegoldstandardtowhichmanyeconomistsclaimtoaspire.
Wewillarguethatthisshouldalsobethecaseinappliedspatialeconomicresearch.
Therelianceonsimplelinearmethodsmayseemastrongfunctionalformassumption.However,the
assumptionofalinearstructuralrelationshiptheConditionalExpectationFunction(CEF)isnotreally
necessaryforacausalinterpretationofregressionAngristandPischke2009(p.69).IftheCEFiscausalthen,
linearregressionisinformativeaboutcausalitybecauseitprovidesthebestlinearapproximationtotheCEF.
6
Sutton(2002)makesasimilarargumentaboutstructuralmodellingwhenmodelsarefarfromreality.Thisis
nottosaythattheoreticalstructurehasnoplaceinempiricalspatialeconomics.Particularlywhengeneral
equilibriumconsiderationsareimportant,theremaybeagreaterrolefortheory(preferablybasedonmicro
economicbehaviouralfoundations).SeeCombes,DurantonandGobillon(2011)fordiscussion.Laterinthe
paperwebrieflydiscusstheuseofspatialeconometricsintheestimationofstructuraleconometricmodels.
7
Forexample,thereislittlediscussionoftheseissuesinArbia(2006)andLeSageandPace(2009).
Therestofthispaperisstructuredasfollows.Section2providesabasicoverviewofstandardspatial
econometricmodels,whilesection3discussesproblemsofidentification.Section4returnstothe
relationshipbetweenthespatialeconometricsandtheexperimentalistparadigm.Section5
concludes.
2.SPATIALECONOMETRICMODELSANDTHEIRMOTIVATION
Thissectionprovidesanintroductiontospatialeconometricmodels,ofthetypepopularisedby
Anselin(1988).Itisnotcomprehensivebutprovidesenoughbackgroundsothatsomeoneunfamiliar
withspatialeconometricsshouldbeabletofollowtheargumentsmadelater.Wegenerallyusethe
modelterminologyofLeSageandPace(2009)andreferthereadertherefordetails.
Todevelopideas,startwithabasiclinearregression:
yi xi' ui
(1)
yi islinearandconstantacrossobservations.Thisgivesthespatialautoregressivemodel(SAR):
yi w i' y xi' ui
(SAR)
Foreaseofexpositionwesubsumetheconstantin xi .
'
LeSageandPace(2009)suggestatimedependencemotivationfortheSARmodel.Assumefixed
acrosstimeexogenousvariables xi determineoutcome yi .Nowassumethatwhendetermining
theirownoutcome,agentstakeintoaccountboththeirowncharacteristicsandrecentoutcomes
forothernearbyagents.Wemightthinkof yi asthepriceofahouse, xi asthefixed
characteristics(numberofrooms)andassumethatwhenagreeingasaleprice,peopleconsiderboth
thecharacteristicsofthehouseandthecurrentsellingpriceofnearbyhouses.Inthiscase
capturesthecausaleffectofhousecharacteristicsandrepresentsthecausaleffectofneighbouring
prices(conditionalonobservedhousingcharacteristics).
Wecoulddroptheassumptionthat yi isaffectedbythespatiallagof yi andinsteadassumethat
yi isaffectedbyspatiallagsoftheexplanatoryvariables.IfXdenotesthematrixofexplanatory
variablesandavectorofparameters,thisgivesthespatial(lagof)Xmodel(SLX):
yi xi' w i' X ui
(SLX)
LeSageandPace(2009)provideanexternalitymotivationforthismodel.Continuingwiththe
housingexample,thisassumesthecharacteristicsofnearbyhouses,e.g.theirsize,directly
determineprices(ratherthanworkingthroughobservedsalesprices).Ofcourse,anexternality
motivationcouldjustifytheSARmodeliftheexternalityworksthroughthespatiallagof yi
Next,droptheassumptionthatoutcomesareaffectedbyspatiallagsoftheexplanatoryvariables
andinsteadassumespatialautocorrelationintheerrorprocess(using,e.g,theSARprocessto
imposestructure).9If u denotesthevectorofresiduals,thisgivesthespatialerrormodel(SE):
yi xi' ui ; ui w i' u vi
(SE)
Finally,combiningtheSARandSLXmodelsgivesustheSpatialDurbinModel(SD):
(SD)
whichassumesdependencebetween yi andthespatiallagsofboththeoutcomeandexplanatory
variables,butdropstheassumptionofspatialautocorrelationintheerrorprocess.Alternatively,the
SDmodelcanbemotivatedbysimplyrearrangingtheSEmodelinaspatialCochraneOrcutt
transformation:
ui yi xi'
(2)
(3)
(4)
Thisideaprovidesanothermotivationforincludingspatiallags,asasolutiontotheomitted
variablesproblem.Seetheappendixforfurtherdiscussion.
Moregeneralspecificationsareavailable,buttheSEmodeldefinedinthetextissufficientforourpurposes.
Thesefiveprocessesarenotexhaustiveofallpossiblemodels,andweconsideraparticularly
importantgeneralisationfurtherbelow,butforthemomenttheyaresufficientforourpurposes.In
thetext,weusetheacronyms(SAR,etc)torefertothespecificationsabove.
EstimationusingOLSgivesinconsistentparameterestimatesifthemodelsincludeaspatiallagof yi
andisnonzero(e.g.theSARandSDmodels).Thisinconsistencyarisesbecauseofamechanical
linkbetween ui and w i' y formostspecificationsof w i .Standarderrorsarealsoinconsistently
estimatedforthesemodels,aswellasformodelsincludingaspatiallagin ui (e.g.theSEmodel).
OLSprovidesconsistentparameterestimatesifthespatialcorrelationoccursonlythroughtheerror
term(SEmodel)orexogenouscharacteristics(SLXmodel).Inbothcasesstandarderrorsare
inconsistent,andOLSestimationoftheSEmodelisinefficient.Incontrast,Lee(2004)showsthat
(quasi)MLestimationprovidesconsistentestimatorsforallthesemodelsconditionalonthe
assumptionthatthespatialeconometricmodelestimatedisthetruedatageneratingprocess.
Alongsidetheoreticaldevelopments,advancesincomputationalpowerandmethodshavemadeML
estimationfeasibleforlargedatasets.10Asaresult,itispreferredinthespatialeconometrics
literature.TheSARandSLXmodelsarenestedwithintheSDmodelandasshowntheSEmodelcan
berearrangedtogivetheSDmodel.ThefactthattheSDmodelnestsmanyoftheothermodels
providesanargumentforestimatingtheSDmodelandthentestingthisagainstthenestedmodels
throughtheuseoflikelihoodtests.Modelcomparisontechniquescanbeusedtocomparemodels
basedondifferentweightmatricesandexplanatoryvariables.Thisistheapproachadvocatedby
LeSageandPace(2009).
Tobegintounderstandtheproblemswiththisapproachitisusefultoseehowthesemodelsare
relatedtoeachother.Considerthereducedform(expressing yi intermsofexogenousfactors)of
theSARmodel.Ifthemodeliscorrect,theonlyexogenousfactorsaffecting yi are xi and ui ,sothe
onlyfactorsaffecting w i' y are w i' X and w i' u .Thespatiallag w i' y alsodependsonthesecond
orderspatiallagof w i' Wy ,thatis,onoutcomesfortheneighboursofmyneighbours.Byrepeated
substitutionthereducedformis:
(5)
10
Theseimprovementsallowmodelsinvolvingsamplescontainingmorethan60,000USCensustract
observationstobeestimatedinonlyafewsecondsondesktop[]computersLeSageandPace(2009,p.45)
(6)
w i' W n 1X forn>1.
Inshort,spatialinteractionin yi ,spatialexternalitiesin xi ,orspatiallyomittedvariablesleadto
differentspatialeconometricspecifications.Thesemodelshavedifferentimplicationsforthe
economicprocessesatwork.However,thereducedformforallthesemodelsis:
(7)
andtheonlydifferencesarisefromhowmanyspatiallagsof xi areincluded,constraintsontheway
theunderlyingparametersdeterminethecompositeparameters ,andwhethertheerrortermis
spatiallycorrelated.Distinguishingwhichofthesemodelsgeneratesthedatathattheresearcherhas
athandisgoingtobedifficultasthespecificationofWisoftenarbitrary,andbecausethespatial
lagsof xi arejustneighbouraveragesthatarealmostalwaysveryhighlymutuallycorrelated.Put
anotherway,thesedifferentspecificationsaregenerallyimpossibletodistinguishwithoutassuming
priorknowledgeaboutthetruedatageneratingprocessthatweoftendonotpossessinpractice.In
shortcontrastingmotivationsleadtomodelsthatcannotusuallybeeasilydistinguished.Itwouldbe
usefuliftheseproblemsweremoregenerallyrecognisedbyallresearchersworkingwithspatial
data,butwethinktheygenerateparticularproblemsforthespatialeconometricsapproachas
outlinedinthissection.Wenowconsiderthesedifficultiesindetail.
3.THEREFLECTIONPROBLEMANDCRITIQUEOFSPATIAL
ECONOMETRICMODELS
Readersfamiliarwiththeneighbourhoodeffectsliterature,willseeimmediateparallelsbetween
thespatialeconometricsmodels(SAR,etc)andlinearinmeansneighbourhoodeffectsmodels.11
TheparallelsbetweenthesefieldshavealreadybeenhighlightedbyLee(2004,2007)andothers.
ThegenericneighbourhoodeffectsmodeldescribedbyManski(1993,2000)andusedbycountless
appliedeconomicspapersonneighbourhoodeffects,takestheform:
yi 1 E[ yi | a ] xi' E[xi | a] vi ,
vi 2 E[vi | a ] ui
(8)
11
Thepointswemakehereapplyequallytothepeereffectsliterature.
(9)
toseethattheparameteron E[ yi | a ] isacompositeofthecausal'endogenous'andincidental
'correlated'effects.Whatthencanbeidentifiedfromtheobservablecharacteristics?Taking
expectationsof(8)andrearranginggivesareducedformversionof(8):
(10)
Equation(10)showsthatthecausaleffectsofneighbourhoodmeanoutcomes( 1 )andof
neighbourhoodmeancharacteristics( ),cannotbeseparatelyidentifiedfromthereducedform
parameters,evenwhenthereisnospatialautocorrellationintheunobservables( 2 0 ).Only
andthecompositeparametervector ( 1 ) / (1 1 ) areidentified.ThisisManski's'reflection
problem':onlytheoveralleffectofneighbourscharacteristicsisidentified,notwhethertheywork
throughexogenousorendogenousneighbourhoodeffects.Theseissuesareintuitive:Howcanyou
distinguishbetweensomethingunobservedandspatiallycorrelateddrivingspatialcorrelationiny
fromthesituationwhereyisspatiallycorrelatedbecauseofdirectinteractionbetweenoutcomes?
Further,howcanyoutellwhetheranindividualisaffectedbythebehaviouroftheirgroup,orbythe
characteristicsoftheirgroupwhengroupbehaviourdependsonthecharacteristicsofthegroup?In
manycircumstancesyoucannot,withoutimposingfurtherrestrictions,eitherassumingawayoneor
moreofthesourcesofneighbourhoodeffects,orbyimposingnonlineartheoreticalrelationships
between E[ yi | a ] and E[xi | a ] (e.g.asinBrockandDurlauf2001,whentheoutcomeisdiscrete,
andtheresearcherimposesalogitorprobitfunctionalform).Insomesituations,however,
experimentaldata(oftherealornaturalvariety)canremovetheneedforsucharbitraryrestrictions.
Wediscussfurtherandprovidesomeconcreteexamplesinsection4.
Inpractice,theneighbourhoodeffectsliteratureusesempiricalcounterpartsto(8)and(10):
(8a)
(10a)
8
where w i asthespatialweightvectorthatcreates'neighbourhoodaverages'asestimatesof
E[ | a ] 12.Forneighbourhoods,wheregroupsarecontiguousandnonoverlapping,thejthelementof
(11)
(12)
wheretheeconometriciancanchooseanarbitrarynumberofspatiallagsof xi asinstruments.13
Sowhydospatialeconometriciansarguethat(8a)isidentified,byvirtueof(11),whenother
economistswouldarguethat(8a)isn'tidentified,byvirtueof(10a)?Thecrucialdifferenceisthat
spatialeconometricsassumesthat W isknownandrepresentsrealworldlinkages.Neighbourhood
effectsresearchersarguethatthetrue W isalmostneverknown,andis,atbest,ameansof
estimating E[ | a ] .Inotherwords,theassumptionofknowledgeabout W iscritical.14
12
ThisinterpretationoftheuseofthespatialweightingvectorhasbeennotedbyPinkseandSlade(2010)
McMillen(2010b)andothers.
13
Interestingly,theideaofputtingmorestructureonneighbourhoodeffects(e.g.byassumingahierarchical
network)hasrecentlybeensuggestedasawayofsolvingtheidentificationproblem.SeeLeeatal(2010).
If W takesthetypicalblockdiagonalstructureusedinneighbourhoodeffectsstudies(or W is
otherwiseidempotent)then(11)collapsesto(10a)andtheparametersarenolongerseparately
identified.Butevenif W isnotidempotent,thereareseriousproblemsinrelyingon(11)and(12)
foridentificationoftheparametersin(8a).Firstlyiftheexactstructureof W isnotknownthen
identificationbreaksdownbecause w i' W , w i' W 2 etc.maybettercapturetheconnectionsbetween
observationianditsneighboursthandoes w i (e.g. xi hasaneffectupto5km,but w i incorrectly
restrictseffectstowithin2km).Ifthehigherorderlags w i' WX , w i' W 2 X affectydirectly,they
cannotprovideadditionalinformationtoidentify 1 .Itiseasiesttothinkaboutthisinthecontextof
thelinear(2SLS)IVestimatorimplicitin(11)and(12)(e.g.KelejianandPrucha1998):Ifthe
exclusionrestrictionson w i' WX , w i' W 2 X etc.in(8a)areinvalid,thenthesespatiallagsarenot
suitableasinstruments,norassourcesofidentificationmoregenerally.
Secondly,evenif W isassumedknown(andisnotidempotent)thereareseriousestimation
problemsbecausethespatiallags w i' X w i' WX , w i' W 2 X etc.arelikelytobehighlycorrelated.In
theIVcontext,thereisaweakinstruments/identificationproblem,becausethereislittle
independentvariation(andhencelittleadditionalinformation)inthehigherorderspatiallagsof xi ,
conditionalon w i' X .SinceBound,JaegerandBaker(1995)andStaigerandStock(1997),applied
researchersworryaboutthestrengthoftheinstrumentsinIVregressions,becauseweak
instrumentscanleadsecondstagecoefficientestimatestobeseverelybiasedandimprecisely
estimated.Thisissuehascertainlybeenrecognisedinthespatialeconometricsliteraturebutthe
profoundconsequencesdonotappeartohavehadmuchinfluenceonappliedresearch.
Intheory,thedegreeofcollinearitybetweenspatiallagsdependsonsamplesize,samplingframe
andhowWchangesasobservationsareadded.15Inpractice,inlargesamples(andusingstandard
15
Intheoreticalanalysisitisusualtodistinguishbetweenincreasingdomainasymptotics(addingobservations
expandsthespaceoverwhichweestimate)andinfillasymptotics(increasingsamplesizemeanssampling
moreobservationsinagivenarea).Thelattermakesmoresenseinmanymicroeconometricsettings,whichis
problematicbecause,generalresultsarenotavailableforMLEunderinfillasymptotics(seeLee(2004)).Our
intuitionisthatconsistencyunderinfillasymptoticswillbedifficulttoderivepreciselybecause,formany
commonWmatrices,infillasymptoticsincreasesthecorrelationbetweenxanditsspatiallags.
16
SeeGibbonsandOverman(2010).
10
additionalinformation(e.g.satisfytheexclusionrestrictionsrequiredtomakethemvalid
instruments).Itisnotclearhowthisassumptioncouldeverbeevaluated.17
Itshouldbecleartherefore,thattherearefundamentalproblemsinusingassumptionsonthe
spatialstructuretogenerateinstruments.Theproblemsareevenmoreprofoundifweallowfor
spatialautocorrelationintheerrorterms,anddroptheassumptionoftheexogeneityof xi togivea
moregeneralspatialmodel:
(13)
ui w i' u vi
(14)
(15)
4.THEEXPERIMENTALISTPARADIGMANDSPATIALECONOMETRICS
Thediscussionsofarhasbeencriticalofthespatialeconometricsapproach,particularlyregarding
thecrucialissueofidentificationofcausalparameters.Othershavemadesimilararguments
17
Statisticaltestsforoveridentifyingrestrictions(e.g.SarganandHansentests,seeAngristandPischke(2009)
p.143146)aregenerallynotveryrobustevidenceforthevalidityofinstruments,unlesssimplyasauxiliary
supportforaconvincingcaseoftheirtheoreticalvalidity.Thesestatistics,ineffect,providetestsforequalityof
theIVestimatesfromeachpotentialinstrument,butwillbepronetoTypeIIerrors(failingtorejectthenullof
thevalidityoftheoveridentifyingrestrictions)insituations,liketheonefacedhere,whereinstrumentsare
eitherweak(becauseIVcoefficientsforeachinstrumentarethenimpreciselyestimated)orendogenousfor
similarreasons(becauseIVcoefficientsforeachinstrumentwillbesimilarlyinconsistent).
11
althoughperhapsnotasforcefully(seeforexampleMcMillen2010aand2010b).Ofcourse,any
alternativeapproachalsohastosolvetheidentificationproblemsthatplaguespatialeconomic
analysis.Ourargumentisthattheseproblemsaresofundamentalthattheymustsitatcentrestage
ofappliedwork,notbeshuntedtothesidelinesthroughtheuseofMLthatassumesknowledgeof
theappropriatefunctionalformsandspatialweights.Inthissectionwearguethatspatialresearch
wouldbebestservedbyturningawayfromtheapplicationofgenericspatialmodelsandattempts
todistinguishbetweenobservationallyequivalentmodelsusingcontestableparameterrestrictions
thatonlyemergefromtheassumedmodel.Instead,weadvocatestrategiesthatarecarefully
designedtoanswerwelldefinedresearchquestionsusinginsightsfromtheexperimentalist
paradigm.
WestartwiththesituationwhereweareinterestedinestimatingparametersinaSARorSD
specificationtotestfordirectspatialinteractionsbetweenoutcomes yi .Itishardtoimagine
situationsinwhichthisisthetruedatageneratingprocessbecausesimultaneousdecisionsbasedon
yi mustrelyonexpectations(asintheneighbourhoodseffectliterature),butletussupposethat
estimationofisthegoal.Asarguedabove,thereisacentralconceptualproblemabout
identificationofthelineardependence yi on w i' y inanSARstylespecification,whichfollowsfrom
the'reflection'problem.Specifically,ifthemodelproperlyspecifieshowyisdetermined,howcan
weinduceexogenouschangein w i' y thatisnotcausedbychangesinelementsofeither w i' X or
w i' u ?MaximumLikelihoodsolutionsseemunconvincingforreasonsdiscussedextensivelyabove.In
somesettings,thespatialeconometricsliteratureoffersinterestinginsightsintothepotentialfor
usingspecificrestrictionson w i toachieveidentification,wheretheserestrictionsarisefromthe
institutionalcontext,forexamplefromthedirectedstructureoffriendshipnetworks,orthespatial
scopeofareatargetedpolicies(e.g.seeCalvoArmengol,PatacchiniandZenou2009).Formost
appliedproblems,however,uncertaintyaboutfunctionalformsandlackofinformationonthetrue
spatialweightsmeanalternativestrategiesaremoreappropriate.
Analternativestrategyistousepaneldataandtodifferencethedataovertimetoprovidefixed
effectsor'differenceindifference'estimates.Differencingremovesunobservablesthatarefixed
overtime,andthattheresearcherconsiderstobepotentialsourcesofendogeneity.Whilethisisa
veryusefulstrategyinmanycontexts(someofwhicharementionedbelow),itdoesnot,onitsown,
offerawaytoidentifythecausaleffectsof w i' y on yi inSAR/SDtypemodelsbecausethe
'reflection'issuessimplytransfertothedifferencedspecification.Thequestionnowbecomeshowto
distinguishchangesin w i' y (i.ein w i' y )thatarenotcausedbyelementsof w i' X and/or w i' u .
Forsimilarreasons,evenrandomisationofferslimitedscopefordistinguishinggroupeffectsarising
fromthespatialinteractioninoutcomes w i' y fromthosearisingfromgroupcharacteristics
anddormmates'groupcharacteristics,togetestimatesofpeergroupeffectsonstudents'college
achievement.DeGiorgioetal(2010)useasimilarstrategyinthecontextofrandomclass
assignments.Althoughbothpapersmakeclaimsthatareprobablytoostrongintermsoftheirability
tosolvethereflectionproblem(becauserandomisationalsochanges w i' X and w i' u asdiscussed
above),randomisationdoesatleastreducetheproblemsinducedbyselfselectionintogroupsand
consequentcorrelationbetweenindividualandgroupcharacteristics.Fieldexperimentsdesignedfor
purposearealsoclearlyveryuseful.However,bigonesliketheMovingtoOpportunityProgramme
(Kling,LudwigandKatz2005,Kling,LiebmanandKatz2007)arerare,costlyandoftensufferfrom
unavoidabledesignflaws,andsmallonessufferfromconcernsaboutexternalvalidity.Itwouldalso
beverydifficulttodesignexperimentstoanswermanyspatialquestionsandwedonotseethisasa
wayforwardformanyproblemsofinterest.
Onealternativeistoreconsiderinstrumentalvariables(IV/2SLS)estimation,eitherincrosssectional
ortimedifferencedspecifications.Asshownabove,iftheSARmodeliscorrectlyspecifiedthen w i' X
providesinstrumentsfor w i' y andthisformsthebasisforthetraditional'spatialIV'method.Most
appliedmicroeconometricresearcherswouldexpectverycarefulargumentstojustifytheexclusion
of w i' X fromtheestimatingequation.18Inpractice,manypapersthatusespatialeconometricsdo
notdothis.Totakejustoneexample,inthetaxcompetitionliteraturecharacteristicsof
neighbouringareas w i' X areoftenusedtoinstrumentforneighbours'taxrates w i' y inaregression
ofowntaxrate yi onneighbourstaxrates.Arunthroughsomeofthereferencesinarecentreview
(Revelli,2005)suggestthattheseexclusionrestrictionsreceivelittle,ifanyconsideration.Besleyand
Case(1995)appeartobeoneofthefirsttoadoptthisstrategybyusingdemographicsofneighbours
toinstrumentforneighbourstaxrate.Theydoprovideabriefdiscussionofwhetherthisrestriction
isvalid,butmostlyrelyonoveridentificationrestrictionsimposedbytheirtheoreticalmodelto
justifythisassumption.BrettandPinske(2000)useasimilarapproachandjustifytheirexclusionsby
noting:Whiletherecouldbereasonswhymunicipalbusinesstaxratesdependonwealthdirectly,
suchreasonsarelessobviousthandependencethroughtheireffectoncapitalbase.(Brettand
Pinske,2000,p.701).Buettner(2001)claimstopaycarefulattentiontowhethertheinstrumentsare
endogenous(byexaminingspatialautocorrelationintheresiduals)buthasnodiscussionofhowthis
relatestothevalidityoftheexclusionrestrictions.HayashiandBoadway(2001)donotinstrumentat
all,insteadusingrestrictionsfromatheoreticalmodeltoachieveidentification.Turningtomore
recentpapers,wefindlittleevidencethatmuchhaschanged.Leprince,MadiesandPaty(2007)has
nodiscussionoftheexclusionrestrictions.CharlotandPaty(2007)useML.EdmarkandAgren(2008)
dodiscussthestrengthoftheirinstruments,butnottheexclusionrestrictions.FeldandReulier
(2009)useIVbutdonotdiscusseitherproblem.Noneofthesepapersdiscusstheproblemspecific
tothespatialsetting,thatspatiallylaggedexogenousvariablesmaybettercapturetheconnections
betweenobservationianditsneighboursthantheincorrectlyspecifiedfirstorderspatiallagof yi .
Thislistofpapersisnotexhaustiveandinclusioninitisnotintendedasaspecificcriticismofthe
18
Ofcourse,if w i' X canbeexcludedthenthissolvestheidentificationproblemforMLaswell.Eveninthis
case,westillthinkthatthecaseforswitchingtoMLisweakbecauseitreliesonpreciseknowledgeof w i .
13
particularpaper(afterall,thesepapershaveallbeenpublishedinrespectablejournalsafterpeer
review).19Butwedothinkthatthelistservestoillustratetheproblemsthatarisewhenthespatial
IV/2SLSapproach,ofAnselin(1988),KelejianandPrucha(1998)andothers,isusedinpractice.The
sameissuesarisewithspatialGMM,whicharesimplyefficientversionsofIVestimators.
SowhatwaysforwardarethereforIVstrategies?Potentially,institutionalarrangementscanprovide
exogenousvariationinone(ormore)elementsof w i' X thathasnodirectinfluenceon yi .20For
instance,aresearchermightarguethattherearenodirectimpactsonoutcomesinadistrictfroma
policyinterventioninneighbouringdistricts(anelementof w i' X ),butthepolicydoeshaveeffects
viaitsinfluenceonneighbouringoutcomes.Asjustdiscussed,thisisthestrategyadoptedbysome
papersinthetaxcompetitionliterature(e.g.BesleyandCase1995andBrettandPinske,2000),
althoughwhetheraresearchercanconvinceothersthattherearenodirecteffectsfromneighbours
policiesdependsonthepolicyinquestion.21Sometimes,however,changinginstitutional
arrangementscanoffermoreconvincingnaturalexperiments.Aparticularlyniceexampleis
providedbyLyytikinen(2011)whoarguesthatchangesinthestatutorylowerlimittopropertytax
ratesinducesexogenousvariationintaxrates,whichcanbeusedtostudytaxcompetitionamong
localgovernmentsinFinland.Specifically,policychangestominimumtaxthresholdsinteractedwith
previoustaxrates,canbeusedtoinstrumentforthechangesintaxratesinneighbouringdistricts.In
thiscase,theexclusionrestrictionsaremoreplausible:adistricttaxauthorityisnotlikelytocare
howchangesintaxthresholdpolicyaffectedneighbours,exceptinsofarasitchangedthese
neighbours'taxrates.Particularlyinteresting,forourpurposes,isthatLyytikinencompareshis
estimatestothosebasedonspatiallagsandtraditionalspatialIVappliedtothesamedata(using
lagsofallthedeterminantsoftaxrates,notjustthepolicyinducedchanges).Whilehefindsno
evidenceofinterdependenceinpropertytaxratesfromhispolicybasedIVresearchdesign(which
contradictsmuchoftheliterature),hisspatialIVestimatesarelargeandsignificant.Heconcludes,
withadegreeofunderstatement,thatthestandardspatialeconometricsmethods[]
overestimatethedegreeofinterdependenceintaxrates.
AnotherinterestingpossibilityforSARtypemodelsemergeswhen w i' y representsexpectations
about yi insomespatialgroup,sincetheexpectationcouldbechangedbyadditionalinformation
about w i' y ,withoutchanging yi itself.Forexample,supposetheregistryofhousepricesbecomes
publicallyavailable,providingindividualswithnewinformationthatallowsthemtoreacttothesale
pricesofnearbyhouses.Orpoliceforcesintroducecrimemapping,whichallowsclosermonitoring
oftheactivitiesofcriminals.Sometimes,however,itmaystillbedifficulttojustifythatother
19
Indeed,oneoftheauthorshasatleastoneolderpaperthatsimilarlyadoptedtheKelejianandPrucha
(1998)IVapproach.
20
formulationwoulduseZtodenotetheseobservablecharacteristicsthathavenodirecteffectonyotherthan
throughtheireffectonX.
21
Thatis,ofcourse,assumingthattheymakeanyattempttojustifytheexclusionrestrictionsatall!
14
appliedcarefullywithattentiontotheidentificationofspecificcausalparametersratherthanan
arbitrarilyspecifiedsystemofequations.ForanIVstrategytogiveconsistentestimatesofthe
parametersofinterestinthesereducedformSLXmodels,instrumentsmustsatisfytheusual
relevanceandexclusionrestrictions.Wehavesaidenoughalreadyaboutexclusionrestrictionsand
weakinstrumentproblemsforittobeclearthatwedonotthinkthatIVstrategiesbasedonusing
higherorderspatiallagsof xi asinstrumentsareaverygoodidea.22Forthesereasons,webelieve
thatstandardIVstrategieswhichpaycarefulattentiontotheomittedvariablesandclearlyjustify
thevalidityofinstrumentsrepresentamoreappropriatewaytoaddresstheproblemofspatially
correlatedomittedvariables.23ManyeffectiveIVstrategiesofthistypemakeuseofpolicydesigns,
institutionalrulesandnaturalenvironmentalfeatures(orevenbetter,changesinthesefactors).
Therearemanyexamplesofthesekindsofwellthoughtthroughinstrumentingstrategiesappliedto
spatialproblems,byresearchersworkingoutsidethetraditionalspatialeconometricmould.For
example,Michaels(2008)notesthattheUShighwaysystemwasplannedonaregularEastWest,
NorthSouthgridconnectingmajorcities,implyingthattownslocateddueEast,West,Northor
Southofamajorcityincidentallyexperiencedlargechangesintransportaccessibilitybyvirtueof
theirpositionrelativetomajorcities.Luechinger(2009)usesthesitesofinstallationofSO2
scrubbersandprevailingwinddirectionstopredictpollutionlevels,inordertoestimatetheeffects
ofpollutiononindividualwellbeing.Theideahereisthatpeoplelivingdownwindofemissions
sourcesexperiencebigimprovementsinpollutionlevelsrelativetothoselivingupwind,when
emissionsreductiontechnologiesareinstalled,butthatthesedirectionsareotherwiseunrelatedto
wellbeing.Gibbons,MachinandSilva(2008)usedistancestoschooladmissionsdistrictboundaries
topredictlevelsofchoiceandcompetitioninschoolmarkets,onthebasisthatstudentsdonot
attendschoolsontheoppositesideofdistrictboundaries,sothenumberofschoolsfromwhich
studentscanchoose(withinagivendistance)istruncated.EarlierexamplesofthecreativeuseofIV
inspatialanalysisarefoundinCutlerandGlaeser(1997)andHoxby(2000),whousethenumberof
riverscuttingacrosscitiesasinstruments.Inbothcases,riversareassumedtobisectcommunities
leadingtogreaterracialsegregationwithincities(CutlerandGlaeser1997)ormoreschooldistricts
andmorechoiceandcompetitioninschoolmarkets(Hoxby2000).
AnotheralternativetoIVisspatialdifferencingtoremoverelevantomittedvariables,e.gthrough
differenceindifference,fixedeffectsorregressiondiscontinuitydesigns.Inthiscase,thefactthat
theunobservedcomponentisspatiallycorrelatedhelpsbecausespatialdifferencing(ofobservations
withtheirneighbours)islikelytobeeffective.Holmes(1998)providesanearlyexample.Gibbons,
MachinandSilva(2009)providemorerecentdiscussion.Otherdifferencingstrategiesdrawingona
casecontrolframeworkmayalsobeappropriate,forexamplethe'milliondollarplant'analysisof
Greenstoneetal(2008)whichcomparestheeffectoflargeplantrelocationsondestination
22
Note,however,thatwhilethespatialstructureofthedatadoesnthelp,neitherdoesitespeciallyhinderthe
searchforasuitableinstrument.Ifwehaveaninstrumentthatisindependentof ui ,thenitisalso
independentof w i u ,(unlesstheweightsareendogenous)sothefactthat xi and ui arebothspatially
correlatedisirrelevant(asidefromtheimplicationsforstandarderrors).
'
23
Controlfunctionapproachesmayalsobeequallyvalid,butrequireinstrumentstoo,andgenerallyrequire
moreassumptionsthanIV.
16
counties,usingtheirsecondrankedpreferencesrevealedinarealestatejournalfeatureasa
counterfactual.BothBusso,GregoryandKline(2010)andKolkoandNeumark(2010)evaluatethe
effectsofspatialpoliciesbycomparingpolicytreatedareaswithcontrolareasthatweretreatedin
laterperiods,asameanstogeneratingplausiblecounterfactuals.Spatialdifferencingcanalsobe
combinedwithinstrumentingasdiscussedin,forexample,Duranton,GobillonandOverman(2011)
andGibbons,McNallyandViarengo(2011).LeeandLemiuex(2010)providefurtherexamplesof
regressiondiscontinuitydesigns,manyofthemrelevanttospatialresearch.
Theexperimentaliststrategiessetoutabovemaynotbewithouttheirproblems(seeforexample
Kean2010),butatleastprovidesomehopeofuncoveringcausalrelationshipsinthespatialcontext,
whichofftheshelfspatialeconometricstechniquesdonot.Onecommoncriticismofthe
experimentaliststyleofresearch,asrepresentedbytheseexamples,concernsthegeneralizabilityof
theestimates.Thesedesignsseekoutsourcesofexogenous,pseudorandomvariationinthe
variablesofinterest,andinevitablyendupestimatingcausaleffectsforsomesubgroupofthe
population,exposedto,andrespondingtoaspecificinterventionordifferenceinenvironment.If
responsesareheterogeneous,thereisthereforenoguaranteethattheestimatedresponseis
representativeofamoregeneralpopulation.ThisissueiswellknowninthecontextofIV,givingrise
toLocalAverageTreatmentEffects(ImbensandAngrist1994).Evenso,wewouldarguethatitis
bettertohaveplausiblycausalestimatesforaspecificgroupinthepopulation,thantohavenon
causalparameterestimatesofunknownmeaning,estimatedbyothermethods.Ourremarksabout
thevalidityofexclusionrestrictionsinspatialIVabovealsoapplytoexperimentalistdesigns,
becausewestillhavetobesurethatwhatevercausesrandomvariationin xi ,affects yi only
through xi andnotthroughotherchannels.Todealwiththesethreats,spatialresearchdesignsin
theexperimentalistmodemustthinkthroughthespecificinstitutionaldetailsverycarefully,and
provideclearstatementsandtestsoftheidentifyingassumptions,andbecautiousingeneralising
resultstoowidely.
Tosummarise,differenteconomicmotivationsleadtospatialeconometricspecificationsthatwillbe
veryhardtodistinguishinpractice.Addtothemixthefactthatin(nearly)allapplicationsweface
uncertaintyabouttheendogeneityof xi ,theappropriatefunctionalformandspatialweightsandit
becomesclearwhymanyappliedresearchersfindMLorIVestimationofsomeassumedspatial
econometricspecificationuninformative.Instead,wesupportafocusonattemptingtosolve
identificationproblemsusingempiricalstrategiesthathavebeencarefullydesignedforthespecific
application.Further,ifempiricalstrategiescannotbedevisedthatsatisfactorilyidentifythecausal
impactofthespatiallagintheendogenousvariable(i.e.manyappliedsituations)thenweadvocatea
reducedformapproachpayingparticularattentiontotheproblemsraisedbyendogeneityofthe xi .
Opponentsofouradvocacyofexperimentalistapproachestospatialquestionsmightarguethat
placesareuniquebecauseoftheiruniquespatialposition,andsonotamenabletothesekindsof
researchdesigns.Theymayarguethatitisinfeasibletofindacomparatorlocationforanylocation,
givendifferencesinspatiallocation.Thispositionissurelytoopessimistic,sinceitrulesoutanyform
ofcausalempiricalanalysisonspatialdata,giventhatnocounterfactualcaneverbeconstructed.On
thecontrary,inalltheseexamplesdiscussedabove,thepurposeinthinkingthroughexperimental
settingstofind'comparator'orcontrolgroupsisnotnecessarilytofindcontrolplacesthatare
17
identicalineverywaytothe'treatment'places.Noristheaimnecessarilytofindsourcesof
variationthatarecompletelyrandom(i.e.instrumentsthatareuncorrelatedwitheveryother
characteristic),althoughthismightbetheideal.Instead,thegoalismoremodesttofind'control'
and'treatment'placesthatarecomparablealongthedimensionsthatinfluencetheoutcomebeing
studied.Similarly,instrumentsshouldbeuncorrelatedwiththeunobservedfactorsthatinfluence
theoutcome.Inshort,evenwhenweareconcernedthatthereareunobservedaspectsofspatial
locationthatdoinfluenceoutcomesandmakeplacesunique,therearestillpotentiallycausally
informativecomparisonsthatcanbemadebetweenneighbouringplaces,which,whilstnotidentical
arepotentiallyverysimilaralongsalientdimensions.
Sofarwehavesaidlittleabouttheroleoftheory.Manyspatialeconometriciansaredefensiveabout
theroletheoryplaysintheconstructionoftheirempiricalmodelsandseecommentsaboutthelack
oftheoryasamisguidedcriticismoftheirwork(e.g.seeCorradoandFingletoninthisjournal
volume).Buttheroleplayedbytheoryisnotourmaincriticism,ratheritisthefailuretoadopta
carefulresearchdesignthatsolvestheproblemsspecifictotheresearchquestionbeingaddressed,
andthelackofattentiontofindingcrediblesourcesofrandomorexogenousvariationinthe
explanatoryvariablesofinterest.Thisisnottosaywedonotthinkthattheoryisveryimportant,or
thatempiricalworkthathasproperbehaviouralstructuralfoundationsisuninformative(seeHolmes
2010forexamplesinregionaleconomics).Theorising,ofaformalisticormoreheuristictype,isof
courseessentialinorganisingthoughtsabouthowtodesignaresearchstrategyandtheoryand
assumptionsatsomelevelarenecessaryforanyempirics.Theoryisalsousefulonceyouhavethese
causalparameterestimatestohand,whenitcomestomakingpredictionsaboutgeneralequilibrium
effects,aslongasitismadeclearthatthesepredictionsarevalidonlyforthattheoreticalviewof
theworld.
Consistentwithouroverallapproach,wearguethattestingtheoriesmeanscorrectlyestimatingthe
coefficientsonspecificcausalvariables(assuggestedbythetheory).Thisprovidesanotherpointof
contrasttomostappliedspatialeconometricswheretheroleoftheoryistoderiveageneric
functionalformwithMLappliedtogivetheparametersthatensurethebestfittoreadilyavailable
data.Forexample,totestthepredictionsofNEGmodels,ourapproachinsistsonaresearch
strategytoidentifywhethermarketpotentialhasacausalimpactonwageswhilerecognisingthat
nomodelisgoingtocompletelyexplainthespatialdistributionofwages.Thiscontrastsstrongly
withtheappliedspatialeconometricsapproachwhichusestheextenttowhichdifferentspatial
econometricmodelsfitthedataasawaytotestcompetingtheories.Thishastheunfortunateside
effectofencouragingtheinclusionofendogenousvariablesinempiricalspecificationsas,for
obviousreasons,thesetendtoincreasethefitofthespatialmodelwiththedata.
Inmanyspatialeconomicproblems,theorymaythusplayanimportantroleinidentifyingvariables
forwhichwewouldliketoknowthecausaleffects.Butempiricalimplementationrequirescareful
researchdesigniftheresultsaretohaveanygeneralscientificcredibilityortobeconsidered
trustworthyforpolicymaking.Itissurelywrongtousespecialisedtheoryalonetoimposespecific
restrictionontheresearchdesign(e.g.byassumingawaypotentiallyconfoundingsourcesof
variation)unlessyouhavereasonableconfidencethatthetheoryiscorrectandthatitis
demonstrablysotoageneralaudience.Unfortunately,thisistheroleplayedbytheoryinmuch
appliedspatialeconometricresearch.Theoryisusedtojustifytheinclusionofaspatiallag,
18
assumptionsaremadeabouttheformofthespatialweightmatrix(possiblyderivedfromtheory),
MLisusedtoachieveidentificationandthenmodelfitisusedasabasisfortestingtheorywhich
justifiedtheinclusionofthespatiallag.Itshouldbeclearbynowthat,formostspatialproblems,we
simplydonotfindthisaconvincingapproach.Withoutwishingtoweighfurtherintothevigorous
debateonstructuralversusexperimentalapproachestoempiricalwork(e.g.JournalofEconomic
Perspectives,Vol24(2)2010)wesimplymakethepointthatwhatevermethodisadopted
(structural,experimental,qualitativeoranyother)anyempiricalresearchthataimstofindoutifx
causesyneedstofindasourceofexogenousvariationinx!
5.CONCLUSIONS
Wehavearguedthatidentificationproblemsbedevilmostappliedspatialeconometricresearch.
Manyspatialeconometriciansaresurelyawareoftheseproblemsbuttheliterature(inadvertently)
downplaystheirimportancebecauseofthefocusonderivingestimatorsassumingthatfunctional
formsareknownandbyusingmodelcomparisontechniquestochoosebetweencompeting
specifications.Whilethisraisesinterestingtheoreticalandcomputationissuesthathavebeenthe
subjectofagrowingandthoughtfulformaleconometricliterature,itdoesnotprovideatoolboxthat
givessatisfactorysolutionstotheseproblemsforappliedresearchersinterestedincausalityandthe
economicprocessesatwork.Itisinthissensethatwecallintoquestiontheapproachofthe
burgeoningbodyofappliedeconomicresearchthatproceedswithmechanicalapplicationofspatial
econometrictechniquestocausalquestionsofspilloversandotherformsofspatialinteraction,or
thatestimatesspatiallagmodelsasaquickfixforendogeneityissues,orthatblindlyappliesspatial
econometricmodelsbydefaultwithoutseriousconsiderationofwhetherthisisnecessarygiventhe
researchquestioninhand.Whilethequestionweposeinthetitletoourpaperisdeliberately
provocativeandtongueincheek,wemaintainthatthismodeofspatialeconometricwork,whilst
maybenotpointless,isoflimitedvaluewhenitcomestoprovidingcredibleestimatesofcausal
processesthatcanguideunderstandingofourworld,andguidepolicymakersonhowtochangeit.
Weurgethoseconsideringembarkingdownthisroutetothinkagain.
Paradoxically,wethinkthatusingthestandardspatialeconometricspecifications(adapted,aswe
havedonethroughoutthetext,toreinforcethefocusonthecausalfactorsthatdriveoutcome yi )
helpsclarifyidentificationproblemsforthoseresearcherswhoareinterestedincausality.In
particular,wethinkthatcloserattentiontomodelspecificationwillbehelpfulinunderstandingthe
exclusionandrelevanceassumptionsthatunderlieIVapproaches.Spatialeconometricsalso
providesimportantinsightsonthecorrectinterpretationofmodelparametersthatmaybe
identifiedfromsomesuitableestimationstrategy.Inshort,therearelessonstobelearntfromthe
spatialeconometricsliteraturebutformostappliedeconomicresearcherstheappropriatestrategy
shouldbebasedontheexperimentalistparadigmwhichputsissuesofidentificationandcausalityat
centrestage.
19
Appendix:Spatialeconometricsasasolutiontotheomittedvariables
problem
Assume yi dependsonexogenous xi andunobservable zi .Thatis:
yi xi zi
(A1)
Further,assumethattheunobservable zi isbothspatiallycorrelatedandpartlydeterminedby xi :
zi wi' z xi vi
(A2)
substituting zi yi xi into(6)andrearranginggives
(A4)
(A5)
(A6)
ThisequationcanbeestimatedconsistentlybyMLornonlinearleastsquares,orestimatesofthe
variousparametersretrievedfromtheOLScoefficients.Althoughendogeneityproblemsofthistype
mightbemitigatedbythisstrategy,itiscertainlynotacompletefix.Toseethis,modifythesetupin
equations(A1)(A3)slightlytocopewithmoregeneralendogeneityinthat xi ispartlydetermined
bytheomittedvariable( fi ).Inthiscasewehave:
yi xi zi
(A7)
20
zi w i' z f i vi
xi fi ui
21
REFERENCES
Ahlfeldt,G.(2011)IfWeBuild,WillTheyPay?PredictingPropertyPriceEffectsofTransport
Innovations,LondonSchoolofEconomics,SpatialEconomicsResearchCentreDiscussionPaper0075
Angrist,JoshuaD.&Krueger,AlanB.(1999)"Empiricalstrategiesinlaboreconomics,"in:O.
Ashenfelter&D.Card(ed.),HandbookofLaborEconomics,edition1,volume3,chapter23,pages
12771366Elsevier.
Angrist,J.andJ.S.Pischke(2009).MostlyHarmlessEconometrics.Princeton,NJ:PrincetonUniversity
Press.
Angrist,J.andJ.S.Pischke(2010).TheCredibilityRevolutioninEmpiricalEconomics:HowBetter
ResearchDesignIsTakingtheConoutofEconometrics,JournalofEconomicPerspectives,24,330.
Anselin,L.(1988)Spatialeconometrics:methodsandmodels.London:Kluwer.
Anselin,L.(2010Thirtyyearsofspatialeconometrics,PapersinRegionalScience,89.
Arbia,G.(2006)SpatialEconometrics:StatisticalFoundationsandApplicationstoRegional
Convergence(AdvancesinSpatialScience),Springer.
Behrens,K.andF.RobertNicoud(2009)KrugmansPapersinRegionalScience:The100dollarbill
onthesidewalkisgoneandthe2008NobelPrizewelldeserved.PapersinRegionalScience88467
489.
Besley,T.andA.Case(1995)DoesElectoralAccountabilityAffectEconomicPolicyChoices?
EvidencefromGubernatorialTermLimits,TheQuarterlyJournalofEconomics110(3):769798
Bogart,W.T.andB.A.Cromwell(2000),HowMuchisaNeighborhoodSchoolWorth?,Journalof
UrbanEconomics,47(2)280305
Bound,J,D.A.JaegerandR.M.Baker(1995)ProblemsWithInstrumentalVariablesEstimationWhen
theCorrelationBetweentheInstrumentsandtheEndogenousExplanatoryVariableIsWeak,
JournaloftheAmericanStatisticalAssociation,90(430)443450
Brett,C.andJ.Pinske(2000),TheDeterminantsofMunicipalTaxRatesinBritishColumbia,
CanadianJournalofEconomics,33,695794.
Brock,W.A.andS.N.Durlauf(2001).Interactionsbasedmodels,in:J.J.Heckman&E.E.Leamer
(ed.),HandbookofEconometrics,edition1,volume5,chapter54,pages32973380Elsevier.
Busso,M.,J.GregoryandP.M.Kline(2010)"AssessingtheIncidenceandEfficiencyofaProminent
Place,"NationalBureauofEconomicResearchWorkingPaper16096
CalvoArmengol,Antoni,E.PatacchiniandY.Zenou(2009)PeerEffectsandSocialNetworksin
Education,ReviewofEconomicStudies,76(4)12391267
22
Charlot,S.andPaty,S.(2007)Marketaccesseffectandlocaltaxsetting:evidencefromFrench
paneldata,JournalofEconomicGeography,7(3),pp.117.
Combes,P.P,GDurantonandLGobillon(2011)Theidentificationofagglomerationeconomies.
JournalofEconomicGeography11(2):253266
Cutler,D.M.andE.L.Glaeser(1997)AreGhettosGoodorBad?TheQuarterlyJournalof
Economics,112(3),82772
Dallerba,S.andJ.LeGallo(2008).RegionalconvergenceandtheimpactofEuropeanstructural
fundsover19891999:Aspatialeconometricanalysis,PapersinRegionalScience,87(2),119244.
Duranton,G.,L.GobillonandH.G.Overman(2011).Assessingtheeffectsoflocaltaxationusing
microgeographicdataEconomicJournal,Forthcoming.
Edmark,Karin&gren,Hanna(2008)."IdentifyingstrategicinteractionsinSwedishlocalincometax
policies,"JournalofUrbanEconomics,Elsevier,vol.63(3),pages849857,
Feld,L.P.andE.Reulier(2009).StrategicTaxCompetitioninSwitzerland:EvidencefromaPanelof
theSwissCantonsGermanEconomicReview10(1),91114,February2009
Fingleton,B.2008.TestingtheNEGModel:FurtherEvidencefromPanelData,SpatialEconomics
ResearchCentreDiscussionPaperno.0005.
Fingleton,B.andLeGallo,J.(2010)EndogeneityinaSpatialContext:PropertiesofEstimatorsin
AntonioPez,RonN.Buliung,JulieLeGalloandSandyDallerbaed.sProgressinSpatialAnalysis:
MethodsandApplicationsAdvancesinSpatialScience,2010,Part1,5973SpringerHeidelberg
DordrechtLondonNewYork
FingletonandCorrado(201x)JRSTOPROVIDEREFERENCETOPUBLISHEDVERSION
Gibbons,S.Lyytikainen,T.,R.SanchisGuarner,H.G.Overman,J.Laird.EvaluatingtheProductivity
ImpactsofRoadTransportSchemes,ReportforDepartmentofTransport,August2010
Gibbons,S.,S.MachinandO.Silva(2008)"ChoiceCompetitionandPupilAchievement",Journalof
theEuropeanEconomicAssociation,6(4)912947
Gibbons,S.,S.MachinandO.Silva(2009)"ValuingSchoolQualityUsingBoundaryDiscontinuities",
LondonSchoolofEconomicsSpatialEconomicsResearchCentreSERCDP0018
Gibbons,S.,S.McNallyandM.Viarengo(2011)DoesAdditionalSpendingHelpUrbanSchools?An
EvaluationUsingBoundaryDiscontinuities,LondonSchoolofEconomics,SpatialEconomicsResearch
CentreDiscussionPaperno.0090.
Gibbons,S.andH.G.Overman(2011)MostlyPointlessSpatialEconometrics,SpatialEconomics
ResearchCentreDiscussionPaperno.0061
Gibbons,S.andS.Telhaj(2008)"PeersandAchievementinEngland'sSecondarySchools",London
SchoolofEconomicsSpatialEconomicsResearchCentreSERCDP0001
23
DeGiorgio,G.,MPellizzariandS.Redaelli(2010)"IdentificationofSocialInteractionsthrough
PartiallyOverlappingPeerGroups,"AmericanEconomicJournal:AppliedEconomics2(2)24175
Greenstone,M,R.Hornbeck,andE.Moretti(2010)"IdentifyingAgglomerationSpillovers:Evidence
fromMillionDollarPlants",JournalofPoliticalEconomy,18(3)536598
Hayashi,M.andR.Boadway2001.AnEmpiricalAnalysisofIntergovernmentalTaxInteraction:The
CaseofBusinessIncomeTaxesinCanada,CanadianJournalofEconomics,34,481503.
Holl,A.(2004)Manufacturinglocationandimpactsofroadtransportinfrastructure:empirical
evidencefromSpain,RegionalScienceandUrbanEconomics,34(3)341363
Holmes,T.(1998)TheEffectofStatePoliciesontheLocationofManufacturing:Evidencefrom
StateBorders,JournalofPoliticalEconomy,106,667705.
Holmes,T.(2010)Structural,ExperimentalistandDescriptiveApproachestoEmpiricalWorkin
RegionalEconomics,JournalofRegionalScience,50(1)522
Hoxby,C.(2000).DoesCompetitionAmongPublicSchoolsBenefiStudentsandTaxpayers?
AmericanEconomicReview,90(5),120938.
Imbens,G.,Angrist,J.(1994)IdentificationandEstimationofLocalAverageTreatmentEffects,
Econometrica62(2)467476
Keane,M.P.(2010)Structuralvs.AtheoreticApproachestoEconometrics,JournalofEconometrics
156(1)320
Kelejian,H.H.andI.R.Prucha(1998)AGeneralizedSpatialTwoStageLeastSquaresProcedurefor
EstimatingaSpatialAutoregressiveModelwithAutoregressiveDisturbances,JournalofRealEstate
FinanceandEconomics,17(1)99121
Kling,J.R.,Ludwig,J.,&Katz,L.F.(2005).NeighborhoodEffectsonCrimeforFemaleandMale
Youth:EvidencefromaRandomizedHousingVoucherExperiment.QuarterlyJournalofEconomics,
120(1),87130
Kling,J.R.,JeffreyB.LiebmanandLawrenceF.Katz(2007)ExperimentalAnalysisOfNeighborhood
Effects,Econometrica,75(1)83119
Kolko,J.andD.Neumark(2010)"Doenterprisezonescreatejobs?EvidencefromCalifornias
enterprisezoneprogram,"JournalofUrbanEconomics68(1)119
Lee,L.(2004).AsymptoticDistributionsofQuasiMaximumLikelihoodEstimatorsforSpatial
EconometricModels,Econometrica,72,18991926.
Lee,L.F.(2007).Identificationandestimationofeconometricmodelswithgroupinteractions,
contextualfactorsandfixedeffects.JournalofEconometrics140(2)33374.
Lee,L.F,X.Liu,andX.Lin(2010)SpecificationandEstimationofSocialInteractionModelswith
NetworkStructures.TheEconometricsJournal13(2)145176
24
Lee,D.S.,andT.Lemieux(2010)"RegressionDiscontinuityDesignsinEconomics."Journalof
EconomicLiterature,48(2)281355.
Leprince,M.&T.Madis&S.Paty(2007)."BusinessTaxInteractionsAmongLocalGovernments:An
EmpiricalAnalysisOfTheFrenchCase,"JournalofRegionalScience,WileyBlackwell,vol.47(3),
pages603621.
LeSage,J.andR.K.Pace(2009)IntroductiontoSpatialEconometrics.NewYork:CRCPress.
Luechinger,S.(2009)"ValuingAirQualityUsingtheLifeSatisfactionApproach,"TheEconomic
Journal,119(536)482515
Lyytikinen,T.(2011)TaxCompetitionAmongLocalGovernments:EvidencefromaPropertyTax
ReforminFinlandSERCDP0082.
Machin,S.andK.Salvanes(2010)ValuingSchoolQualityViaSchoolChoiceReform,LondonSchool
ofEconomics,CentreforEconomicsofEducationDiscussionPaperCEEDP113
Manski,C.(1993)IdentificationofEndogenousSocialEffects:TheReflectionProblem,TheReview
ofEconomicStudies,60(3),531542.
Manski,C.(2000)EconomicAnalysisofSocialInteractions,TheJournalofEconomicPerspectives,
14(3),115136
Martin,R.(1999)Criticalsurvey.Thenew'geographicalturn'ineconomics:somecritical
reflectionsCambridgeJournalofEconomics,23,6591.
McMillen,D.P.(2010a)"IssuesinSpatialDataAnalysis,"JournalofRegionalScience,50(1),119141
McMillen,D.P(2010b)PerspectivesonSpatialEconometrics,unpublishedmanuscript,University
ofIllinois
Michaels,G.(2008)"TheEffectofTradeontheDemandforSkill:EvidencefromtheInterstate
HighwaySystem,"ReviewofEconomicsandStatistics,90(4)683701.
Ord,J.K.(1975)EstimationMethodsforModelsofSpatialInteraction,JournaloftheAmerican
StatisticalAssociation,70,120126.
Pinske,J.AndM.E.Slade(2010)TheFutureofSpatialEconometrics,JournalofRegionalScience,
50(1),103117.
Redding,S.J.andD.Sturm.(2008)"TheCostsofRemoteness:EvidencefromGermanDivisionand
Reunification,"AmericanEconomicReview,98(5)176697
Revelli,F.(2005)OnSpatialPublicFinanceEmpirics,InternationalTaxandPublicFinance12,475
492.
Sacerdote,B.(2001)"PeerEffectswithRandomAssignment:ResultsforDartmouthRoommates,"
QuarterlyJournalofEconomics,116(2)681704
25
StaigerandStock(1997)"InstrumentalVariablesRegressionwithWeakInstruments,"Econometrica,
65(3),557586.
Sutton,J.(2002)MarshallsTendencies:WhatcanEconomistsKnow?MITPress,
Todd,P.andK.Wolpin(2003)OntheSpecificationandEstimationoftheProductionFunctionfor
CognitiveAchievement,EconomicJournal,113,F3F33.
26