FINANCIALCONSTRAINTSAND INTANGIBLEINVESTMENT By WilliamGrieser hi hi ADISSERTATION Submittedto MichiganStateUniversity inpartialentoftherequirements forthedegreeof BusinessAdministration-Finance-DoctorofPhilosophy 2015 ABSTRACT FINANCIALCONSTRAINTSANDINTANGIBLEINVESTMENT hi By hi WilliamGrieser Thisdissertationiscomposedofthreeessaysconcerningtheroleofncialcon- straintsinainvestmentpolicy.Intheessay,Iconsidertheroleofthe marketforcreditdefaultswaps(CDS)inrelaxingaconstraintsand spurringinvestmentinintangibleassets.IthatCDStradingisassociatedwith anincreaseinresearchanddevelopment(R&D)spendingandpatentingactivity. ThisisconsistentwithachannelinwhichCDStradingprotectscreditorswhenbor- rowers'assetshavelimitedcollateralvalue.Inmysecondessay,Ipresentevidence thatthepresenceofconstraintsforacompetitorstendstospurinvest- mentspending.IuseataxlawchangetoidentifyexogenouschangesinFinancial Constraints(FC)toestablishcausality.Thisevidencesuggeststhatproductmarket considerationsinteractwithconstraintsinthedeterminationofboth levelandindustry-levelinnovation.Mythirdessayisamethodologicalinvestigation pertainingtoallstudiesthatusepaneldatatechniques,includingstudiesof cialconstraintsandinvestmentdecisions.IinvestigatewhethertheStrictExogeneity (SE)assumption,anecessaryconditionforconsistentparameterestimatesusingcom- monestimationtechniques,holdsincanonicalpaneldataregressions.I thatthiskeyassumptionisviolatedquitefrequentlyandcanleadtosubstantivebi- asesinparameterestimates.Isuggestageneralapproachforideninthese situationsthatexploitsindustry-timevariationinthevariableofinterest. DedicatedtomywifeAnne,andtomyparentsRobertandKathyGrieser iii ACKNOWLEDGEMENTS hi IamgratefultoAndresAlmazan,DamienBrooks,JohnathanCohn,CesareFracassi, CharlieHadlock(Co-Chair),ZackLiu,GonzaloMaturana,JordanNickerson,Joerg Picard,JoshPollet,AvishaiScAndreiSimonov(Co-Chair),ParthVenkat, Wooldridge,HayongYun(CommitteeMember),andMiriamSchwartz-Ziv(Commit- teeMember)forusefulcommentsandguidance. iv TABLEOFCONTENTS LISTOFTABLES :::::::::::::::::::::::::::::::: vii CHAPTERITheRealofCreditDefaultSwaps ...... 1 1.1Introduction........................... 1 1.2DataandSampleConstruction................. 7 1.3EmpiricalMotivationandDiscussionofResults........ 10 1.3.1CreditDefaultSwapsandFirmInnovation..... 10 1.3.2AccesstoUnsecuredDebt.............. 12 1.3.3EndogeneityandPotentialSolutions......... 16 1.3.4LenderAnalysis.................... 19 1.4Conclusion............................ 20 CHAPTERIIPeerofFinancialConstraintsonInnovation 22 2.1Introduction........................... 22 2.2DataandSummaryStatistics.................. 27 2.2.1FirmRelationshipswithPatentCitations. 29 2.2.2PatentingResponsestoFinanciallyConstrainedPeers 30 2.2.3Iden...................... 32 2.2.4PatentPortfolioSimilarity.............. 34 2.2.5SelfCitations..................... 35 2.3Conclusion............................ 36 CHAPTERIIIPanelDataEstimationinFinance ........... 40 3.1Introduction........................... 40 3.2Priorliteratureandempiricalstrategy............. 45 3.2.1Outliningtheproblem................. 45 3.2.2Recent/currentpractice................ 47 3.2.3Priorworkinthataccountsforviolationsof strictexogeneity.................... 48 3.3Testingforstrictexogeneity.................. 49 3.3.1Identifyingasetofcanonicalregression....... 49 3.3.2Testingforstrictexogeneity............. 51 3.3.3Insightsonmagnitudes................ 53 3.3.4Hopingfora1/Tsave................. 55 3.3.5Robustnessandextensions.............. 56 3.4Usingindustry-yearvariationforidenincertainsettings 57 3.5Conclusion............................ 59 APPENDIX :::::::::::::::::::::::::::::::::::: 61 v BIBLIOGRAPHY :::::::::::::::::::::::::::::::: 89 vi LISTOFTABLES TableA.1SummaryStatistics......................... 62 TableA.2CreditDefaultSwapsandInnovation............... 63 TableA.3UnsecuredandSecuredLoans:MultinomialLogitEstimates.. 64 TableA.4CDSandSecuredvs.UnsecuredFinancing............ 65 TableA.5CDSandUnsecuredFinancing:LinearProbabilityEstimates.. 66 TableA.6insfor2003ISDAprovisions........ 67 TableA.7ImpactofCDSonUnsecuredDebt:ParameterEstimates.... 68 TableA.8ImpactofCDSonUnsecuredDebt:Marginal...... 69 TableA.9UnsecuredFinancingandFirmTangibility............ 70 TableA.10LenderSizeandtheImpactofCDSonAccesstoCapital.... 71 TableA.11SummaryStatistics......................... 72 TableA.12FinancialConstraintsOLSRegressions.............. 73 TableA.13AJCATaxHoliday......................... 74 TableA.14AJCA2............................... 75 TableA.15JunkBond.............................. 76 TableA.16JunkBond2............................. 77 TableA.17SelfCitationPercentage...................... 78 TableA.18SelfCitationPerPatent...................... 79 TableA.19MahalanobisDistance....................... 80 TableA.20MarketValue............................ 81 TableA.21LargeMovesinFC......................... 82 TableA.22ModelSp........................ 83 TableA.23FullSampleTests.......................... 83 TableA.24Sub-sampleTests.......................... 84 TableA.25SignBetweenFDandFE............... 85 vii TableA.26FE i correlations10yearsub-periods.............. 86 TableA.27FE i correlations5yearsub-periods............... 87 TableA.28FE i correlations5yearsub-periods............... 88 viii CHAPTERI TheRealofCreditDefaultSwaps 1.1 Introduction Docreditdefaultswaps(CDS)haverealeconomicIfmarkets arecomplete,thencreditdefaultswapsareredundantsecurities,makingtheirexis- tenceinconsequential.Inthepresenceofmodestfrictionshowever,CDShavebeen arguedtoreduceconstraintsforlendersandincreasecreditsupply[ Hirtle (2009),and Jarrow (2011)].Furthermore,evidencesuggeststhattheserelaxedcreditconstraints canleadtochangesinthecapitalstructureofwithaCDStradingontheirdebt [ Saretto (2013)].WhileCDSmarketsmaycreditsupplyand,inturn,a capitalstructure,itisunclearwhethertheseultimatelyencecorporate investment. Inthispaper,Iexploitcross-sectionalvariationinaabilitytopledgeassets ascollateraltoexaminetheofCDStradingonborrowingandinvestment decisions.Spy,Ishowthatswithanintangibleassetbaseseeimproved accesstounsecuredcingafteraCDSstartstradingontheirdebt.Interestingly, thisimprovedaccesstounsecureddebtisfollowedbyanincreaseininnovationand investmentspendingonintangibleassets,measuredviapatentingactivityandR&D spending. Iarguethatthelikelychannelthroughwhichcreditdefaultswapshaveanim- pactoncorporateandinvestmentpolicyisthealleviationofcollateralcon- straints.BothleverageandinvestmentaremoreresponsivetoCDStradingfor investinginintangibleassets.Theseresultsareconsistentwiththetheoreticalwork 1 of BoltonandOehmke (2011),whoarguethatCDScanrelaxfrictionswhen haveimperfectaccesstocapitalduetoalimitedcommitmentproblem.While previousliteraturehasidenedchannelsthroughwhichCDScanimpact frictionsonthesupplysideofcreditmarkets,Iprovidetheevidencethatthese interactwithdemandsideconstraintstocorporateinvestment. ThisevidencesuggeststhatthepresenceofanactiveCDSmarketmitigatescollateral constraintsandhasarealeconomicimpactbyfacilitatinginnovationandinvestment forcompanieswithalimitedabilitytopledgefuturecashws. Alargebodyofliteraturehasarguedthatcollateralisanimportantdeterminant ofdebtbecauseitprovidescreditorstherighttoliquidateassetsinthe eventofdefault[e.g. BoltonandScharfstein (1990); HolmstromandTirole (1997)]. Itisgenerallyperceivedthatthefunctionoftransferringliquidationrightsistopro- tectlenderswheninformationalasymmetriespreventobservation(vofa borrowersperformance[ HartandMoore (1994)].Collateralconstraintscanoccurif ahasalreadypledgedallofitsassetsonpriorloansorifaassetshave limitedcollateralvalue.Assetintangibilityisbelievedtonegativelya capacitytocollateralize[e.g. AlmeidaandCampello (2007)].Intangibleassetsmay requirespecialmanagement,becostlytoevaluate,andbetoredeploywhen comparedtotangibleassets.Thesecharacteristicsresultinlessvalueforcreditors indefaultstates. 1 Ifcollateralcircumventsinformationalasymmetriesthatrestricta accesstocapital,collateralconstraintscouldleadtong,andthus underinvestment,forinvestinginintangibleassets. Unlikecollateral,CDScanincreasetherangeofprojectsthatreceivecing, evenforwithalimitedabilitytocollateralizeassets.Creditdefaultswapscan serveafunctionsimilartothatofcollateralbytransferringbargainingpowertocred- itorsduringdebtrenegotiations,raisingaborrowerspledgeableincome[ Boltonand 1 see ParsonsandTitman (2009)foranindepthsummary 2 Oehmke (2011)].Theabilityofcreditorstoinitiate(purchase)aCDScontracton adebtisdependentonthedepthoftheCDSmarketforthatandisnot directlyrelatedtotheunderlyingabilitytopledgecashws. 2 Consequently, creditdefaultswapsmaypartiallyalleviatetheandunderinvestment problemthatisduetocollateralconstraints.Cross-sectionalvariationinthepledge- abilityofaassetsisthensuggestiveofvariationintheimpactthatCDStrading shouldhaveonrelaxingaconstraints. AlthoughrelaxingcollateralconstraintsisaclearchannelthroughwhichCDScan impactcorporateinvestment,theempiricalliteraturehasyettodocumentthis tothebestofmyknowledge.Fillingthisvoidistheprimaryobjectiveofmyanalysis. Ifollowtherelatedworkof AshcraftandSantos (2009)and Saretto (2013),usingthe introductionofCDSmarketsasatreatmenttostudytheimpactofCDSon leveloutcomes. IanalyzeprivateloansandthataftertheinceptionofCDStrading,are 7.62%morelikelytoobtainaloan.Interestingly,thiscorrespondstoa9.08%increase intheuseofunsecureddebtanda2.51%decreaseintheuseofsecureddebtfor withanactiveCDSmarket.Theseresultsareconsistentwithpriorevidenceby Hirtle (2009)and Saretto (2013)thatCDScanimproveaccesstodebtbut ithighlightsthatthemaybeparticularlyrelevantforwithaninabilityto postcollateral.Toverifythatthisisthecase,Ishowthatthemagnitudesofthe areincreasinginmeasuresofinabilitytocollateralize,suchasassetintangibility,R&D intensity,andTobinsQ.Asanadditionalmeasure,Iestimatetheunusedpledgeable assetbaseofaandshowthattheimprovedaccesstounsecureddebtismore pronouncedforthathaveexhaustedtheirpledgeableassetsonpreviousloans. OnemightbeconcernedthatthesearedrivenbythepotentialthatCDS tradeonwithlessinformationasymmetry,sincesucharelikelytoobtain 2 TheinabilityofatocollateralizeassetscaninthedepthoftheCDSmarkettothe extentthatitencouragesspeculation. 3 unsecureddespitehavingtheabilitytopledgecollateral.Iwouldargue thatthisisincontrastwiththebeinglargerforwithalesstangibleasset baseandforR&DintensiveNonetheless,toalleviatetheconcernIcontrolfor measuresofinformationasymmetrysuchasanalystforecastdispersion,PIN,implied volatility,S&P500inclusion,andcreditrating.Furthermore,theresultsholdfor withinestimates,suggestingthatwithinthesameaccesstounsecureddebt increasesaftertheinceptionofCDStrading. Mostimportantly,Ithatrespondtorelaxedconstraintsby increasingR&Dspending(innovationinput)andpatentingactivity(innovationout- put).Sp,thepresenceofanactiveCDSmarketleadstoa24.94%increase inR&Dspendingasapercentageofsales(inthousands),correspondingtoa1.3% increaseinR&Dasapercentageofassets.Inaddition,ageneratesroughly 2additionalpatentsperemployeeand1.4additionalpatentcitationsperemployee aftertheintroductionofaCDSontheirdebt.Theresultsareobtainedforboth estimation(i.e.andbetwestimationwithinan industry(CDS-yearstonon-CDSearswithinanindustry). AnotherpotentialconcernisthattheemergenceofaCDSmarketissimultaneously determinedwithaleverageandinvestmentdecisions.Whilecreditdefault swapstradeover-the-counter(OTC),andhaveessentiallynocontrolwhethera CDSistradingontheirdebt,aCDSmarketcouldmaterializeinresponsetochanges inam'sriskAnychangesincapitalstructureorinvestmentpolicycould ultimatelychangetheriskofaandsimultaneouslyencouragespeculationinthe CDSmarketforthat 3 Iemploythreemethodstoalleviatethisconcern.First, followingtheapproachusedby AshcraftandSantos (2009)and Saretto (2013),I exploitthetimingoftheintroductionofCDSbyexaminingaccesstounsecureddebt intheyearafteraCDSbeginstotrade.ForR&DspendingandpatentingactivityI 3 OehmkeandZawadowski(2014)documentthatspeculationisoneofthetwoprimaryfunction ofCDSmarkets. 4 usea1-5yearwindowsinceR&Dandpatentingareinherentlyslowerprocessesthan theprocessofincreasingdebt.ItmightstillbethecasethatCDSmarketsareforward lookingandbegintradingwellinadvanceofashiftsinleverageorinvestment decisions.Usingtheof Hilscher (2014),whoshowthatinformationinequity marketsleadsinformationinCDSmarkets,Iincludeforwardlookingimpliedvolatility asacontrolvariableinallregressions. Asanaddedmeasure,Iuseainanalysiswiththe2003and 2009ISDAprovisionsasatreatmentTheseprovisionsdrasticallyenhanced theliquidityofaCDScontract,andthereforelikelihoodthataCDSwouldtrade onagivenAdditionally,itishardtoimaginethattheseprovisionsarerelated toaR&Dspendingorpatentingactivity,otherthanthroughtherelaxationof frictionsbymakingCDScontractsmorestandardizedandmoreliquid. Myanalysisdrawsontwodistinctandgrowingliteratures,theliteratureexamin- ingtheimpactoffrictionsoninnovation,andtheliteratureonunderstand- ingtheimpactofcreditderivatives. Intheliteratureonginnovation, Manso (2011)arguesthatinnovation requirestolerance(orevenreward)forshorttermfailureandanemphasisonlong termgrowth.Alongtheselines, Holmstrom (1989)suggeststhatpublicmay innovationbyover-emphasizingshorttermgainsattheexpenseoflongrunde- velopment.However, RajanandZingales (2003)arguethatadverseselection,moral hazard,andbanks'inabilitytounderstandinnovativemakeinnova- tionthroughprivatedebtmoreproblematicthaninnovationthroughpublic markets[see Atanassov (2014)].Mypaperaddstotheseargumentsbyprovidingem- piricalevidencethatanactiveCDSmarketcanmitigatefrictionsduetoasymmetric informationandallowbankstotoinnovativeMythat CDSresultinimprovedaccesstoprivatedebtforinnovativewhichisfollowed byincreasedpatentingactivityandR&Dspending,lendssupportingevidencetothe 5 argumentsof Holmstrom (1989)and Manso (2011),thatshorttermoriented islessconducivetoinnovation. Theliteratureaimedatunderstandingtheimpactofcreditderivativeshasseen tremendousgrowthinrecentyears.Jarrow,2013describesconditionsinwhichcredit derivativesarewelfareimproving. Hilscher (2014)showthatinformationinequity marketsleadsthatofCDSmarkets. AshcraftandSantos (2009)thatCDShave littleonthecostofborrowingforwithanactiveCDSmarketontheirdebt. Twoarticlesinthisstrandofliteratureareparticularlyrelatedtomyanalysis,an empiricalarticlebySarettoandTookesandatheoryarticlebyBoltonandOehmke. Saretto (2013)provideempiricalevidencethatwithaCDStradingontheir debtareabletoincreaseleverageandlengthentheirdebtmaturity.Ibuildontheir inthreeways.First,whiletheiranalysisemphasizessupplysideof CDSsuchasrisksharing,creditrisktransfer,andcapitalrequirementreductionsto studycapitalstructuredecisions,Ifocusondemandsideconstraintstoseeif CDSmarketshaveimplicationsforinvestmentdecisions.Sp,Ishow thatCDScanhaveimplicationsoninnovation.Second,Iexploitcross-sectional variationinrms'abilitytopledgecollateraltotestpredictionsonaccesstounsecured ByexaminingaccesstounsecureddebtinadditiontototalleverageIcan providemoredetailedanalysisastowhicharebyCDStradingandhow. MyevidencethatCDSimproveaccesstounsecureddebtandthesubsequentincrease inspendingonintangibleinvestmentsuggeststhatCDSrelaxcingfrictionsfor constrainedratherthansimplychangingtherelativecostsofsources foralreadyatoptimalinvestmentlevels.Tothebestofmyknowledge,my resultsprovidetheempiricalevidencethatCDSmarketscanhavereal oninvestment.Finally,IcorroboratingevidenceforSaretto'sandTookes' resultsontotalleverageusingalargersample. BoltonandOehmke (2011)developatheoreticalmodelinwhichcreditdefault 6 swapstransferbargainingpowertocreditors.Theirmodelgeneratesempirically testableimplications.Ononehand,thetransferofbargainingpowertocreditors canincreasethepledgeableincomeofborrowersandincreasetherangetheprojects thatreceiveancing.Ontheotherhand,thiscanleadtoanemptycreditorproblem inwhichcreditorsoverinsure,makingthemexcessivelytoughduringdebtrenego- tiations.Inthelattercasemaychoosetodecreaserisk,ratherthanincrease investment,inordertoavoidexcessivelyharshcreditorsindefaultstates.Theempty creditorproblemisthefocusofBolton'sandOehmke'spaper. Myanalysiscanlargelybeviewedasanempiricaltestoftheimplicationof Bolton'sandOehmke'smodel,butwithafewadditionalfeatures.First,Iarguethat thecommitmentchannelofCDSisparticularlyrelevantforwithaninabilityto pledgecollateral.Furthermore,Ibringtogethertheliteraturesontheofcredit derivativesandinnovation.Sinceinnovativeoftenfacethestrictest collateralconstraints,CDScanpartiallymitigateproblemsininnovation. Thisofcoursehasrealeconomicconsequencessinceinnovationcanbequite t[see HallandLerner (2010)foranoverview]. 1.2 DataandSampleConstruction IuseMarkitgroupdataforcreditdefaultswapspreadsandtradingdatesforthe period2001-2012.Markitgroupprovidesauniqueidenforthereferenceentities ofaCDScontractwhichcanthenbemappedtotheCRSPandCompustatmerged database. 4 Dataforimpliedvolatility,historicalvolatility,andputoptionopenin- terestcomefromOptionMetrics.DataoncreditratingscomefromtheCompustat ratingsmarginaltaxratescomefromtheCompustattaxrateinformation onarm'slendersandcovenantdatacomefromReuter'sDealscan,andinformation 4 Imergeby6digitcusipnumberandthenverifythatthecompanynamesmatch.Ifacusip matches,butthecompanynamedoesntIhandchecktheanddiscardtheonesforwhichIam unsureifthematchisvalid. 7 onbondissuancesandunderwriterscomefromMergentFISD. Idropallineandutilitysectors(i.e.siccodes6000-6999and4900).I alsodropthatarenotinthesampleforatleast2years,thathavenegative bookequityormarkettobookratio,withcodeotherthan\USA",and thathavemissingvaluesforthemainvariablesofinterest(withtheexceptionofthe patentvariablesdescribedbelow).IalsodropaifIobserveCDStradingon itsdebtwithinthetwomonthsof2001sinceIcannotaccuratelyidentifythe introductionofCDSforsuch ForagivenICDSintroductionastheyearthataquotefora CDSwith5yearmaturityisobservedinthedata.TRADINGisadummyvariable equaltooneforallearsinwhichquotesona5yrCDSisobserved.TRADEDis adummyvariableequaltooneifa5yrCDSquoteiseverobservedduringthesample periodforagivenThesevariablefollowthethoseof Ashcraftand Santos (2009),and Saretto (2013). Inadditiontodata,IusepatentdatafromtheNationalBureauofEco- nomicResearch(NBER)patentdataproject.Inajointontheproject,the UnitedStatesPatentandTrademark(USPTO)and Halletal. (2001)(hence- forthHJT)havecollecteddataonover3millionpatentsand16millioncitations.The projecthasrecentlybeenupdatedtoincludedatafrom1976-2006.Foreachpatent Iobservethepatentstechnologicalcategory,applicationdate,grantdate,thelistof citedpatents,andinformationaboutthepatentsassignees(inventors). Halletal. (2001)havealsograciouslyprovidedamatchbetweenthepatentas- signeesandCOMPUSTATthatdynamicallytrackstheownershipofeachpatent. ForthisstudyIammostlyconcernedwithpatentapplicationdates,butIdomake useofthedynamicmatchforsomeofthevariablesthatIdetailbelow.Withthe matchprovidedIamabletolinkinformationtothepatentsusingthe CRSP-COMPUSTATmergedlinktable. 8 Itiswelldocumentedthatpatenting(citing)propensitiesexhibittremendoushet- erogeneityacrosspatenttechnologyclassesandthroughtime.Tocomparethevalue ofpatents(citations)inttechnologyclassesattpointsintime,HJT developareducedformapproachandastructuralapproachtoadjustpatentand citationcounts.InthispaperIfollowrelatedliteratureandemploythere- ducedformapproach,whichinvolvessortingpatentsinto6technologicalclasseswith 36totalsubcategories.Eachpatent(citation)isthenscaledbythetotalnumber ofpatents(averagecitationsperpatent)ineachsubcategory-year.Theseadjusted patents(citations)arethenaggregatedattheearlevel,creatingaweightedsum ofeachpatents.Inaddition,Iincludeindustryandtimecontrolsinregressions tomitigateanyremainingheterogeneityproblems. Thedependentvariablesthatincludepatentsarebasedonpatent applications thatareeventuallygranted.Theimportantthingtonoteisthatthepatentsshow upintheyearofapplicationhowever,andnottheyearthatthepatentisgranted. Thisisconsistentwithpriorliteratureandhighlightsthe decision toincrease innovationthatisnotdependentonthepatentgrantingprocess. ControlvariablesincludeMarket-to-Book(MTB),logoftotalassets(LNTA),log oftotalmarketcapitalization(size),leverage(lev),researchspendingdivided bysales(RDS),cashholding(cashholdings),ytangibility(tan- gibility),adummyforwhetherahasapublicdebtrating(rated),andimplied volatility(impvol).Markettobookisusedtocontrolforchangesinain- vestmentopportunities.Leverage,cashholdings,y,andtangibilityare standardlevelcontrols(foradetaileddiscussionseeSeru).Impliedvolatility isusedtocontrolforCDStradingthatmightoccurforspeculativereasons.For industryIusetheFamaandFrench48industryportfolios.Detailed descriptionsofvariableconstructionareprovidedintheappendix. Thesamplerunsfrom2001to2012andconsistsof4,904or16,151 9 ears.613haveaTRADEDCDSatsomepointduringthesampleperiod. Thisincludes5,924TRADEDearsand4,357TRADINGears. 5 Once aCDSmarketemergesonadebtittendstoremainforthedurationofthe sampleperiod.However,111ofthe613shaveaCDSmarketthatbecomes inactive. 6 1.3 EmpiricalMotivationandDiscussionofResults 1.3.1 CreditDefaultSwapsandFirmInnovation Itisacommonlyheldviewthatinthepresenceofincompletecontracting anduncertainty,inginnovativeprojectsthroughexternalsourcescanbepro- hibitivelymorecultthanroutineprojectsthathavemorecertaincash ws.Forinnovativeprojects,moralhazardandtheinabilitytocollateralizehuman capitalcanmakedebtacostlysourceof[e.g. RajanandZingales (2003); Atanassov (2014)].Furthermore,recentworksuggeststhatequitymarketsmaynotbe aviablealternativebecauseof1)ashorttermemphasisthatdisincentivizesmanagers fromexploringlong-term,innovativeprojectsand2)aninabilitytovalueresearch anddevelopment(R&D)atthebeginningofaproject,despiteastrongpersistence inR&Dsuccess[e.g. Cohenetal. (2013); ChavaandRoberts (2008)]. Thesefrictionscanresultinasubstantialunderinvestmentininnovation ifwithinnovativeopportunitieshaveexhaustedinternalfunds.Thatis,when additionalinvestmentrequiresexternalmaychoosetoforgonewin- vestmentsaltogetherorchoosetosubstituteinnovativeprojectsforlessnovelprojects 5 Thedatausedinthisarticlearenotquiteascomprehensiveasthoseof(Tang,etal.,forthcoming) whosupplementMarkitdatawithdatafromCreditTradeandGFIgrouptoidentify901with atradedCDS.Iimplicitlyassumingthatthethatwecannotmatchbutthathaveatraded CDSwillnotsystematicallybiasourresults.Resultsobtainbothwithinthesetofiden asCDSandinthefullsetofthusthisdoesnotappeartobeatenuousassumption. 6 Thepatentsampleendsin2005becausethedatastopin2006andthereisatruncationbias sinceweonlyobservepatentapplicationsthatareeventuallygranted. 10 attheexpenseoflongtermgrowth.Understandingwhetherinnovationscan alleviatetheseconstraintsinanimportantissue. InthissectionIinvestigatethepotentialroleofcreditdefaultswaps(CDSs)as ashocktothesupplyofexternalforinnovation.Buildingonthe theoreticalworkof BoltonandOehmke (2011),Iclaimthatalikelychannelthrough whichCDSscanrelaxfrictionsisatransferofbargainingpowertocreditors whenhaveimperfectaccesstocapitalduetoalimitedcommitmentproblem. Iftrue,theintroductionofCDSwouldparticularlyimportantforinnovative sinceinnovationtypicallyreliesonintangibleassetsthatarenotpledgeable,suchas highlyspecializedhumancapital. Sp,innovationisbecauseinnovativerelyonsoft assetsthatarenotpledgeableascollateral.Assuch,creditorsarewearyofinnovative forfearofdefaultstateswherethecreditorhasnobargainingpowerbecause theborrowersassetsareoflittlevaluetothecreditor. TableA.2presentsOLSestimatesoftheimpactofCDStradingoninnovation. IfollowCohenet.al.,2012andR&Dscaledbysalesasmyprimaryinnovation inputandpatentsscaledbythenumberofemployeesasmyprimarymeasureof innovationoutput.Theresultsareobtainedfromthefollowingregression: Y i;t = ind=firm + t + 1 TRADING i;t 1 + 2 TRADED i + Controls i;t + i;t WhereTRADINGisadummyvariableequaltooneinrmyearswithanactively tradedCDSonadebtandTRADEDisadummyvariableequaltooneifa CDSevertradesonadebtduringthesampleperiod.Thehypothesisisthat 1 >0inallspNotethattheTRADINGvariableislaggedoneperiod,so wearelookingR&Dandpatentapplicationsintheyearfollowingtheintroduction ofaCDS.Dependentvariablesofinterestincludetheadjustednumberofpatent applications(npatents),theadjustednumberofpatentapplicationsscaledbythe numberofemployeesinthousands(patents/emp),andresearchanddevelopment 11 expensesscaledbysales(R&D/sales). Controlsincludesize,MTB,impliedvolatility,leverage,CF,andadummyvariable forwhetherthehasacreditrating.Standarderrorsareclusteredatthe level.Columns1,2,and4includeindustrylevelinterceptsandcolumns3and4 includemlevelintercepts.Allcolumnsincludeyeardummies.Thesampleonly includesthathavepositiveR&Dinatleastoneearduringthesample period.Thepatentdataonlycovers2001-2005andisthereforehasamorelimited sampleperiodthantheanalysisofR&D. WecanseethatincreaseR&D/salesby(.13%-.25%)whenaCDSisactively tradedontheirdebt,correspondingtoa1.4%increaseinR&Dasapercentageof assets.Theseestimatesareinlinewiththeresultsfoundin[ ? ]wholookat constraintsoninnovative.Wealsoseethatinnovationoutputincreases byroughly1.42-2.21patentapplicationsperthousandemployees. 7 Incolumn1the resultsontherawnumberofpatentsist.Noticehoweverthatwhenwe scalepatentsbynumberofemployeestheresultsbecometandthesignon sizechangessign.Thiswouldsuggestthatrawpatentcountispickingupsome non-linearityinsizethatscalingmythenumberofemployees 1.3.2 AccesstoUnsecuredDebt Ifcreditdefaultswapsserveasasubstituteforcollateral,thenweshould seegreateraccesstounsecureddebtforaftertheinceptionofCDStrading. IntableA.3Ipresenttheestimatesfromamultinomiallogitestimation(columns 1and2)onthedecisionoftoobtainanunsecuredorsecuredloanwithno loanconsideredasthebasecase.Column3presentsresultsfromabinomiallogit estimationforasimplebinarychoicedependentvariable:loanvs.noloan,where loanincludesbothunsecuredloansandsecuredloans.Columnthreeispresentedin 7 Recall,thesearepatentapplicationsthatareeventuallygranted. 12 conjunctionwithcolumns1and2toillustratethattheextensivemarginforobtaining aloaningeneralisalsoincreasing. Aswecanseeincolumn3,are7.62%morelikelytoobtainaloanwhena CDSistradingontheirdebt.Interestingly,thiscorrespondstoa9.08%increasein theuseofunsecureddebtanda2.51%decreaseintheuseofsecureddebtfor withanactiveCDSmarket.InunreportedanalysesIperformthesametestsunder probitandlinearprobabilitymodelspandsimilarresults.Forbrevity Ionlyincludetheresultsforthemultinomiallogitestimation. OnemightbeconcernedthatthesearedrivenbythepotentialthatCDS tradeonwithlessinformationasymmetry,sincesucharelikelytoobtain unsecureddespitehavingtheabilitytopledgecollateral.IfCDSserveas aconduitforwithlimitedpledgeabilitytoobtainexternalthenwe shouldseelargershockstounsecuredforwithlesspledgeableassets. IntableA.4Iuseasimplesortingproceduretoillustratewhichareimpactedthe mostbytheintroductionofCDSs.Inthefullsample,Isortintoquintilesbased ontangibility,MTB,andR&Dintensity.Ithencomparethepercentofunsecured loansbyquintilefor-yearswithatradedCDSandearswithoutanactively tradedCDS. Boththerelativeandabsolutechangesinaccesstounsecureddebtareincreasing inpledgeabilityaccordingtotheR&Dandtangibilityproxies.AccordingtoMTB, theabsolutechangeisincreasing,buttherelativechangeisslightlydecreasing.In nonCDSrmyears,thelowestquantileoftangiblesecures82%ofloanswhile thelowestquantilesecuresonly36%aftertheintroductionofaCDS.However,s inthehighesttangibilityquantileonlydecreasefromsecuring71%innonCDS yearstosecuring40%inyearsafterintroduction. TheresultsintableA.4donottakeintoaccountothervariablesthatdetermine thedemandoraccesstounsecuredIntableA.5Icontinuetheanalysis 13 bylookingatinteractionsinalinearprobabilitymodel.Icreatetwoadditional variables, p and .Thesearedummyvariablesequaltooneifa belongsinthetop40%ofrankedonpatentingintensity(R&Dintensity), thenumberofpatents(amountofR&D)scaledbythenumberofemployees(total assets)inagivenyear.Asdescribedin TitmanandWessels (1988),R&Dintensity providesaroughproxyfortheuniquenessofaprojects.Firmsthatinvest heavilyinR&Dusuallyinvestinhighlyspecializedprojectsthatrelymoreheavily oninalienablehumancapitalwhichisnothighlypledgeable.Thus,withhigh R&Dintensityareprobablymorelimitedintheirabilitytopledgecollateral.Itake patentingintensitytohaveasimilarinterpretation. TheinteractionofinterestisTRADING paten(TRADING where wecanseethattheimpactofTRADINGonaccesstounsecureddebtisincreas- ing(.0844)forwithalimitedcommitmentproblem.Asonewouldexpect,the estimatesareremarkablysimilar(.0844vs..0909)forthetwooflim- itedpledgeability.InunreportedresultsIchecktheestimateswithprobitandlogit spbutonlypresentthelinearprobabilityestimatesbecauseoftheability toincludeinalinearprobabilitymodel.Theincidentalparameter problemandnon-convergentestimatespreventmefromusinginthe probitandlogitsp ToalleviateconcernsthatCDSaremerelycorrelatedwithapropensityto increaseleverage,Irunsub-sampleanalysisintablesA.7andA.8onthesubsample ofearsinwhichanewloanisoriginated.Iagainseparatethesampleevenfur- therandconsideronlyearsinwhichanewloanisoriginatedonthesubsample ofTRADEDThetablepresentsprobit,logit,andlinearprobabilitymodel spthatincludecontrolvariablestoisolatetheofanactivelytraded CDSmarketonunsecuredTableA.7presentstheparameterestimates, andtableA.8presentsthemoremeaningfulaveragemarginalAllsp 14 tionsincludeindustryyearandcreditrating AccordingtotableA.8,themarginalofaCDSonaccesstounsecuredcapital rangesfrom3.25%(column3)to6.26%(column5).Theunconditionalprobabilityof obtainingunsecuredinthesampleis.48.Sotheseestimatestranslateinto a6.7%-13%increaseinaccesstounsecureddebtrelativetothemean. TableA.9providesasimlinearprobabilitymodelwithaninteraction foranalternativeonofrmpledgeability.Theinteractionincludesa dummyvariable, tang ,whichisequalto1ifaisinthetoptwoquintilesofthe fullsamplewhensortedontangibility.Similarly, intang isadummyvariableequalto 1ifaisinthetoptwoquintilesofthefullsamplewhensortedonintangibility. Tangibility isaelmeasureofexpectedassetliquidationvaluesthatborrows fromBergeretal.(1996). 8 Firmswithgreaterexpectedliquidationvalueswill beabletoobtaingreateraccesstosincecreditorscanrecoveragreater amountinbankruptcy.Indeterminingwhetherinvestorsrationallyvaluetheir abandonmentoption,Bergeretal.gatherdataontheproceedsfromdiscontinued operationsreportedbyasampleofCOMPUSTATmsoverthe1984-1993period. Theauthorsthatadollarofbookvalueyields,onaverage,72centsinexitvalue fortotalreceivables,55centsforinventory,and54centsforassets.Following theirstudy,weestimateliquidationvaluesfortherm-yearsinoursampleviathe computation: tangibility = : 715 receivables + : 547 inventory + : 535 ppegt 9 Intangibility isdasoneminuspropertyplantandequipmentasapercentage ofassets.Propertyplantandequipmentareassetsthatcanbepledgedascollateral. Thegreaterpercentageofarm'sassetsthatconsistofPP&E,themoretangible 8 AlmeidaandCampello (2007)usethismeasure. 9 Thisoftangibilityisusedinunreportedresults,butgeneratesresultsthatarenot materiallytfromthealternativelistedbelow. 15 assetbasetheyhave.ToconvertthemeasuretoameasureofintangibilityIsimply takeoneminustheratio. Columnone(two)presentsresultsfromaLPmodelbasedonthesampleof with tang =0(1).TheinterestisonthebetweenthecotsonTRAD- INGbetweenthetwosubsamples.Columnthreepresentsresultsonthefullsample withaninteractiontermbetweenTRADINGand tang .Columnfoure)presents resultsfromalinearprobabilityregressiononthesubsampleofwith intang = 0(1),andcolumnsixpresentstheresultsfromthefullsamplewithaninteraction termbetweenTRADINGand intang .TheresultssuggestthatCDSTRADINGhas agreaterbonaccesstounsecureddebtforintangibleThetranges from.0561%to10.36%.Thesearesimilarinmagnitudetothe p and above. 1.3.3 EndogeneityandPotentialSolutions Theexistingempiricalliteratureoncreditdefaultswapsincorporate isfocusedonthesupplysideofdebtTheprimaryendogeneityconcernin CDSanalysisiswhethertheemergenceofCDSmarketsissimultaneouslydetermined withaleverageandmaturitydecisions.Forthesupplyside,onecaninstrument withbank'suseinforeignexchangederivatives.Thisinstrumentwasusedin Saretto (2013)andismeanttocontrolforthepropensityofcertainbankstohedgein awaythatisnotnotjointlydeterminedbytheborrowingchoicevariables.This instrumentworkswellforthesupplysidesincebankstypicallyhavemanyborrowers andabanksgeneralhedgingpropensityshouldnotbeheavilydependentonanysingle borrower.However,onthedemandside,ifselectionoflendersisdeterminedbyfactors correlatedwithainvestmentdecisions,thenanyinstrumentrelatedtoabank's typecouldalsobepickingupinthetypeofainvestments.Assuch, Iexploreanalternativeapproachtoidentifyacausallink. 16 Iuseaanalysiswherethetreatmentisthe2003 enforcementoftheISDAprovisionsthatgreatlyenhancedtheliquidityofCDScon- tracts.TheInternationalSwapsandDerivativesAssociation(ISDA)isatradeor- ganizationofparticipantsinthemarketforcreditdefaultswapsandpublishesthe ofcrediteventsforCDScontracts.In2001theISDAadoptedanewset ofprovisionsmeanttoaddressrestructuringandtherangeofdeliverableobligations thatcouldbeincludedinacreditdefaultswap'smasteragreement.Restructuring wastodisallowtheapplicabilityofrestructuringasacrediteventwhere restructuringislimitedtobilateralloans.Theprovisionsalsolimitedtherangeof deliverablesecuritiesthatcouldbeincludedinacontract,reducinguncertaintyand ambiguityinpricingthederivatives.Bothprovisionswerenotputintoctuntil 2003,buttheygreatlyenhancedthestandardizationofCDScontracts,makingthem muchmoreliquidandeasytotrade. TheenforcementoftheISDAprovisionsthatgreatlyenhancedtheliquidityof CDScontracts,representingapositiveshocktothedemandforCDSinstruments. ThusthereisintuitivereasontosuspecttheISDAprovisionisrelevantfor thepropensityofcreditdefaultswapstoradeonadebt.Futhermore,the2003 ISDAprovisionsshouldbeunrelatedtoaborrowingchoicevariablesotherthan throughapotentialimpactonfrictions. TableA.6presentstheresultsofaanalysis.Thetreatment isthe2003ISDArevision,andthetreatmentgroupisthesetofthathadaCDS marketinitiatedonitsdebtduring2003.Thecontrolgroupincludesallthat hadaCDStradingontheirdebtpriorto2003.Allthateitherdonothavea CDStradingatanypointduringthesampleperiod,ornoCDStradinguntilafter 2003aredroppedfromthisanalysis.InunreportedanalysisIincludeallthat donothaveaCDSbegintotradein2003asthecontrolgroupwith However,Ibelievetheapproachusedinthepresentedanalysisamore 17 conservativeapproachtohandlingtheselectionproblem.Iusethesetofwith CDStradingpriorto2003asthecontrolgrouptocontrolforanybetween thatultimatelyhaveaCDStradingandthosethatdonot. TheimportantaspectofthisanalysisisthattheintroductionofaCDSin2003is notlikelytobeassociatedwiththeselectionproblemsdiscussedabove,butinstead aresultoftheenhancementoftheCDSmarketasawhole.Ofcourse,theISDA provisionsalsotheliquidityofCDScontractsforwithexistingCDS marketspriorto2003.However,Icannotruleouttheselectionfortheexistence ofaCDSmarketonsuchFurthermore,mswithCDStradingpriorto2003 arealreadylikelytohavebfromtheofCDStradingifsuchexist. ItisalsoimportanttonotethatanypositivebofCDStradingthatstemfrom increasedliquidityforthatalreadyhaveactiveCDSmarketsontheredebtwill onlydiminishthepowerofthistestandnotthevalidity. Irestrictthesampletotwoyearsbeforeandaftertheevent(i.e.2001-2005).Iuse thetreatmenttotesttheofCDSonalloftheprimaryvariablesofinterestinthe previoustables:namely,therawnumberofadjustedpatentapplications(npatents), numberofpatentsperemployee(pat/emp),researchanddevelopmentexpensesas apercentageofsales,(R&D/sale),leverage,andwhetheranewloanisunsecured (unsecured).Theresultsareallconsistentwiththeprevioustests. Asanadditionalmeasure,Iincludeaplacebotestwhereall withaCDStradingpriorto2005aredroppedfromthesampleandwithaCDS marketinitiatedafter2005areusedasthetreatmentgroup(TREATED). InthissettingIhaveidenasetofthathavecharacteristicsassociatedwith CDSmarketstradingontheirdebt,butthatdonotexperienceanyduring2003. Asweshouldexpect,thecotsarealltontheofTRADING onthevariablesofinterest.Iinterpretthisascorroboratingevidencefor1)thetiming ofCDStradingapparentlyhasanonandinvestmentdecisions 18 and2)evenwhentheintroductionofCDSmarketsisexogenoustoinvestment decisions,thereisstillatonbothncingandintangibleinvestment. 1.3.4 LenderAnalysis Ifcreditdefaultswapsalleviatesupplysidefrictionsfordebtthen thereshouldbealargerforsmallbankswhodonothaveasmuchbargaining powerindebtrenegotiationsaslarger,divbanks.Smallerbanksareless divthanlargebanksandtendtobeheavilyconcentratedinaparticularregion orindustry.Becauseoftheirregionalconcentration,smallbanksareusuallybetter suitedtomonitornearbywithhighlyspecializedassets.Thus,onaverage, thatrelyonfromrelativelysmallbanksaremorelikelytobethat investinlesspledgeableassets.(insertcitation) TableA.10presentsresultsfortheimpactofCDStradingonaccesstounsecured debtasitrelatestolendersize.Todlendersize,banksintheDealscandatabase aremergedtothefederalreservecallreportsonWRDS.Abankisas largeifitisoneofthe15largestbanksintheDealscandatabasesortedontotal grossloansoutstandingforthebankholdingcompany.Thevariable biglender isa dummyvariableequalto1ifalargebankisincludedontheloanfacilityofa accordingtoDealscan. InthesampleofwithatradedCDSatsomepointduringthesampleperiod, thereareonly203loanfacilitiesoriginatedwithoutalargelenderasasyndicate member.YettheimpactofCDSsonaccesstounsecuredcapitalisbothstatistically andeconomicallymorepronouncedfortheseColumnone(two)oftableA.10 includestheregressionrunonlyonwithout(with)alargelenderontheloan syndicate biglender =0(1).Columnsthreeandfourpresentresultsonthefullsample withaninteractionterm TRADING biglender .Columnthreeincludesindustry andcolumnfourincludes 19 ItisimportanttonotethatalloftheidenintableA.10comesfrom withinthesampleofthathaveaCDStradingatsomepointduringthesample period.Thebofthisapproachisthatwedonothavetobeconcernedwithany non-static,systematicbetweenCDSandnon-CDSthatcannotbe controlledforwithdummyvariables.Oneconcernisthatwiththemostlimited assetpledgeabilitytendtobeyoung,smallwhilecreditdefaultswapstendto existmostlyformoremature,largeWhilethisisaconcern,thereisenough variationintheTRADEDsampletotesttherelationshipbetweenassetpledgeability andaccesstounsecureddebt.Furthermore,thismaysuggestthatexpandingtheset offorwhichaCDSistradedwouldleadtogreatereconomicgrowth. 1.4 Conclusion Doesthepresenceofamarketforcreditdefaultswaps(CDS)haveanon corporateand/orinvestmentdecisions?Inaworldwithcomplete markets,CDSareredundantsecurities,inwhichcasewewouldexpecttheirpresence tobeinconsequentialtotheunderlyings.However,asillustratedtheoreticallyby the BoltonandOehmke (2011),inaworldcharacterizedbyincompletemarketsand frictions,itispossiblethatthepresenceofCDSsecuritiescanease constraintsfortheunderlyingalongsomedimensions. Recentempiricalworkhasprovidedsomeinterestingevidencewithregardtothis issue.Inparticular,Hirtle(2009)reportsevidenceindicatingbanks'willingnessto extendcreditincreaseswhentheyareabletousecreditderivativestohedgerisk.It appearsthatthisshiftincreditsupplyisexploitedbyclients,as Saretto (2013) reportthatthepresenceofactiveCDStradingonadebtisassociatedwith anincreaseinleverageandalongeraveragedebtmaturity.Whilethisevidenceon isinterestingandimportant,therelatedquestionofwhethertherelaxation ofconstraintsarisingfromCDStradingisultimatelytransmittedto 20 investmentdecisionshasnotyetbeenfullyanswered.Iprovidetherstevidencethat thesupplysidedocumentedinpreviousliteratureinteractwithdemandside constraintstocorporateinvestment. IclaimthatalikelychannelthroughwhichCDSsrelaxfrictionsisa transferofbargainingpowertocreditorswhenhaveimperfectaccesstocapital duetoalimitedcommitmentproblem.Iarguethatthatanysuchroleismostlikely tomanifestitselfininnovativesinceinnovationtypicallyreliesoninvestment inintangibleassetsthatarenotpledgeable.Idemonstratethatpatentingactivity andR&Dspendingincreaseatboththeintensiveandextensivemarginfollowingthe introductionofCDSmarketsonadebt.Ishowthatseeimprovedaccess tounsecuredhighlightingthatthismaybeparticularlyrelevantfor projectswithlimitedcollateralvalue.Tosubstantiatethisclaim,Iexploit crosssectional-variationinaabilitytopledgecollateral,demonstratingthe isstrongerforinnovativewithanintangibleassetbase.Toestablisha causallinkbetweencreditdefaultswapsandalleviationofconstraints,I corroboratingevidencebyexploitingthe2003ISDAprovisionsasanexogenousshock tothelikelihoodthatCDSmarketsareinitiatedonadebt. Collectively,theevidenceprovidedinthispapersuggeststhatthepresenceofan activeCDSmarketmitigatescollateralconstraintsandhasarealeconomicimpact byfacilitatinginnovationandinvestmentforcompanieswithalimitedabilityto pledgefuturecashws.Moregenerally,myslendsupporttheargumentthat innovationisimportantforeconomicgrowth. 21 CHAPTERII PeerofFinancialConstraintsonInnovation 2.1 Introduction Alargeandgrowingliteratureinaimstounderstandhowialcon- straintscorporateinvestmentdecisions.Mostexistingresearchinthisarea implicitlyassumesthatadecisionsdependonitsownconstraintsbutarein- dependentoftheconstraintsofcompetitorsorpeers.Incontrast,otherliteratures havereasonedthatactionsarenotalwaystakenindependently.Forexample, empiricalevidencesuggeststhatpleverageacapitalstruc- ture,pricing,andlocationdecisions[ LearyandRoberts (2014), ShleiferandVishny (1992),and ? ].Inasimilarfashion,wearguethatthatpeerconstraints haveadirectimpactontheinvestmentopportunitiesanddecisionsofcompetitors. Thegoalofthispaperistoprovideempiricalevidencethataltertheirinvest- mentdecisionsaccordingtopeers'constraints.Financialconstraintscause toforgopositiveNetPresentValue(NPV)investments.Thiscanmanifestin decreasedcompetitionforrivals,whichincreaseprojectedcashwsonsubstitute goodsandservices.Furthermore,constrainedarelessabletoadapttonew productsandservicesbycompetitors.Thisopensupthepotentialforuncon- strainedcompetitorstoestablishtbarriersforfutureproductdevelopment. Allelseequal,itfollowsthatainvestmentopportunitysetincreasesasitspeers becomeconstrained. Weusepatentcitationsofpubliccorporationstoidentifyinnovationnetworks andstudycompetingoversimilartechnologies.Wethencreateameasureof 22 competitorconstraintsbytakingtheaverageconstraintsacross atechnologicalpeers.Usingthismeasure,wethatincreasetheir patentingactivityinresponsetopeerbecomingconstrained. Tofurtherexaminethestrategicimplicationswecreateameasureofpatentport- foliosimilaritybetweenWethatshifttheirpatentingportfoliosinthe directionofconstrainedpeersandawayfromlessconstrainedpeers.Thismovement turnsouttobeshortlivedandwethatnewpatentsgeneratedwhilemovingcloser toconstrainedpeersreceivefewerselfcitationswhencomparedtoaprevious patents.Weinterpretthisasevidencethatmsusepatentstoproducebarriersfor futureinnovationwhilepeersareconstrained. Empiricallytestinghypothesesaboutinteractionsbetweenpconstraints andinvestmentdecisionsischallengingforseveralreasons.First,limiteddatamakes identifyinginrelationshipsTheusualresponsetothisproblemis tocontrolforunobservedheterogeneitywithindustryorbutthese approachesoftenignorerelevantinformationinrelationshipsthatcanhelp usunderstandthetrueimpactofconstraintsoninvestment. Second,informationisusuallyaggregatedatthelevelwhichlimitstheability tostudyspinvestmentdecisions.Finally,itistoidentifychangesin constraintsthatarenotrelatedtosimultaneouschangesininvestment spending. Patentingisatypeofinvestmentthatisparticularlywell-suitedforstudying strategicresponsestopinvestmentconstraints.Patentsarerequiredtoin- cludedetailedself-descriptionsandtocitepatentsonrelatedwork(whicharev byareviewer).Thismeansthatpatentcitationscanbeusedtoidentifylinksbe- tweenthathavetheabilitytoapplyspecializedknowledgeuniquetoaparticular technologyspaceorproductmix.Theabilitytoidentifymeaningfulcon- nectionsallowsustostudyinteractionsbetweeninvestmentdecisionsandp 23 constraints. Thepatentmarketalsoprovidesbovertheproductmarketwhenstudying constraints.Asconstrainedsareforcedtocutprojectswemightnatu- rallyexpectthemtoabandonlessdevelopedworkratherthanmorestableprojects thatarelikelyeasiertoWhilemostproductmixincludesbothyoung andoldinvestments,patentsbyrepresentnewideas.Furthermore,a forcedtocutprojectsinthedevelopmentstagemaynothavesecuredthenecessary patentstoprotectaprojectfrombeingtakenbycompetitors. Thereareseveralwaysthatapatentingbehaviormightchangeaspeers becomeconstrained.Whilecompetitorsareexposedtheymaynothave theresourcesnecessarytotpatentinfringementsortocontinuepublishingpatents vitaltocontinuingaparticularproject.Amayinthegap"ofacompetitor's unprotectedintellectualpropertybypublishingpatentscriticaltotheproduction process.Anunconstrainedcanthenstealtheprojectbygainingmonopoly rightstoclaimaproductiontechnology. Alternatively,amaymovecloserintoacompetitor'stechnologicalnicheand producepatentsneverintendedfordirectuse,butforthesolepurposeofblocking acompetitor'sfutureabilitytopatentinthearea.Thiscanproveebecause theconstrainedpeerwillhavetomaneuveraroundthesepatents,oravoidinvesting insimilartechnologiesaltogether. Itisnottrivialthatacompetitor'sconstraintsshouldhaveapositive oninnovation.Alargeportionofinnovationoccurscollectivelywhencom- petingshareknowledgethroughmimickingandsharedlaborpools.Avastlit- eratureineconomics,boththeoreticalandempirical,claimthatknowledgespillovers areacrucialcomponentofsuccessfulinnovation( Bloometal. ,2013).Asconstrained areforcedtocutinnovativeprojectsthepotentialforknowledgespilloversis diminished.Ifthereductionintheknowledgepoolistenoughitmayhinder 24 ainnovation. Onemaybeconcernedthatainnovation(quality)andtheirpropensityto becomeconstrainedarejointlydetermined.Forexample,goodmanage- mentmaysuccessfullymotivateR&Dcreativityandsecurestrongcingrelation- ships.Alternatively,onemightbelievethatinnovationqualityispersistent,leading withhighpatentqualitytodrivetheirpeersintogreaterialconstraints. Toalleviatetheseconcerns,weexploittwoshockstonancialconstraintsthat areunrelatedtopatentingactivityotherthanthroughchangesinacashw. First,weusetheAJCArepatriationtaxholidayin2004asapositivecashshock towithaninternationalpresence.Thislawchangeprovidesarelaxationof constraintsforswithsigtinternationalcashws,butitdoes notimpactthecashwofpurelydomesticWeatdecrease thestrategicmovementformswhosepeersbfromtheAJCAtaxholiday. Furthermore,theisisolatedtowho'speerswereconstrainedbeforethe taxholiday. Asasecondtreatmentweusethejunkbondcrisisof1989asanegativeshock tofundraisingforwithajunkbondcreditrating.Westrongevidenceof strategicpatentingactivityforwithcompetitorsthatwereadversely bythejunkbondcrisis.Sp,thathavepeerswithjunkbondratings exhibitatincreaseinpatentingactivityandmovepatentportfoliosinthe directionoftheirjunkratedpeersafterthecrisis. Ourstudydrawsonthreedistinctandgrowingliteratures.First,wecontributeto theliteratureontheimpactofconstraintsoninvestment. Fazzarietal. (1988)wroteoneoftheandmostuentialpapersonthistopic,highlighting thatconstraintscanhaveaprofoundlyoninvestment.A hostofliteraturefollowedsimilarresults,buthighlightingmanytiesin 25 estimatingconstraintandinvestmentinteractions. 1 However,theempiricalliteratureonconstraintsisquitescarcewhenit comestoinrelationships.Thiscontraststhevoluminouscorporateand industrialorganizationliteraturesondistress,whichdirectlyacknowledge thatpeerrmdistresscanincentivizepredatorybehaviorforcash-rich [cites]. Theseliteraturestypicallyrelyonthepresenceof"prey"nearbankruptcy andondrivingthepreytoexitastheprimaryobjectiveforthepredator[Tirole 2006].Thissettingmaynotbeappropriatewhenstudyingthestrategicinteractionof publiccorporationsthatmaybeconstrained-unabletoinvestinallposi- tiveNPVprojectsbecauseoffrictions-butnotdistressed(near bankruptcy).Ifcompetitorbankruptcyisunlikely,thentypicalpredationstrategies suchaspricecuttingandoverproductionmaynotberelevantandmayneedto focusstrategiesonmoreincrementalgoalsforcapturingmarketshare. Wealsodrawonrecentliteratureoninnovationinbyexamininghow peers'lconstraintspatentingdecisions.Ourpaperiscloselyrelatedto theworkof Almediaetal. (2014)whothatdistresscanforcemanagers toweedouttprojects. Manso (2011)showsthatlong-termcontractsare optimaltoincentivizemanagerstoundertakeinnovativeprojectsinsteadofshort- term,saferprojects. Brownetal. (2012)thatR&Dinvestmentisexceptionally sensitivetoconstraints,leadingtogreaterunderinvestment.Weaddto thisliteraturebyshowingthatinnovatestrategicallywhencompetitorsare constrained.Thiscanhavewelfareimplicationsasconstraints canultimatelyleadtofewermsholdinglargerpiecesofasmallerinnovationpie.(i knowthissoundsterrible,butitsagoodpointweneedtomakemoreclearly). Finally,wecontributetotheliteratureonpeerinOneofthe 1 Forexample,Whited92,AlmeidaCampello2007,Alti2003 26 problemsinstudyingpeeristheidenofpeerWhensimple areused,suchasindustrycategoriesorsortingoncharacteristics, theproblem(see Manski (1993))canmakeidennearlyimpossible. Furthermore,linksaccordingtothesemethodsmaynotalwaysrepresentmeaningful relationshipsforthequestionsbeingstudied.Ourpaperprovidesanovelapproach tocircumventingtheseproblems. 2.2 DataandSummaryStatistics WeusepatentdatafromtheNationalBureauofEconomicResearch(NBER) patentdataproject.Inajointontheproject,theUnitedStatesPatentand Trademark(USPTO)and Halletal. (2001)havecollecteddataonoverthree millionpatentsand16millioncitations.Theprojecthasrecentlybeenupdatedto includedatafrom1976-2006.Foreachpatent,weobservethepatent'stechnological category,applicationdate,grantdate,thelistofcitedpatents,andinformationabout thepatent'sassignees(inventors). Halletal. (2001)(henceforthHJT)havegraciouslyprovidedamatchbetweenthe patentassigneesandCOMPUSTATthatdynamicallytrackstheownershipofeach patent.Withthematchprovidedweareabletolinkm'sinformationto thepatentsusingtheCRSP-COMPUSTATmergeddatabase. Itiswell-documentedthatpatenting(citing)propensitiesexhibittremendoushet- erogeneityacrosspatenttechnologyclassesandthroughtime.HJTdevelopastruc- turalapproachandareducedformapproachtoadjustpatentandcitationcounts.In thispaperwefollowrelatedliteratureandemploythereducedformapproach [e.g. Seru (2012)].Theprocedureinvolvessortingpatentsintosixtechnological classeswith36totalsubcategories.Eachpatent(citation)isthenscaledbythetotal numberofpatents(averagecitationsperpatent)ineachclass-year. 2 Theseadjusted 2 Weranourresultswithboththe6categoryandthe36categoryadjustmentsandqual- 27 patents(citations)arethenaggregatedattheearlevel,creatingaweightedsum ofeachpatents. Formeasuresofconstraints,weusethe WhitedandWu (2006) distressindex(WW)andthesize-agedistressindex(SA)developedby Hadlockand Pierce (2010).Thesemeasuresservetwopurposesinourregressionsp First,weusethesevariablestobuildconstraintindicesofacompeti- tors.Weprovideadetaileddescriptionofthisprocessbelow.Second,weuseSA andWWasacontrolvariablesforaownconstraints.Thesecontrol variablesareimportantbecauseperformanceislikelytobehighlycorrelatedfor innovatinginthesametechnologyspace,andwedonotwantourcompetitor constraintindicestobecapturingthiscorrelation. OthercontrolvariablesincludeMarket-to-Book(MTB),researchanddevelopment spendingdividedbysales(RDS),cashholding(cashholdings),y andtangibility(tangibility).MTBisusedtocontrolforchangesinainvest- mentopportunitiesthatareunrelatedtopconstraints.Cashhold- ings,y,andtangibilityareadditionalcontrolsrelatedtoa slackandinvestmentopportunitiesandarecommonlyusedcontrolsforregressions onpatentvariables(foradetaileddiscussionsee Seru (2012)).Forindustry cationsweusetheFamaandFrench17industryportfolios. Inoursample,wehave18,232uniqueearobservations.Thesummary statisticsarepresentedinTable1. [InsertTable1HERE] Allvariablesarewinsorizedatthe5%level(2.5%levelineachtail).Thisen- suresthatextremeobservationsdonotdrivetheresults.Indeed,ourresultsarenot sensitivetothelevelofwinsorization.Additionally,theconstraintvariables WW;SA;WW cited ,and SA cited arenormalizedtohavezeromeanandastandard itativelysimilarresults.Resultspresentedinthismanuscriptareobtainedwiththe36category adjustmentaswebelieveittobethemoreaccurateapproach. 28 deviationofone.Thistheinterpretationofthebetacotsandallows foreasiercomparisonacrossthesp 2.2.1 FirmRelationshipswithPatentCitations TheNBERdatasetincludesdetailedinformationonpatent-to-patentcitations. Patentsarerequiredbylawtocitetialandrelatedwork.Thus,patentcitations representalinkbetweencloselyrelatedpatentsanddirectlytracktheevolutionof innovationbothwithinandbetweenBoththepatentapplicantandpatent reviewer,assignedbytheUSPTO,arepermittedtoaddtothelistofpatentcitations. 3 Thishelpstoensurethatallrelevantpatentsarecited.[cite]and[Liet.al.,2015] areamongotherstudiestousepatentclassestocategorizerelationships. Inoursample,AisasapeertoBifrmBcitesAduring thepreviouseyears.Indeterminingthelengthoftimethatremainlinked, wefaceabetweenhavingmorelinksinoursampleandhavingthelink representameaningfulconnection.Forexample,twothatcitedeachotheron patentsdeveloped20yearspriorwithoutanysubsequentcitationsmightnotrepresent ameaningfulrelationshipbecausethemighthavedrasticallychangedtheir researchanddevelopmentfocus.Ontheotherhand,shorteningthewindowtoo muchwouldunnecessarilyruleoutmeaningfulrelationshipsanddecreasethe powerofourtests.Webelieveusingae-yearwindowasourbaselineprovidesa good,althoughadmittedlysubjectivebalancebetweenthetwos.Wecheck therobustnessofourresultsusingtwo-yearandseven-yearlinksanddsimilar results. Thelinksarethenusedtoconstructourmainindependentvariablesofinter- est.Wecompetitor'sdistressasthesimpleaverageoftheWWindex (SAindex)foreachofalinkedcompetitorsbyear. 3 Patents.Entrepreneur'sGuideToUSPatentsAndPatentApplications.Ozluturk, Kimmelblatt,andPatel(2013) 29 FC cited i;t = P j 2 C t ; FC j;t num ( C t ) ; where C t isthelistofcitedbyiand num ( C t )isthenumberofin C t .Wenormalizetheseconstraintvariablestohavemeanzeroandstandard deviationonetoaidininterpretationoftheregressionresults.Althoughthevariables arenormalized,theyexhibitsignitvariationinboththecross-sectionandthe time-series. Inasimilarfashionweconstructpvariablestocontrolforaverageresearch spending(RDS cited)andmarket-to-book(MTB cited).Wecontrolforthesevari- ablesbecausewewanttoisolatetheonpmialconstraintsholding constantthelevelofpeers'investmentopportunitiesandR&Dspending.Thisaids inthecross-sectionalinterpretationofourresults....why? 2.2.2 PatentingResponsestoFinanciallyConstrainedPeers InTable2,weestimateourmostbasicelregressionsofpatentoutcomeson peerconstraints(FC cited)laggedbyoneperiod.PanelApresentsre- sultsforregressionswithtransformationsandyeardummyvariables. PanelBrepeatstheanalysisincludingindustry-times-yeardummyvariablesandno yeardummies.Incolumns1and3ofpanelAweestimatethefollowingregression: adjpatents = i + FCcited i;t 1 + controls i;t 1 + t + i;t andincolumns2and4weestimatethefollowingregression: adjcitations = i + FCcited i;t 1 + controls i;t 1 + t + i;t WhereFCcitedistheconstraintmeasure.Incolumns1and2weusethe WhitedWuindextobuildFCcitedandincolumns3and4weusetheSAindex.A detaileddiscriptionforhowwecalculatethismeasurecanbefoundinsection!!The 30 vectorofcontrolsincludesaownmarkettobookratio(MTB),researchand developmentexpendituresscaledbysales(RDS),cashholdings,andtangibility aswellaspeermarkettobookratiosandresearchanddevelopmentexpenditures. Alldependentvariablesarelaggedbyoneperiod.Weexpect tohaveapositivesign sincewearecontrollingforaownconstraints,whichshouldmitigate anyspilloverduetoindustryspcollateralshocksorotherindustryshocks tofundraising. Itisnottrivialthat shouldbepositive.Knowledgespilloversareconsideredan importantdeterminantofpatentactivityintheeconomicsliteratureoninnovation [citations].Onceaisforcedtocutprojectsbecauseofcialconstraintsthere islesspotentialforknowledgespilloverstorelatedIfknowledgespillovers dominateinoursettingwewouldexpectpeerconstraintstohavea negativeonpatentingactivity( < 0). Ontheotherhand,ifcompetitiondominatesinoursetting,theninnovation shouldbwhenpeersbecomeconstrained.Therearetwoprimary waysthatacantakeadvantageofconstrainedrivalsbyusingpatents.First, acandirectlytakeonpositiveNPVprojectsthatconstrainedcompetitorsare forcedtocut.Ifconstrainedchoosetocutlongerterm,lessdevelopedprojects thenweshouldexpecttoseealotofthisactivitytakeplaceinthetechnology spaceratherthanintheproductmarket.Afteraconstrainedstopsdevelopinga positiveNPVproject,acompetitorcouldbeginpatentingtechnologycriticaltothe productionprocessandstealtheproject. Second,cantakeadvantageofconstrainedpeers'inabilitytotpatents inthelegalsystem.Theycandothisbyodingthemarketwithpatentstoprevent futuredevelopmentbyrivalsintargetedtechnologicalareas.Wecannotdistinguish betweenthesetwoalternativesintheregressionsreportedintable2,butitisimpor- tanttonotethatbothscenarioswouldleadtoapositivesignon . 31 Turningtothemagnitudeoftheresultsincolumns1and2,aonestandarddevia- tionincreaseintherivals'constraintindexleadsto0.388(7.359)additional adjustedpatents(adjustedcitations).Theconstraintvariablesincolumns1and2 areconstructedusingtheWhitedWuindex.Incolumns3and4weusetheSA indexasanalternativeinputtoconstructFCcited.Usingthecoentestimates fromcolumns3and4,aonestandarddeviationincreaseinrivals'con- straintsleadsto0.305(4.397)additionalpatents.Theseestimatesrepresent20-30% increasesrelativetotherespectivemeansofthedependentvariables.Thepositive signontheadjustedpatentcotssuggestthatmsincreasepatentingactiv- ityascompetitorsbecomemoreconstrainedandthepositivesignontheadjusted citationcotssuggestthatcitationsalsoincrease. Patentingactivitytendstobeclusteredintimeduetotechnologicalbreakthroughs andthecompoundingofknowledgespillovers.Althoughweadjustourpatent measurestoaccountfortimeandindustryvaryingpatentpropensities,thesead- justmentsaren'tperfectanddon'taccountfortimeandindustryvaryingshocksto constraints.Forthisreason,weincludeyeardummiesintheregressionsrep- resentedinpanelAtocontrolforanyaggregateshocksthatarepotentiallyrelatedto constraintsandpatentingactivity.Wealsoreplaceyeardummieswith industry-times-yeardummyvariablesinpanelbetoallowforaggregatetimeshocks tobeheterogenousacrossindustries. 2.2.3 Iden Therearesomepotentialendogeneityconcernswithourbaselineregressions. First,onecouldbeconcernedaboutreversecausality-successfulinnovation couldbedrivinguprivals'constraintsratherthanrivals'constraints patentingactivity.Ifinnovativecapturefuturemarketsharebyse- curingcrucialpatentsinatechnologicalarea,thencreditorscouldbewearytolend 32 moretorivalsinnovatinginthesametechnologyspace. Second,ourconstraintmeasurescouldpotentiallyexhibittmea- surementerror. 4 Innovativetendtoholdlesscashandhavelowerbookvalues becausetheyhavealesstangibleassetbaseand....Itmightbethatinnovatives appearconstrainedinoursamplewhiletheysimplyhavetcharacteristics. Toaddresstheseconcernsweexploittwoexogenousshocksthatareonlyrelatedto patentingactivitythroughtheironpeerconstraints.Sp, weusetheAJCAtaxholidayasapositiveshocktothecashholdingsof constrainedwithtinternationaloperations.Asasecond,negative shocktofundraising,welookatwithjunkbondstatusbeforeandafterthe junkbondcrisisin1989. InTableX,wereporttheresultsoftheinregressionsusing theAJCAtaxshockasatreatmentvariable.Wecomparetheinpatenting activityofwho'srivalshavetoverseasrevenuetothosethatdonot beforeandafterthepassageoftheAJCA.Thisallowsustoidentifyhowaloosening ofrivalconstraintsstrategicpatentingbehavior.Themain variableofinterestintheindicatorfortreatedinteractedwiththeindicatorforpost. ThecontrolsinthisregressionarethesameasthoseusedinTable2. Looseningofpconstraintsresultsin0.813(12.89)fewerpatent applications(citations).Thisisatrepresentingapproximately35- 50%relativetomeanchangeinpatentingandcitationactivity.Weseethelargest impactforwithrivalsthatwereconstrainedbeforetheeventand thatreceivedalooseningoftheirconstraintsafterthetaxholiday. Toseethis,lookattheinteraction(treated post)intable3.Weinclude thesamecontrolvariablesasbefore.ThevariableofinterestisFC cited 1 [ Treat Post ].Thiscottellsusthechangeinpatenting(citation)behaviorforan1 4 SeveralpapersweshouldciteonFCmeasurementerror 33 standarddeviationinrivalsconstraintwho[Needmorehere] Incolumns1and2,theconstraintvariablewefocusonistheWhited WuIndex.Thecotforpatents(citations)is-0.916(-17.50),antatthe 1%level.Incolumns3and4,theconstraintvariablewefocusonistheSize andAgeIndex.Thecotfor(citations)increasesby-0.508(6.818),t atthe1%level. Inthistable,welookattheresultsrelatedtothejunkbondcrashof1989.Firms whoissuedcheapjunkbonddebtpriorto1989werenotabletorollovertheirdebt in1989andearly1990.Thiscausedtheircostofcapitaltoincreasedrastically.This negativeancialshockwasharmfultotheirbusinessandtheircompetitorscould takeadvantageoftheweakenedstateofthese InTableJunk,wereporttheresultsoftheinregressionsof thisjunkbondshockonpatentingactivity.Tobespwecomparethe inpatentingactivityofwhosrivalsissuedtamountsofjunkbondbefore andafterthejunkbondmarketcrashin1989.Thisallowsustoidentifyhowa tighteningofrivalconstraintsstrategicpatentingbehavior.The mainvariableofinterestintheindicatorfortreatedinteractedwiththeindicatorfor post.ThecontrolsinthisregressionarethesameasthoseusedinTable2. 2.2.4 PatentPortfolioSimilarity Measuringthesimilaritybetweenpatentportfoliosallowsustotestwhether shifttheirpatentingportfolioinresponsetocompetitors'lconstraints. Asmentionedinthedatasection,theUSPTOcpatentsintooneof36technol- ogycategories.Itisrareforpubliccorporationstopatententirelywithinaparticular category. 5 Bymeasuringtheoverlapbetweentwopatentingportfolioswecan getasenseoftheirtechnologicalsimilarity.TheMahalanobisDistancebetween 5 fewerthan27%ofinoursamplepatentexclusivelyinonetechnologyclass. 34 iandjattimetisas: MD ij;t = p ( P i ; t P j ; t 1 ) COV 1 ( P i ; t P j ; t 1 ) where P i isa36 1vectorrepresentingmi'ssuccessfulpatentapplicationsinyear t.Eachelementof P i containsthepercentageofatotalpatentsproducedwithin anindexedtechnologyclass. COV 1 isthevariance-covariancematrixofyear-level patentportfoliosaggregatedacrossThemoretendtopatentinthesame technologyclassthesmallertheirMahalonobisdistancewillbe(themoresimilar theirportfolioswillbe).Firmsthatpatentintheexactsameproportionsineach technologyclasswillhaveadistanceofzero. Therearemanypotentialmeasuresforestimatingpatentportfoliosimilarity.How- ever,Mahalanobisdistancehastheaddedadvantagethatitdoesnottreattechnology classesasorthogonal.Thevariance-covariancematrixexplicitlyaccountsforthefact thatsometechnologyclassesaremorerelatedthanothersbyweightingobservations accordingtocross-categorypatentingpropensities.UndertheMDmetric,a patentingentirelyincomputerhardwareisclosertoapatentingentirelyincom- putersoftwarethanitistoapatentingentirelyinautomobiles. 6 Thisb characteristicoftheMDmetricisdiscussedindetailin[Bloomet.al.,2013]. 7 2.2.5 SelfCitations Firmsmayshifttheirpatentportfoliosinthedirectionofconstrainedpeersfortwo reasons.First,thecouldbepickingupprojectscutbycompetitors.Ifthisisthe case,thenweshouldexpectthetocontinueproducingfollowuppatentsinthe sametechnologicalareatomaintainmonopolyrightsastheprojectdevelops.These 6 Underametricthattreatstechnologyclassesasorthogonal,apatentingsolelyincomputer hardwarewouldbecompletelyunrelatedtoapatentingentirelyincomputersoftware.While thisispossible,itisunlikelythatthereislittleoverlapbetweenthetwoclasses. 7 NotethatifwecalculatedMDusingthesametimeperiodforeachpatentportfoliothe measurewouldbesymmetric,whichmeansthatAwouldbethesamedistancefromBas BwouldbefromA.Inoursettingthiswillnotbethecasebecauseoftheoneyearlagbetween thepatentportfoliosbeingmeasured. 35 follow-uppatentswillciteprecedingpatentsontheproject.Therefore,weshould expectnewpatents,meanttohelpcaptureprojectscutbycompetitors,tocontinue toreceiveselfcitationsatasimilarratetootherprojectsdevelopedbya(what ifthenewprojectsaretlybetterorworsethanaexistingprojects? Canweclaimthat,onaverage,newprojectsshouldbesimilartooldprojects?- Sincewearecomparinginthecross-section,Ithinkso,butweneedabetter Second,amaychoosetopatentinthesametechnologyspaceasaconstrained peerinordertomakeitmoreforthepeertocontinuedevelopingaparticular technology.Awillfacelessoppositionwhenpublishingpatentswhilepeersare tooconstrainedtotpatentinfringementortocontinuethepaceofinnovation requiredtomaintaintheappropriatelegalrightsonaparticulartechnology.Ifa ispublishingpatentsforthesolepurposeofblockingpeerinnovation,or forcingpeertomaneuveraroundpatents,thenwehavenoreasontobelieveself citationswillkeeppacewithapreviouspatents. Webuildtwomeasuresofselfcitations:selfcitationsasapercentageoftotal citationsonnewpatents,andselfcitationspernewpatentapplication.Ifrmsare capturingprojectscutbycompetitors,weshouldexpectbothmeasurestostaythe sameorincreaseaspeerbecomeconstrained.Ifareissuingpatentsto blockpinnovation,thentheshouldbenegative. 2.3 Conclusion Alargeliteratureincorporatenanceconsiderstherelationbetweenarms nancialconstraintsanditsinvestmentbehavior.Ifconstraintslimita abilitytoinvest,thiscouldhaveimportantimplicationsforothermarketparticipants whoarebytheactions.Inparticular,acompetitorsmayseize upontheopportunitybythepresenceoftheconstraintsandaltertheir 36 owninvestmentbehavior.Thus,constraintscouldhaveasubstantiveonin- dustrycompetitivedynamics,industrylevelinvestment,andalsothecompositionof thatinvestment.Whilethisfeedbackfromatocompetitordecisions hasbeenrecognizedinthecontextofcapitalstructure,pricing,andlocationdeci- sions[LearyandRoberts(2014),ShlieferandVishny(1992),andChevalier(2015)], thepresenceofarelationbetweenaconstraintsandcompetitorinvest- mentdecisionshasnotbeenwidelyexplored. Inthispaperwedirectlyexaminetherelationbetweenconstraintsand competitorinvestmentdecisions.Toconductpowerfultests,wefocusonasp typeofinvestment,namelyinvestmentinintellectualpropertyintheformofpatented innovations.Therearetwoprincipleadvantagestofocusingonthistypeofinvest- ment.First,itprovidesuswithverymicro-levelinformationonthenatureofa investmentdecisionsintermsofthetypeofinvestmentmadeandtheultimatesuccess oftheinvestmentasindicatedbysubsequentpatentcitations.Second,bylookingat cross-referencingofpatentsandpatentportfoliosimilarity,wecancarefullymeasure thedegreeofproduct-marketoverlapoftheotherentitiesthatcompetewithany given UsingalargesampleofCompustat-listedfrom1976-2006,westrong initialevidenceofapositiverelationbetweenaconstraintsandcom- petitorpatent-relatedinvestmentdecisions.Inparticular,usingtheWWandSA indexesofconstraints,wethatanincreaseinconstraintsisassociated withincreasesinpatentsbycompetitors.Thisresultappearsrobusttoavarietyof tmodelspandchoices.Thisinitialevidencesuggeststhatcon- straintsnotonlylimitaowninvestmentdecisions,butalsotheyspurothers inthesamecompetitivespacetoinvestmoreheavily.Ifthisisindeedthecase,the consequencesofbeingconstrainedmaybeevenlargerthaniscommonly recognized. 37 Whileourinitialevidenceiscompelling,thereistheusualvexingissueofcausality. Inparticular,onemaybeconcernedthatcompetitorconstraintsandthetechnology governingapatent-relatedinvestmentopportunitiescouldbepositivelycorre- latedduetoomittedfactorsthatourempiricalmodelscannotperfectlycontrolforor, perhaps,becauseofreversecausality.Toaccountforthesepossibilities,weidentify twotexogenousshocksthatshouldcompetitorconstraintswhilenotaf- fectingaownpatentingopportunitiesindependentofthecompetitivefeedback thatweseektoidentify.TheoftheseistheAJCArepatriationactof2004 thatelyrelaxedconstraintsforwithoverseasoperationsbyloweringthe etaxrateoncashharvestedfromoverseasoperations.Thesecondshockis the1989junkbondmarketcollapsethatelyincreasedconstraintssharplyfor thatreliedonthismarketasamajorsourceof Whenweusethetimevariationinconstraintsdrivenbytheseexogenousevents, ourresultsareverysimilartoourinitialevidence.Inparticular,wecompelling evidencethatincreases(decreases)incompetitorconstraintsareassociatedwithmore (less)patent-relatedinvestmentactivity.Theseresultsaresigntinbothaneco- nomicandstatisticalsense,andtheyincreaseourthatwehaveiden anunderlyingpositivecausalrelationbetweencompetitorconstraintsandpatent- relatedinvestmentspending. Toaugmentthisevidenceonthelevelofpatentingactivity,wealsoconsidershifts inthetypeofactivitypursuewhenacompetitorbecomesmoreconstrained. Interestingly,weevidencethatarelativeshiftinconstraintsacrossaport- folioofcompetitorsresultsinashiftofaspendingtowardsthemore/increased constrainedandawayfromtheless/decreasedconstrainedThisisan interestingasitsuggeststhatboththelevelandtypeofinvestmentareaf- fectedinsubstantivewaysbyinterplaybetweencompetitivedynamicsand constraints. 38 Takenasawhole,ourprovideinterestingnewevidencethatcon- straintsnotonlylimitaownspending,butalsoinviteaggressivecompetition intheformofincreasedspendingbyanciallystrongerinthesamecompet- itivespace.Forthetypesofwestudy,inwhichintellectualpropertyisakey ingredientforlong-runsuccess,thedamagedonebythiscompetitivefeedback couldbequitesevere.Ifthisisindeedthecase,ourresultssuggestthatinthese environmentsshouldhaveanaturaltendencytomakeandorganizational choicesthatrelaxconstraintstothehighestdegreepossible. 39 CHAPTERIII PanelDataEstimationinFinance 3.1 Introduction Theuseofpaneldataisextremelycommoninresearch.Animportant bofthepanelstructureisthatitallowsresearcherstocontrolforomitted unit-levelfactorsthatdonotvaryovertimebutmaybearbitrarilycorrelatedwith explanatoryvariablesofinterest.Inacommonpanelsetting,theearis theunitofobservationandpaneldataestimationtechniquesareintendedtocontrol forthepresenceoftime-invariantThemostcommonpaneldata estimatorintherecentliteratureistheestimator.However,other cousins(bothcloseanddistant)ofthisestimatorarealsooccasionallyused. Animportantstreamofrecentresearchhighlightssomeoftheerrorsthatappear intheliteratureinstudiesthatrelyontheestimatorapplied topaneldata.First,ashasbeenhighlightedby GormleyandMatsa (2014),many researchersoftendonotcalculatethesestimatorcorrectlybecauseofer- rorsintheprocessoftransformingthevariablesthatentertheregressionequation. Second,ashighlightedby Petersen (2009)and Thompson (2011),researchersoften usestandarderrorsthatdonotadequatelyadjustforthetypesoferrorcorrelation structuresandheteroscedasticitythatareubiquitousinsettings.Thereare standardsolutions/approachestotheseproblemsthatarewellknownintheecono- metricsliterature(see Wooldridge (2010).Thesesolutionscanbeimplementedby transformingthedatainawaythatisconsistentwiththeunderlyingmodelofinterest andbyusingtheappropriateestimationcommandsandoptionsinstandardmicroeo- 40 conometricsoftwarepackagessuchasStata(see CameronandTrivedi (2010)).In somecases,forexamplealongmultipledimensions,directlyprogramming tocreateanestimatormaybenecessary(see GormleyandMatsa (2014)). Whilethisrecentliteraturemakesnumeroususefulandimportantpoints,itdoes notemphasizeafundamentalassumptionthatmustbetruefortheFixed estimator,oritscousin,theFirstestimator,tohaveanyhopeofconsis- tentlyestimatingthecotsofinterest.Theuseoftheseestimatorstoderive consistentestimatesrequiresrelianceonaStrictExogeneityassumption.Thisisa muchstrongerrequirementthanthetypicalnotionofcontemporaneousexogeneity, which(loosely)onlyrequiresalackofcontemporaneouscorrelationbetweentheer- rortermandtheexplanatoryvariables.Inparticular,asarticulatedbyWooldridge (2010),strictexogeneityvelyrequirestheretobenofeedbackfromthedepen- dentvariabletofuturevaluesoftheindependentvariable.Evenacursorylookatthe variablesusedinresearchsuggeststhatthisassumptionisusuallyviolated. Manyofthedependentvariablesofinteresttoeconomists,forexample performance/returns,leverage,andcompensationarealmostsurelyrelatedtothe subsequentevolutionoftheexplanatoryvariablesofinterestsuchassize,risk,or governancecharacteristics.Infact,manydynamictheoreticalmodelspositexactly thistypeoffeedbackprocess. Inthispaperweexaminethestrictexogeneityassumptioninasetofcanonical panel-dataregressionmodelsselectedfromtheexistingliterature.Foreach ofthesemodelswe:(a)formallytestwhetherthestrictexogeneityassumptionholds, and(b)explorewhetherfailuresinthestrictexogeneityassumptionarelikelyto leadtosubstantiveinconsistenciesincommonestimators.Wepresentoverwhelming evidencethatthestrictexogeneityassumptionis,infact,quitecommonlyviolated. Infact,whenweuselargesamples,wecanrejectthevalidityofthestrictexogeneity assumptioninvirtuallyallofthecanonicalregressionmodelsweconsider.Thus, 41 thereislittlehopethatthecommonFE(orFD)estimatesthatappearinmuchof theliteratureareconsistentlyestimatingtheparameterofinterest.Ifthe estimatescannotbeexpectedtoconvergetotheirtruevalueswhenthenumberof cross-sectionalunits(N)growswithoutbound,anyconcernsaboutthenuancesof thestandarderrorcalculationwouldappeartobearelativeside-show. Withregardtowhetherfailureofthestrictexogeneityassumptionleadstosub- stantiveestimationerrorswithmeaningfuleconomiccontent,itistomake strongstatementswithoutknowingmoreabouttheunderlyingstructuraldynamics. However,douncoverseveralfactsthatsuggestthatthisproblemcanhaveameaning- fuloneconomicinferencesinsettings.First,wenotethattheproblem ofinconsistencyintheFEestimatorisknowntobeontheorderof1/T,whereT isthenumberoftimeperiods,suggestingthattheproblemmaybecomesmallifT islarge(see Nickell (1981)).Unfortunately,thisresultdependsonthepresenceof stable(i.e.,time-invariant)Asweshowbelow,thereissubstantialevi- dencethatintypicalsettingsunit-leveldoappeartochangeover time,perhapsbecauseofoccasionaldiscretechangestothemanagement,ownership, orgovernanceofThus,intheminorityofsettingsinwhichalargenumberof timeperiodsareevenavailable,itappearsunlikelythatthe1/Tresultwillsolvethe problem. Togaugethepossiblemagnitudeofinferenceerrorswhenamaintainedassump- tionofstrictexogeneityisviolated,weconsidertherelativevariationintheFEand FDestimateswhenappliedtocommondatasets.Understrictexogeneity,thesetwo estimatesasymptoticallyconvergetothesametrueunderlyingparametervalue.If strictexogeneityisviolated,asfrequentlyappearstobethecase,theseestimators havetprobabilitylimits,neitherofwhichisthetrueparametervalueofinter- est.Inthesettingsweexamine,wethattheencebetweentheFEandFD estimatorscanbequitelarge,withontheorderof100%beingrelatively 42 common.Moreover,therearesomecasesinwhichtheseestimatorsaret andofoppositesign.Thesepathologicalcasesareontheorderof20timesmore prevalentthanwouldbesuggestedbychanceundersomeconservativeassumptions. Takenasawhole,ourevidencesuggeststhatalargeportionofpriorresearchuses anestimationmethodthatleadstoinconsistentestimatesandthisinconsistencycan besubstantive.Thus,evenifaresearchercarefullyfollowstherecommendationof therelatedrecentliterature,inmanycasestheywilllikelyestimatesomethingthat non-triviallyfromtheparameterofinterest. OurleadtoachallengeastheysuggestthatsimpleFEorFDpaneldata estimatorsareinmanycasesnotthecorrecttoolstouseinresearchinsettings thatincludethepresenceofunit-levelAttheveryleast,ourevidence suggeststhatoneshouldtestthestrictexogeneityassumptioninallsettingsbefore proceedingwiththeseestimators.If,thesetestsreject,asappearstocommonlybe thecase,onecaneithersettleforaninconsistentestimator,anunappealingoption, orturntoalternativeestimatorsusingeitherthesamedataathand(theinternal option)and/ortechniquesthatexploitadditionaloutsideinformation(theexternal option). Withregardtotheinternaloption,thereareavarietyoftestimators, mostlyoftheGMMvariety.Whileitisbeyondthescopeofourpapertocomment onthesespestimators,inthespiritofouranalysisitisworthnotingthat commonGMMestimatorsinnance(i.e., ArellanoandBond (1991), Blundelland Bond (1998))alsorelyontestableassumptionsrelatedtothesuitabilityofthose methods.Inparticular,thosemethodsusuallymaintainanassumptionofnoserial correlation,anassumptionthatcanbetested.Arecentpaperby DangandShin (2015)demonstratesthatthesetestsquitefrequentlyreject,atleastinthecontext ofdynamiccapitalstructureresearch.Thissuggeststhattheinternaloptionmay unfortunatelyattimesnoimprovementoverthemoretraditionalalternatives. 43 Theexternaloptionwherenewinformationisbroughtintoapaneltoidentifythe ofinterestisofcourseusuallythemostdesiredcourseofaction.Thechallengeis inidentifyinggoodinstrumentsthatbothsatisfytheexclusionaryrestrictionandthe relevancycondition.Whilesomenotablesuccessesalongtheselineshaveappeared intheliterature,thereislittleofgeneralitythatcanbesaidaboutthis approachasitusuallyreliesonspecial(oftenone-time)eventssuchasexogenous shockstoaeconomic,legal,regulatory,ortaxenvironment.Inmanycontexts, shocks/instrumentsofthistypearenotreadilyapparent. Inantosomeconstructiveguidancetoresearchersgiventhe econometricchallengeswehighlight,weconsiderasystematic(quasi)externalap- proachtoidenthatexploitsindustry-yearvariationintheexplanatoryvari- ableofinterestasininstrumentfor-levelvariation.Recentresearchhasempha- sizedindustry-yearvariationaseitherapotentialnuisancefactortobecontrolledfor ( GormleyandMatsa (2014)),orasasourceofvariationtotestwhethertheoretically irrelevantfactorsmayeconomicdecisions(e.g., JenterandKanaan (2014)). Wesuggestthatinsomecasesindustry-yearvariationisnotanuisance,butrather canbeviewedasusefulandtheoreticallyrelevantifexploitedasaninstrumentfor theunderlyinglevelexplanatoryvariableofinterest.Ofcoursetheusefulness ofthisvariationasapotentialinstrumentwilldependonthecontext,anissuewe discussatlengthbelow. Toillustratethepotentialofthisapproach,weconsidertheroleofrisk indeterminingalevelofmanagerialownership.Thisisacontextinwhich thedataclearlyrejectthestrictexogeneityassumption.Moreover,eventheweaker requirementofcontemporaneousexogeneityishighlyquestionablegiventhepossible contemporaneousfeedbackefromownershiptorisktaking(see Tufano (1996)). Aswewouldexpectifriskispartiallydrivenbyindustryshocks,weastrong positiverelationbetweeninnovationsinindustryriskandisk,wheretheis 44 excludedintheindustrycalculation.Thus,industry-riskinnovationscertainlyappear tosatisfytherelevancycondition.Wearguethattheexclusionaryrestrictionisalso verylikelytoholdinthiscontext,asfeedbackfrominnovationsinownershipand/or risktoindustryriskwouldappeartobenegligible,atleastwhenweexclude largeinconcentratedindustries. Whenweproceedtoinstrumentforinnovationswithindustry-riskin- novations,ourpreliminaryevidenceindicatesatpositiveroleforriskin thedeterminationofmanagerialownershiplevels.Whilethisresultisinteresting initsownright,forourpurposesthemoreimportantpointisthatitsuggeststhat exploitingindustry-levelinnovationsinexplanatoryvariablesofinteresttoachieve convincingideninpaneldatacontextsinmaybeaproductivestrat- egy.Givenourevidencethatmanyofthewidelyusedpanel-dataapproachesrely onassumptionsthatarerejectedbythedata,thiswouldappeartobeaparticularly usefulstrategytoexploitinsomesettings,particularlywhentheotheriden approachesdiscussedby RobertsandWHited (2012)arenotfeasible. 3.2 Priorliteratureandempiricalstrategy 3.2.1 Outliningtheproblem Whilethestrict(alsocalledstrong)exogeneityassumptionisdiscussedintext- booktreatmentsofpaneldata,withacoupleofnotableexceptions,thisassumption isalmostneveracknowledgedoraddressedinpanel-dataapplications.Given thislackoffamiliaritytoaudiences,weoutlinetheissueherewitha speyetowardsapplications.Thereaderisreferredtotextbooktreat- mentsformoreofthetechnicaldetails(e.g., CameronandTrivedi (2005), Wooldridge (2010)). Weconsiderasimpleregressionmodelwithadependentvariabley,asingleinde- 45 pendentvariableofinterestx,andanassumedmodelinwhich y it = i + x it + it , whereidenotesanarbitrarycross-sectionalunit(from1toN)andtdenotesanar- bitrarytimeperiod(from1toT).SinceinapplicationsNisalmostalways muchlargerthanT,allasymptoticswillbeforNapproachingy. Following Wooldridge (2010),wewillrefertotheassumption E ( it j x it ; i )=0as thecontemporaneousexogeneityassumptionand E ( it j x is ; i )=0foralltandsas thestrictexogeneityassumption.Assumingcontemporaneousexogeneityholds,and recognizingthatlaggedcontrolvariablescanalwaysbeaddedtothemodel,weare concernedprimarilywithviolationsofstrictexogeneityinwhich E ( it j x is ; i ) 8 s> t .Toseehowthisassumptionmaybeviolated,considerthecaseinwhichhigher realizationsofthedependentvariableattimet(sayperformance)haveapositive onsubsequentlevelsoftheexplanatoryvariable(saymanagerialownership).In thiscase,strictexogeneitywouldbeviolatedbecause E ( it x i ( t +1) j i ) > 0highervalue ofthisyearsperformanceareassociatedwithhigherlevelsofnextyearownership. TounderstandtheresultingbiasintheFEandFDestimator,considerthesimple caseinwhichwehavetwotimeperiods(callthem1and2),perhapsseveralyears apart,sothattheFEandFDestimatorsarenumericallyidenticalandtheresulting parameterestimatefor isderivedfromasimplelinearregressionofchangesiny (inourexampleperformance)onchangesinx(inourexampleownership).Suppose alsothatthetrue is0inwhichcasethereisnocausalofxony.Ifweregress y = y 2 y 1 )on x = x 2 x 1 ),theonlysystematicvariationinthedatawillarise fromhigh(low) y 1 valuestendingtofeedbacktohigh(low) x 2 values.Thiswillresult inanapparentnegativecorrelationbetween y and x andwillyield(asymptotically withprobability1)aspuriousnegativeestimatedcot.Extensionsofthis argumentapplytoupwardanddownwardinconsistenciesinparameterestimatesthat dependonthesignoftheactualcot(when 6 =0)andthesignofthedynamic feedbackeClearlymultivariatemodelsandmultipletimeperiodsmakeitmore 46 tosignandunderstandtheresultingbias. Giventhepotentialseriousnessofthisissue,itisnotsurprisingthat Wooldridge (2010)andothereconometrictreatmentsemphasizetheimportanceoftestingfor strictexogeneitybeforerelyingofFEorFDestimationprocedures.However,with theexceptionofmodelsthatincludelaggeddependentvariables,andacoupleof otherrareexceptionsdiscussedbelow,researchersinrelyingonFEorFD estimationnevertestforstrictexogeneity.Unfortunately,weshowthatthesetests willverycommonlyrejectthestrictexogeneityassumption,inwhichcasethereported estimateswillbeinconsistentandshouldthereforebeviewedwithsuspicion. 3.2.2 Recent/currentpractice Tocharacterizecurrentpractice,wesearchthrougheveryissueoftheJournalof Finance,JournalofFinancialEconomics,andReviewofFinancialStudiesfrom2006 to2013forthementionoftheword(orasynonym).Wequicklyscan eacharticletodeterminewhetherthepaperfeaturesanempiricalmodelwith unit-level(e.g.,bank,person)ratherthansolelytime(e.g.,year, quarter,etc.)Weplacedeachpaperintonon-mutuallyexclusivecategories basedonwhethertheauthorsreport(a)traditionalFEestimates,(b)traditional FDestimates,and/or(c)someversionofadynamicpanelGMMestimate.Wedo notcategorizemodelsthatrelyonexternalinstrumentsornaturalexperiments,as ourfocusisonevaluatingmodelsinwhichthistypeofexternalinformationisnot exploited. Ourprocedure251articlesthatreportunit-levelFE(222)andFD(47)es- timates,and17reportGMMestimates.Ifweexcludemodelsthatincludelagged dependentvariables,thecorrespondingnumbersare240,216,44,and6respectively. ClearlytheseindicatethatFEisthemostpopularestimationprocedure.In allofthepapersthatreportsolelyFEestimates,only3mentionthewordstrictor 47 strongexogeneity,andoftheseonly1actuallytestthestrictexogeneityassumption. Clearlythehaseithernotwidelyrecognizedthisissue,orperhapstheiscol- lectivelyhopefulthatanyresultinginconsistenciesarenotlargeenoughinmagnitude tosubstantivelychangetheeconomicinferencesofinterest. 3.2.3 Priorworkinthataccountsforviolationsofstrictexogene- ity Asiswellknown,anypaneldataanalysisthatincludesalaggeddependentvariable asacontrolvariablemustviolatethestrictexogeneityassumption(i.e.,thereisno needtotestforstrictexogeneity,itisassumedtobeviolatedintheunderlyingthe model).Themostprominentareaininwhichthisisrecognizedisinthe dynamiccapitalstructureliterature,ascurrentleverageisusuallyassumedtobe partiallygovernedbypastleverage.SinceconventionalFEandFDestimatorsare inconsistentinthiscontext,alargeliteraturehasappearedusingvariantsofthe GMMapproachandtidentifyingassumptionstoestimatetheparametersof interest,withmuchdebateonthemeritsofrentestimationapproaches.We havelittletohere,excepttonotethattestingtheunderlyingassumptionsina GMMestimationisalsocalledforwheneverfeasible.Recentevidenceby Dangand Shin (2015)suggestthattheseassumptionsareoftenrejectedinthedynamiccapital structureframework. Aswediscussearlier,thestrictexogeneityissueisalmostentirelyunacknowledged inpaneldatamodelsthatdonotincludealaggeddependentvariable.Theone notableexceptionistheimportantworkof WintokiandNetter (2012).Thoseauthors highlighttheimportanceofthestrictexogeneityissueinonespsetting,namely theofboardstructureonperformance.Theytestforstrictexogeneity inthiscontextandrejectthevalidityofthisassumption,leadingthemtoquestion priorworkonthisissuethatrelieson(inconsistent)FEorFDestimators.They 48 proceedtouseaGMMframeworkandshowthattheassumptionsunderlyingthe GMMestimationarenotrejectedinstandardtests,althoughtheydocautionthe readerwithregardtotestingpowerandotherpotentiallimitations. Ourpaperismostsimilarto WintokiandNetter (2012)whoexplorethisissue inanimportantspcontext.Thedistinguishingfeatureofourstudyis thatwehighlightthatthisissueappliestoalargesetofempiricalmodelsin andshowthatthestrictexogeneityassumptionisroutinelyrejectedinpanel datamodelsevenwhenthereisnolaggeddependentvariable.Thus,theconcern raisedby WintokiandNetter (2012)turnsouttobeonlythetipintheiceberg. WealsoevidenceonthepotentialmagnitudesoftheinconsistenciesinFEand FDestimatorswhenthestrictexogeneityissueisignored.Finally,wesuggestan alternativesystematicapproachtoexternalideninlargepanelswhichhas thepotentialinsomecasestobemoreconvincingthaninternalidenvia GMM,andmorewidelyapplicablethanthemagicbulletstrategyofhopingto auniqueeconomic,tax,legal,orregulatoryeventthatperturbstheexplanatory variableofinterest. 3.3 Testingforstrictexogeneity 3.3.1 Identifyingasetofcanonicalregression Inordertoexploretheseissuesinamanageablesetofwell-knowncontexts,we identifyasetofcanonicalpanel-dataregressionmodelsfromtherecentliter- ature.Todothis,weassigneachstudyinourearlier literaturesearchintooneofabroadsetofmutuallyexclusivecategoriesbasedonthe maindependentvariableofinterest.Thesixlargestcategorieshavedependentvari- ablesofthefollowingtype:(a)ainvestmentlevel,(b)aleverage/capital structure,(c)aCEOscompensationlevelorownershipposition,(d)acash 49 holdings,(e)aannualfundraisingchoice,and(f)aperformanceincluding stockreturns,accountingreturns,orsalesgrowth.Intotal,59%ofthepublished panelstudieswithunit-levelectsintheelitesetofthreejournalssearched canbeplaced(usingsomesubjectivejudgment)intooneofthesecategories. Foreachofthesesixdependentvariablecategories,weidentifyasmallsetofspe- variables(bothdependentandindependent)thatareusedmostfrequentlyinthe regressionmodelsideninourliteraturesearch,subjecttotheconstraintthatthe variablescanbeconstructedfromstandarddatasets.Ourchoiceistoconstruct variablesandmodelsthatcorrespondtothechoicesby GormleyandMatsa (2014),as theysomethoughtfulhe-shelfspthatarealsoinformedbythelit- erature.Forvariables/modelsnotincludedintheirstudy,weusevariable thatappeartobemostcommonintheliterature,withsomesubjectivejudgementon ourpartingroupingsimilarvariablestogether.Forallsixdependentvariablecate- goriesweidentifyone(ormore)commondependentvariableconstructions.Wethen modelthesedependentvariablesasafunctionofallindependentvariablesthat(a) eitherappearinthecorresponding GormleyandMatsa (2014)model,or(b)appear inatleast20%ofallassociatedpublishedstudiesinourliteraturesearch. Foreaseofexposition,wewillrefertotheselecteddependentvariablesasDepvar1 throughDepvar6.Insomecasesweusemultipleconstructsofagiventypeofdepen- dentvariable,forexamplebothmarketandaccountingmeasuresofperformance.In thesecasesweaddletterstotheendoftheDepvarnotation.Thus,forexample,De- prvar1aandDepvar1bwouldrefertotwotmeasuresofinvestmentspending (saycapitalspending/assetsandR&D/assets). Ratherthandiscusseachoftheindependentvariablesindetail,wereportinTable 1asummaryofthedependentvariablesandassociatedindependentvariablesinmod- elspredictingeachdependentvariable.Theactualconstructionprocedure/technical ofeachofthesevariablesisrelegatedtoanonlineappendix.Ourhopeis 50 thatnoneofourchoicesarecontroversial.Wearesimplytryingtocollectandcharac- terizealargeandvariedliteratureinareasonableandsuccinctway.Thenumberof independentvariablesvariesdependingonthedependentvariable,withamaximum numberofsix.AttimeswerefertotheseasIndvar1throughIndvar8,wherethe mappinginTable1canbeusedtorecovertheactualeconomicvariableinquestion. 3.3.2 Testingforstrictexogeneity Foreachdependentvariableandasingleassociatedexplanatoryvariable,wecon- ductthestrictexogeneitytestsoutlinedby Wooldridge (2010),onebasedontheFE transformationandtheotherbasedontheFDtransformation.Eachtestisfora modelinwhichthedependentvariableisalinearfunctionoftheunit-level(i.e., ,yeardummies,andtheselectedexplanatoryvariable.Wealsoreporttests foramodelinwhichallindependentvariablesareincludedtogether.Teststatistics arecalculatedwithstandarderrorsclusteredattheunit-leveltoallowforarbitrary heteroskedasticityandserialcorrelation.Thesetestsessentiallyincludefuturevalues oftheindependentvariableintotheregression.Understrictexogeneity,theco cientsonthesefuturevaluesshouldbe0.Thus,evidenceofanon-zerocot(or asetofoneormorenon-zerocotsinthecaseofmultipleexplanatoryvariables) istakenasevidenceagainstthestrictexogeneityassumption. Thep-valuesforthesetestsusingtheentireuniverseofavailableCompustatdata from1950to2012arereportedinTable2.Thedependentvariableforeachmodel islistedintheleftcolumn,andtheothercolumnheadingsindicatethetestthatis conducted,withFE1forexampleindicatingtheversionofthe Wooldridge (2010)testforstrictexogeneityinamodelwithIndvar1asthesoleindependent variable(inadditiontotheyearandThejointtestdesignationsare forthepooledversionoftheFEandFDtestsfrommodelsthatincludeallofthe explanatoryvariablestogether. 51 Astheintable2indicate,thevastmajorityofthep-valuesarebelow.01, indicatingthatinmostcasesstrictexogeneitycanberejectedwithahighdegreeof Infact,ofthe86testsconductedonindividualexplanatoryvariables,we canrejectthestrictexogeneityassumptionatthe1%levelinmorethan3/4ofthe reportedmodels.Moreover,inthejointteststhatincludeallexplanatoryvariables, thestrictexogeneityassumptionisrejectedatthe1%levelinallbutonemodelandat the10%levelinallmodels.Clearlythisindicatesthatviolationsofstrictexogeneity areextremelycommon,andtheof ? wouldappeartoextendmuchmore broadlytopaneldatastudies.Thisisnotsurprisingifonebelievesthat choices/outcomes,performance,andincentives(thedependentvariables) oftenhaveanonthefuturedeterminantsofthesechoices(theexplanatory variables),forexampleaasset,growth,orgovernancecharacteristics. Tofurtherexploretheseresultsinmorehomogeneoussettings,webreakthesam- pleintoshortertimeperiods,or,alternatively,intoindustry-basedsubsamples.In particular,weconducttheprecedinganalysisfortheentiresampleofsre- strictedto10yearsamplesub-periods,andthenweconductananalysisforthesample ofallyearsbutrestrictedtoindustry-basedsubsamplesbasedona1-digitSIC code.Sincewenowhavemultiplep-valuesonteststatisticsforeachdependent variable-independentvariablecombination(oneforeachsubsample),wetabulatethe medianp-valuefortheassociatesetofteststatistics.Theseguresarereportedin Table3,withpanelAreportingyearsubsampleresultsandpanelBreportingindustry subsampleresults. Aswewouldexpectgiventhesmallersamplesizesinvolvedinthesetest,the p-valuesinbothpanelsofTable3aregenerallysomewhathigherthanthelarger samplepooledtestsinTable2.However,itisquiteremarkablethatthemedian p-valuesarefrequentlybelow.10and.05,suggestingthatinmorethanhalfofallof thesesubsamplesthereissubstantialevidenceofarejectionofstrictexogeneity.If 52 wecoupletheseobservationswithstrongaprioritheoreticalreasonstobelievethat strictexogeneitywillbeviolated,thegeneralcaseforsuspectingthatmostFEand FDestimatesinthesetypesofcanonicalmodelsareinconsistentseemscompelling. 3.3.3 Insightsonmagnitudes ThefactthatFEandFDestimateswillgenerallybeinconsistentinmanyormost paneldatasettingsisconcerning.However,iftheinconsistencyissmall,itis possiblethatconclusionsregardingthemagnitudeofacotofinterestoratest ofwhetherthecotistfromzeromaybeatleastapproximatelyvalid.It istomakemoreprecisestatementswithoutspecifyingthepossibledynamic structureoftheunderlyingmodel.However,somepotentiallyusefulinformationcan beinferredbycomparingtheFEandFDestimates,aslargebetweenthese estimateswouldsuggestaproblemofsubstantialmagnitude. Toinvestigate,wecomparethemagnitudesandsignsofthecorrespondingFE andFDcotestimatesforthemodelsinTables2and3.Wecollectallofcoef- tpairsandcalculatethepercentageofpairsinwhichtheFEandFDcot estimatesareofoppositesign,andalsothepercentageofcasesinwhichbothcoef- tsaretatthe10%levelorhigherandarealsoofoppositesign.We alsocalculatethemedianratioofthelargerofthetwocotsinmagnitudeto thesmallercotinthesubsetofcasesinwhichbothcotshavethesame sign.ThesearereportedinTable4(PanelsA,B,andC).Thecolumn ofthetableindicatesthenumberofpairsusedtomakethecalculations,whichin allcasesisequalto32x(#ofindependentvariables),asthereare16tsam- ples/subsamples(allobservations,10industrysubsamples,5tenyearsubsamples) and2waystoestimateanFEorFDcotonagivenexplanatoryvariable(as thesoleindependentvariableinthemodelorwithalloftheindependentvariablesin themodel). 53 Theincolumn1indicatethatFEandFDestimatesarenotinfrequently ofoppositesign.Thisisconcerning,sincewewouldexpecttwoestimatorsthatare cousinsofoneanotherandareappliedtothesamedatatousuallybeofthesamesign. As Wooldridge (2010)notes,substantialbetweenFEandFDestimators areoftenanindicationofaviolationofstrictexogeneity.Itwouldbeparticularly concerningifthesetwoestimatorsyieldtcotsofoppositesign.Ifa givencotis0,thelikelihoodofobservingtwocotsthataret atthe10%levelandofoppositesignis.5%ifweassumeindependenceofthetwo estimators.Ingeneralwewouldexpecttheestimatorstobepositivelycorrelated,and alsothetruecotwilloftennotbe0.Bothoftheseconsiderationswilltend tolowerthelikelihoodofobservingthistandofoppositesignphenomenon understandarddistributionalassumptions.TheinTable4,panelsAandB, indicatethatthisbehaviorisnotnearlyasrareaswouldbeexpected,withrates inallcasesfarabovethe.5%threshold.Weinterpretthisasadditionalevidence ofpotentiallymisleadinginferencesbeingdrawnfromFEandFDestimates,either becauseofafailureofstrictexogeneityorothermodelmissp IfwerestrictourselvestocasesinwhichtheFEandFDestimatesareatleastof thesamesign,wereportinpanelCthatthemedianratioofthelargermagnitude estimatetothesmallermagnitudeestimateisfrequentlyquitelarge.Poolingacross alldependentvariables,themedianofthesemediansis1.53,indicatingthat encesofmagnitudeontheorderof53%arequitecommon.Thisisofcourseafter alreadyexcludingthesubstantialnumberofcaseswherethepointestimateshave oppositesigns.Certainlythisdoesnotinspireintheimplicitassumptions underlyingtheFE/FDestimationapproachinthesecanonicalpaneldata models. 54 3.3.4 Hopingfora1/Tsave WhiletheFEandFDestimatorsarebothinconsistent,thedegreeofinconsistency oftheFEestimatormaybesmallerinalongpanelbecausetheFEestimatoruses ofvariablesfromtheirmeanswhiletheFDestimatorusefrom theadjacentperiods.Intuitionsuggeststhatfeedbackwillbemoretial whendirectlycomparingadjacentperiods,andthisnotionisformallycapturedbythe factthattheinconsistencyoftheFEestimatorisontheorderof1/Twhiletheincon- sistencyoftheFDestimatorisindependentofT(see Nickell (1981)).Thus,onemight hopethatalongpanel,whenitisavailable,wouldrendertheFEcotstobe relativelyinformative.Forthis1/Tresulttobepotentiallyuseful,a forthedependentvariableofinterestneedstobestableoveranentiresampleperiod. Unfortunately,giventheoccasionaldiscretechangesthatoccurovertimetoatypical management,ownership,andassetbase(viamergers/acquisitions/divestitures), theassumptionofastableunit-leveloveralongsampleperiodmaynot bevalid. Toinvestigate,wecalculatecorrelationsinestimatedcots derivedfrommodelsusingtsubperiods.Inparticular,foreachdependent variableweestimateaFEmodelusingtheentiresetofassociatedindependentvari- ablesfornon-overlapping5-yearand10yearsubperiodsstartingwiththemostrecent observationyearandrollingbackwards.Wethencollecttheestimated andcorrelatetheseestimatesfortlags. InTable5wereportthepairwisecorrelationsofthedacross10-year periodsforeachofthecanonicalmodels.Astheindicates,forallofthemodels thereisstrongevidenceofamonotonicdeclineincorrelationofthe estimatesastheestimationtimeperiodsgetfurtherapart.Themostdistantlags (i.e.corr( 1 ; 5 )),mostofthecorrelationsarebelow.50.Certainlythispreliminary evidencedoesnotstrongsupportforstabilityoftheunderlyingunit-level 55 InTable6wepresentthesamecorrelations,butfor5-yearestimationperiods. Again,asthetimelaggetsfurtherthesecorrelationsdropfairlysharply,withcor- relationsfallingsteadilyaswemovefromadjacentperiodstoperiods15-30years apart.Inuntabulatedresults,wethatthesecorrelationsdieoutfurtherwhenwe considerevenmoredistantlags.Toassurethattheseresultsarenotsptoinclu- sionofalloftheindependentvariables,wereplicatethecolumn6correlationusing modelsthatonlyincludeasingleexplanatoryvariableforeachdependentvariable. Thesereportedintable7tellthesamebasicstory.Aspanelsgetlonger,the implicitassumptionofaconstantunit-levelbecomesquitequestionable. Thus,unfortunately,the1/TsavetothepossibleinconsistencyofFEestimatesin thepresenceofaviolationofstrictexogeneitymaynotbeveryuseful. 3.3.5 Robustnessandextensions Themodelswepresentinthetablesareintendedtobeboilerplateanduncontro- versial.Thebasicpointthatemergesisthat:(a)therearegoodtheoreticalreasons toexpectthatfuturevaluesofmanycommonindependentvariablesarecorre- latedwiththedependentvariableevenafterpartialingoutcontemporaneouscontrol variables,(b)formalteststhiswithaveryhighdegreeofand(c) standardFEandFDestimatesareinconsistentwhenthisstrictexogeneityassump- tionisviolated,and(d)therearereasonstosuspectthatthisinconsistencycanat timesbelargeinmagnitude. Whiletheevidencewepresentseemsquitestrong,onemaybeconcernedthat somepeculiarityinourmodelingorsamplingchoicesmaybedrivingtheresults.To investigate,wehaveexperimentedwith(a)amoreaggressivewinsorizationatthe 5%tailsratherthan1%,(b)trimming(dropping)the1%and5%tailobservations ratherthanwinsorizing,and(c)completelyeliminatingallwinsorizationofthedata. 56 Wehavealsoexperimentedwithrestrictingthesampletothepost1970,post1980, andpost1990timeperiods,andwehavealsoestimatedthemodelsononlythe industriesthatwereexcludedfromtheinitialsample(utilitiesandFinally, inallcasesinwhichthereisanalternativepopulardofan independentvariable,wehavesubstitutedthemostpopularalternativeinplaceof thevariablethatweuse.Inallcasestheresultswiththesemodelorsamplealterations aresubstantivelyunchangedfromwhatwereportinTable2.Thus,itseemsthatthe evidenceagainststrictexogeneityisquiterobust. Theevidenceagainststrictexogeneityseemssostrongthatonemightquestion whetheratestonalargesamplewouldevernotreject.Toinvestigate,weselect asanindependentvariableameasurethatisveryhardtopredictastock returnandasadependentvariable,aseeminglyinnocuousconstructthatislikely todependontheindependentvariableandalsotohaveacomponent theratioofareceivablestopayables.IfweconducttheTable2testsina modelusingthisdependentvariable/independentvariablecombination,thep-value oftheFE(FD)testforstrictexogeneityis.77(.91).Thus,itisnotthecasethatthe strictexogeneitytestalwaysrejects.Itwouldthusappearthatitrejectsinmostofthe modelsweconsiderbecauseoftheeconomicfeedbackbetweenallofthevariablesthat entermanycanonicalpaneldataregressions.Ifthisisthecase,traditional FEandFDestimatesarenotconsistent. 3.4 Usingindustry-yearvariationforidenincertain settings IfconventionalFEandFDestimatorsareinconsistentinagivensetting,are- searcherleftwithasubstantialchallengeintheirgoalofidentifyingaparameterof interest.Asdiscussedearlier,aninternalidenapproachusingavariantof 57 GMMmaybesuitableinsome,butcertainlynotall,settings.Externaliden exploitingexogenouschangesinvariouseconomicparametersofinterestisofcourse alwaysdesirable,butinmany(perhapsmost)settings,isnotfeasible. Insomecases,wesuggestthattimevaryingindustryshocksmaybesuitableas instrumentsforlevelinnovationsintheexplanatoryvariableofinterest.Itwill almostalwaysbethecasethatindustryinnovationsinavariableofinterestwillbe correlatedwithelinnovations,sowesuspecttherelevancyconditionusingthis strategywillgenerallybeeasytoestablish.Ifindustryshockscaptureinputsinto shocksthatarenotdrivenbydecisions,theexclusionaryrestrictionwillalso inmanycasesbedefensible.Inmeasuringindustryshocks,itwillbeimportantto excludetheinanycalculationsoastopurgeanyendogeneityarisingfrompurely elvariation.Inaddition,ifaislargeoranindustryisconcentrated,the exclusionaryrestrictionwillbelesslikelytohold,astheremaybecontemporaneous feedbackfromachoicestootherindustryparticipants. Weexplorethepotentialusefulnessofthisapproachforunderstandingtherelation betweenariskandinsideownership.Sinceownershipcaneithercurrent orfuturerisktaking,boththecontemporaneousandstrictexogeneityassumptions areconcerninginthiscontext.Inuntabulatedresultswehavetestedthestrictex- ogeneityassumptionusingthetestswerecommendabove,andtheassumptionis soundlyrejected.Ifweregressriskonindustrymeasuresofrisk(alongwithyear dummiesandwedetecttheexpectedtpositiverelation betweenindustryriskandriskindicatingthattherelevancyconditionis Sinceindustry-riskshouldbeunrelatedtoaownershiporindividualrisktaking decisions,atleastincasesinwhichtheindustryisnotconcentrated,theexclusionary restrictionwouldalsoseemtobequitereasonableinthissetting.Whenweundertake apanel 2SLS! ( 2SLS! )analysisusingindustryriskasaninstrumentforrisk, wedetectapositiverelation.Ifweexclude4-digitindustrieswithaindex 58 inthetopquartile,oralternativelywithsalesthataremorethan10%ofthe 4-digitindustrytotal,theseresultsaresubstantivelyunchanged.Whilethisinitial evidenceispreliminaryinnature,itdoessuggestsubstantialpromiseforthisgeneral approachtoideninsomepanelsettings. 3.5 Conclusion Severalrecentarticlesintheliteraturehaveinvestigatedissuesrelatedto properlyconstructingconventionalpanel-dataestimatorsandtheirassociatedstan- darderrorsgiventhetypesofpaneldatabasesthatarecommonlyusedinIn thisstudyweaskthepreliminaryquestionofwhethertheseconventionalestimators (i.e.,(FE)or(FD)estimates)arelikelytobeinformative inthesensethattheywillyieldconsistentestimates.Wehighlightthefactthat consistencyofFEandFDestimatesreliesonastrictexogeneityassumptionthatis bothmuchstrongerthanthetypicalnotionofcontemporaneousexogeneityandis alsotestable. Weproceedtoconductthesetestsinasetofstandardpanel-datamodels idenfromtherecentliteratureto.Perhapsnotsurprisinglygivenvariousdy- namictheoriesofnancialchoices,weshowthatstrictexogeneitycanberejectedin essentiallyallofthesemodels.Inourviewthisevidenceindicatesthatconventional FEandFDestimates,whichweshowarequitecommoninnceresearchare,in mostcases,inconsistentestimatorsoftheparameterofinterest.Attheveryleast, ourevidencesuggeststhatresearchersshouldaddressthestrictexogeneityassump- tionfromatheoreticalperspective,andalsotestthisassumption,beforetheyeven considerusingconventionalestimators. Inantogaugehowmisleadingconventionalpaneldataestimatorsmaybein typicalresearchsettings,weexaminebetweenFEandFDestimates derivedforthesamemodeloverthesamesample.Weshowthattheseestimates 59 frequentlydivergesubstantially,suggestingthatthemagnitudesofinconsistencies arisingfromtheseestimationprocedurescanbe,atleastinsomecases,substantial. WealsocautionthatrelianceonalongpaneltominimizetheinconsistencyintheFE estimatormaynotbeparticularlyuseful,asunitlevelintypical researchsettingsdonotappeartobestableoverlongsampleperiods. Ourresultsarechallengingastheysuggestthatresearchersfrequently mustturntolessconventionalapproachestoestimateparametersofinterest.GMM estimators,whichhavebeenincreasinginpopularity,areonesuchapproach.How- ever,incasesinwhich(a)theorysuggeststhattheassumptionsunderlyingGMM areviolated,(b)thetestableassumptionsunderlyingGMMarerejected,or(c)the samplepropertiesofGMMarepoorlybehaved,thisisunlikelytobeasuitable alternative. GiventheselimitationstoGMM,wesuggestthepossibilitythatinsomesettings thepresenceofindustry-yearshockstoexplanatoryvariablemaybeausefulidenti- strategy.Inparticular,ifthereisantime-varyingindustrycomponenttoan explanatoryvariableofinterestthattheexclusionaryrestrictionconditional onunit-levelandsample-wideyearthisvariationcouldbequite informativeandcouldleadtoasystematicapproachto(quasi)externaliden tion.Wedemonstratethepotentialusefulnessofthisapproachbyusingindustry innovationsinriskasaninstrumentforriskinanexaminationoftheof riskoninsideownership.Inthisspcontextouridenapproachleads tofairlypreciseestimatesthatsubstantivelyfromnaveestimatesthatignore endogeneityinthepanel.Whileexploratoryinnature,thispreliminaryevidenceap- pearsquitepromisingintermsofrecommendingthisasastrategytobeconsidered inothercontexts. 60 APPENDIX 61 APPENDIX TableA.1SummaryStatistics ThistableprovidessummarystatisticsforbothCDS(PanelA)andnon-CDS (PanelB). leverage isbookleverageaslongtermdebtplusshorttermdebtover totalbookassets(dltt+dlc)/at, mktlev ismarketleverageaslongtermdebtplus shorttermdebtdividedbytotalassetslesscommonequityplusmarketvalueofcommon stockplusdtaxes(dltt+dlc)/(at-ceq+mktcap+txditc), MTB ismarkettobookratio astotalassetsplusmarketvalueofcommonequityminusbookequityallovertotal assets(at+mktcap-ceq)/at, FixedAssets isasgrosspropertyplantandequipment scaledbytotalassets(ppegt/at), pr isasoperatingincomedividedbylagged assets(oibdp/L.at), lnsales isthenaturallogofsaleslog(sale+1), lnsize isthenaturallog ofthemarketvalueofcommonstock, hvol ishistoricalvolatility(5yr)fromOptionMetrics, mtaxrate isthemarginaltaxratebeforeinterestdeductions, R&D isresearchanddevelop- mentexpensesscaledbytotalassets(xrd/at), capex iscapitalexpendituresscaledbytotal assets(capx/at),WWistheWhitedandWu(2006)constraintindex.Allvariables arewinsorizedatthe1%level. CDSFIRMSNONCDSFIRMS meansdminp50maxmeansdminp50max leverage0.300.160.000.290.760.160.190.000.090.76 mktlev0.220.140.000.190.650.110.140.000.050.65 MTB1.670.890.551.398.392.161.480.551.668.39 FixedAssets0.640.400.010.622.840.450.400.000.326.24 0.150.09-0.340.140.500.080.23-1.000.120.50 size8.351.140.008.3210.395.911.800.006.1710.39 lnsize8.371.322.298.3110.716.481.291.386.4810.71 mtaxrate0.330.050.000.340.380.270.110.000.330.39 hvol0.430.180.140.401.610.630.240.150.593.83 RND0.020.050.000.000.680.080.130.000.020.68 INV0.060.060.000.040.410.060.070.000.030.41 WW-0.360.07-0.48-0.36-0.00-0.230.09-0.48-0.230.08 N=4,496N=11,505 62 TableA.2CreditDefaultSwapsandInnovation ThistablepresentsOLSestimatesfortheofcreditdefaultswapsoninnovation input(R&D/sales)andinnovationoutput(patentvariables).Variableare listedintheappendixanddescribedintheempiricalmotivationsection.Thesample includesyears2001-2005incolumns1-3,andincludesyears2001-2012incolumns4-5. Columns1-3onlyincludethatpatentatleastoncefrom2001-2005andcolumns 4-5onlyincludethathavepositiveR&Dexpensesinatleastoneyearfrom2001- 2012.Columns1,2and4includeFama-French48industryctsandcolumns3 and5includeAllcolumnsincludetimedummies.Standarderrorsare clusteredatthelevelinallcolumns. npatentsPatents/empPatents/empR&D/salesR&D/sales TRADING2.10191.4212 2.2185 0.2494 0.1316 (1.33)(2.34)(4.18)(5.45)(2.95) TRADED1.29750.01140.2544 (1.10)(0.02)(4.68) size2.2765 -1.0815 -2.7146 -0.4274 -1.2127 (8.78)(-3.47)(-1.88)(-10.90)(-8.68) rated0.38460.40571.31190.3489 0.3735 (0.52)(0.81)(1.06)(6.68)(4.12) impvol6.8358 3.1008 7.2214-0.8689 -0.8911 5.47(2.43)(1.24)(-6.61)(-4.53) DE-0.9529 -0.9083 -1.2104 -0.0383 -0.0067 (-3.08)(-4.47)(-3.51)(-2.34)(-0.29) MTB0.8056 0.5169 -0.03390.0780 0.0082 (4.71)(2.29)(-0.08)(3.06)(0.28) lnkl0.8082 2.1784 2.9316 0.02360.1033 (3.24)(6.17)(2.31)(1.17)(1.37) RDS0.1944 0.4497-1.2667 (1.99)(1.65)(-1.74) CF-2.5768 -2.8789-0.1632-2.8287 -1.0330 (-2.91)(-1.49)(-0.05)(-12.16)(-4.94) FixedIndustryIndustryFirmIndustryFirm YearFEYESYESYESYESYES nobs5153515351531615116151 R20.27250.18300.79220.40530.8225 t statisticsinparentheses p<: 01, p<: 10, p<: 05, 63 TableA.3UnsecuredandSecuredLoans:MultinomialLogitEstimates Thistablepresentsmultinomiallogitestimates(columns1-2)fortheof CDStradingontheuseofsecuredandunsecureddebtwithno-loanasthebase case.Column3presentstheestimatesfromabinomiallogitregressionwhere thedependentvariableisadummyvariableequaltooneiftheobtained anewloan,securedorunsecured,andzeroifthedidnotinitiateanew loan.AllcolumnsincludeyeardummiesandFama-French48industry Controlvariablesincludetotalleverage,markettobookratio,size, adummyforwhethertheisrated,researchanddevelopmentspending scaledbysales,Cashwscaledbyassets,impliedvolatility,anddebtrating categorydummy.Allcontrolsexceptfortheratingdummiesarelaggedone period.standarderrorsarecalculatedusingtheHuber/Whitemethod. MultinomialLogitBinaryLogit SecuredUnsecuredLoan NOTTRADING.1344***.3191***.4528*** (37.42)(70.13)(96.39) TRADING.1092***0.4099***.5291*** (14.23)(34.00)(37.21) -.0251***.0908***.0762*** (-2.58)(6.15)(4.33) FixedIndustryIndustryIndustry YearFEYESYESYES nobs161511615116151 64 TableA.4CDSandSecuredvs.UnsecuredFinancing Thistablepresentsamoredetailedbreakdownofuseofsecuredvs.unsecured forearswithandwithoutanactivelytradingCDS.Firmsaresortedintoquintilesbased onthevariable tangibility inpanelA, MTB inpanelB,and R&D inpanelC.Sortsarebased onthefullsampleofbutthistableonlyincludesearsinwhichanewloanfacility wasoriginatedfrom2001-2012. CDSNotTradingCDSTrading PanelA:TangibilityQuantiles QuantileUnsecuredSecuredUnsecuredSecured Low12517.63%58482.37%7163.96%4036.04% 219021.25%70478.75%11669.46%5130.54% 321822.78%73977.22%12764.80%6935.20% 427127.26%72372.74%19657.82%14342.18% High32128.08%82271.92%26159.32%17940.68% PanelB:MTBQuantiles QuantileUnsecuredSecuredUnsecuredSecured Low12412.06%90487.94%11348.09%12251.91% 225122.21%87977.79%21053.71%18146.29% 326825.35%78974.65%19867.81%9432.19% 427929.49%66770.51%17471.90%6828.10% High20337.87%33362.13%7681.72%1718.28% PanelC:R&DQuantiles QuantileUnsecuredSecuredUnsecuredSecured Low42722.72%145277.28%32762.52%19637.48% 213925.50%40674.50%9953.23%8746.77% 328427.00%76873.00%20060.98%12839.02% 420025.94%57174.06%12570.22%5329.78% High7516.67%37583.33%2873.68%1026.31% Total1,12523.95%3,57276.05%77161.53%48238.47% FullSample CDSNotTradingCDSTrading AllTraded Unsecured1,12523.95%16034.33%77161.53% Secured3,57276.05%30665.67%48238.47% Total4,6974661,253 65 TableA.5CDSandUnsecuredFinancing:LinearProbabilityEstimates Thistablepresentslinearprobabilitymodelestimatesforthebreakdownof accesstounsecuredforearswithandwithoutanactivelytrading CDS.Thedependentvariable, unsecured ,isadummyvariableequaltoonefor thatoriginatedanunsecuredloanfacilityaccordingtoDealscanfrom2001-2012. Patenisadummyequaltoforyearswithpatentingactivity (R&Dspending)thatisgreaterthanthemedianforthatyear.Thethirdcolumn islimitedbytheavailabilityofpatentdatawhichstopsin2006andistruncatedto 2005.Controlvariablesonpageoneoftheresultssection)areCF,MTB, size,rated,andimpvol.Industry,year,andratingcategorydummiesarealso included. t-statistics arecomputedusingHuber-Whitestandarderrors unsecuredunsecuredunsecuredunsecured TRADING0.1944 0.1422 0.1659 0.2010 (4.49)(3.40)(4.78)(3.38) TRADED0.3286 0.0697 0.0603 0.0048 (7.94)(4.09)(2.14)(0.18) paten0.0104 (0.63) TRADED paten-0.1584 (-3.41) TRADING paten0.0844 (2.33) 0.0120 (0.65) TRADED -0.1539 (-3.41) TRADING 0.0909 (2.11) N 1615116151384216151 R 2 0.1610.2870.3390.263 FixedYESYESYESYES YearYESYESYESYES RatingNOYESYESYES ControlsNOYESYESYES t statisticsinparentheses p<: 10, p<: 05, p<: 01 66 TableA.6inesfor2003ISDAprovisions PanelApresentstheresultsofaDIDanalysisusingthe2003ISDAprovisionsasatreatment Thesampleincludesyears2001-2005.Thetreatmentgroupincludesallthat hadaCDS begin tradingduring2003. TREATED isadummyvariableequalto1ifa belongstothisgroup.IfahadaCDSstarttradingafter2005orisanon-CDS thentheisdroppedfromthesampleinPanelA.Thecontrolgroupincludesthat hadaCDStradingpriorto2003. PT isadummyvariableequalto1during2003-2005. TRADING isadummy=1inearsforwhichaCDSisactivelytradingforinthe treatmentgroup.Forthecontrolgrouphowever,thisvariableissetto0.Inthis analysis TRADING isequivalenttointeractiontermofinterest(posttreatment treatment). PanelBpresentsresultsfromatestinwhichallwithaCDStradingprior to2005aredroppedandallwithaCDSthatstartedtradingafter2005areusedas thetreatmentgroup(TRADED)andthecontrolgroupbecomesallnon-CDSThe interaction(TRADING)issetto1during2003-2005forthisgroup.All spincludecontrols,industryandstandarderrorsclusteredatthe level. Spforcolumns1-3: y i;t = ind + 1 pt + 2 treated + 3 pt treated + 1 Q i;t 1 + 2 lnsize i;t 1 + 3 ROA i;t 1 + 4 Profit i;t 1 + 5 CF i;t 1 + 6 rated i;t + 8 hvol i;t 1 + " i;t Spforcolumns4-5: y i;t = ind + 1 pt + 2 treated + 3 pt treated + 1 Q i;t + 2 lnsize i;t + 3 ROA i;t + 4 Profit i;t + 5 CF i;t + 6 rated i;t + 7 mtaxrate i;t 1 + 8 hvol i;t 1 + " i;t PanelA:in npatentspat/empR&D/saleLeverageUnsecured TRADING2.6364 0.5848 0.05598 0.0302 0.047 (PT TREATED)(2.09)(1.66)(1.76)(3.12)(2.04) PostISDARevision-6.64320 -0.0262-0.038279 -0.0183 0.00333 (PT)(-3.62)(-0.72)(-2.07)(-2.9)(1.01) TREATED-4.902 0.08313-0.1185 0.02221 -0.00942 (-3.71)(0.25)(-1.88)(2.05)(-0.65) N 18271827237323732373 PanelB:FalsiTest npatentspat/empR&D/saleLeverageUnsecured TRADING0.817-0.057460.00420.00390.0057 (PT TREATED)(0.99)(-0.22)(1.21)(1.01)(1.39) PostISDARevision-2.64320 -0.589 -0.0149 0.0005-0.00798 (PT)(-2.12)(2.36)(-2.65)(1.43)(-1.19) TREATED-1.902 -0.0262-0.0222 0.0146 0.0003 (-1.91)(-0.06)(-2.08)(2.81)(0.59) N 34273427368836883688 67 TableA.7ImpactofCDSonUnsecuredDebt:ParameterEstimates Thistablepresentsprobit,logit,andlinearprobability(LP)estimatesfortheimpactof creditdefaultswapsonaccesstounsecuredThesampleconsistsofloan-level observationsform-yearsinwhichanewloanwasoriginatedaccordingtoDealscan.Inde- pendentvariablesareintable1andintheappendix.Theindependentvariableisa dummyvariableequaltooneifaobtainsanew,unsecuredloaninagivenyear.The unconditionalprobabilityofobtainingunsecuredinthesampleis.37. ProbitLogitLPProbitLogitLP TRADING0.1792 0.2907 0.0325 0.2308 0.4009 0.0433 (2.71)(2.52)(2.03)(3.28)(3.28)(1.81) MTB0.1440 0.2388 0.0414 0.01290.00010.0045 (5.85)(5.55)(4.39)(0.27)(0.00)(0.29) FixedAssets0.1586 0.2380 0.0263-0.0810-0.1643-0.0341 (2.44)(2.09)(1.05)(-0.84)(-0.99)(-0.73) 0.4537 0.9278 0.07840.7067 1.3707 0.1765 (1.96)(2.21)(1.00)(1.79)(1.95)(1.41) size0.0464 0.0765 0.0128-0.1198 -0.2017 -0.0279 (2.26)(2.13)(1.03)(-3.81)(-3.70)(-1.58) rated-0.6394-1.29090.9113 -0.8794 -1.73350.6610 (-1.36)(-1.25)(27.44)(-1.66)(-1.54)(8.31) mtaxrate2.3382 4.5415 0.3438 -0.7156-1.0644-0.1162 (4.31)(4.29)(2.72)(-0.77)(-0.64)(-0.37) hvol-1.4610 -2.6135 -0.2221 -0.6437 -1.1676 -0.0672 (-9.18)(-9.20)(-6.40)(-2.60)(-2.59)(-2.29) TRADED-0.1621 -0.2693 -0.0321 (-2.23)(-2.10)(-1.72) nobs779277927822379837983861 R2/PR20.20420.20500.19250.19970.20070.2052 IndustryFEYESYESYESYESYESYES RatingFEYESYESYESYESYESYES YearFEYESYESYESYESYESYES ClusterNONOYESNONOYES SampleFULLFULLFULLTRADEDTRADEDTRADED t statisticsinparentheses Allvariableswinsorizedatthe1%level 68 TableA.8ImpactofCDSonUnsecuredDebt:Marginal Thistablepresentsprobit,logit,andlinearprobability(LP) averagemarginalcts esti- matesfortheimpactofcreditdefaultswapsonaccesstounsecuredestimatedin theprevioustable.Independentvariablesareintable1andintheappendix.The dependent variableisadummyvariableequaltooneifaobtainsanew,unsecuredloan inagivenyear.Thesampleinthistableonlyincludesearsinwhichanewloans havebeenoriginatedaccordingtoDealscan.Theunconditionalprobabilityofaloanbeing unsecuredinthissampleis.37. ProbitLogitLPProbitLogitLP TRADING0.0441 0.0415 0.0325 0.0611 0.0626 0.0433 (2.66)(2.48)(2.03)(3.36)(3.38)(1.81) MTB0.0347 0.0335 0.0414 0.003490.00002300.0045 (5.89)(5.59)(4.39)(0.27)(0.00)(0.29) FixedAssets0.0382 0.0333 0.0263-0.0219-0.0262-0.0341 (2.44)(2.08)(1.05)(-0.84)(-1.00)(-0.73) 0.109 0.130 0.07840.1910.219 0.1765 (1.96)(2.21)(1.00)(1.79)(1.96)(1.41) size0.0112 0.0107 0.0128-0.0324 -0.0322 -0.0279 (2.26)(2.13)(1.03)(-3.83)(-3.72)(-1.58) rated-0.154-0.1810.9113 -0.237-0.2770.6610 (-1.36)(-1.25)(7.44)(-1.66)(-1.55)(8.31) mtaxrate0.564 0.636 0.3438 -0.193-0.170-0.1162 (4.33)(4.30)(2.72)(-0.77)(-0.64)(-0.37) hvol-0.352 -0.366 -0.2221 -0.174 -0.186 -0.0672 (-9.29)(-9.31)(-6.40)(-2.60)(-2.60)(-2.29) TRADED-0.0391 -0.0377 -0.0321 (-2.23)(-2.11)(-1.72) N 779277927822379837983861 R2/PR20.20420.20500.19250.19970.20070.2052 IndustryFEYESYESYESYESYESYES RatingFEYESYESYESYESYESYES YearFEYESYESYESYESYESYES ClusterNONOYESNONOYES SampleFULLFULLFULLTRADEDTRADEDTRADED t statisticsinparentheses Allvariableswinsorizedatthe1%level 69 TableA.9UnsecuredFinancingandFirmTangibility Thistablepresentsestimatesoftheimpactofcreditdefaultswapsonaccesstounsecured asitrelatestotangibility. tang isadummyvariable=1(0)ifaisinthe top(bottom)2quintilesofrankedontangibilityinthefullsample.Similarly, intang isadummy=1ifaisinthetop2quintilesofrankedonintangibility(measured asR&D+advertisingexpensescaledbytotalassets).Allothervariablesareasin table1.The(second)columncontainstheestimatesofalinearprobabilitymodelfor onlythosemyearsinwhichtang==0(1).Columnthreecontainsestimatesforall. Similarly,Thefourthcolumncontainstheestimatesofalinearprobabilitymodelfor onlythosemyearsinwhichintang==0(1).Finally,columnsixalsocontainsestimates forallms. FullSampleTRADEDsample tang=0tang=1interactionintang=0intang=1interaction TRADING0.1092 0.0850 0.1607 0.0833 0.1101 0.0783 (4.33)(2.13)(4.57)(3.31)(3.01)(5.05) tang0.0347 (1.70) TRADING-0.1036 tang(-2.41) intang 0.0092 (0.93) TRADING0.0561 intang(2.90) MTB0.0431 0.0396 0.0497 0.0729 0.0283 0.0508 (2.42)(3.15)(5.10)(4.86)(2.51)(5.18) 0.1836 -0.05430.06010.01010.15330.0178 (2.49)(-0.45)(0.63)(0.10)(1.39)(0.70) size0.00420.0593 0.02850.02280.02740.0392 (0.25)(3.57)(1.54)(1.28)(1.65)(6.04) rated-0.0057-0.1394 -0.0636-0.0393-0.1113 -0.0647 (-0.13)(-6.00)(-1.67)(-0.89)(-3.77)(-4.02) mtaxrate0.2382 0.06000.11290.4033 -0.08940.4594 (1.90)(0.35)(0.79)(2.79)(-0.38)(5.71) hvol-0.4826 -0.3293 -0.4336 -0.4134 -0.4666 -0.4564 (-12.10)(-4.20)(-8.42)(-8.55)(-5.61)(-8.24) TRADED-0.0665 0.0426-0.0281-0.0376 0.0031-0.0048 (-2.53)(1.24)(-1.18)(-1.98)(0.06)(-0.34) nobs399524687822420516487822 R20.12140.16640.12400.12580.14860.1168 FixedIndustryIndustryIndustryIndustryIndustryIndustry YearYESYESYESYESYESYES ClusterYESYESYESYESYESYES SampleFULLFULLFULLFULLFULLFULL 70 TableA.10LenderSizeandtheImpactofCDSonAccesstoCapital Thistablepresentstheresultsofaanalysisofthe impactofaCDSonleverageasitrelatestolendersize.Thedependentvariableinthe regressionsisbookleverage. biglender isadummyvariable=1ifahasabankthat isoneofthe15largestcapitalprovidersasasyndicatememberonaloan. biglender =0if allofam'ssyndicatememberslieoustsidethethe15largestcapitalproviders.Allother variablesareintable1.Thecolumnpresentsregressionresultsobtainedonly onthesetofwithoutabiglenderonitsloanfacility.Columnsthreeandfourpresent resultsfromthefullsetofwiththeinteractionterm. biglender=0biglender=1AllFirmsAllFirms TRADING0.0821 0.0344 0.0481 0.0252 (2.12)(3.61)(4.32)(3.63) biglender0.0499 0.0171 (3.75)(2.91) TRADING biglender-0.0186 -0.0062 (-2.47)(-1.00) MTB-0.0331-0.0356 -0.0318 0.0018 (-1.14)(-3.90)(-3.40)(0.31) FixedAssets0.1297 0.03650.0368-0.0056 (2.34)(1.65)(1.62)(-0.23) -0.1129-0.0012-0.0143-0.0180 (-0.31)(-0.01)(-0.14)(-0.54) lnsize-0.0613 -0.0477 -0.0491 -0.0474 (-4.75)(-6.96)(-7.36)(-8.15) SPrating-0.0010-0.00030.0001-0.0054 (-0.23)(-0.20)(0.07)(-5.87) mtaxrate-0.1376-0.1063-0.0990-0.1518 (-0.24)(-0.89)(-0.81)(-2.06) nobs203698371867186 R20.63100.33540.35050.8281 FEIndustryIndustryIndustryFirm YearFEYESYESYESYES ClusterYESYESYESYES SampleTRADEDTRADEDTRADEDTRADED t statisticsinparentheses p<: 10, p<: 05, p<: 01 71 TableA.11SummaryStatistics (1)(2)(3)(4)(5)(6)(7)(8) VARIABLESNmeansdp10p25p50p75p90 adj patent19,8822.0395.834000.2150.9744.738 adj cites19,88235.01108.9002.44315.9574.74 lnSize19,8755.6932.1403.0114.1655.5857.1018.524 MTB19,8822.2892.1110.9421.1081.5432.5224.430 RDS19,8820.4471.8330.008030.02050.05830.1470.441 SA19,882-1.31e-071.000-1.430-0.6580.04380.6931.227 WW19,882-2.16e-081.000-1.374-0.7360.04770.7331.247 SA cited19,8824.30e-081.000-1.295-0.6900.01150.6721.269 WW cited19,8821.17e-071.000-1.239-0.674-0.03170.6211.285 RDS cited19,8820.1340.2950.03330.04390.06360.09810.187 MTB cited19,8821.9290.9481.2261.4251.7592.2102.838 tangibility19,8820.4840.2110.1890.3390.4990.6300.727 cashholdings19,8820.2150.2240.01300.03850.1280.3280.570 19,8820.07120.283-0.1940.03000.1250.1980.280 hi hi hi hi hi 72 TableA.12FinancialConstraintsOLSRegressions PanelA WWSA patentscitespatentscites FC cited0.593***13.27***0.749***15.54*** (0.0684)(2.002)(0.101)(3.077) MTB0.0452***1.247***0.0639***1.486*** (0.0132)(0.308)(0.0135)(0.311) RDS0.0341***0.676***-0.0101-0.124 (0.0112)(0.237)(0.00668)(0.147) FC-0.623***-11.30***-0.906***-12.70*** (0.0777)(1.744)(0.201)(4.85) MTB cited0.03790.9480.04351.098 (0.0426)(0.908)(0.0438)(0.941) RDS cited0.431**9.997***0.404**9.880*** (0.175)(3.244)(0.175)(3.211) -0.0759-1.099-0.0370.156 (0.0886)(1.999)(0.091)(2.078) cashholdings-0.338**-10.79***-0.312*-11.28*** (0.167)(3.628)(0.166)(3.395) Observations19882198821988219882 R-squared0.8520.7890.8520.789 FirmFEYESYESYESYES YearFEYESYESYESYES PanelB WWSA patentscitespatentscites FC cited0.568***12.71***0.708***15.09*** (0.0675)(1.926)(0.0982)(2.978) MTB0.0372***1.135***0.0520***1.348*** (0.0126)(0.307)(0.013)(0.314) RDS0.0305***0.607**-0.00894-0.131 (0.0114)(0.249)(0.00754)(0.174) FC-0.551***-10.33***-0.692***-10.54** (0.071)(1.695)(0.208)(5.213) MTB cited0.02971.3610.0351.494 (0.0399)(0.994)(0.0412)(1.035) RDS cited0.385**10.03***0.370**9.963*** (0.151)(3.057)(0.152)(3.009) -0.0456-0.1910.007391.236 (0.0829)(1.882)(0.0883)(2.097) cashholdings-0.430***-12.64***-0.439***-13.50*** (0.163)(3.452)(0.168)(3.467) Observations19882198821988219882 R-squared0.860.7970.860.796 FirmFEYESYESYESYES IndxYearFEYESYESYESYES 73 TableA.13AJCATaxHoliday WWSA citespatentscitespatents TreatedxPost-0.670***-11.55***-0.656***-11.39*** (0.148)(3.432)(0.142)(3.419) FC0.05171.458**-0.817***-9.141** (0.034)(0.665)-0.269-4.07 MTB0.0324**0.641*0.0442**0.764** (0.0145)(0.327)(0.0173)(0.338) RDS-0.0207**-0.385**-0.0208**-0.339* (0.00858)(0.19)(0.00969)(0.174) MTB cited-0.436*-11.52*-0.402*-11.13* (0.238)(5.846)(0.236)(5.909) RDS cited0.2736.3710.276.309 (0.215)(4.931)(0.219)(5.011) 0.0488-1.142-0.0908-2.841 (0.124)(3.111)(0.113)(2.801) cashholdings-0.518*-11.96-0.329-9.56 (0.269)(7.703)(0.275)(8.017) tangibility-0.622-15.22-0.222-10.08 (0.48)(13.45)(0.485)(14.06) Observations5047504750475047 R-squared0.6730.5050.6740.506 FirmFEYESYESYESYES YearFEYESYESYESYES 74 TableA.14AJCA2 WWSA patentscitespatentscites FC cited-0.0000593-0.486-0.081-3.190* (0.0664)(1.283)(0.0734)(1.621) FC cited 1 [ Post ]-0.129***-1.687**-0.222***-3.283*** (0.0445)(0.745)(0.0614)(1.154) FC cited 1 [ Treated ]0.346**8.917**0.08652.502 (0.169)(3.703)(0.127)(2.329) FC cited 1 [ Treated Post ]-1.005***-19.34***-0.728***-14.87*** (0.274)(6.968)(0.196)(5.362) 1 [ Treated Post ]-0.772***-13.43***-0.883***-16.07*** (0.191)(4.332)(0.245)(5.591) MTB0.01060.2480.03050.493* (0.0166)(0.247)(0.0185)(0.262) RDS-0.00158-0.0252-0.0146-0.213 (0.00882)(0.174)(0.00963)(0.166) FC0.03571.204**-0.692***-6.941* (0.0296)(0.559)(0.236)(3.768) MTB cited-0.193-7.091-0.266-8.752 (0.181)(4.849)(0.187)(5.23) RDS cited0.1153.0030.1734.671 (0.173)(3.272)(0.176)(3.562) 0.0551-1.141-0.0595-2.311 (0.104)(2.61)(0.0968)(2.295) cashholdings-0.588*-13.27-0.531*-13.37 (0.303)(8.547)(0.274)(8.54) tangibility-0.34-10.180.0418-5.002 (0.44)(12.38)(0.408)(11.94) Observations5047504750475047 R-squared0.7120.5520.7040.543 FirmFEYESYESYESYES IndxYearFEYESYESYESYES 75 TableA.15JunkBond WWSA patentscitespatentscites JunkBondPct 1 [ Post ]1.218**37.96***1.215**36.70*** (0.532)(9.746)(0.522)(9.574) MTB0.0102-0.5380.0137-0.523 (0.0285)(0.557)(0.0269)(0.557) RDS0.01330.457-0.00310.531 (0.0298)(0.472)(0.0381)(0.476) FC-0.161-3.294**-0.1792.766 (0.0957)(1.438)(0.164)(2.597) MTB cited0.205*-0.1010.188-0.195 (0.112)(1.613)(0.118)(1.655) RDS cited-0.565-1.638-0.551-0.866 (0.416)(8.895)(0.431)(9.012) 0.153-7.792**0.174-7.928** (0.272)(3.777)(0.275)(3.709) cashholdings-1.132***-18.37***-1.159***-19.84*** (0.393)(6.084)(0.403)(6.427) tangibility-0.834-24.08**-0.871-27.38** (0.584)(11.45)(0.643)(12.3) Observations1856185618561856 R-squared0.980.9770.980.977 FirmFEYESYESYESYES YearFEYESYESYESYES 76 TableA.16JunkBond2 WWSA patentscitespatentscites 1 [ Treated Post ]0.0805.570***0.0815.464*** (0.0648)(1.11)(0.0649)(1.102) MTB0.0136-0.3040.0163-0.294 (0.0242)(0.481)(0.023)(0.489) RDS0.02590.6960.01680.847** (0.0309)(0.425)(0.0379)(0.42) FC-0.112-2.297-0.09313.327 (0.0961)(1.501)(0.148)(2.388) MTB cited0.125-0.4930.114-0.505 (0.101)(1.433)(0.105)(1.438) RDS cited-0.336-0.171-0.3240.504 (0.42)(8.974)(0.431)(9.062) 0.111-8.424**0.123-8.629** (0.293)(3.8)(0.296)(3.769) cashholdings-1.214***-20.57***-1.234***-21.84*** (0.366)(5.744)(0.375)(6.008) tangibility-1.104*-26.08**-1.144*-29.05** (0.56)(11.31)(0.615)(12.14) Observations1856185618561856 R-squared0.9810.9780.9810.978 FirmFEYESYESYESYES YearFEYESYESYESYES 77 TableA.17SelfCitationPercentage WWSA FC cited-0.00509***-0.00480***-0.00778***-0.00753*** (0.00113)(0.00112)(0.00144)(0.00142) FC0.00563***0.00625***0.0105***0.0140*** (0.00181)(0.00189)(0.00391)(0.00398) MTB0.0000.0010.0001470.000237 (0.000377)(0.000378)(0.00039)(0.000384) RDS0.0010.0010.00164**0.00146* (0.000808)(0.00076)(0.000808)(0.000752) MTB cited-0.001-0.001-0.000859*-0.000692 (0.000517)(0.000499)(0.000516)(0.000495) RDS cited-0.00831***-0.00692***-0.00742***-0.00612*** (0.0018)(0.00192)(0.00186)(0.00198) 0.0010.0020.000850.00179 (0.00307)(0.00316)(0.00314)(0.00321) cashholdings0.0156***0.0164***0.0147***0.0146*** (0.00522)(0.00525)(0.00514)(0.00512) tangibility-0.00948*-0.00911*-0.0109*-0.0117** (0.00555)(0.00546)(0.00559)(0.00549) 1 [ Patents> 0]-0.0242***-0.0241***-0.0244***-0.0243*** (0.00231)(0.00228)(0.00229)(0.00228) citations4.99e-05***4.82e-05***5.07e-05***4.91e-05*** (0.0000135)(0.0000139)(0.0000135)(0.0000141) Observations19882198821988219882 R-squared0.4760.490.4760.491 FirmFEYESYESYESYES YearFEYESNOYESNO IndxYearFENOYESNOYES 78 TableA.18SelfCitationPerPatent WWSA FC cited-0.130**-0.118*-0.289***-0.280*** (0.064)(0.0669)(0.0569)(0.0573) FC0.432***0.426***1.405***1.389*** (0.101)(0.102)(0.38)(0.392) MTB0.0000.002-0.0326***-0.0308*** (0.0104)(0.0107)(0.0103)(0.0106) RDS-0.010-0.0120.02060.0182 (0.016)(0.0153)(0.0128)(.0003) MTB cited-0.0369**-0.0399**-0.0382**-0.0404** (0.0171)(0.0179)(0.0172)(0.0177) RDS cited-0.192***-0.177***-0.137**-0.126** (0.057)(0.0615)(0.0569)(0.0601) 0.0990.0950.187**0.190** (0.118)(0.12)(0.0926)(0.0924) cashholdings0.378*0.464**0.140.207 (0.209)(0.213)(0.214)(0.217) tangibility-0.336**-0.284*-0.656***-0.619*** (0.151)(0.148)(0.211)(0.214) 1 [ Patents> 0]-0.162***-0.174***-0.186***-0.196*** (0.0447)(0.0444)(0.0462)(0.0463) citations-0.00121**-0.00115*-0.00110*-0.00105* (0.000579)(0.000586)(0.000594)(0.000607) Observations19882198821988219882 R-squared0.3780.3850.3880.393 FirmFEYESYESYESYES YearFEYESNOYESNO IndxYearFENOYESNOYES 79 TableA.19MahalanobisDistance WWSA MdMdMdMd FC -0.0901***-0.118***-0.0336***-0.0712*** (0.0076)(0.004)(0.00827)(0.00234) RDS 8.68e-05**0.000108***-0.00001931.26e-05* (0.0000347)(0.0000229)(0.0000165)(0.00000725) FC-0.0940***-0.0469*** (0.0128)(0.0136) RDS0.00853***-0.00107 (0.00181)(0.00137) MTB0.00828**0.00361 (0.0035)(0.00398) junkpost junk pifopct pifopost Observations83906810494728569711056654 R-squared0.1460.5850.0320.499 Firm-YearFENOYESNOYES 80 TableA.20MarketValue WWSA ln ( Size ) ln ( Size ) ln ( Size ) ln ( Size ) FC cited0.0120.0030.0370.030 (0.494)(0.127)(1.327)(1.126) FC cited 1 [ HighCites ]0.01510.0437** (0.726)(2.177) 1 [ HighCites ]0.119***0.118*** (6.227)(6.053) FC cited 1 [ HighPatents ]0.02450.0494*** (1.3)(2.609) 1 [ HighPatents ]0.140***0.140*** (7.99)(7.793) RDS0.0422***0.0419***0.01310.0132* (4.909)(0.975)(0.986)(1.001) FC-0.414***-0.407***-0.745***-0.729*** (-11.99)(7.23)(6.696)(6.71) RDS cited0.04310.04260.04340.0431 (0.97)(2.439)(2.983)(3.018) 0.813***0.814***0.796***0.797*** (7.218)(-15.05)(-14.91)(-14.95) cashholdings0.234**0.239**0.298***0.300*** (2.377)(7.99)(7.793) tangibility-1.831***-1.827***-1.831***-1.828*** (-15.01)(90.59)(71.43)(71.64) Observations19875198751987519875 R-squared0.9220.9220.9210.921 FirmFEYESYESYESYES YearFEYESYESYESYES 81 TableA.21LargeMovesinFC citespatents FC cited move139.6***5.527*** (5.936)(5.698) FC-52.53***-2.622*** (-9.267)(-10.15) MTB4.585***0.196*** (3.103)(2.952) RDS2.115***0.0839*** (2.926)(2.613) MTB cited13.26***0.740*** (3.2)(4.334) RDS cited0.2230.0205 (0.575)(1.164) -25.78***-1.632*** (-3.863)(-5.493) cashholdings51.39***2.077*** (4.215)(3.753) tangibility42.01***2.074*** (3.749)(4.035) Constant-117.1***-3.439*** (-4.348)(-3.071) Observations1988219882 R-squared0.1940.237 FirmFEYESYES YearFEYESYES 82 hi hi hi hi hi TableA.22ModelSps DepVarX1X2X3X4X5X6 bookleverageQlogSaleROAz-scoremcaptangibility debtissQlogSaleROAz-scoremcaptangibility equityissQlogSaleROAz-scoremcaptangibility capex/atQlogSalez-scoreCFcash R&D/atQlogSalez-scoreCFcash ROAQlogSaleR&D/at ActRec/ActPayQROAStockReturn ownershipQROAStockReturn compensationQROAStockReturn QR&D/atlogSaleROA TableA.23FullSampleTests DepVarFE1/FD1FE2/FD2FE3/FD3FE4/FD4FE5/FD5FE6/FD6Joint leverage.00/.20.00/.00.02/.00.00/.00.31/.00.00/.14.00/.00 debtiss.00/.00.00/.00.00/.00.00/.00.00/.00.72/.53.00/.00 equityiss.00/.00.00/.35.00/.00.00/.00.00/.48.00/.00.00/.00 capex/at.00/.00.00/.00.00/.34.00/.00.00/.00.00/.00 R&D/at.00/.00.00/.00.42/.11.13/.00.00/.00.00/.00 ROA.91/.59.00/.00.22/.17.00/.00 Rec/Pay.00/.00.00/.02.77/.90.12/.04 ownership.28/.00.07/.05.00/.00.01/.00 comp.00/.00.00/.31.00/.00.00/.00 Q.23/.96.00/.00.00/.01.00/.00 83 TableA.24Sub-sampleTests PanelA:Medianp-valueby10-yearsub-periods DepVarFE1/FD1FE2/FD2FE3/FD3FE4/FD4FE5/FD5FE6/FD6Joint leverage.127/.525.007/.000.566/.000.000/.001.184/.035.277/.258.000/.000 debtiss.006/.053.000/.001.051/.000.000/.000.404/.000.000/.045.000/.000 equityiss.000/.000.000/.117.316/.001.002/.056.000/.298.000/.155.000/.000 capex/at.000/.493.001/.000.219/.058.016/.000.000/.000.000/.000 R&D/at.014/.120.037/.362.422/.335.187/.363.000/.199.000/.074 ROA.286/.007.000/.012.685/.289.000/.000 Rec/Act.312/.000.075/.172.485/.917.020/.024 ownership.188/.113.135/.259.229/.079.000/.010 comp.524/.280.007/.340.033/.000.000/.000 Q.187/.277.000/.002.000/.065.000/.000 PanelB:Medianp-valueby1digitsicindustrysub-samples DepVarFE1/FD1FE2/FD2FE3/FD3FE4/FD4FE5/FD5FE6/FD6Joint leverage.086/.512.086/.000.253/.000.000/.339.271/.118.008/.069.006/.000 debtiss.001/.017.000/.009.406/.002.000/.000.109/.023.121/.053.000/.001 equityiss.000/.000.000/.340.242/.129.011/.129.000/.090.000/.001.000/.016 capex/at.001/.301.039/.000.313/.060.312/.000.000/.000.000/.001 R&D/at.063/.387.196/.191.428/.191.304/.279.331/.279.257/.372 ROA.047/.020.000/.044.108/.274.000/.274 Rec/Pay.579/.225.045/.043.298/.651.548/.206 ownership.346/.090.284/.472.303/.041.210/.232 comp.214/.236.177/.493.071/.000.244/.036 Q.409/.399.000/.045.006/.405.000/.039 84 TableA.25SignBetweenFDandFE %regressionswithsign DependentVariablex1x2x3x4x5x6 bookleverage.062.500.000.000.000.005 debtiss.000.222.171.000.500.280 equityiss.000.500.112.390.000.171 capex/at.000.222.444.000.000 R&D.567.112.000.444.171 ROA.000.000.000 ActRec/ActPay.281.000.062 ownership.062.281.171 compensation.062.000.000 Q.333.062.171 %regressionswithsignsigtat10%level DependentVariablex1x2x3x4x5x6 bookleverage.000.171.000.000.000.005 debtiss.000.171.062.000.390.171 equityiss.000.390.112.112.000.062 capex/at.000.112.390.000.000 R&D.281.062.000.000.062 ROA.000.000.000 ActRec/ActPay.000.000.000 ownership.000.000.062 compensation.062.000.000 Q.110.005.062 medianratioofcots:max( fe ; fd )/min( fe ; fd ) DependentVariablex1x2x3x4x5x6 bookleverage1.582.191.281.511.521.18 debtiss1.161.991.281.571.392.15 equityiss1.132.961.733.414.422.50 capex/at1.461.952.131.341.52 R&D2.201.211.241.431.56 ROA1.241.701.11 ActRec/ActPay1.531.441.74 ownership2.242.432.85 compensation1.361.341.20 Q1.761.411.74 85 TableA.26FE i correlations10yearsub-periods Model1 alpha1alpha2alpha3alpha4alpha5 alpha11.00000.70060.33800.23390.0995 alpha20.70061.00000.53940.31030.1066 alpha30.33800.53941.00000.59900.2479 alpha40.23390.31030.59901.00000.5656 alpha50.09950.10660.24790.56561.0000 Model2 alpha11.00000.69590.29610.31450.1201 alpha20.69591.00000.43190.33400.1004 alpha30.29610.43191.00000.31690.1368 alpha40.31450.33400.31691.00000.2035 alpha50.12010.10040.13680.20351.0000 Model3 alpha11.00000.7250-0.4637-0.2369-0.4584 alpha20.72501.0000-0.1509-0.0583-0.3113 alpha3-0.4637-0.15091.00000.36750.3652 alpha4-0.2369-0.05830.36751.00000.3909 alpha5-0.4584-0.31130.36520.39091.0000 Model4 alpha11.00000.61010.38320.44740.3231 alpha20.61011.00000.57700.53410.5205 alpha30.38320.57701.00000.63990.4704 alpha40.44740.53410.63991.00000.6985 alpha50.32310.52050.47040.69851.0000 Model5 alpha11.00000.82650.68980.68160.6110 alpha20.82651.00000.84780.72820.6921 alpha30.68980.84781.00000.80790.7077 alpha40.68160.72820.80791.00000.8636 alpha50.61100.69210.70770.86361.0000 Model6 alpha11.00000.78290.72160.59470.5896 alpha20.78291.00000.71580.63470.6514 alpha30.72160.71581.00000.65330.6526 alpha40.59470.63470.65331.00000.6121 alpha50.58960.65140.65260.61211.0000 Model7 alpha11.00000.61890.29020.31130.3609 alpha20.61891.00000.44110.25570.2798 alpha30.29020.44111.00000.56990.4191 alpha40.31130.25570.56991.00000.5088 alpha50.36090.27980.41910.50881.0000 86 TableA.27FE i correlations5yearsub-periods leverage = Q + logSale + ROA + zscore + mcap + tangibility + i:year alpha1alpha2alpha3alpha4alpha5alpha6alpha7alpha8alpha9 alpha11.0000.7740.5050.3760.2660.2190.2240.0970.072 alpha20.7741.0000.7370.4680.2880.2930.1780.1370.095 alpha30.5050.7371.0000.6670.3800.3600.1970.1020.029 alpha40.3760.4680.6671.0000.6110.4540.2870.1860.099 alpha50.2660.2880.3800.6111.0000.6290.3930.2610.233 alpha60.2190.2930.3600.4540.6291.0000.6270.3610.312 alpha70.2240.1780.1970.2870.3930.6271.0000.6410.419 alpha80.0970.1370.1020.1860.2610.3610.6411.0000.720 alpha90.0720.0950.0290.0990.2330.3120.4190.7201.000 debtiss = l:Q + logSale + ROA + l:zscore + l:mcap + l:tangibility + i:year alpha1alpha2alpha3alpha4alpha5alpha6alpha7alpha8alpha9 alpha11.0000.7340.5480.5580.2030.5430.1390.3360.268 alpha20.7341.0000.6940.5790.2850.5330.0700.3720.248 alpha30.5480.6941.0000.5480.3530.4630.0530.3300.153 alpha40.5580.5790.5481.0000.3860.5430.1430.3590.319 alpha50.2030.2850.3530.3861.0000.3860.2030.2680.219 alpha60.5430.5330.4630.5430.3861.0000.2760.4050.379 alpha70.1390.0700.0530.1430.2030.2761.0000.1980.189 alpha80.3360.3720.3300.3590.2680.4050.1981.0000.295 alpha90.2680.2480.1530.3190.2190.3790.1890.2951.000 equityiss = l:Q + logSale + ROA + l:zscore + l:mcap + l:tangibility + i:year alpha1alpha2alpha3alpha4alpha5alpha6alpha7alpha8alpha9 alpha11.0000.5360.432-0.521-0.1130.054-0.201-0.288-0.433 alpha20.5361.0000.323-0.1070.1240.1840.023-0.018-0.231 alpha30.4320.3231.000-0.1540.0640.083-0.007-0.116-0.154 alpha4-0.521-0.107-0.1541.0000.2970.2140.4440.4520.577 alpha5-0.1130.1240.0640.2971.0000.4400.3940.3230.224 alpha60.0540.1840.0830.2140.4401.0000.4810.3750.153 alpha7-0.2010.023-0.0070.4440.3940.4811.0000.5660.400 alpha8-0.288-0.018-0.1160.4520.3230.3750.5661.0000.488 alpha9-0.433-0.231-0.1540.5770.2240.1530.4000.4881.000 87 TableA.28FE i correlations5yearsub-periods capex=at = i + l:Q + logSale + cf + l:zscore + l:cash + i:year alpha1alpha2alpha3alpha4alpha5alpha6alpha7alpha8alpha9 alpha11.0000.5590.5400.2730.3630.4440.3640.3220.238 alpha20.5591.0000.6970.6310.4770.5040.5490.5500.508 alpha30.5400.6971.0000.5740.4910.4990.4880.4910.461 alpha40.2730.6310.5741.0000.5820.5010.4750.4200.456 alpha50.3630.4770.4910.5821.0000.6680.5430.5320.467 alpha60.4440.5040.4990.5010.6681.0000.7010.6090.562 alpha70.3640.5490.4880.4750.5430.7011.0000.7270.658 alpha80.3220.5500.4910.4200.5320.6090.7271.0000.793 alpha90.2380.5080.4610.4560.4670.5620.6580.7931.000 R & D=at = i + l:Q + logSale + cf + l:zscore + l:cash + i:year alpha1alpha2alpha3alpha4alpha5alpha6alpha7alpha8alpha9 alpha11.0000.8540.7830.7110.6780.6910.6500.5950.559 alpha20.8541.0000.9130.8170.7840.7490.6910.6920.681 alpha30.7830.9131.0000.8840.8030.7670.7150.6690.643 alpha40.7110.8170.8841.0000.8930.7930.6460.6760.681 alpha50.6780.7840.8030.8931.0000.8750.7350.7440.721 alpha60.6910.7490.7670.7930.8751.0000.8750.8150.797 alpha70.6500.6910.7150.6460.7350.8751.0000.8700.808 alpha80.5950.6920.6690.6760.7440.8150.8701.0000.919 alpha90.5590.6810.6430.6810.7210.7970.8080.9191.000 ROA = i + l:Q + logSale + R & D + i:year alpha1alpha2alpha3alpha4alpha5alpha6alpha7alpha8alpha9 alpha11.0000.8280.7770.7140.7030.6260.6330.6450.531 alpha20.8281.0000.8550.7080.7550.6730.6610.6840.602 alpha30.7770.8551.0000.7350.7650.7140.7140.7220.678 alpha40.7140.7080.7351.0000.6940.6600.6670.6990.721 alpha50.7030.7550.7650.6941.0000.7290.7050.6820.674 alpha60.6260.6730.7140.6600.7291.0000.7150.6860.643 alpha70.6330.6610.7140.6670.7050.7151.0000.6950.642 alpha80.6450.6840.7220.6990.6820.6860.6951.0000.746 alpha90.5310.6020.6780.7210.6740.6430.6420.7461.000 Q = i + R & D + logSale + ROA + i:year alpha1alpha2alpha3alpha4alpha5alpha6alpha7alpha8alpha9 alpha11.0000.4250.4640.2200.3280.3310.2960.4000.316 alpha20.4251.0000.5480.5110.4400.4270.4720.3700.350 alpha30.4640.5481.0000.4250.3100.2770.1320.3600.173 alpha40.2200.5110.4251.0000.5970.4370.4380.3620.280 alpha50.3280.4400.3100.5971.0000.6170.3810.4160.349 alpha60.3310.4270.2770.4370.6171.0000.5850.3990.358 alpha70.2960.4720.1320.4380.3810.5851.0000.5050.375 alpha80.4000.3700.3600.3620.4160.3990.5051.0000.594 alpha90.3160.3500.1730.2800.3490.3580.3750.5941.000 88 BIBLIOGRAPHY 89 BIBLIOGRAPHY Almedia,H.,P.-H.Hsu,andD.Li(2014),Whenlessismore:Financialconstraints andinnovative, WorkingPaper . Almeida,H.,andM.Campello(2007),Financialconstraints,assettangibility,and corporateinvestment, TheReviewofFinancialStudies ,pp.1429{1460. Arellano,M.,andS.Bond(1991),Sometestsofspforpaneldata:Monte carloevidenceandanapplicationtoemploymentequations, TheReviewofEco- nomicStudies , 58-2 ,277{297. Ashcraft,A.,andJ.Santos(2009),Hasthecdsmarketloweredthecostofcorporate debt, JournalofMonetaryEconomics ,pp.514{523. Atanassov,J.(2014),Arm'slengthandinnovation:Evidencefrompublicly traded ManagementScience(forthcoming) ,pp.1{49. Bloom,N.,M.Schankerman,andJ.V.Reenen(2013),Identifyingtechnology spilloversandproductmarketrivalry, Econometrica . Blundell,R.,andS.Bond(1998),Initialconditionsandmomentrestrictionsindy- namicpaneldatamodels, JournalofEconometrics , 87 ,115{143. Bolton,P.,andM.Oehmke(2011),Creditdefaultswapsandtheemptycreditor problem, TheReviewofFinancialStudies ,pp.2616{2655. Bolton,P.,andD.S.Scharfstein(1990),Atheoryofpredationbasedonagency problemsincialcontracting, TheAmericanEconomicReview . Brown,J.R.,G.Martinsson,andB.C.Peterson(2012),Doconstraints matterforr&d?, EuropeanEconomicReview . Cameron,A.C.,andP.K.Trivedi(2005),Microeconometrics:Methodsandappli- cations, CambridgeUniversityPress,NewYork:NY . Cameron,A.C.,andP.K.Trivedi(2010),Microeconometricsusingstata,revised edition, StataPress,CollegeStation:Texas . Chava,S.,andM.Roberts(2008),Howdoesgimpactinvestment?therole ofdebtcovenants, TheJournalofFinance , 58 ,2085{2121. Cohen,L.,K.Diether,andC.Malloy(2013),Misvaluinginnovation, TheReviewof FinancialStudies ,p.hhs183. Dang,M.K.,VietAnh,andY.Shin(2015),Insearchofrobustmethodsfordynamic paneldatamodelsinempiricalcorporate JournalofBankingandFinance . 90 Fazzari,S.M.,R.G.Hubbard,andB.C.Peterson(1988),Financingconstraintsand coporateinvestment, BrookingsPapersonEconomicActivity . Gormley,T.A.,andD.A.Matsa(2014),Commonerrors:Howto(andnotto)control forunobservedheterogeneity, ReviewofFinancialStudies , 27(2) . Hadlock,C.J.,andJ.R.Pierce(2010),Newevidenceonmeasuringcon- straints:Movingbeyondthekzindex, TheReviewofFinancialStudies . Hall,B.H.,andJ.Lerner(2010),Theingofr&dandinnovation, Tech.rep. Hall,B.H.,A.B.andM.Trajtenberg(2001),Thenberpatentcitationsdata Lesson,insightsandmethodologicaltools, Tech.rep. ,NBERWorkingPaper Series. Hart,O.,andJ.Moore(1994),Atheoryofdebtbasedontheinalienabilityofhuman capital, TheQuarterlyJournalofEconomics , 109 (4),841{879. Hilscher,P.J..W.M.,J.(2014),Arecreditdefaultswapsasideshow?evidencethat informationwsfromequitytocdsmarkets, JournalofFinancialandQuantitative Analysis . Hirtle,B.(2009),Creditderivativesandbankcreditsupply, JournalofFinancial Intermediation ,pp.125{150. Holmstrom,B.(1989),Agencycostsandinnovation, JournalofEconomicBehavior andOrganization , 12 ,305{327. Holmstrom,B.,andJ.Tirole(1997),Financialintermediation,lonablefunds,and therealsector, TheQuarterlyJournalofEconomics ,pp.663{691. Jarrow,R.(2011),Theeconomicsofcreditdefaultswaps, AnnualReviewofFinance andEconomics , 3 ,235{257. Jenter,D.,andF.Kanaan(2014),Ceoturnoverandrelativeperformanceevaluation, TheJournalofFinance(forthcoming) . Leary,M.T.,andM.R.Roberts(2014),Dopeercorporate policy?, TheJournalofFinance ,(69),129{178. Manski,C.(1993),Idennofendogeneoussocialct:Theprob- lem, TheReviewofEconomicStudies . Manso,G.(2011),Motivatinginnovation, TheJournalofFinance , 66 ,1823{1869. Nickell,S.A.(1981),Biasesindynamicmodelswith Econometrica , 49 (6),1471{1426. Parsons,C.A.,andS.Titman(2009),Empiricalcapitalstructure:Areview, Foun- dationsandTrendsinFinance , 3 . 91 Petersen,M.A.(2009),Estimatingstandarderrorsinpaneldatasets:Com- paringapproaches, TheReviewofFinancialStudies . Rajan,R.G.,andL.Zingales(2003),Thegreatreversals:Thepoliticsof developmentinthetwentiethcentury, JournalofFinancialEconomics , 69 (1),5{ 50. Roberts,M.R.,andT.M.WHited(2012),Endogeneityinempiricalcorporate nance, G.Constantinides,M.Harris,andR.Stulzeds.HandbookoftheEconomics ofFinanceVolume2 . Saretto,H.,A.&Tookes(2013),Corporateleverage,debtmaturity,andcreditsupply: Theroleofcreditdefaultswaps, TheReviewofFinancialStudies ,pp.1190{1247. Seru,A.(2012),Firmboundariesmatter:Evidencefromconglomeratesandr&d activity, JournalofFinancialEconomics . Shleifer,A.,andR.W.Vishny(1992),Liquidationvaluesanddebtcapacity:A marketequilibriumapproach, TheJournalofFinance , 47 (4),1343{1366. Thompson,S.B.(2011),Simpleformulasforstandarderrorsthatclusterbyboth andtime, JournalofFinancialEconomics , 99 ,1{10. Titman,S.,andR.Wessels(1988),Thedeterminantsofcapitalstructurechoice, The JournalofFinance , 43 (1),1{19. Tufano,P.(1996),Whomanagesrisk?anempiricalexaminationofriskmanagement practicesinthegoldminingindustry, TheJournalofFinance , 51 (4),1097{1137. Whited,T.,andG.Wu(2006),Financialconstraintsrisk, TheReviewofFinancial Studies . Wintoki,L.J.S.,BabajideM.,andJ.M.Netter(2012),Endogeneityandthedy- namicsofinternalcorporategovernance, JournalofFinancialEconomics , 105 (3), 581{606. Wooldridge,J.M.(2010),Econometricanalysisofcrosssectionandpaneldata,2nd edition, MITPress,Cambridge:Mass . 92