Why Do Innovative Firms Hold So Much Cash? Evidence from Changes in State R&D Tax Credits

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Why Do Innovative Firms Hold So Much Cash? Evidence from Changes in State R&D Tax Credits Antonio Falato and Jae Sim 2014-72 NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Why Do Innovative Firms Hold So Much Cash? Evidence from Changes in State R&D Tax Credits AntonioFalato JaeW.Sim FederalReserveBoard1 FederalReserveBoard ThisDraft: May2014 1Correspondingauthor:AntonioFalato,FederalReserveBoard,20thst&ConstitutionAvNWWashington DC20511Phone: (202)452-2861. Email: antonio.falato@frb.gov. WethankKaiLi, MitchellPetersen, Gordon Phillips,DavidScharfstein,andseminarparticipantsattheFederalReserveBoardforhelpfulcommentsand discussions. SpecialthankstoDarrellAshtonforhishelpwithparsingtheSECfilings,toTimLoughranand JayRitterformakingtheirsampleofseasonedequityofferingsavailabletous,andtoDanielWilsonandBob Chirinko for kindly sharing their hand-collected data on state-level investment tax credits. Suzanne Chang providedexcellentresearchassistance. Allremainingerrorsareours.

Abstract ThispaperusesthestaggeredchangesofR&DtaxcreditsacrossU.S.statesandovertimeasaquasinatural experiment to examine the impact of innovation on corporate liquidity. By generating plausibly independent variation in firms’ incentive to invest in R&D, we are able to assess the empirical importanceofspecifictheoriesofthelinkbetweeninnovationandcorporateliquidity. Firmsincrease (decrease) their cash to asset ratios by about one and a half percentage point when their home state increases(cuts)R&Dtaxcredits. Thesebaselinedifference-in-differencesestimatesholduptoabatteryofvalidation,falsification,androbustnesschecks,whichcorroboratetheirinternalandexternal validity. The treatment effect of R&D tax credits increases monotonically with several specific proxies for debt and equity financing frictions. Increases (cuts) in tax credits also lead to increases (decreases)intheratiosofcashtobanklinesofcreditandtobookequity,andtodecreases(increases)in bank debt, secured debt, and overall net indebtness, supporting debt and equity financing channels through which innovation impactsthe demandfor cash. We also find support for a product market competition channel, and assess repatriation and agency explanations. Overall, our analysis offers endogeneity-free evidence that innovation is a first-order driver of corporate liquidity management decisions.

1 Introduction That the top U.S. cash holders are innovative corporations which rely heavily on R&D investments iswell-knownandtobeexpectedbasedonfirstprinciplesincorporatefinance(forexample,Aghion andTirole(1994),HartandMoore(1994)). Infact,innovativecorporationsheld,onaverage,asmuch as a third of their assets in cash or near-cash instruments over the last decade, and the five largest cashholderswereApple,Microsoft,Google,Cisco,andPfizer,whoheldmorethanonequarterofthe total cash of the U.S. corporate sector. Especially since U.S. corporate cash holdings hit new record highs after the financial crisis, the subject of cash holdings of innovative firms has made headlines inthepressandreceivedattentionbypolicymakersandinstitutionalshareholderactivists,withthe likesofCarlIchanandDavidEinhornpressuringinnovativefirmstodisgorgetheircash.1 However,whatisrelativelylesswellunderstoodandremainsactivelydebatedisthequestionof why innovative firms hold so much cash. Is it because innovation matters for cash, or is it because innovativefirmsarejustfundamentallydifferentfromtheaverageU.S.corporationalongobservable and unobservable dimensions that are hard to control for? In addition, if innovation matters, how does it matter – i.e., through which channels does it impact cash? Despite the small yet growing literature on the financial economics of innovation, which has been generally focused on external ratherthaninternalfinancing(Brownetal(2009; 2013), Chavaetal(2013); HallandLerner(2010)is asurvey),andthevastliteratureonthedeterminantsofcash,which,startingfromOpleretal(1999), hasgenerallyincludedR&Dasacovariateforalargecross-sectionoffirms,todatetherehasbeenno attemptatempiricallyidentifyingtheimpactofinnovationoncashandatsystematicallyevaluating thechannels. Inordertofillthegapintheliterature,wedesignaquasi-naturalexperimentalsettingthatovercomes the empirical identification challenge of finding plausibly exogenous sources of variation in innovationbyexploitingstaggeredchangesinR&DtaxcreditsacrossU.S.statesandovertime. One ofthechallengeswithcross-sectionalestimatesisthatmeasuresofinnovationarealsocorrelatedwith proxies for why innovation matters, which complicates the examination of how innovation impacts corporateliquidity. Bygeneratingvariationinfirms’incentivestoinvestinR&Dthatisplausiblyunrelatedtofirm-levelcharacteristics,ourquasi-experimentaldesignalsohelpstoovercomethisissue. In summary, we offer well-identified estimates of the impact of innovation on cash, direct evidence that innovation is a first-order determinant of corporate liquidity, and a comprehensive assessment 1See,forexample,"TooMuchCashIsn’tGoodforApple,"BusinessWeek,February26,2013. 1

ofthespecificchannelsthroughwhichinnovationimpactscash. We start with presenting descriptive evidence that there is a positive cross-sectional relation betweenmeasuresoffirminnovationbasedonR&Dexpendituresandcashholdingsrelativetoassets as well as external liquidity (bank lines of credit) in a standard panel of U.S. firms between 1986 and2011. Thisevidence,whichreplicatesinoursamplethefindingsofOpleretal(1999)andBates, Kahle,andStulz(2009)forcashandSufi(2009)forlinesofcredit,isimportant,butitsinterpretation is limited by a variety of well-known endogeneity concerns. In particular, when measuring the effect of innovation using comparisons across firms, there will always be a suspicion that the controls included in the analysis are not exhaustive. Such concerns are mitigated when the specification includes firm fixed effects. Nevertheless, if there are time-varying omitted firm characteristics, such as future profit growth opportunities, that affect both the firm’s incentives to invest in R&D and its cash holdings decisions, then the estimates would not have a causal interpretation. As such, the cross-sectional evidence lays out the foundation for our effort to take a first step toward identifying thecausallinkbetweenR&Dandcash,andassessingtheempiricalrelevanceofthetheoriesthatcan explainthelink. Toachieveidentification,weexploitplausiblyexogenoustime-seriesvariationinfirms’incentives to innovate that arises due to changes in state R&D tax credits.2 By way of validation, we show that R&D expenses respond strongly to changes in state R&D tax credits, which is in line with the existingevidenceontheeffectivenessoftaxincentivesforR&D(seeHallandVanReenen(2000)for a survey). In our baseline specification, we use a standard difference-in-difference (DID) approach to derive estimates of the effect of changes in state R&D tax credits on cash holdings. To insure that therearenosystematicdifferencesbetweentreatedandcontrolfirmsthatarerelevantfortheircash holding decisions, which is our key identifying assumption, we include an extensive set of controls for unobserved, time-invariant heterogeneity across firms; unobserved, time-varying heterogeneity acrossindustries("industryshocks");andfirmlevelchangesinotherdeterminantsofcashholdings.3 Theinclusionofthesecontrolsinsuresthatourestimatesareidentifiedbycomparingthedifferential (within-firm) response of cash holdings in the same industry-year for similar firms located in stateyears when there is a change in the R&D tax credit (the ’treatment group’) relative to those that are 2Over the last two decades, nearly every U.S. state has enacted some type of tax incentive for R&D and subsequentlyrepealed,reduced,orexpanded. 3To minimize the risk of biases arising from the inclusion of potentially endogenous variables as per the "badcontrols"problemdiscussedinAngristandPischke(2009),weincludethepre-treatment(lagged)values ofthefirm-levelvariables. 2

not(the’controlgroup’). Anyresidualidentificationconcernabouttheinternalvalidityofourbaselineestimatescanonly comefromresidualunobserved,time-varyingheterogeneityacrossstatesthatisnotcapturedbythe industry-year effects, which may bias our estimates if it coincides with changes in state R&D tax creditsandisrelevantforcashholdingdecisions.4 Wetakeseveralstepstobuildconfidencethatour estimates are well-identified. First, we implement validation tests of the main residual threat to our identificationbyexaminingthedeterminantsofstates’likelihoodofchangingtheirR&Dtaxcredits, which we allow to be a function of state-level variables that, on an a-priori basis, we expect may be relevantforcashholdingdecisions,suchasstatebusinesscycleandpoliticalconditions,labormarket forces,andotherstatetaxes. TheresultsofthesevalidationtestsindicatethattheintroductionofR&D taxcreditsistheonlytypeofchangethatissystematicallyrelatedtostateeconomicconditionsand, in general, to state-level variables, while increases or cuts of R&D tax credits subsequent to their introduction appear generally unresponsive to state-level variables. Based on these results, we take a conservative approach in our baseline analysis and only include in the treatment group changes instateR&Dtaxcreditssubsequenttotheirintroduction.5 Throughouttheanalysis,wetakeseveral additionalstepstoaddresspotentialstate-levelconfoundsandbolsteracausalinterpretationofour DIDestimates,whichincludefalsificationandabatteryofsensitivitytests. Our baseline DID estimates indicate that there is an increase (decrease) in cash to book assets of about2(1.7)percentagepointsafteranincrease(decrease)inR&Dtaxcredits,aresultwhichisrobust to using alternative definitions of cash ratios.6 To build confidence that these baseline estimates are well-identified, we implement two batteries of falsification tests.7 First, we repeat our analysis for two different sub-sample splits of the data based on proxies for the institutional design of the R&D tax credits that should affect their effectiveness because of the way they affect the opportunity-cost of R&D. The estimates indicate that there are strong and significant effects only for firms that are 4Thereasonwhythisissueisapotentialchallengeforouridentificationstrategyisthatitpotentiallythreatens our key identifying assumption of "parallel trends" by which we are effectively assuming that absent a changeinthestateR&Dtaxcredit,treatedfirms’cashwouldhaveevolvedinthesamewayasthatofcontrol firms. Thus,weneedtoaddresspotentialstatelevelconfoundsthataffectfirms’demandforcashirrespective oftheR&Dtaxcredits’changes. 5Since we recognize that tax credit introductions, though less suitable for standard DID research design, arepotentiallyinterestingdiscreteevents,weexaminetheirimpactinanextensivesetofdedicatedrobustness tests. 6There is no evidence of either firm anticipation or pre-event trends, while we also present evidence that largertaxcreditchangestriggerlargerresponsesincash. 7The specifications for both tests are all estimated in first differences to remove firm fixed effects in the levelsequations,andlaggedchangesinthefirm-levelcontrolsofthebaselinespecification(1)aswellasyearindustry(48-FF)dummiesareincluded. 3

subjecttoastrongertreatment. Second,weimplementfalsificationteststhatexploitchangesinstate investmenttaxcredits,whicharestate-leveltaxincentivesprogramsotherwiseanalogoustotheR&D tax credits except for the fact that they give tax relief for capex – i.e., for investment in PPE, such as plants,machinery,andbuildings.8 Theresponseofcashtochangesininvestmenttaxcredits,though somewhat weaker in magnitude, is qualitatively the exact opposite of the response to R&D credits, withcashdecreasing(increasing)inresponsetoincreases(cuts)ininvestmenttaxcredits. Weimplementanextensivesetofsensitivityteststofurthercorroborateboththeinternalandexternal validityof our baseline estimates. First, we estimate versions of our baseline DID regressions thatusealternativestandarderrors9 andaddmorefirm-,state-,andindustry-levelcontrolvariables, such as state fixed effects to control for unobserved time-invariant differences among states. Second,inanattempttosharpenouridentification,werefinethecontrolgrouptoincludeonlyfirmsin neighboring counties across border (see Homes (1998) for a similar approach) for which we can exploitthenaturalgeographicdiscontinuityoftaxpolicy(i.e.,thefactthatitstopsatthestate’sborder) todifferenceoutunobservedvariationinlocaleconomicconditions.10 Third,weconsiderrobustness to alternative estimators and asses the external validity of our baseline estimates. One of the alternative estimators we consider is a matched-sample DID estimator (Heckman, Ichimura, and Todd (1997))that replicates the DID tests after matching treated and control firms based on pre-treatment characteristicsthatincludestatebusinesscycleconditions,whichfurtheraddressestheconcernthat differencesbetweentreatedandcontrolfirmsmayinvalidateourparallel-trendassumption.11 Toexamine external validity, we include R&D tax credit introduction events in the treatment group. Our estimates remain stable across different specifications, estimators, and types of treatments. Overall, the results of our falsification, sensitivity, and external validity tests support a causal interpretation oftheobservedresponseofcash. Inthesecondpartofouranalysis,weinvestigatethedriversofthetreatmenteffectbyexamining 8ChirinkoandWilson(2008)hand-collectedstate-leveldataontheseprogramsandshowevidenceoftheir effectiveness at stimulating physical capital expenditures. We use their data to construct indicator variables forchangesinstateinvestmenttaxcredits,whichconsistof18introductionevents,27increases,and8cuts. 9Specifically,weclusterstandarderrorsatthestatelevel,toaddresstheconcernthatclusteringatthefirm levelmayleadtoinflatedstandarderrorssincethekeysourceofvariationintheanalysisisatthestatelevel (Bertrand,Duflo,andMullainathan(2004)) 10We do so by estimating a version of our baseline specification that adds county-pairs fixed effects and is estimated only for firms that have a neighboring county across the border. The resulting estimates are identifiedwithincontiguouscountydyadsacrossstateborders. 11 We also consider the system IV-GMM of Blundell and Bond (1998) to address potential biases from including a lagged dependent variable; and an IV-DID (2-SLS) estimator that uses R&D tax credit changes as instruments for changes in R&D capital, to offer well-identified estimates of the impact of R&D on cash (see Bloom,Schankerman,andVanReenen(2013)forasimilarapproach). 4

itsvariationwithseveralvariablesthat,basedontheory,weexpectshouldaffectthedemandforcash byinnovativefirms. Westartbytestingwhetherthetreatmenteffectdisplayssystematicheterogeneityacrossvariablesthatcapturefinancingfrictions.12 ThefindingsindicatethatR&Dtaxcreditshave astrongerimpactoncashforfirmsthathaverelativelylessaccesstoeitherdebtorequityfinancing.13 To the extent that physical assets are relatively more suitable than intangible assets to support debt financing,theevidencefromourfalsificationtestssupportstheoriesofcashbasedondebtfinancing frictionssuchasRampiniandViswanathan(2010)andFalato,KadyrzhanovaandSim(2013). Tooffer amoredirectassessmentoffinancingfrictions,wealsoexaminetheimpactofchangesinstateR&D tax credits on corporate liquidity structure, debt, and equity financing. Increases (cuts) in R&D tax creditsleadtoincreases(decreases)incashrelativetobanklinesofcredit,andtodecreases(increases, thoughestimatesoftheseincreasesarenoisy)inbankdebt,secureddebt,andoverallnetindebtness. They also lead to increases (decreases) in net equity issuance, though cash increases also relative to bookequity. Inall,thisevidenceindicatesthatexternalfinancingfrictionsareanempiricallyrelevant explanationforwhyinnovativefirmsholdsmorecash,andthatthisisduenotonlytodebtfinancing frictions,whichhavebeenthetraditionalfocusoftheexistingliteraturesurveyedinHallandLerner (2010),butalsotoequityfinancingfrictions,whoseimportancehasbeenrecentlyemphasizedbydynamic corporate finance models such as Bolton, Chen, and Wang (2011), Eisfeldt and Muir (2013), andWarusawitharanandWhited(2013). Wealsotestforvariationinthetreatmenteffectbyproxiesthatcaptureotherimportanttheories, including agency frictions, repatriation, and product market competition. Specifically, we run our baseline DID regressions for splits of the data based on ex-ante proxies for repatriation incentives and product market competition,14 as well as of the severity of agency frictions faced by the firm.15 12Specifically,wesplitthesamplebetweenaboveversusbelowmedian(pre-treatment)valuesofthefollowing variables: firm size, Kaplan and Zingales (1997) KZ-Index, Whited and Wu (2006) WW-Index, firm cash flows, firmage, firmmarket-to-bookratio, firmcashflowvolatility, andbydividendpayerstatus, aswellas severalotherproxiesfordebtandequityfinancingfrictions. 13To assess debt frictions, we run our baseline DID regressions of cash holdings for three different subsamplesplitsofthedatabasedonex-anteproxiesfortheseverityofthedebtfinancingfrictionsfacedbyfirms: Berger et al. (1996) asset liquidation value, Balasubramanian and Sivadasan (2009) index of industry asset redeployability, and based on whether the firm has a long-term debt rating. For equity frictions, the sample issplitbasedonthetotaldollarvolumeofSEOsinafirm’s(48-FF)industryinanygivenyear,theaverage(0, +1)cumulativeabnormalreturn(CAR)ofSEOsannouncementsinafirm’s(48-FF)industryinanygivenyear, which are commonly employed proxies for hot vs.cold equity markets (e.g., Korajczyk and Levy (2003), and thestandarddeviation(dispersion)ofanalysts’EPSforecastsfromIBES. 14Thesampleissplitbetweenaboveversusbelowmedian(pre-treatment)valuesoftheforeigntaxburden measureofFoley,Hartzell,andTitman(2007),basedonwhetherthefirmreportsforeignincomeinagivenyear, and based on above vs. below median (pre-treatment) values of the (48-FF) industry Herfindahl-Hirschman Index(HHI)ofsales. 15We report results for the sub-sample of firms in relatively concentrated industries (those with abovemedian HHI) and further stratify the sample based on above vs. below median (pre-treatment) values of 5

TheresponseofcashtochangesinstateR&Dtaxcreditsisweakenedbyfirms’multinationalstatus, suggesting that the effect of increasing domestic financing frictions is empirically the driving factor oftherepatriationchannel,whiletheresponseismagnifiedbytheintensityofindustrycompetition, whichsupportsrecenttheoriessuchasMorellec,Nikolov,andZucchi(2013)andMa,Mello,andWu (2013)wherecashholdingsstrengthenthecompetitivepositionoffirmsvisavistheirindustryrivals. Theresponseisalsomagnifiedbyproxiesforagencyissuesbetweenmanagersandshareholders,but only in relatively concentrated industries. This evidence is in line with the findings of significant interactioneffectsbetweengovernanceandindustry(GiroudandMueller(2011)),andsuggeststhat inrelativelylesscompetitiveindustriesclassicalagencyconsiderationsofthetypehighlightedinthe empiricalliteraturebyDittmarandMahrt-Smith(2007)andGao,Harford,andLi(2013)arelikelyto bemostimportantforinnovativefirms. Insummary,ourpapermakesthreemaincontributions. First,wedocumentthefirstevidenceof a causal relation between innovation and corporate liquidity. Our evidence shows that innovation is a first-order driver of corporate liquidity decisions. As such, it complements existing findings in the literature on the financial economics of innovations (e.g., Brown et al. (2009, 2013), Chava et al. (2013)), which has so far mostly focused on external finance. It also complements the earlier crosssectionalfindingsofOpleretal(1999)andBates,Kahle,andStulz(2009)forcashandSufi(2009)for linesofcreditbyindicatingthatacausalinterpretationofthesecorrelationsiswarranted. Second,we clarify how innovation impacts cash. By identifying the channels through which innovation affects firms’ corporate liquidity decisions, we are able to make progress on the question of why innovation matters for corporate liquidity, which is challenging in a standard cross-sectional setting since measuresofinnovationtendtobecorrelatedwithproxiesforwhyinnovationmatters. Ourfindingshaveimportantimplicationsfortheongoingpolicydebate,andarehelpfultoidentify when in the cross-section shareholder activists’ pressure on innovative firms to disgorge their cash is more likely to be beneficial. In particular, they indicate that financing frictions and competitivepressureleadtosignificanttrade-offsforsuchshareholders’efforts. Ontheotherhand,strategies targeting the cash holdings of entrenched firms in relatively less competitive sectors are likely to be value enhancing. Finally, at a more methodological level, our paper joins a recent literature that exploitsstateandcountrylevelchangesintaxestoachieveidentification(HeiderandLjungqvist(2012), Panier,Pérez-González,andVillanueva(2012),FaccioandXu(2011)). Ourcontributiontothisliterastandard metrics for the likelihood of managerial entrenchment, which include board size, board independence,andwhetherthefirmhasaclassifiedboardofdirectors. 6

tureistoexamineR&Dtaxcreditsandcorporateliquidity,whichbroadensthefocusofexistingwork on corporate profit taxes, debt tax shields, and capital structure. In doing so, and by documenting a stark contrast in the response of cash to changes in tax credits for R&D versus those for physicalinvestment, wealsocontributetothepolicyevaluationliteraturethatexaminestheeffectiveness of R&D tax credits as incentives for innovation (see Hall and Van Reenen (2000) for a survey, and Chirinko and Wilson (2008) for evidence on the effectiveness of investment tax credits), which had notexaminedcorporatefinanceissues. 2 Theoretical Framework and Descriptive Evidence Inthissection,wefirstprovideanoverviewofthemaintheoriesthathelptopindownthechannels throughwhichinnovationshouldmatterforfirms’cashholdingdecisions. Wethenofferdescriptive evidence that in our sample replicates the well-known finding that high R&D firms hold more cash – i.e., that there is a positive correlation between R&D and corporate cash holdings. This evidence suggeststhatinnovationisapotentiallyimportanteconomicdeterminantofcorporatecashmanagement policies, but its interpretation remains yet unclear. As such, it lays out the foundation for our efforttotakeafirststeptowardidentifyingthecausalinterpretationofthecorrelationbetweenR&D andcash,andassessingtheempiricalrelevanceofthealternativetheoriesthatcanexplaintheimpact ofinnovationoncash. 2.1 WhyShouldInnovativeFirmsHoldMoreCash? AnOverviewoftheTheory There are several reasons to expect that, based on first principles in corporate finance theory, innovationshouldleadtohighercashholdings. First,acentralinsightoffinancialcontractingand,more broadly, capital structure theories in corporate finance is that external finance frictions give rise to a precautionary or hedging motive to hold cash. This insight, which dates back to Keynes (1936), wasdevelopedintheinfluentialstudyofFroot,Scharfstein,andStein(1993)16 andhasbeenrecently explored quantitatively within dynamic corporate finance models (Riddick and Whited (2009) and Bolton,Chen,andWang(2011)). Ademandforinternalfundsarisesbecausefinancialfrictionslimit the firm’s ability to raise external finance. Due to such frictions, cash flow shortfalls might prevent 16See Almeida, Campello, and Weisbach (2004) for supporting evidence based on the cash-flow sensitivity ofcash;andBonaime,Hankins,andHarford(2013)formorerecentevidenceofsubstitutionbetweenhedging andpayoutdecisions. 7

firms from investing in profitable projects if they do not have liquid assets. Thus, firms may find it profitable to hold internal finance (cash) in order to mitigate costs of financial distress and preserve investmentopportunities. TherearetwodistinctfinancingchannelsthroughwhichR&Dmayincreasecashbyraisingexternalfinancecosts. Thereisadebtfinancingchannel,throughwhichR&Dlimitsfirms’abilitytoraise debtsinceknowledgeassetshavelimitedcollateralvalue(e.g.,HartandMoore(1994),Rampiniand Viswanathan(2010),andFalato,Kadyrzhanova,andSim(2013)). Basedonthisliterature,R&Dmay limit external debt financing because it complicates contractibility problems by lowering the value that can be captured by creditors in default states. In addition, debt finance can lead to problems of financialdistressthatmaybeparticularlysevereforfirmsthatrelyheavilyonR&D,sincethedesign of standard debt contracts does not work well for investments characterized by a high probability of failure and some chance of extremely large upside returns (Opler and Titman (1994)). The debt financingchannelalsohasdirectimplicationsforliquiditystructure,anditimpliesthatR&Dshould reducefirmrelianceonexternalsourcesofliquidity,suchasbanklinesofcredit.17 There is also an equity financing channel. Basic considerations from information theory suggest thatR&Dinvestmentsarepronetopotentiallysevereinformationproblemsthatarelikelytoincrease the cost of raising external equity. Bolton, Chen, and Wang (2011) show that costly equity financing cangiverisetoaprecautionarymotivetoholdcash. Whileexternalequitymayhaveadvantagesover debtforfinancingR&D,18 internalandexternalequityfinancearenotperfectsubstitutessincepublic stockissuesincursizeableflotationcostsandrequirea“lemonspremium”duetoasymmetricinformation (e.g., Myers and Majluf (1984)). Information asymmetries are likely to make outside equity financing more expensive when firms rely more on R&D, because due to the inherent uncertainty associatedwithR&Dinvestmentoutsideinvestorshavemoredifficultydistinguishinggoodprojects from bad compared to investments in more low-risk projects such as those in capital expenditures (LelandandPyle(1977)).19 17See Mann (2013) and Chava, Oettl, Subramanian and Subramanian (2013) for recent evidence on debt financingandinnovation.HallandLerner(2010)isasurvey. 18For example, see Brown, Martinsson, and Petersen (2013), who document in a cross-section of countries thatshareholderprotectionsandbetteraccesstostockmarketfinancingleadtohigherR&Dinvestment,particularlyinsmallfirms,butareunimportantforfixedcapitalinvestment. 19A final financing aspect of R&D that favors cash is its inflexibility, which is due to the fact that R&D investmententailslargeadjustmentcosts. Forexample,R&DinvestmentsincludesettingupandrunningR&D labs with highly skilled workers, whose firing can result in large hiring and training costs as well as the unwanteddisseminationofproprietaryinformationoninnovationefforts,makingitveryexpensiveforfirmsto cutdownR&Dinresponsetotemporarychangesintheavailabilityofexternalfinance. Consistentwiththis reasoning, Brown and Petersen (2011) document that financially constrained firms use cash to smooth their R&Dexpenditures. Relatedly,MacKay(2003)documentsthatthereisapositiverelationbetweenleverageand 8

Second, the higher uncertainty and lack of verifiability associated with R&D may lead to an agencychannelthroughwhichinnovationimpactscash. ThisisthecaseifR&Dmakesinsiders’decisionshardertomonitorbyoutsideshareholders,thuseffectivelyexacerbatingagencycosts. Thelink between agency costs and cash is well understood at least since Jensen (1986), who emphasized the conflictbetweenmanagersandshareholdersoverinternalfunds.20 Relatedly,thereisalsoarepatriationchannel. Thefactthatmultinationalshaveamotiveforholdingcashbecausetheirunrepatriated foreignearningsaretaxadvantagediswellestablished(seeFoley,Hartzell,Titman,andTwite(2007) and Falkeunder and Petersen (2012)). However, innovation may either lower multinational firms’ incentives to repatriate foreign income, by increasing financing costs at home, or it could heighten the repatriation motive by increasing the firm’s flexibility to shift profits abroad. Thus, the overall effectofinnovationoncashisambiguousbasedonrepatriation. Thirdandfinal,thereisaproductmarketcompetitionchannel. Agrowingtheoryliteratureconsiders optimal cash holding decisions in a setting where firms compete dynamically over time (see, for example, Morellec, Nikolov, and Zucchi (2013), Ma, Mello, and Wu (2013); and Hoberg, Phillips andPrabhala(2014)forsupportingevidence). LyandresandPalazzo(2014)considerasettingwhere in addition to competing in the product market, firms also make innovation decisions. A common prediction across these theories is that the demand for cash should be increasing in the intensity of product market competition, and especially so for innovative firms. The basic intuition is that productmarketinteractionsgiverisetoastrategicmotiveforholdingcash, bywhichcashholdings effectively improve the competitive prospects of a firm vis-a-vis its industry rivals.21 Thus, based on the product market channel, we expect that industry competition should increase the impact of innovationoncashholdings. Insummary,thetheoryliteratureleadstothefollowingmaintestablehypotheses: investmentflexibility. 20Inacross-countrystudy,Dittmar,Mahrt-Smith,andServaes(2003)documentthatinlowinvestorprotectioncountriesfirmsholdmorecash. Harford(1999)documentsevidencethatacquisitionsbycash-richfirms arevalue-decreasing,whichisconsistentwithprivatemotivestoholdcash. DittmarandMahrt-Smith(2007) find that the value of cash is significantly lower at poorly governed firms. More recently, Gao, Harford, and Li(2013)comparethecashpoliciesofpublicly-tradedvs. privately-heldfirms,andarguethatthelargedifferencesincashholdingsbetweenthesetypesoffirmscanbeattributedtothemuchhigheragencycostsinpublic firms. 21He(2014)documentsevidenceinsupportoftheproductmarketchannel,withR&Dfirmsincreasingtheir cashholdingsinresponsetotariffreductionsthatleadtointensifiedforeigncompetition,butnon-R&Dfirms not doing so. While informative about the importance of product market competition, this evidence still reliesonacross-sectionalcomparisionbetweenR&Dandnon-R&Dfirms,whichmaydifferalongoservableas well as unobservable characteristics that are hard to control for. As such, it is less informative about the impactofinnovationsinceinterpretationissubjecttothesametypeofendogeneitylimitationsasstandardOLS estimates. 9

Innovationshouldhaveapositiveimpactoncashholdings,butnotonexternal(bank)liquidity. • The positive impact of innovation on cash holdings should be stronger for financially con- • strainedfirms(debtandequityfinancingchannels),forfirmswithmoreentrenchedmanagers (agencychannel),andthoseinmorecompetitiveindustries(productmarketchannel). 2.2 DataandDescriptiveEvidence Our primary data is standard accounting information from Compustat for all nonfinancial firms22 incorporated in the U.S. between 1986 and 2011, which yields a starting panel of 124,504 firm-year observations for 11,091 unique firms. 1986 is the earliest year for which we can retrieve historical informationonfirmheadquarterlocation,whichisnecessarytoimplementouridentificationstrategy as described in detail in the next section. We complement this sample with detailed information on firmliquidityanddebtstructurefromCapitalIQ,whichisavailableforasub-setof23,086firm-year observations for 2,866 unique firms in the 2002 to 2011 period (see Sufi (2009) for an early study of liquidity structure that uses hand-collected information from SEC filings, and Colla, Ippolito and Li, (2013) for a recent study of debt structure that uses Capital IQ). In order to better isolate the causal impact of innovation, it is important to minimize the risk that our estimates are driven by spurious correlation from comparing innovative to non-innovative firms, which differ along many dimensions that are hard to control for. In addition, tax incentives for R&D are clearly not relevant forfirmsthatarenotactiveinR&D,whoseinclusionwouldreducethepowerofourtests. Thus,we take a conservative approach and exclude firms that are never active in innovation throughout our sampleperiod.23 Table1providessummarystatisticsforourresultingfinalsampleofinnovativefirms,whichconsistsof6,058(1,798)uniqueCompustatfirms(withCapitalIQinformation)thatreportnon-zeroR&D expendituresinatleastoneyearoverthe1986-2011periodforatotalof72,587(14,504)firm-yearobservations. Wereportmeans,medians,andstandarddeviationsofourmaindependentvariable,the ratio of cash holdings to book assets, as well as of liquidity structure variables, such as the ratio of bankliquiditytototalliquidity. Wealsotabulatecommonlyemployedmeasuresoffirminnovation, R&D expenditures and the stock of R&D assets,24 and our firm- and industry-level controls, which 22Asitisstandardintheliterature,weexcluderegulatedUtilities(SIC4900-4999)andfirmswithmissingornon-positive bookvalueofassetsinagivenyear. 23Theresultsarerobusttoaddingthemoreaggressivefilterofexcludingallfirm-yearswithzeroR&D. 24Since under current accounting principles R&D assets do not appear in firm balance sheet, we follow an approach whichisstandardintheliteratureontheeconomicsofinnovation(Corrado,Hulten,andSichel(2009)andCorradoand 10

includethosegenerallyusedintheliterature. DetailedvariabledefinitionsareinAppendixA.Overall, our starting Compustat sample is comparable to those used in previous studies, such as Opler, Titman, and Stulz (1999), with firms holding on average 18% of their total (balance sheet) assets in cash. In our final sample, innovative firms hold on average 24% of their total assets in cash, which isconsistentwiththewell-replicatedfindingofapositivecorrelationbetweencashandR&D(Bates, Kahle,andStulz(2009)andHallandLerner(2010)). However,innovativefirmsclearlydifferfromthe average U.S. firm along most characteristics, including that they are generally smaller and younger, havehighergrowthopportunities,andarelessreliantondebtandlesslikelytopaydividends,which illustratesoneaspectoftheidentificationchallengeinvolvedinderivingwell-identifiedestimatesof theimpactofR&Doncash. Before proceeding to our main analysis, Table 2 summarizes descriptive evidence on the crosssectionalrelationbetweenstandardmeasuresoffirminnovationandcashholdings,aswellasliquiditystructure. Specifically,wereportresultsofOLSregressionsofcash-to-bookassetsratios(Columns 1 to 4) and ratios of external liquidity-to-total liquidity (the sum of cash and external liquidity) (Columns5to10)on(year-prior)R&D(expendituresorcapital),whilecontrollingforstandardcrosssectionalcovariatesofcashholdings(year-priorcashflowvolatility, market-to-bookratio, firmsize, cashflow,capex,acquisitionexpenditures,andadummyforwhetherthefirmpaysdividendinany given year) in addition to year and 48-FF industry effects. In line with previous findings (Bates, Kahle, and Stulz (2009), Lyandres and Palazzo (2014), He (2014) for cash, and Sufi (2009) for liquidity) and with the evidence surveyed in Hall and Lerner (2010), the coefficient on either measure of R&Disstatisticallyandeconomicallysignificantlypositive(negative)inthecash(externalliquidity) regressions,25 aresultthatholdsinboththeoverallCompustatsample(PanelA)andourinnovative firmssample(PanelB).Thus,consistentwiththeory,innovativefirmstendtoholdarelativelylarger fractionofboththeirassetsandtheirtotalliquidityinternallyintheformofcash. Hulten (2010); see Eisfeldt and Papanikolau (2011;2014) and Falato, Kadyrzhanova, and Sim (2013) for recent papers in financethathaveusedasimilarapproachtoconstructmeasuresofintangibleassets)andconstructthestockofR&Dassets by capitalizing R&D expenditures using the perpetual inventory method as follows: G = (1 δ )G +R&D whereGt istheend-of-periodstockofR&Dcapital, R&D it isthe($1990real)expenditure it sonR& − Dd R u & r D ingt it h−e 1 year,and it δ = 15%followingHall,Jaffe,andTrachtenberg(2001). IfR&Dexpendituresareconstant(inrealterms),thestockof R&D y R e & a D rd c i a v p id it e a d li b s y G t t he = d ∑ ep ∞ s= re 0 c ( ia 1 t − ion δ) r s a R te & δ Dt − s . = Ina R δ d . d W it e io s n e , t w th e e in in t i e t r i p al o s la to te ck m t i o ss b in e g eq v u al a u l e t s o o t f h R e & R D &D fol e l x o p w e i n n d g it H u a re ll s ( i 1 n 99 th 3) e w fir h s o t R&D showsthatthisresultsinanunbiasedmeasureofR&Dcapital.TheR&Dstockisscaledby($1990real)bookassets. 25Coefficientsoncontrolvariablesarealsoinlinewiththepreviousliterature,withlargefirmsandfirmsthatpaydividendsholdinglesscash, andfirmswithhighercashflowvolatilityandmarket-to-bookholdingmore. Thecoefficients oncapitalexpendituresandacquisitionsarenegativeandsignificant, consistentwithfirmsusingtheircashholdingsto pursueinvestmentopportunities. 11

3 Exploiting Changes in State R&D Tax Credits: Baseline Estimates of the Impact of Innovation on Cash Thecross-sectionalevidenceontherelationbetweenR&DandcashreplicatedinTable2isimportant, butitsinterpretationislimitedbyavarietyofwell-knownendogeneityconcerns. Inparticular,when measuring the effect of innovation using comparisons across firms, there will always be a suspicion that the controls included in the analysis are not exhaustive. Such concerns are mitigated when the specificationincludesfirmfixedeffects. Nevertheless,iftherearetime-varyingomittedfirmcharacteristics,suchasfutureprofitgrowthopportunities,thataffectboththefirm’sincentivestoinvestin R&Danditscashholdingsdecisions,thentheestimateswouldnothaveacausalinterpretation. Our main contribution is to build on the existing evidence and identify the causal impact of innovation oncashbyexploitingplausiblyexogenoustime-seriesvariationinfirms’incentivestoinnovatethat arises due to changes in state R&D tax credits. Over time, state policy makers have used legislation and public subsidies to influence the costs of R&D, altering firms’ incentives to make these investments. Inthissection,weuseastandarddifference-indifferencesapproachandexaminetheimpact ofchangesinstateR&Dtaxcreditsoncorporatecashpolicies. 3.1 InstitutionalBackground Most U.S. states offered an R&D tax credit as of December 2011. These state-level tax credits apply to R&D activities incurred within the state borders and are typically credits for the state corporate incometax.26 ThebasicstructureofthestatecreditsisgenerallydesignedtomimicthefederalR&D taxcredit27 andstipulatesastatutory(percentage)taxcreditrateforqualifiedresearchexpenditures (QRE)incurredinthecurrentyearinexcessofabaseamount,whichistypicallyafunctionofaverage QREs incurred in up to the prior three tax years. Most states use the federal definition of QRE from theInternalRevenueCode,Section41,28withamodificationtoincludeonlyexpensesincurredwithin 26Only3statesofferasalestaxcredit. Firmsaresubjecttostateincometaxesiftheyhaveaneconomic"nexus"withthe state,whichisgenerallydeterminedbasedonwhethertheyderiveincomefromsalesinthestate,haveemployeesorown leasepropertyinthestate(seeHeiderandLjungqvist(2012)foradditionaldetails). 27ThefederalR&Dtaxcreditwasintroducedin1981andappliesinadditiontothestateones. 28“Qualifiedresearch”isidentifiedasresearchundertakenforthepurposeofdiscoveringinformationthatistechnological in nature and the application of which is intended to be useful in the development of a new or improved business component,aswellasalloftheactivitiesofwhichconstituteelementsofaprocessofexperimentationforaneworimproved function, performance, reliability, or quality. There is also a list of research activities that do not qualify for the credit,suchascomputersoftwareorsocialsciences. Finally,QREsaretheamountpaidforwagesofemployeesengaging inqualifiedresearch,suppliesusedforqualifiedresearch,and65percent(higherforcertaintypeofentities)ofcontract researchexpensespaidtooutsideentitiestoperformqualifiedresearch. 12

the state borders.29 There is a lot of variation across states in the details of the definition of the base amount,butafeaturethatholdsrobustlyacrossstatesisthatthebasedefinitionisintendedtocapture R&Dintensityofthefirmbasedonprior-years’QREs.30 Appendix B contains the full detailed list of state R&D tax credit changes, with their respective effective dates – i.e., the first year when the firm can claim the credit. For example, California increaseditsR&Dtaxcreditfrom12percentto15percent(topstatutoryrate)effectivefromfiscalyear 2000. In2011,topstatutorycreditpercentageratesrangedfrom2percentinMichiganto100percent for the "super" credit in Wisconsin, with 2 to 5 percent being relatively low credits rates, 15 to 20 percentbeingrelativelyhighrates, and10percentbeingthemostcommonrate.31 Overthelasttwo decades,nearlyeveryU.S.statehasenactedsometypeoftaxincentiveforR&Dandsubsequentlyrepealed,reduced,orexpandedit.32 Itisthisvariationovertimethatweexploittogenerateexogenous variationinfirms’incentivestoinnovate. 3.1.1 Intuition Inordertobuildintuitiononhowouridentificationstrategyworks,itisusefultoconsiderhowstate R&Dtaxcreditsaffectfirms’decisionstoinvestinR&D.Followingthestandardapproachintheeconomicsofinnovationliterature(see,forexample,HallandLerner(2010)),thestandardbenchmarkto evaluatefirms’R&Dinvestmentdecisionsisthe’neo-classical’marginalprofitconditionforoptimal investmentwhichisbasedontheoriginalexpressionforthemarginalcostortheso-called’usercost’ ofphysicalcapitalofHallandJorgenson(1967)). Generalizingthisexpression, wecanderiveanexpressionforthefirm(after-tax)marginalcostofR&Dcapital(perdollarofinvestment)asafunction ofthestateR&Dtaxcreditbutalsoofalltheotherrelevanteconomicandinstitutionaldeterminants ofR&Dinvestmentdecisions.33 29Onlyfivestates(Colorado,Kentucky,NewHampshire,Washington,andWestVirginia)departfromthefederalQRE definition. Somearenarrower;NewHampshireandWashingtononlyincludecertainindustries,whilesomearebroader; WestVirginia’sstatutorydefinitionincludesmanyexpensesnoteligibleforthefederalcreditandKentucky’screditapplies tothecostsofconstructing,equipping,orexpandingfacilitiesusedforR&D. 30EighteenstatesusetheFederalSection41definitionwithanadjustmenttoapplytoin-stateexpenses.Sevenstatesuse someformofaprioryear(s)movingaveragebase. Anotheralternative,usedbyfourstates(Alaska,Delaware,Nebraska, andNewYork),istoallowtaxpayerstoclaimsomepercentageoftheirfederalcredit. Fourstates(Connecticut,Delaware, Oregon,andWestVirginia)employadualbase,whereeitherdifferentratesapplytotwodifferentbaseamountsortaxpayerscanelecttoclaimthegreateroftwomethodsforcomputingthevalueoftheircredit.Finally,onlytwostates(Kentucky andNorthCarolina)havenon-incrementalcreditsandconsequentiallydonotdefineabaseperiod. 31Somestatesoffertieredrates,withthepercentagedecreasingatsomedollaramountofQRE. 32MinnesotawasthefirststatetoenactaR&Dtaxincentivein1981,followedbyIndianain1984andIowain1985. All otherstates’creditenactmentsandchangesarewithinoursampleperiod. 33Forsimplicity,hereweabstractfromthefederalR&Dtaxcredit,whichwouldenterasanadditionallineartermatthe numeratorofourexpression,thusleavingourmainconclusionsunchanged. 13

This derivation (see Appendix A in Hall and Van Reenen (2000) for full details) assumes that at time t = 0 the firm maximizes its market value, which is defined as the discounted present value of future dividends, with the discount factor implied by the real interest rate, r , and with the R&D t capital stock defined using the same perpetual inventory method we used to define our empirical δ measure with depreciation rate . The resulting expression for firms’ marginal cost of R&D investmentisgivenby: 1 state statutory creditrate f(R&D base(R&D )) DA MP i R t &D = − it × 1 τ it − t − n − it [r t +δ] it − τ where denotes the (effective) corporate income tax rate, DA is the NPV of depreciation alit it lowances,andtheeffectofthestateR&Dtaxcreditisgivenbytheproductofthestatutorycreditrate andthepartofR&Dthatqualifiesforpreferentialtaxtreatment,whichisgivenbythequalifiedR&D expendituresinexcessofthebaseamount. Itisimmediatefromthisexpressionthat ∂MP i R t &D < 0- ∂CreditRate i.e.,anincrease(decrease)inthestatutoryR&Dtaxcreditratereduces(increases)the(after-tax)cost ofanextradollarinvestedinR&D.Thus,changesinstateR&Dtaxcreditratesareaplausibleshock tofirms’incentivestoinvestinR&D. While parsimonious enough to derive transparent intuition for our identification strategy, this expression offers a relatively rich characterization of the determinants of the decision to invest in R&D.Inparticular,thefollowingimplicationscanbeusedtoderivefalsificationtestsforasub-setof firmsthatareunlikelytobedirectlyaffectedbythechangesintaxcredits: Comparativestatic1, variation with τ : ∂2MP i R t &D < 0 - i.e., any given increase (decrease) in it ∂CreditRate∂τ thestatutoryR&Dtaxcreditratereduces(increases)the(after-tax)costofanextradollarinvestedin R&D by proportionally more in states with higher corporate income tax rates. Intuitively, since tax credits are generally not refundable, in any given year when the firm has no tax liabilities it cannot claimacredit. Whilethecreditcanbecarriedforwardoverthenextyears(generallyupto10),clearly creditshavemorebiteinstateswithhighercorporateprofittaxes. Comparativestatic2, variation with f(R&D base(R&D )) : ∂2MP i R t &D < 0 - i.e., any it − t − n ∂CreditRate∂f( · ) given increase (decrease) in the statutory R&D tax credit rate reduces (increases) the (after-tax) cost of an extra dollar invested in R&D by proportionally more for firms that experience higher growth in R&D expenditures relative to their pre-treatment levels. Intuitively, this is the case since the base amount is an increasing function of past R&D intensity, which effectively operates as a ’claw-back’ provision in terms of current tax credit. Thus, for any given amount of current R&D investment, 14

higherpre-treatmentR&Dexpendituresreducetheeffectivenessoftaxcredits. 3.2 EmpiricalFramework To examine the effect of changes in state R&D tax credits on cash holdings, we use the following baselinedifference-in-difference(DID)regressionspecification: Δ Cash ijst = β 1× Δ RDTC s + t +β 2× Δ RDTC s−t +δ Δ X +δ Cash +α +ε (1) 1 it 1 2 ijst 1 jt ijst × − × − Δ where i, j, s, and t index firms, industries, states, and years. denotes the first-difference operator, Cash is the ratio of cash holdings to book assets, Δ RDTC+ and Δ RDTC are indicators that equal − one if any given state increases or cuts its R&D tax credit in any given year, respectively. The latter are contemporaneous changes, because our timing convention is to assign to each tax credit change itsyeareffective–i.e.,thefirstyearwhenthefirmcanclaimthecredit,whichisgenerallyeitherthe sameyearortheyearsubsequentthepassageofthechange. Inadditiontothisbaselinespecification, we also estimate a more inclusive specification that adds leads and lags of the tax credit changes to examine the timing of the response of cash holdings.34 X are firm-level controls for standard covariatesofcash,35 and α areindustry-yearfixedeffects. Weevaluatestatisticalsignificanceusing jt robustclusteredstandarderrorsadjustedfornon-independenceofobservationswithinfirms.36 The β β null hypothesis is that the coefficients of interest, and , which capture the effect of tax credit 1 2 changesoncashholdings,areequaltozero. β β ThekeyidentifyingassumptionbehindourDIDestimates, and ,isthattherearenosystem- 1 2 atic differences between treated and control firms that are relevant for their cash holding decisions. Ourbaselinespecificationrulesoutthreemainpotentialsourcesofsuchdifferencesbetweentreated and control firms: first, we control for unobserved, time-invariant heterogeneity across firms by estimating our DID regression in first differences; second, we control for unobserved, time-varying heterogeneity across industries ("industry shocks") by including industry-year fixed effects; and fi- 34Inadditionalrobustnesschecks,weverifiedthatourbaselineestimatesarerobusttousingaspecificationwithonly laggedchangesinR&Dtaxcredits(resultsavailableuponrequest). 35Specifically, our baseline specification controls for cash flow volatility, market-to-book, firm size, cash flow, capital expenditures,(cash)acquisitionsexpenditures,andadummyforwhetherthefirmpaysdividend. 36In robustness analysis (Table 6, Panel A), we verify that the results are robust to adjusting the standard errors for clustering at the state level, to address the concern that the key source of variation in the analysis is at the state level (Bertrand,Duflo,andMullainathan(2004)).Thiscorrectionrelaxestheassumptionthatfirmobservationsareindependent withineachstate. 15

nally, we control for firm level changes in performance or other determinants of cash holdings by including standard firm level determinants of cash. To minimize the risk of biases arising from the inclusionofpotentiallyendogenousvariablesasperthe"badcontrols"problemdiscussedinAngrist andPischke(2009),weincludethepre-treatment(lagged)valuesofthefirm-levelvariables. The inclusion of an extensive set of controls for (observable and unobservable) potential confounds insures that our DID estimates are identified by comparing the differential (within-firm) responseofcashholdingsinthesameindustry-yearforsimilarfirmslocatedinstate-yearswhenthere is a change in the R&D tax credit (the ’treatment group’) relative to those that are not (the ’control group’). Anyresidualthreattoidentificationandtotheinternalvalidityofourbaselineestimatescan onlycomefromresidualunobserved,time-varyingheterogeneityacrossstates,suchas,forexample, state business cycle or political conditions or changes in other state taxes, that is not captured by theindustry-yeareffects,whichmaybiasourestimatesifitcoincideswithchangesinstateR&Dtax creditsandisrelevantforcashholdingdecisions.37 Given the importance of these potential state-level confounds, we take several steps to address the issue: first, for our baseline estimates we take a conservative approach and only include in the treatment group changes in state R&D tax credits subsequent to their introduction, which, based on a probit analysis, appear not to be systematically related to (observable) state-wide confounds;38 second, we implement falsification tests that exploit the institutional features of the R&D tax credit programs; third, we implement an extensive set of sensitivity tests, which include adding controls forpotentialstate-levelconfounds;fourth,werefinethecontrolgrouptoincludeonlyfirmsinneighboring counties across border (see Homes (1998) for a similar approach) for which we can exploit the natural geographic discontinuity of tax policy (i.e., the fact that it stops at the state’s border) togetherwiththefactthatthesefirmsarelikelytosharecommoneconomicfundamentalstodifference out unobserved variation in local economic conditions; and fifth, we replicate the DID tests after matchingtreatedandcontrolfirmsbasedonpre-treatmentcharacteristicsthatincludestatebusiness cycle conditions. This matched-sample DID estimator (Heckman, Ichimura, and Todd (1997)) further addresses the concern that differences between treated and control firms may invalidate our 37Thereasonwhythisissueisapotentialchallengeforouridentificationstrategyisthatitpotentiallythreatensourkey identifyingassumptionof"paralleltrends"bywhichweareeffectivelyassumingthatabsentachangeinthestateR&Dtax credit,treatedfirms’cashwouldhaveevolvedinthesamewayasthatofcontrolfirms.Thus,weneedtoaddresspotential statelevelconfoundsthataffectfirms’demandforcashirrespectiveoftheR&Dtaxcredits’changes. 38Inrobustnessanalisys(Table5,PanelD),weassesstheexternalvalidityofourbaselineestimatesbyincludingalsothe introductionofstateR&Dtaxcreditsinthetreatmentgroup. 16

parallel-trendassumption.39 Inrobustnessanalysis, wealsoaddresspotentialconcernswiththefactthatourspecificationincludes a lagged dependent variable, which controls for imperfections in cash rebalancing or partial adjustment in cash ratios (there is a vast literature on partial adjustment in leverage ratios – e.g., Lemmon, Roberts, and Zender (2008); see Dittmar and Duchin (2010) for recent evidence of partial adjustment for cash).40 Allowing for partial adjustmentof cash is important in light of the evidence onthepersistentresponseofR&Dtotaxcreditsinthepreviousliterature. Thelaggeddependentalso controlsfor"Ashenfelterdip"typefactors,thatmaybeaconcernincasethestatesarechangingR&D tax credits in response to potentially unobserved factors related to firm cash holdings (Ashenfelter δ (1978)). Since OLS estimates of may be biased in small-T unbalanced dynamic panels (Nickell 2 (1981), Arellano and Bond (1991)), we do not emphasize this coefficient estimate and focus our discussiononthebaselineestimatesoftheimmediateimpactofR&Dtaxcreditchangesor"short-term" elasticities, β and β .Asarobustnesstest,however,were-estimatespecification (1) usingasystem 1 2 IV-GMMestimator(BlundellandBond(1998)),41 whichisdesignedtoaddresstheeconometricconcernsassociatedwithestimatingdynamicpaneldatamodelsinthepresenceoffirmfixedeffects.42 Finally, in order to estimate (1), we remedy the measurement issue that Compustat’s location informationisoftenincorrectbyhand-collectinghistoricalheadquarterstatesandzipcodesforeach firm-yearinoursamplefromSECfilings. Compustatreportsthecurrentaddressofafirm’sprincipal executive office, not its historic headquarter location, which is an issue since firms relocate their headquarters. Dealingwiththisissueisnecessarysincefirmslocationinformationinourexperiment is crucial to sort firms into the treatment vs. control groups, and incorrect location information is a source of measurement error that would likely bias our estimates in favor of a false null. For each fiscalyear,weuseaPerlprogramand,foreachfirminoursample,searchforlocationinformationin thatyearwithintheuniverseofitsSECfilings(Def14A,10-Q,10K).43 Asanadditionaldatacheckonthisprocedure,wecross-checkourhand-collectedlocationinfor- 39Wedonotexplorerobustnesstoanalternativespecificationthatwouldaddresspotentialconfoundsbyincludingthe meanofthegroup’sdependentvariable–i.e.,meancashholdingsacrossfirmsinagivenstate-year–asacontrol,sinceit hasbeenshowntoleadtoinconsistentestimates(GormleyandMatsa(2014)). 40Allowing for partial adjustment is important also in light of the evidence that financial frictions lower the speed of adjustmentofleverageand,thus,maybeexpectedtoalsoincreaseadjustmentcostsofcash(seeFaulkenderetal. (2012) forevidenceonleverage). 41Specifically,thesystemIV-GMMapproachincludeslaggedvariablelevelsanddifferencesintheinstrumentsettoaddresstheproblemofpersistentregressors,which,whendifferenced,containlittleinformationforparameteridentification. 42Inadditionalrobustnesschecks,weverifiedthatourbaselineestimatesarerobusttousingaspecificationthatexcludes thelaggeddependentvariable(resultsavailableuponrequest). 43WeretrievethebulkoftheproxyfilingsfromtheCompactDisclosuredisksuntiltheyareavailable(2005)andsupplementthemwithfilingsfromEdgarOnlinewhenevernotavailable. 17

mationwithhistoricaleventinformationfromCapitalIQKeyDevelopmentsdatabase,whichtracks headquarter relocation announcements starting from 2001.44 We find a high degree of overlap betweenlocationinformationfromCapitalIQandtheoneweretrievedmanually, withourprocedure "missing" location changes in Capital IQ for less than 1/2% of firm-years observations in the sample.45 By contrast, Compustat state information is incorrect in 11.5% of the firm-years observations, affecting17.1%ofinnovativefirmsinthesample. 3.2.1 ValidationofR&DTaxCreditChanges: DoTheyCoincidewithOtherState-LevelChanges thatarePotentiallyRelevantforCash? And,DoTheyMatterforInnovation? For our baseline DID estimates to be well-identified, the key assumption of parallel-trends between treatedandcontrolfirmshastobevalid,becauseitisunderthisassumptionthatwecanbeassured that absent any R&D tax credit changes, treated firms’ cash holdings would have evolved in the same way as that of control firms. In addition, changes in state R&D tax credits must be effectively "innovation shocks," because if they did not matter for innovation, that would cast doubt on the interpretationofourcoefficientestimatesonchangesinR&Dtaxcreditsasreflectiveoftheimpactof innovationoncash. Inordertoofferascreeningoftheinternalvalidityofourestimates,weexamine thesetwoissuesinturn. First,weimplementvalidationtestsofthemainresidualidentificationconcernaboutourbaseline estimates. Aswehaveemphasizedinthedescriptionofourbaselinespecification,potentialidentificationconcernscanonlyarisebecauseofresidual,time-varyingheterogeneityacrossstatesthatisnot captured by the industry-year effects, which may bias our estimates if it coincides with changes in stateR&Dtaxcreditsandisrelevantforcashholdingdecisions. Thus,weexaminethedeterminants ofstates’likelihoodofchangingtheirR&Dtaxcredits,whichweallowtobeafunctionofstate-level variablesthat,onana-prioribasis,weexpectmayberelevantforcashholdingdecisions. Thesevariables include proxies for state business cycle conditions, since there is evidence that firm financial policies (Korajczyk and Levy (2003)) and cash holdings (Eisfeldt and Muir (2013), Warusawitharan 44Capital IQ Key Developments database consists of information gathered from a wide variety of sources, including publicnews,companypressreleases,regulatorylings,calltranscripts,investorpresentations,stockexchanges,regulatory andcompanywebsites.CapitalIQanalystsfilterthedata,linkittostandardcompanyidentifiers(gvkey),andthencategorizeitbasedonthetypeofeventinvolved. Weretainonlythe"AddressChange"announcementcategory,whichcontains about1,500announcementsinvolvingCompustatfirmsinthe2001-2011period. 45Welookedupahandfulofsuchcases, andtheyallcorrespondedtoaddresschangesthatwereannounced(and, as such, recordedintheCapitalIQdatabase), butnotcompleted. Thus, weoptedforkeepingourlocationinformationin the few instances when Capital IQ records a relocation announcement that is not recorded in the regulatory filings. In additionalrobustnesstests(availableuponrequest),weverifiedthatourresultsarelittlechangedifwealternativelyuse theCapitalIQlocationinformationinthesecases. 18

and Whited (2013)) are responsive to the business cycle; for political and labor market forces, since thereisevidencethatpoliticaluncertaintyheightensfirms’precautionarymotive(Baker,Bloom,and Davis(2013))andthatbargainingwithunionsmattersforleverage(Matsa(2010))andcash(Schmalz (2013)); andforothertaxes, since, forexample, corporatetaxescouldhaveanindependenteffecton cashforreasonsrelatedtothetaxbenefitsofdebt(seeHeiderandLjungqvist(2012)forevidenceon corporatetaxesandleverage). The results of this first set of validation tests are reported in Panel A of Table 3, which tabulates estimates from linear-probability regression analysis of the likelihood that a state changes its R&D taxcreditinagivenyearfordifferenttypesofchangesinturn: taxcreditintroductionorsubsequent increases (Column 1), introduction only (Column 2), increases subsequent to introduction (Column 3), and cuts subsequentto introduction (Column 4). The introduction of R&D tax credits is the only typeofchangethatissystematicallyrelatedtostateeconomicconditionsand,ingeneral,tostate-level variables,withtheWaldtestunabletorejectthenullthatthesevariablesarejointlyinsignificantatthe 5%confidencelevel. Bycontrast,increasesorcutsofR&Dtaxcreditssubsequenttotheirintroduction appeargenerallyunresponsivetoeconomicconditionsandunrelatedtootherstate-levelvariables. In addition,asshowninFigure1,thetime-seriesofthesesubsequentchangesisindeedstaggeredover timebystateandtheirgeographicdistributionisfairlyspreadoutacrosstheU.S.map. Basedonthese results,wetakeaconservativeapproachandforourbaselineestimatesonlyincludeinthetreatment groupchangesinstateR&Dtaxcreditssubsequenttotheirintroduction. Sincewerecognizethattax creditintroductions,thoughlesssuitableforstandardDIDresearchdesign,arepotentiallyinteresting discreteevents,weexaminetheirimpactinanextensivesetofdedicatedrobustnesstests. Asasecondvalidationtestofourquasi-experimentalsetting,PanelBofTable3reportsparameter estimates from DID regressions of changes in innovation on changes in state R&D tax credits. Column1referstoR&DexpendituresandColumn2toR&Dcapital, while,forthesakeoffalsification, Column 3 refers to Capex and Column 4 to property, plants, and equipment (PPE).46 The estimates showthat,onaverageandrelativetootherfirmsinthesameindustrythatarenotsubjecttoR&Dtax credit changes in their headquarter state-year, there is a strongly statistically significant 0.3% (0.2%) increase(decrease)inR&DexpendituresfollowingR&Dtaxcreditincreases(cuts). Theseeffectsare economically significant, at about 10% of their average pre-treatment levels.47 The impact on R&D 46All specifications are estimated in first differences to remove firm fixed effects in the levels equations, and lagged changesinthefirm-levelcontrolsofthebaselinespecification(1)aswellasyear-industry(48-FF)dummiesareincluded. 47AsshowninTable1,averageR&Dexpendituresintheinnovativefirmssampleare3%ofbookassets. 19

capital is even larger, with an 8.8% (7.4%) increase (decrease), which are about 1/4 of their average levels in the sample. Capex and PPE show no significant response. These findings confirm the existingevidence that R&D expensesrespond strongly to tax incentives and their impact is long-lived (seeHallandVanReenen(2000)forasurvey). BywayofsummaryoftheexistingevidenceontheeffectivenessofR&Dtaxcredit,previousstudies find a considerable response of R&D to tax credits in the U.S. For example, several studies show evidencethatthefederalR&Dtaxcreditintroducedin1981producedroughlyadollar-for-dollarincrease in R&D investment (Hall (1992), Berger (1993), and Hines (1993)). Another main finding is that the R&D response tends to increase over time and is ultimately sizable in the long term. Wilson (2009) documents roughly similar estimates for R&D tax credits at the state rather than at the federal level. There is also growing evidence from cross-country studies that supports the overall effectivenessoftaxcreditsatstimulatingR&D,thoughpointestimatesdifferdependingonthespecific country and period, the type of R&D tax credit, as well as the specifics of its implementation. For example, Bloom, Griffith, and Van Reenen (2002) document evidence of tax credits leading to a dollar-for-dollar increase in R&D investment for a panel of nine OECD countries over the 1979 to 1997 period. In a recent study, Mulkay and Mairesse (2013) document a bit smaller (three quarters ofadollar-for-dollarincrease),butstilloveralllargeandreliablysignificantestimatesoftheeffectof the2008R&DtaxcreditinFrance. 3.3 BaselineDIDEstimates Our baseline estimates from DID regressions of changes in cash holdings on changes in state R&D taxcredits forseveraldifferent specificationsanddefinitions ofcashratios arereportedin Table4.48 Columns1to3reportresultsforcashtobookassetratios,inturnforthebaselinespecification(1),as well as for a more inclusive specification with leads and lags of R&D tax credit changes to address pre-trends,andforaspecificationwherethepercentagechangeintheR&Dtaxcreditisusedinstead of an indicator variable to examine the question of whether the size of the R&D tax credits matters. Robustlyacrossthesespecifications, theestimatesindicatethatafterstateschangetheirtaxcodesto increase(decrease)R&Dtaxcredits,firmsincrease(decrease)theircashholdingsrelativetootherwise similarfirmsinthesameindustrythatarenotsubjecttoR&Dtaxcreditchangesintheirheadquarter state-year. The baseline estimates in Column 1 show that there is an increase (decrease) in cash to 48All specifications are estimated in first differences to remove firm fixed effects in the levels equations, and lagged changesinthefirm-levelcontrolsofthebaselinespecification(1)aswellasyear-industry(48-FF)dummiesareincluded. 20

book assets of about 2 (1.7) percentage points after an increase (decrease) in R&D tax credits. These effectsareeconomicallysignificant,atabout9%(7%)oftheiraveragepre-treatmentlevels.49 Thelack of significance and the relatively small size of the estimates for leads and lags indicate that there is no evidence of either firm anticipation or pre-event trends. Finally, larger tax credit changes trigger largerresponsesincashholdings. Nonparametric analysis of average annual within-firm changes in cash to book asset ratios in eventtimeleadingtoandaftertheyearwhenastateincreases(PanelAofFigure1)orcuts(PanelB ofFigure1)itsR&Dtaxcreditconfirmsthatthereisasharpandstatisticallysignificantchangeincash holdingsintheyearoftreatment(t = 0),whichisnotreversedinsubsequentyears.50 Theremaining columns of Table 4 report DID estimates for the baseline specification using alternative definitions of cash ratios, with the goal of addressing measurement issues with our dependent variable. We experiment with several alternative definitions of cash ratios, which include cash to book assets net of cash (Column 4) and cash to market value of assets (Column 5), which address the concern that changesinthedenominatorofourdependentvariablemaybedrivingtheresult,aswellaswithcash minus net income to book assets (Column 6) and cash to lagged cash plus net income (Column 7). Theselattermeasures,eitherbynettingincomeoutofcashorbyconsideringthemarginalpropensity to retain cash out of cash flow, address the concern that we may be hard-wiring the result if cash changes are simply due to mechanical retention of, say, potentially higher realizations of after-tax cashflowsduetofirmscashinginonhigherR&Dtaxcredits. Theresultsareremarkablystableacross all of these measures, indicating that changes in firm cash holding decisions (the numerator) are drivingthebaselineestimatesandthatthesechangesarenotpurelydrivenbymechanicalretention. 3.3.1 FalsificationTests AreourbaselineDIDestimateswell-identified? Weaddressthisquestionwithtwobatteriesoffalsificationtests.51 First,werepeatouranalysisofchangesincashholdingstobookassetsafterchanges in state R&D tax credits for two different sub-sample splits of the data based on proxies for the in- 49Specifically,asshowninTable1,theaveragecashtobookassetratiointheinnovativefirmssampleis24%. Thus,our baselineDIDestimateisabout9%ofthesamplemean(0.021/0.24=0.0875)forR&Dtaxcreditincreasesandabout7%of thesamplemean(-0.017/0.24=0.071)forR&Dtaxcreditcuts. 50Plottedchangesareinexcessofcontemporaneouscashchangesinthefirm’s48-FFindustry,toremovetheinfluence oftime-varyingchangesinindustryconditions, andofpredictedchangesbasedonpre-treatmentcashlevels, tocontrol for partial adjustment of cash. In the years prior to treatment (t = 2, 1), cash changes are close to zero on average, − − confirmingtheregressionresultthatthereisnoevidenceofpre-eventdifferentialtrendsbetweentreatedandcontrolfirms. 51The specifications for both tests are all estimated in first differences to remove firm fixed effects in the levelsequations,andlaggedchangesinthefirm-levelcontrolsofthebaselinespecification(1)aswellasyearindustry(48-FF)dummiesareincluded. 21

stitutionalfeaturesoftheR&Dtaxcreditsthatshouldaffecttheireffectiveness. Theintuitionbehind these falsification tests is to exploit the non-linearities from the comparative static properties of the opportunity-cost of R&D that we derived above from the institutional design of the state R&D tax credit programs. If cash holdings are responding to unobserved factors, such as changes in local conditions (orthogonal to our state-level controls) rather than to R&D tax credit changes, then we should see no variation across groups when we split the sample by proxies for the intensity of the treatment. The estimate, which are presented in Columns 1 to 4 of Table 5, show a starkly different pattern. Robustly across either of our two sample splits, there are strong and significant effects for firmsthataresubjecttoastrongertreatment,whichareproxiedbyabovemedian(year-prior)values ofstatecorporatetaxrate(Column1)andofgrowthinfirmR&Dexpenditureswithrespecttotheir pre-treatment level (Column 3). By contrast, the effects are weak and not statistically significant for firmsthatunlikelytobeaffectedbythetaxcreditbasedoneithermetric(Columns2and4). Second,weimplementanothersetoffalsificationteststhatexploitsstatetaxcreditsforadifferent typeofinvestment. Specifically, weconsidertheresponseofcashtochangesinstateinvestmenttax credits, which are state-level tax incentives programs otherwise analogous to the R&D tax credits exceptforthefactthattheygivetaxreliefforcapex–i.e.,forinvestmentinPPE,suchasplants,machinery,andbuildings. ChirinkoandWilson(2008)hand-collectedstate-leveldataontheseprograms and show evidence of their effectiveness at stimulating physical capital expenditures. We use their data to construct indicator variables for changes in state investment tax credits, which consist of 18 introductionevents,27increases,and8cuts.52 Ifourestimatesaresimplypickingupeithermechanic retention from cashing in on tax incentives or unobserved state-wide factors, then the nature of the tax incentives should not matter and we should expect a response of cash to changes in investment taxcreditsthatisqualitativelysimilartotheonewedocumentedforR&Dtaxcredits. However,this is not what the estimates presented in Columns 5 to 6 of Table 5 indicate: the response of cash to changesininvestmenttaxcredits,thoughsomewhatweakerinmagnitude,isqualitativelytheexact opposite of the response to R&D credits, with cash decreasing (increasing) in response to increases (cuts)ininvestmenttaxcredits. Inadditiontoofferingausefulfalsificationtest,theevidenceofadifferentialresponseofcashto tax credits for R&D versus investment in physical assets supports theories of cash based on debt financingfrictionssuchasRampiniandViswanathan(2010)andFalato,KadyrzhanovaandSim(2013), 52WereferthereadertoChirinkoandWilson(2008)forfurtherdetailsonthisdata. 22

since cash is accumulated in response to higher incentives to invest in intangible assets that cannot beeasilyfinancedwithdebtduetothelackofpledgeabilityoftheseassets. Thus,whilebothstatetax credits for capital expenditures and those for R&D constitute incentives for higher investment, only R&Dtaxcreditsleadtocashaccumulation,whichisconsistentwiththenotionthatdebtisrelatively more suitable to finance capital expenditures. In the next section, we will return to the question of debtfinancingfrictionsandofferadditional,moredirectevidenceontheirroleinouranalysisofthe channelsthroughwhichinnovationimpactscash. 3.3.2 RobustnessandExternalValidityTests Next, we run a battery of sensitivity tests to further corroborate the internal validity of our baseline estimates. Table 6 tabulates the results. We consider three different sets of sensitivity tests.53 First, we estimate versions of our baseline DID regressions that use alternative standard errors and add more firm-, state-, and industry-level control variables (Panels A and B). Row 1 shows results for a specification that clusters standard errors at the state level, to address the concern that the key sourceofvariationintheanalysisisatthestatelevel(Bertrand,Duflo,andMullainathan(2004)),and addsstatefixedeffectstocontrolforunobservedtime-invariantdifferencesamongstates,sothatthe resultingestimatesareidentifiedusingwithin-firm,within-industry-year,andwithin-statevariation. Rows2to8addcontrolsforfirm-andindustry-level,aswellastime-varyingstate-levelconfounds. Second, we address the concern that local economic conditions could contaminate our inference by refining our control group (Panel C, Row 9). Specifically, we refine the control group to include only firms in neighboring counties across border (see Homes (1998) for a similar approach). The intuition for why this approach is effective at differencing out local economic conditions is that tax policiesdisplayanaturalgeographicdiscontinuityattheborder(i.e.,theystopatthestate’sborder), whileotherspuriousfactorssuchaslocaleconomicconditionsplausiblydonot. Thus,bycomparing neighboring county pairs across state borders, we can difference out unobserved variation in local economicconditions. Wedosobyestimatingaversionofourbaselinespecificationthataddscountypairs fixed effects and is estimated only for firms that have a neighboring county across the border. The resulting estimates are identified within contiguous county dyads across state borders, which sharpensouridentification. Overall,ourDIDestimatesarestableacrossthesetwosetsofsensitivity tests. 53All specifications are estimated in first differences to remove firm fixed effects in the levels equations, and lagged changesinthefirm-levelcontrolsofthebaselinespecification(1)aswellasyear-industry(48-FF)dummiesareincluded. 23

Third,weconsiderrobustnesstoalternativeestimatorsandassestheexternalvalidityofourbaseline estimates. We consider three alternative estimators: the system IV-GMM of Blundell and Bond (1998) to address potential biases from including the lagged dependent variable (Panel C, Row 10); an IV-DID (2-SLS) estimator that uses R&D tax credit changes as instruments for changes in R&D capital, to offer well-identified estimates of the impact of R&D on cash (see Bloom, Schankerman, and Van Reenen (2013) for a similar approach; estimates of the second stage regression with instrumented R&D stock are reported in Panel C, Row 11); and a matched-sample DID estimator (Heckman,Ichimura,andTodd(1997))thatmatchestreatedandcontrolfirmsbasedonpre-treatmentsize, industry, and performance to addresses the concern that these differences between treated and controlfirmsmayinvalidateourparallel-trendassumption. To examine external validity, we include R&D tax credit introduction events (Panel D, Rows 13 and15)andasubsetofthemostdiscretesuchevents(topquartileofthestatutorycreditrate;PanelD, Rows14and16)inthetreatmentgroup. Givenourfindingthattaxcreditintroductionsaresystematicallyrelatedtostateeconomicconditions,forbothcaseswereportmatched-sampleDIDestimates basedonpre-treatmentstateeconomicconditions.54 Ourestimatesremainstableacrossdifferentestimatorsandtypesoftreatments. Overall,resultsofthefalsification,sensitivity,andexternalvalidity testssupportacausalinterpretationoftheobservedresponseofcash. 4 Why Does Innovation Matter? Assessing the Channels Inthesecondpartofouranalysis,weinvestigatethedriversofthetreatmenteffectbyexaminingits variationwithseveralvariablesthat,basedontheory,weexpectshouldaffectthedemandforcashby innovativefirms. Wetestwhetherthetreatmenteffectdisplayssystematicheterogeneityacrossfour setofvariablesthatcapturethemaintheoriessummarizedinourtheoryoverviewsection: financing frictions,agencyfrictions,repatriation,productmarketcompetition. Wealsoofferadirectassessment offinancingfrictionsbyexaminingtheimpactofchangesinstateR&Dtaxcreditsonfirms’liquidity structure (cash vs. lines of credit), as well as debt and equity financing decisions. This analysis providesanempiricalassessmentofthedifferentexplanationsforwhyinnovationhasanimpacton cash. Inaddition,itoffersevidenceonthedeterminantsofcashholdingand,morebroadly,financing decisionsforinnovativefirms. 54WhilewereportvanillaDIDestimatesforreferenceinRows13and14,weemphasizethattheyarebestinterpretedas descriptiveevidence. 24

4.1 AssessingDebt&EquityFinancingFrictions If innovative firms hold more cash because of financing frictions, then the treatment effect of state R&D tax credits on cash should be stronger among firms that face more severe financing frictions. Table 7 reports results from running our baseline DID regressions in several different sub-sample splitsof thedatabased onbroad ex-anteproxiesfor theseverityof financialfrictionsfaced byfirms (PanelA),aswellasfortheintensityoftheprecautionarymotivetoholdcash(PanelB),anapproach which is standard in the literature (e.g., Hennessy and Whited (2007)).55 Specifically, we split the samplebetweenabovevs. belowmedian(pre-treatment)valuesofthefollowingvariables:firmsize (Column 1), Kaplan and Zingales (1997) KZ-Index (Column 3), Whited and Wu (2006) WW-Index (Column 4), firm cash flows (Column 5), firm age (Column 6), firm market-to-book ratio (Column 7), firm cash flow volatility (Column 8), and by dividend payer status (Column 2). Robustly across these proxies, the results indicate that the response of cash to changes in state R&D tax credits is magnifiedbyfirms’financingconstraintstatus,withfirmsthathavegreaterfinancialslackshowing a muted response, and by the intensity of their precautionary motive, with the effect concentrated amongrelativelyyoungerandriskierfirms,thosewithlowcashflowsandgreatergrowthpotential. Next,weexaminewhichfinancingfrictionsmatter. Toassessdebtfinancingfrictions,wetaketwo different approaches, with results summarized in Table 8. First, we examine the impact of changes in state R&D tax credits on corporate liquidity structure and debt financing (Panel A). We consider probitregressionsofthelikelihoodofhavingbankliquidity(Column1),DIDestimatesofchangesin unusedbankliquidity(unusedlinesofcredit)tototalliquidity(Column2),innetdebtissuance(Column3),andinnetbookleverage(Column4),andestimatesfromprobitregressionsofthelikelihood of having a debt structure that is specialized in bank debt (Column 5) and in secured debt (Column 6). Inallthesetests,weuseaspecificationanalogoustoourbaseline(1). Second,werunourbaseline DID regressions of cash holdings for three different sub-sample splits of the data based on ex-ante proxies for the severity of the debt financing frictions faced by firms (Panel B): Berger et al. (1996) assetliquidationvalue(Columns1and2),BalasubramanianandSivadasan(2009)indexofindustry assetredeployability(Columns3and4),andbasedonwhetherthefirmhasalong-termdebtrating (Columns 5 and 6). Increases (cuts) in R&D tax credits lead to increases (decreases) in cash relative to bank lines of credit, and to decreases (increases, though estimates of these increases are noisy) in 55All specifications are estimated in first differences to remove firm fixed effects in the levels equations, and lagged changesinthefirm-levelcontrolsofthebaselinespecification(1)aswellasyear-industry(48-FF)dummiesareincluded. 25

bank debt, secured debt, and overall net indebtness. These results, together with the evidence that R&D tax credits have a stronger impact on cash for firms that have relatively less access to debt financing, are all consistent with a debt financing channel through which innovation increases firms’ demandforcashbyreducingtheirdebtcapacity. Wefollowananalogousstrategytoassessequityfinancingfrictions. Theresultsaresummarized in Table 9, which reports parameter estimates of the impact of changes in state R&D tax credits on equity financing (Panel A), and DID estimates of our baseline cash regressions for several different sub-samplesplitsofthedatabasedonex-anteproxiesfortheseverityoftheequityfinancingfrictions faced by firms (Panel B). Specifically, in Panel A we report DID estimates of changes in net equity issuance(Column1)andcashtobookequityratio(Column3),andestimatesfromprobitregressions ofthelikelihoodofasecondaryequityissue(SEO)(Column2). InPanelB,thesampleissplitbased on the total dollar volume of SEOs in a firm’s (48-FF) industry in any given year (Columns 1 and 2), the average (0, +1) cumulative abnormal return (CAR) of SEOs announcements in a firm’s (48- FF) industry in any given year (Columns 3 and 4), which are commonly employed proxies for hot vs. coldequitymarkets(e.g., KorajczykandLevy(2003), andthestandarddeviation(dispersion)of analysts’EPSforecastsfromIBES(Columns5and6). Table 9 shows that increases (cuts) in R&D tax credits lead to increases (decreases) in net equity issuance. However, cash increases also relative to book equity, suggesting that, even though innovative firms are well-recognized to be relatively more reliant on equity, external financing is only an imperfect substitute for internal funds. Also consistent with an equity financing channel, R&D tax credits have a stronger impact on cash at times when firms have relatively less access to equity markets. In all, this evidence indicates that external financing frictions are an empirically relevant explanationforwhyinnovativefirmsholdsmorecash,andthatthisisduenotonlytodebtfinancing frictions,whichhavebeenthetraditionalfocusoftheexistingliteraturesurveyedinHallandLerner (2010) (see Chava et al. (2013) and Mann (2013) for recent studies), but also to equity financing frictions, whose importance has been recently emphasized by dynamic corporate finance models such asBolton,Chen,andWang(2011),EisfeldtandMuir(2013),andWarusawitharanandWhited(2013) (seeBrown, Martinsson, andPetersen(2013) forrecent cross-countryevidence onthe importanceof equityfrictionsforR&Dfinancing). 26

4.2 AssessingRepatriation,IndustryCompetition,andAgency In our final set of tests, we run our baseline DID regressions of cash holdings for three types of sub-samplesplitsofthedatabasedonex-anteproxiesforrepatriationincentivesandproductmarket competition(PanelAofTable10),aswellasoftheseverityofagencyfrictionsfacedbythefirm(Panel B of Table 10).56 In Panel A, the sample is split between above vs. below median (pre-treatment) values of the foreign tax burden measure of Foley, Hartzell, and Titman (2007) (Columns 1 and 2), based on whether the firm reports foreign income in a given year (Columns 3 and 4), and based on above vs. below median (pre-treatment) values of the (48-FF) industry Herfindahl-Hirschman Index (HHI) of sales (Columns 5 and 6). The results indicate that the response of cash to changes in state R&D tax credits is weakened by firms’ multinational status, suggesting that the positive effect of innovation on domestic financing frictions is empirically the driving factor of the repatriation channel, while the response is magnified by the intensity of industry competition, which supports recent theories such as Morellec, Nikolov, and Zucchi (2013) and Ma, Mello, and Wu (2013) where cashholdingsstrengthenthecompetitivepositionoffirmsvisavistheirindustryrivals. In Panel B of Table 10, we examine the agency channel. We report results for the sub-sample of firms in relatively concentrated industries (those with above-median HHI) and further stratify the sample based on above vs. below median (pre-treatment) values of standard metrics for the likelihoodofmanagerialentrenchment,whichincludeboardsize(Columns1and2),boardindependence (Columns 5 and 6), and whether the firm has a classified board of directors (Columns 3 and 4). Robustly across these metrics, the results indicate that R&D tax credits have a stronger impact on cash for firms that are relatively more likely to face agency issues between managers and shareholders. By contrast, when we replicate these agency splits either in the overall sample or in the sub-sample of firms in relatively less concentrated industries (those with below-median HHI), we find mixed results.57 This evidence is consistent with the view that there is no "one size fits all" governance for innovative firms that face intense competitive pressure (Coles, Daniel, and Naveen (2008)). By contrast,andinlinewiththefindingsofsignificantinteractioneffectsbetweengovernanceandindustry (GiroudandMueller(2011)),inrelativelylesscompetitiveindustriesclassicalagencyconsiderations ofthetypehighlightedintheempiricalliteraturebyDittmarandMahrt-Smith(2007)andGao,Harford,andLi(2013)arelikelytobeimportantfactorsforinnovativefirms. 56All specifications are estimated in first differences to remove firm fixed effects in the levels equations, and lagged changesinthefirm-levelcontrolsofthebaselinespecification(1)aswellasyear-industry(48-FF)dummiesareincluded. 57Whichwedonotreportforbrevity. 27

5 Conclusions Does innovation matter for corporate cash holdings, and why? That innovation should matter for a variety of explanations, including financing frictions, agency problems, product market competition, and repatriation is well-understood. Yet, it is challenging to provide well-identified estimates of the impact of innovation on cash and to understand which explanations are empirically relevant. One big challenge is that commonly-employed variables associated with innovation, such as R&D expendituresorcapital,arealsorelatedtofuturebusinessopportunities,whichcomplicatesacausal interpretation of existing cross-sectional evidence. In an attempt to advance the empirical literature on this important yet challenging question, we have exploited time-series variation in firms’ incentivestoinnovatearoundchangesinstateR&Dtaxcredits. In particular, we have taken several steps to design a quasi-natural experimental setting that is successful at overcoming the empirical identification challenge of finding plausibly exogenous sourcesofvariationintheinnovation-drivendemandforcash. Wehaveusedourquasi-experimental setting to assess specific channels through which innovation matters. Our evidence indicates that innovation is a first-order determinant of corporate liquidity, and clarifies when innovation matters inthecross-section. Specifically,innovationmattersnotonlyforcashholdingdecisionsoffirmsthat face more severe debt financing frictions, but also at times when firms have less access to equity financing. Inaddition,itmattersmoreinrelativelymorecompetitiveindustries. Finally,wealsofind evidencesupportingtheimportanceofagencyissuesforinnovativefirms,whichmatterempirically butonlyinrelativelylesscompetitiveindustries. 28

References [1] Aghion, P. and J. Tirole, 1994„ "The Management of Innovation," Quarterly Journal of Economics,109(4),1185-1209. [2] Almeida, H., M. Campello, and M. S. Weisbach, 2004, "The Cash Flow Sensitivity of Cash," JournalofFinance,59(4),1777-1804. [3] Angrist, J. D. and J. Pischke, 2009, Mostly Harmless Econometrics, Princeton University Press, NewJersey. [4] Arellano,M.andS.Bond,1991,"Sometestsofspecificationforpaneldata: MonteCarloevidence andanapplicationtoemploymentequations,"TheReviewofEconomicStudies,58,277–297. [5] Ashenfelter,O.,1978,"EstimatingtheEffectofTrainingProgramsonEarnings,"ReviewofEconomicsandStatistics,60,47-57. [6] Baker,S.R.,N.Bloom,andS.J.Davis,2013,"MeasuringEconomicPolicyUncertainty,"mimeo, StanfordUniversity. [7] Bates T. W., K. M. Kahle, and R. M. Stulz, 2009, "Why Do U.S. Firms Hold So Much More Cash thanTheyUsedTo?,"JournalofFinance,64(5),1985-2021. [8] Berger,P.G.,1993,"ExplicitandImplicitTaxEffectsoftheR&DTaxCredit,"JournalofAccountingResearch,31(2),131-171. [9] Berger, P., E. Ofek, and I. Swary. 1996, "Investor Valuation and Abandonment Option," Journal ofFinancialEconomics,42,257–87 [10] Bertrand,M.,E.Duflo,andS.Mullainathan.2004.“HowMuchShouldWeTrustDifferences-in- DifferencesEstimates?”TheQuarterlyJournalofEconomics,119(1),249-275. [11] Bloom, N., R. Griffith, and J. Van Reenen, 2002, "Do R&D tax credits work? Evidence from a panelofcountries,1979–1997,"JournalofPublicEconomics,85,1–31. [12] Bloom, N., M. Schankerman, and J. Van Reenen, 2013, "Identifying Technology Spillovers and ProductMarketRivalry,"Econometrica,81(4),1347–1393. [13] Bolton, P., H. Chen, and N. Wang, 2011, "A unified theory of Tobin’s q, corporate investment, financing,andriskmanagement,"JournalofFinance,66(5),1545–1578. [14] Bonaime, A. A., K. W. Hankins, and J. Harford, 2013, "Financial Flexibility, Risk Management, andPayoutChoice,"ReviewofFinancialStudies,Forthcoming. [15] Blundell,R., andS.Bond,1998, “InitialConditionsandMomentRestrictionsinDynamicPanel DataModels,”JournalofEconometrics,87,115-143. [16] Brown, J. R. and B. C. Petersen, 2011, "Cash Holdings and R&D Smoothing," Journal of CorporateFinance,17(3),694-709 [17] Brown, J. R., S. M. Fazzari, and B. C. Petersen, 2009, "Financing Innovation and Growth: Cash Flow,ExternalEquity,andthe1990sR&DBoom,"JournalofFinance,64(1),pp.151-185. [18] Brown, J. R., G. Martinsson, and B. C. Petersen, 2013, "Law, Stock Markets, and Innovation," JournalofFinance,68(4),pp.1517-1549. 29

[19] Chava, S., A. Oettl, A. Subramanian and K. Subramanian, 2013, “Banking Deregulation and Innovation”,JournalofFinancialEconomics,109(3),759-775. [20] Chirinko,R.andD.Wilson,2008,"StateInvestmentTaxIncentives: AZero-SumGame?"Journal ofPublicEconomics,92(12),2,362-2,384. [21] Coles,J.L.,D.N.Daniel,andL.Naveen,2008,"Boards: Doesonesizefitall?"JournalofFinancial Economics,87(2),329-356. [22] Colla,P.,F.Ippolito,andK.Li,2013,"DebtSpecialization,"JournalofFinance,68,2117-2141. [23] Corrado, CarolA., andCharlesR.Hulten, 2010, "HowDoYouMeasurea"TechnologicalRevolution"?"AmericanEconomicReview,100(2),99–104. [24] Corrado, C., C. Hulten, and D. Sichel, 2009, "Intangible Capital and US Economic Growth," ReviewofIncomeandWealth,55(3),661-685 [25] Dittmar,A.andR.Duchin,2010,"TheDynamicsofCash,"mimeo,UniversityofMichigan [26] Dittmar,A.andJ.Mahrt-Smith,2007,"CorporateGovernanceandtheValueofCashHoldings," JournalofFinancialEconomics,83(3),599–634. [27] Dittmar, A., J. Mahrt-Smith, and H. Servaes, 2003, "International Corporate Governance and CorporateCashHoldings,"JournalofFinancialandQuantitativeAnalysis,38(1),111-133. [28] Eisfeldt,AandT.Muir,2013,"AggregateIssuanceandSavingsWaves,"mimeo,UCLA. [29] Eisfeldt,AandDPapanikolaou,2011,"OrganizationCapitalandtheCross-SectionofExpected Returns,"JournalofFinance,Forthcoming. [30] Eisfeldt, A. and D. Papanikolaou, 2014, "The Value and Ownership of Intangible Capital," The AmericanEconomicReview,104(5),189-194. [31] Faccio,M.,andJ.Xu,2011,“TaxesandCapitalStructure.”WorkingPaper,PurdueUniversity. [32] Falato,A,D.Kadyrzhanova,andJ.Sim,2013,"RisingIntangibleCapital,ShrinkingDebtCapacity,andtheUSCorporateSavingsGlut,"mimeo,FederalReserveBoard. [33] Faulkender,M.,M.J.Flannery,K.WatsonHankins,J.M.Smith,2012,"CashFlowsandLeverage Adjustments,"JournalofFinancialEconomics,103(3),632-646. [34] Faulkender, M. and M. Petersen, 2012, "Investment and Capital Constraints: Repatriations UndertheAmericanJobsCreationAct,"ReviewofFinancialStudies,25(11),3351-3388. [35] Flannery,M.,andK.Rangan,2006,"Partialadjustmenttowardtargetcapitalstructures,"Journal ofFinancialEconomics79,469–506. [36] Foley, F., J. Hartzell, S. Titman, and G. Twite, 2007, "Why Do Firms Hold So Much Cash? A Tax-BasedExplanation,"JournalofFinancialEconomics,86,579-607. [37] Froot, K. A. and Scharfstein, D. S. and Stein, J. C., 1993, "Risk Management: Coordinating CorporateInvestmentandFinancingPolicies,"TheJournalofFinance,48(5),1629-1658 [38] Gao H., J. Harford, and K. Li, 2013, Determinants of Corporate Cash Policy: Insights from PrivateFirms,"JournalofFinancialEconomics,109,623–639. [39] Giroud,X.,andH.M.,Mueller,2011,"CorporateGovernance,ProductMarketCompetition,and EquityPrices,"TheJournalofFinance,66(2),563-600. 30

[40] Gormley,T.A.andD.A.Matsa,2014,"CommonErrors: Howto(andNotto)ControlforUnobservedHeterogeneity,"ReviewofFinancialStudies,27(2),617-61. [41] Graham,J.R.,1996,"Proxiesforthecorporatemarginaltaxrate,"JournalofFinancialEconomics 42,187-221. [42] Hall, B. H., 1992, “R&D Tax Policy during the 1980s: Success or Failure?" Tax Policy and the Economy,7,1-36 [43] Hall,B.H.,1993,“TheStockMarket’sValuationofR&DInvestmentduringthe1980s,”American EconomicReview,83(2),259–64 [44] Hall, B.H., A. B. Jaffe, and M. Trajtenberg, 2001, "The NBER Patent Citation Data File: Lessons, InsightsandMethodologicalTools,"NBERWP#8498. [45] Hall, B.H. and J. Lerner, 2010, "The Financing of R&D and Innovation," in Hall, B. H. and N. Rosenberg(eds.),HandbookoftheEconomicsofInnovation,Elsevier-NorthHolland. [46] Hall,B.H.andJ.VanReenen,2000,"HowEffectiveareFiscalIncentivesforR&D?AReviewof Evidence,"ResearchPolicy,29(4-5),449–469. [47] Hall,R.,2001,“TheStockMarketandCapitalAccumulation,”AmericanEconomicReview91(5), 1185–1202 [48] Hall, R. E. and D. W. Jorgenson, 1967, "Tax Policy and Investment Behavior,” American EconomicReview,57(3),391-414. [49] Harford, J, 1999, "Corporate Cash Reserves and Acquisitions," The Journal of Finance, 54(6), 1969-1997 [50] Hart, O and J. Moore, 1994, "A Theory of Debt Based on the Inalienability of Human Capital," TheQuarterlyJournalofEconomics,109(4),841-79. [51] He, Z., 2014, "The Cost of Innovation: R&D and High Cash Holdings in U.S. Firms," mimeo, UniversityofKansas. [52] Heckman,J.J.,H.IchimuraandP.E.Todd,1997,“MatchingasanEconometricEvaluationEstimator: EvidencefromEvaluatingaJobTrainingProgramme,”TheReviewofEconomicStudies, 64,605-654. [53] Heider,F.andA.Ljungqvist,2012,“AsCertainasDebtandTaxes: EstimatingtheTaxSensitivity of Leverage from Exogenous State Tax Changes.” Working Paper, European Central Bank and NewYorkUniversity. [54] Hennessy, C. A., and T. M. Whited, 2007, "How Costly Is External Financing? Evidence from a StructuralEstimation,"TheJournalofFinance,62(4),1705-1745. [55] Hines, J. R., 1993, "On the Sensitivity of R&D to Delicate Tax Changes: The Behavior of U.S. Multinationalsinthe1980s,"StudiesinInternationalTaxation,EditorsA.Giovanni,R.G.Hubbard,andJ.Slemrod,UniversityofChicagoPress,149-194. [56] Hirsch,B.T.,andD.A.Macpherson,2003,"Unionmembershipandcoveragedatabasefromthe CurrentPopulationSurvey: Note,"IndustrialandLaborRelationsReview,56(2),349-354. [57] Hoberg,G.,G.PhillipsandN.Prabhala,2014,"ProductMarketThreats,Payouts,andFinancial Flexibility,"JournalofFinance,69(1),293-324. 31

[58] Holmes, T. J., 1998, "The Effect of State Policies on the Location of Manufacturing: Evidence fromStateBorders,"JournalofPoliticalEconomy,106(4),667-705. [59] Jensen, M., 1986, "AgencyCostsofFreeCashFlow, CorporateFinance, andTakeovers,"AmericanEconomicReview,76(2),323-329. [60] Kaplan,S.,andL.Zingales,1997,"DoFinancingConstraintsExplainWhyInvestmentisCorrelatedwithCashFlow?"QuarterlyJournalofEconomics,112,269–215. [61] Korajczyk,R.A.,andA.Levy,2003,"CapitalStructureChoice: MacroeconomicConditionsand FinancialConstraints,JournalofFinancialEconomics,68,75-109. [62] Leary, M., and M. R. Roberts, 2005, "Do firms rebalance their capital structures?," Journal of Finance60,2575–2619 [63] Leland,H.E.,Pyle,D.H.(1977).“InformationalAsymmetries,FinancialStructure,andFinancial Intermediation,”JournalofFinance32,371-87. [64] Lemmon,M.,M.Roberts,andJ.Zender,2008,"Backtothebeginning: persistenceandthecrosssectionofcorporatecapitalstructure,"JournalofFinance,63,1575-1608. [65] Linck,J.,J.Netter,andT.Yang,2008,"TheDeterminantsofBoardStructure,"JournalofFinancial Economics,87(2),308–328. [66] LoughranT.andJ.Ritter,1995,"TheNewIssuesPuzzle,"JournalofFinance,50,23-51. [67] Lyandres, E. and D. Palazzo, 2014, "Cash Holdings, Competition, and Innovation," mimeo, BostonUniversity. [68] Ma, L., A. S. Mello, and Y. Wu, 2013, "Industry Competition, Winner’s Advantage, and Cash Holdings,"mimeo,UniversityofWisconsin. [69] MacKay, P,2003, "RealFlexibilityandFinancialStructure: AnEmpiricalAnalysis,"TheReview ofFinancialStudies,16(4),1131-1165. [70] Mann,W,2013,"CreditorRightsandInnovation: EvidencefromPatentCollateral,"mimeo,The WhartonSchool,UniverstiyofPennsylvania. [71] Matsa,D.A.,2010,"CapitalStructureasaStrategicVariable: EvidencefromCollectiveBargaining,"JournalofFinance,65,1197-1232. [72] Morellec,E.,B.Nikolov,andF.Zucchi,2013,"Competition,CashHoldings,andFinancingDecisions,"SwissFinanceInstituteResearchPaperNo.13-72. [73] Mulkay,B.andJ.Mairesse,2013,"TheR&DtaxcreditinFrance: assessmentandexanteevaluationofthe2008reform,"OxfordEconomicPapers,OxfordUniversityPress,65(3),746-766. [74] Nickell,S.,1981,"Biasesindynamicmodelswithfixedeffects,"Econometrica,49,1417–1426. [75] Nikolov,B.andT.Whited,2013,"AgencyConflictsandCash: EstimatesfromaDynamicModel," JournalofFinance,Forthcoming. [76] Opler, T., L. Pinkowitz, R. Stulz, and R. Williamson, 1999, “The Determinants and Implications ofCorporateCashHoldings,”JournalofFinancialEconomics52,3-46. [77] Opler, T., Titman, S., 1994. Financial distress and corporate performance. Journal of Finance 49 (3),1015–1040 32

[78] Panier, F., F. Pérez-González, and P. Villanueva, 2012, "Capital Structure and Taxes: What HappensWhenYou(Also)SubsidizeEquity?"mimeo,StanfordUniversity [79] Rampini, A. and S. Viswanathan, 2010, "Collateral, Risk Management, and the Distribution of DebtCapacity,"TheJournalofFinance,65(6),2293–2322. [80] Riddick L. and T. Whited, 2009, “The Corporate Propensity to Save,” Journal of Finance, 64, 1729–1766. [81] Schmalz,M.C.,2013,"ManagingHumanCapitalRisk,"RossSchoolofBusiness,MichiganUniversity,WorkingPaperNo.1215. [82] Shleifer, A.andVishny, R.W., 1992, "LiquidationValuesandDebtCapacity: AMarketEquilibriumApproach,"JournalofFinance,47(4),1343-66. [83] Stein, J., 2003, "Agency, Information and Corporate Investment," in G.M. Constantinides, M. HarrisandR.Stulz,eds.: HandbookoftheEconomicsofFinance(Elsevier,Amsterdam). [84] Sufi, A, 2009, "Bank Lines of Credit in Corporate Finance: An Empirical Analysis," Review of FinancialStudies,22(3),1057-1088. [85] Warusawitharan, M.andT.Whited, 2013, "EquityMarketMisvaluation, Financing, andInvestment,"mimeo,UniversityofRochester. [86] WhitedT.andGWu,2006,"FinancialConstraintsRisk,"ReviewofFinancialStudies19,531-559. [87] Wilson, D., 2009, “Beggar thy Neighbor? The In-State, Out-of-State and Aggregate Effects of R&DTaxCredits,”ReviewofEconomicsandStatistics,91(2),431-436. 33

Appendix A. Variable Definitions Thevariablesusedintheanalysisaredefinedasfollows,withtheirrespectivedatasources: CorporateLiquidityVariables: CashtoBookAssetsisdefinedascashandmarketablesecurities(Compustatdataitem#1)dividedby • bookassets(#6). CashtoNetBookAssetsiscashandmarketablesecurities(#1)dividedbybookassets(#6)minuscash • andmarketablesecurities(#1). Cash to Market Value of Assets is cash and marketable securities (#1) divided by long-term debt (#9) • plusdebtincurrentliabilities(#34)plusmarketvalueofequity,whichisconstructedascommonshares outstanding(#25)*price(#199). CashtoBookEquityiscashandmarketablesecurities(#1)dividedbybookequity,whichisdefinedas • bookvalueofcommonequity(#60)pluspreferredstock(#10). NetCashtoBookAssetsiscashandmarketablesecurities(#1)minusnetincome(#172)dividedbybook • assets(#6). DollarCashiscashandmarketablesecurities(#1)in1990dollars(usingCPI). • Bank Liquidity to Total Liquidity is the ratio of total line of credit (drawn plus undrawn, Capital IQ • data)tothesumoftotallineofcredit(drawnandundrawn,CapitalIQdata)pluscashandmarketable securities(#1). UnusedBankLiquiditytoTotalLiquidityistheratioofunusedlineofcredit(undrawn,CapitalIQdata) • tothesumofunusedlineofcredit(undrawn,CapitalIQdata)pluscashandmarketablesecurities(#1). Innovation&StateR&DTaxCredits: R&DExpendituresistheratioofR&Dexpenditures(#46)tobookassets(#6). • R&D Capital is constructed by capitalizing R&D expenditures using the perpetual inventory method • as follows: G = (1 δ )G + R&D where G is the end-of-period stock of R&D capital, it R&D it 1 it t R&D isthe($1990rea − l)expenditu − resonR&Dduringtheyear,and δ = 15%followingHall,Jaffe, it R&D and Trachtenberg (2001). If R&D expenditures are constant (in real terms), the stock of R&D capital is G t = ∑∞ s=0 (1 − δ)s R&D t − s = R δ .Wesett δ heinitialstocktobeequaltotheR&Dexpendituresinthe first year divided by the depreciation rate . In addition, we interpolate missing values of R&D R&D following Hall (1993) who shows that this results in an unbiased measure of R&D capital. The R&D stockisscaledby($1990real)bookassets(#6). R&DTaxCreditIncreaseisadummyvariablethatequalsoneifastateeitherincreasesorre-introduces • itsR&Dtaxcreditinagivenyear. DetailsondatasourcesandthefulllistofeventsareinAppendixB. R&DTaxCreditCutisadummyvariablethatequalsoneifastateeithercutsitsR&Dtaxcreditorletsit • expireinagivenyear. DetailsondatasourcesandthefulllistofeventsareinAppendixB. R&DTaxCreditIntroductionisadummyvariablethatequalsoneifastatethatneverhadanR&Dtax • creditintroducesitinagivenyear. DetailsondatasourcesandthefulllistofeventsareinAppendixB. BaselineFirmControls: Cash Flow Volatility is the standard deviation of cash flow to book assets. Standard deviation of cash • flowtobookassetsiscomputedforeveryfirm-yearusingdataovertheprevioustenyears. Market-to-Book ratio is the ratio of the book value of assets (#6) minus the book value of equity (#60) • plusthemarketvalueofequity(#199*#25)tothebookvalueofassets(#6). FirmSizeisthenaturallogarithmofbookassets(#6)in1990dollars(usingCPI). • Cash Flow is earnings after interest, dividends, and taxes before depreciation divided by book assets • ((#13–#15–#16–#21)/#6). 34

CapitalExpendituresistheratioofcapitalexpenditures(#128)tobookassets(#6). • Dividend is a dummy variable equal to one in years in which a firm pays a common dividend (#21). • Otherwise,thedummyequalszero. AcquisitionExpendituresistheratioofacquisitions(#129)tobookassets(#6). • Channels&AdditionalControls: Debt&EquityFinancingChannels: Net-Leverage is the ratio of long-term debt (#9) plus debt in current liabilities (#34) minus cash and • marketablesecurities(dataitem#1)tobookassets(#6). Leverageistheratiooflong-termdebt(#9)plusdebtincurrentliabilities(#34)tobookassets(#6). • BankDebttoTotalDebtistheratiooftotalbankdebt(lineofcreditplustermloan,CapitalIQdata)to • totaldebt(CapitalIQdata). Bank90 is a dummy that takes value of one for firm-year observations that are in the top decile of the • distributionofbankdebttototaldebt(CapitalIQdata). Secured Debt to Total Debt is the ratio of total secured debt (Capital IQ data) to total debt (Capital IQ • data). Secured90isadummythattakesvalueofoneforfirm-yearobservationsthatareinthetopdecileofthe • distributionofsecureddebttototaldebt(CapitalIQdata). Net Debt Issuance is long term issuance (#111) minus long term debt reduction (#114) plus changes in • currentdebt(#301),dividedbybookassets. Net Equity Issuance is Sale of Common and Preferred Stock (#108) minus purchase of common and • preferredstock(#115),dividedbybookassets. SEO Volume is the total dollar value of SEOs in a firm’s (48-FF) industry in a given year, and is con- • structedfromthedatabaseof4,291SEOsofLoughranandRitter(1995)(asupdatedbytheauthors)for the1973-2001period,andfromSDCforthe2002-2011period. SEOCostsistheaverage(0,+1)cumulativeabnormalreturn(CAR)ofSEOsannouncementsinafirm’s • (48-FF) industry in a given year, and is constructed from the database of 4,291 SEOs of Loughran and Ritter (1995) (as updated by the authors) for the 1973-2001 period, and from SDC for the 2002-2011 period. AnalystForecastDispersionisthestandarddeviationofanalysts’EPSforecasts(IBESdata). • KZ-IndexisbasedonKaplanandZingales(1997)andisasfollows: • KZ=-1.002*Cashflow+0.283*Q+3.139*Leverage-39.368*Dividends-1.315*Cash, where Q is the market-to-book ratio, Dividends is the ratio of total dividends to book assets, and the othervariablesareasdefinedabove. WW-IndexisbasedonWhitedandWu(2006)andisasfollows: • WW=-0.091*CashFlow-0.062*Dividend+0.021*Leverage -0.044*Size+0.102*IndustryGrowth-0.035*Growth, whereIndustryGrowthisthe4-SICindustrysalesgrowth,Growthisown–firmrealsalesgrowth,and theothervariablesareasdefinedabove. Resale Value is based on Berger et al. (1996) and is the sum of 0.715*Receivables (#2), 0.547*Inventory • (#3),and0.535*Capital(#8). Asset Redeployability Index is based on Balasubramanian and Sivadasan (2009) and is the fraction of • totalcapitalexpendituresinanindustryaccountedforbypurchasesofused(asopposedtonew)capital, computedat4-digitSIClevelandconstructedusinghand-collectedUSCensusBureaudata. Sincethese dataareavailableonlyonceevery5yearsandnotformorerecentyears, wecomputeatime-invariant indexbyaveragingtheavailablequinquennialindexesatthe4-SIClevel. Thismeasureisonlyavailable forarestrictedsampleofmanufacturingfirms. 35

Firm Age is the number of years since IPO, with information on IPO year gathered from Compustat, • whenavailable,andotherwisefromSDCandJayRitter’sIPOdatabase. Debt Rating is a dummy that equals one if the firm has a debt rating from S&P in a given year, and is • constructedbasedonCompustathistoricaldebtratinginformation. Agency,ProductMarket,andRepatriationChannels: MarginalTaxRateistheafter-interestmarginaltaxratefromGraham(1996). • ForeignTaxBurdenisconstructedfollowingFoley,Hartzell,Titman,andTwite(2007)asthemaximum • between zero and foreign pre-tax income (#273) times the marginal tax rate minus foreign taxes paid (#64). ForeignIncomeistheratioofforeignpre-taxincome(#273)tobookassets(#6). • IndustryCompetitionistheHerfindahl-HirschmanIndex(HHI),whichisconstructedasthesumofthe • squaresoftheindividualcompanymarketsharesforallthefirmsinagiven(48-FF)industry-year. High-Tech industries are defined following Loughran and Ritter (2004) as SIC codes 3571, 3572, 3575, • 3577, 3578, 3661, 3663, 3669, 3674, 3812, 3823, 3825, 3826, 3827, 3829, 3841, 3845, 4812, 4813, 4899, 7370, 7371,7372,7373,7374,7375,7378,and7379. Board Size is the total number of directors on the board in a given firm-year, and is constructed using • informationfromproxyfilingsextractedfromCompactDisclosureforthe1986-2005period(asinLinck, NetterandYang(2008)),andusingCapitalIQforthe2006-2011period. ClassifiedBoardisadummyvariablethatequalsoneifthefirmhasaclassifiedboardofdirectors,and • isconstructedusingIRRCdataforthe1992-2005period,andCapitalIQforthe2006-2011period. Board Independence is the ratio of the number of non-executive directors to the overall number of di- • rectors in a given firm-year, and is constructed using information from proxy filings extracted from CompactDisclosureforthe1986-2005period(asinLinck,NetterandYang(2008)),andusingCapitalIQ forthe2006-2011period. AdditionalStateControls: StateGDPgrowthistherealannualgrowthrateingrossstateproduct(GSP),fromBEAdata. • StateUnemploymentRateisthestateunemploymentrate,fromBLSdata. • VoteShareofRepublicanPresidentialCandidateisthestate’sshareofthevotecastfortheRepublican • candidateinthelastpresidentialelection,fromthedataoftheAmericanPresidencyProjectatUCSanta Barbara. Union Membership is the fraction of private-sector employees who belong to a labor union in a state- • year,fromtheHirschandMacpherson(2003)database. ChangeinCorporateTaxesisacategoricalvariablethatequalsoneifastateincreasesitscorporatetax • rate in a given year, and minus one if a state cuts its corporate tax rate in a given year. Data is from AppendixAandAppendixBofHeiderandLjungqvist(2012). ChangeinPersonalIncomeTaxesisacategoricalvariablethatequalsoneifastateincreasesitspersonal • incometaxrateinagivenyear,andminusoneifastatecutsitspersonalincometaxrateinagivenyear. The personal income tax rate is the maximum state tax rate on wage income, from the NBER Taxsim data. StateInvestmentTaxCreditIntroductionsisadummyvariablethattakesthevalueofoneifastatethat • neverhadaninvestmenttaxcreditintroducesitinagivenyear,andisfromthedatabasehand-collected byChirinkoandWilson(2008)andthereindetailed. State Investment Tax Credit Increases (Cuts) is a dummy variable that takes the value of one if a state • increases (cuts) its investment tax credit in a given year, and is from the data base hand-collected by ChirinkoandWilson(2008)andthereindetailed. 36

Appendix B. List of Changes in State Tax Credits This table details all the 73 state R&D tax credits’ changes (35 introductions and 38 subsequent changes, of which 24 are increases and 14 are cuts) between tax years 1987-2011 used in the DID analysis. We handcollectedthisinformationforallU.S.statesusingavarietyofsources. TheStateTaxHandbook(variousyears) andWilson(2009)providedthestartingpointtogetasummarylistofR&Dtaxcreditsavailableinanygiven state at a particular point in time. We then cross-checked this information by verifying the passage of state tax bills and obtaining information on the year when a given tax credit becomes effective through keyword searches in Lexis-Nexis (State Net) and online searches of state code legislation and tax forms on the states’ DepartmentofRevenuewebsites. Year State Description Effective 1987 Minnesota CutsR&Dtaxcreditfrom6.25%to2.5% 1988 California Introducesataxcreditof8%ofqualifiedincrementalR&Dexpenditures 1988 Kansas Introducesataxcreditof6.5%ofqualifiedincrementalR&Dexpenditures 1988 NorthDakota Introducesataxcreditof4%ofqualifiedincrementalR&Dexpenditures 1989 Colorado Introducesataxcreditof3%ofqualifiedincrementalR&Dexpenditures 1989 Oregon Introducesataxcreditof5%ofqualifiedincrementalR&Dexpenditures 1990 Illinois Introducesataxcreditof6.5%ofqualifiedincrementalR&Dexpenditures 1991 Massachusetts Introducesataxcreditof10%ofqualifiedincrementalR&Dexpenditures 1993 Arizona Introducesataxcreditof20%ofqualifiedincrementalR&Dexpenditures 1993 Connecticut Introducesataxcreditof6%ofqualifiedincrementalR&Dexpenditures 1993 NewHampshire Introducesataxcreditof7.5%ofqualifiedincrementalR&Dexpenditures 1994 Missouri Introducesataxcreditof6.5%ofqualifiedincrementalR&Dexpenditures 1994 NewHampshire IncreasesR&Dtaxcreditfrom7.5%to15% 1994 NewJersey Introducesataxcreditof10%ofqualifiedincrementalR&Dexpenditures 1994 RhodeIsland Introducesataxcreditof5%ofqualifiedincrementalR&Dexpenditures 1995 NewHampshire LetsR&Dtaxcreditexpire 1995 Washington Introducesataxcreditof7.5%ofqualifiedincrementalR&Dexpenditures 1996 Maine Introducesataxcreditof5%ofqualifiedincrementalR&Dexpenditures 1996 NorthCarolina Introducesataxcreditof5%ofqualifiedincrementalR&Dexpenditures 1997 California IncreasesR&Dtaxcreditfrom8%to11% 1997 Pennsylvania Introducesataxcreditof10%ofqualifiedincrementalR&Dexpenditures 1998 Georgia Introducesataxcreditof10%ofqualifiedincrementalR&Dexpenditures 1998 RhodeIsland IncreasesR&Dtaxcreditfrom5%to16.9% 1998 Vermont Introducesataxcreditof10%ofqualifiedincrementalR&Dexpenditures 1999 California IncreasesR&Dtaxcreditfrom11%to12% 1999 Montana Introducesataxcreditof5%ofqualifiedincrementalR&Dexpenditures 1999 Utah Introducesataxcreditof6%ofqualifiedincrementalR&Dexpenditures 2000 California IncreasesR&Dtaxcreditfrom12%to15% 2000 Delaware Introducesataxcreditof10%ofqualifiedincrementalR&Dexpenditures 2000 Hawaii Introducesataxcreditof20%ofqualifiedincrementalR&Dexpenditures 2000 Maryland Introducesataxcreditof10%ofqualifiedincrementalR&Dexpenditures 2000 NewMexico Introducesataxcreditof4%ofqualifiedincrementalR&Dexpenditures 2001 Arizona CutsR&Dtaxcreditfrom20%to11% 2001 Idaho Introducesataxcreditof5%ofqualifiedincrementalR&Dexpenditures 2001 Ohio Introducesataxcreditof7%ofqualifiedincrementalR&Dexpenditures 37

List of Changes in State Tax Credits (Continued) Year State Description Effective 2001 SouthCarolina Introducesataxcreditof2.5%ofqualifiedincrementalR&Dexpenditures 2001 Texas Introducesataxcreditof4%ofqualifiedincrementalR&Dexpenditures 2002 Kentucky Introducesataxcreditof5%ofqualifiedincrementalR&Dexpenditures 2002 SouthCarolina IncreasesR&Dtaxcreditfrom2.5%to5% 2002 Texas IncreasesR&Dtaxcreditfrom4%to5% 2003 Illinois LetsR&Dtaxcreditexpire 2003 Indiana IncreasesR&Dtaxcreditfrom5%to10% 2003 Louisiana Introducesataxcreditof8%ofqualifiedincrementalR&Dexpenditures 2003 WestVirginia ExtendsR&Dtaxcreditof10%fromonly5high-techsectorstoallsectors 2003 Arkansas Introducesataxcreditof20%ofqualifiedincrementalR&Dexpenditures 2004 Illinois Re-introducesataxcreditof6.5%ofqualifiedincrementalR&Dexpenditures 2005 Indiana IncreasesR&Dtaxcreditmiddle&lowbracketsto15%and12.5% 2005 Missouri LetsR&Dtaxcreditexpire 2006 Indiana CutsR&Dtaxcreditmiddle&lowbracketsto10% 2006 Nebraska Introducesataxcreditof3%ofqualifiedincrementalR&Dexpenditures 2006 NorthCarolina LetsR&Dtaxcreditexpire 2006 Oregon IncreasesR&Dtaxcreditcapfrom$0.5Mto$2M 2007 NewHampshire Re-introducesataxcreditof10%ofqualifiedincrementalR&Dexpenditures 2007 NorthCarolina Re-introducesataxcreditof3.25%ofqualifiedincrementalR&Dexpenditures 2007 NorthDakota IncreasesR&Dtaxcreditfrom4%to20% 2007 Vermont LetsR&Dtaxcreditexpire 2007 Washington IncreasesR&Dtaxcreditfrom7.5%to15% 2008 Indiana IncreasesR&Dtaxcreditfrom10%to15% 2008 Michigan Introducesataxcreditof2%ofqualifiedincrementalR&Dexpenditures 2008 Texas LetsR&Dtaxcreditexpire 2008 Utah CutsexistingR&Dtaxcreditto5%,butintroducesanadditional newR&DtaxcreditonallqualifiedR&Dexpenditures(ambiguous) 2009 Utah IncreasesR&Dtaxcreditfrom5%to6% 2010 Arizona IncreasesR&Dtaxcreditfrom11%to13% 2010 Minnesota IncreasesR&Dtaxcreditfrom6.25%to10% 2010 Utah IncreasesR&Dtaxcreditfrom6%to9% 2011 Arizona IncreasesR&Dtaxcreditfrom13%to15% 2011 Hawaii LetsR&Dtaxcreditexpire 2011 Michigan LetsR&Dtaxcreditexpire 2011 Montana LetsR&Dtaxcreditexpire 2011 NewYork Introducesataxcreditof3%ofqualifiedincrementalR&Dexpenditures, buteligibilityisconditionalonjobcreationcovenants 2011 NorthDakota CutsR&Dtaxcreditfrom20%to8% 2011 Utah LetsR&Dtaxcreditexpire 2011 Vermont Re-introducesataxcreditof6%ofqualifiedincrementalR&Dexpenditures 2011 Wisconsin Introducesadditionalsupercreditof100%ofthequalifiedR&Dexpenditures thatexceed1.25timestheaverageofR&Dexpendituresinthe3prioryears 38

Table1: SummaryStatistics ThestartingsampleconsistsofallUSnonfinancialfirms(excludingUtilities)inCompustatfrom1986to2011, whichyieldsapanelof124,504observationsfor11,091uniquefirms. Thesampleofinnovativefirmsconsists of the sub-sample of all US nonfinancial firms (excluding Utilities) in Compustat that report non-zero R&D expenditures inat least one yearover the same period, which yields apanel of 72,587 observationsfor 6,058 uniquefirms. Thetablereportssummarystatisticsofthedependentvariablesandmainexplanatoryvariables. Themaindependentvariableistheratioofcashholdingstobookassets. Cashholdingsarethesumofcash and short-term marketable securities. We also present results for variables related to liquidity structure (the ratiosofbankandunusedbankliquiditytototalliquidity),whichareconstructedusingdatafromCapitalIQ (available for the 2002-2011 period, 23,086 (14,504) observations for 2,866 (1,798) unique (innovative) firms). DetaileddefinitionsforallvariablesareprovidedinAppendixA.ThedetailedlistofchangesinstateR&Dtax creditsisinAppendixB. CompustatSample InnovativeFirmsSample CorporateLiquidityVariables: Mean Median StdDev Mean Median StdDev Cash/BookAssets 0.18 0.08 0.22 0.24 0.14 0.25 Cash/NetBookAssets 0.29 0.08 0.50 0.39 0.14 0.54 Cash/MarketValueofAssets 0.14 0.07 0.20 0.16 0.09 0.21 DollarCash($M,1990real) 52.0 5.4 169.7 59.3 6.2 186.6 BankLiquidity/TotalLiquidity 0.29 0.02 0.37 0.23 0 0.34 UnusedBankLiquidity/TotalLiquidity 0.08 0 0.22 0.07 0 0.21 Innovation&StateR&DTaxCredits: R&DExpenditures/BookAssets 0.01 0 0.14 0.03 0.00 0.20 R&DCapital/BookAssets 0.15 0.01 0.28 0.30 0.19 0.33 StatesAdoptingR&DTaxCredit(%obs) 3.3% StatesIncreasingR&DTaxCredit(%obs) 5.1% StatesCuttingR&DTaxCredit(%obs) 0.8% BaselineFirmControls: CashFlowVolatility 0.37 0.37 0.11 0.39 0.41 0.10 Market-to-Book 1.96 1.47 1.44 2.20 1.63 1.63 Size(log(BookAssets),1990real) 4.78 4.64 2.15 4.49 4.27 2.28 CashFlow/BookAssets 0.04 0.06 0.14 0.02 0.06 0.16 Capex/BookAssets 0.06 0.04 0.07 0.05 0.04 0.06 Dividenddummy 0.25 0 0.43 0.23 0 0.42 Acquisitions/BookAssets 0.02 0 0.06 0.02 0 0.06 Channels&AdditionalControls(selected): NetLeverage 0.09 0.11 0.43 -0.01 0.01 0.46 BankDebt/TotalDebt 0.39 0.22 0.40 0.35 0.12 0.41 SecuredDebt/TotalDebt 0.45 0.33 0.43 0.42 0.22 0.44 DebtRatedDummy 0.19 0 0.39 0.16 0 0.36 NetDebtIssuance/BookAssets 0.02 0 0.17 0.01 0 0.16 NetEquityIssuance/BookAssets 0.05 0 0.16 0.06 0 0.18 Cash/BookEquity 0.35 0.21 0.39 0.41 0.28 0.42 FirmAge(years) 15 10 15 14 9 14 ForeignTaxBurden/BookAssets 0.001 0 0.001 0.001 0 0.001 ForeignIncomeDummy 0.30 0 0.45 0.38 0 0.48 IndustryCompetition(HHI) 0.07 0.05 0.06 0.05 0.04 0.05 HighTechIndustryDummy 0.25 0 0.43 0.38 0 0.48 BoardSize 7.45 7 2.83 7.31 7 2.79 ClassifiedBoardDummy 0.57 1 0.49 0.57 1 0.50 BoardIndependence(%OutsideDirectors) 0.65 0.66 0.20 0.66 0.71 0.20 39

smriFevitavonnIfoerutcurtSytidiuqiLdnasgnidloHhsaCetaroproCnoecnedivEevitpircseD :2elbaT evitavonnifoelpmasehtelihw,)AlenaP(1102ot6891morftatsupmoCni)seitilitUgnidulcxe(smrfilaicnanfinonSUllafostsisnocelpmastatsupmoCehT raeyenotsaeltaniserutidnepxeD&Rorez-nontropertahttatsupmoCni)seitilitUgnidulcxe(smrfilaicnanfinonSUllafoelpmas-busehtfostsisnocsmrfi gniylrednu eht fo noitaived dradnats eht yb delacs ,setamitse retemarap stroper elbat eht ,selpmas owt eseht fo hcae roF .)B lenaP( doirep emas eht revo ,noitavonni fo serusaem level-mrfi owt no )01-5 snmuloC( erutcurts ytidiuqil dna )4-1 snmuloC( sgnidloh hsac fo snoisserger SLO lenap morf ,elbairav ,wofl hsac ,ezis mrfi ,oitar koob-ot-tekram ,ytilitalov wofl hsac( hsac fo stnanimreted dradnats rof gnillortnoc elihw ,latipac D&R dna serutidnepxe D&R ,dedulcni osla era ,raey nevig yna ni dnedivid syap mrfi eht rehtehw rof ymmud a dna ,serutidnepxe noitisiuqca ,xepac ;detroper era stneicfifeoc esohw ahtiwdetaicossaelbairavtnednepedehtniegnahcehtsitneicfifeocdetroperhcaefonoitaterpretniehT .)ytiverbrofdettimoerasetamitsetneicfifeoctub niesaercni%4.5ahtiwdetaicossasiserutidnepxeD&Rniesaercninoitaiveddradnats-enoa,elpmaxeroF.tnanimretedehtniegnahcnoitaiveddradnats-eno hcnerF-amaF-84dnaraeY .dohtemyrotnevnilauteprepehtgnisudetcurtsnocserutidnepxeD&RtsapfokcotsehtsilatipacD&R .)1nmuloC(sgnidlohhsac ,%1ehttaecnacfiingisgnitoned*dna,**,***htiw,levelmrfiehttaderetsulcera)sesehtnerapni(seulav-p .snoissergerllanidedulcnieraseimmudyrtsudni .AxidneppAnidedivorperasnoitinfiedelbairaV .ylevitcepser,level%01dna,%5 elpmaStatsupmoC:AlenaP erutcurtSytidiuqiLetaroproC sgnidloHhsaCetaroproC ytidiuqiLknaBdesunU otytidiuqiLknaB fotiborP /hsaC /hsaC ytidiuqiLlatoTot ytidiuqiLlatoT ytidiuqiLknaB stessAkooBteN stessAkooB )01( )9( )8( )7( )6( )5( )4( )3( )2( )1( ***600.0- ***120.0- ***591.0- ***530.0 ***450.0 serutidnepxED&R )200.0( )400.0( )070.0( )400.0( )300.0( ***310.0- ***440.0- ***681.0- ***470.0 ***080.0 latipaCD&R )200.0( )400.0( )330.0( )400.0( )300.0( 200.0- 200.0- **700.0- **700.0- ***450.0- ***460.0- ***720.0 ***720.0 ***610.0 ***710.0 ytilitaloVwolFhsaC )200.0( )200.0( )300.0( )300.0( )110.0( )110.0( )200.0( )200.0( )100.0( )100.0( ***800.0- ***600.0- ***830.0- ***230.0- ***110.0- ***900.0- ***350.0 ***640.0 ***720.0 ***420.0 kooB-ot-tekraM )200.0( )100.0( )400.0( )400.0( )300.0( )300.0( )300.0( )300.0( )200.0( )100.0( 100.0 *200.0 300.0- 000.0 ***420.0 ***920.0 ***220.0- ***720.0- ***710.0- ***120.0eziSmriF )100.0( )100.0( )200.0( )200.0( )400.0( )300.0( )300.0( )300.0( )200.0( )200.0( ***500.0 ***500.0 200.0 400.0 100.0 810.0 **600.0 *400.0 ***400.0- 200.0stessAkooB/wolFhsaC )200.0( )200.0( )400.0( )400.0( )910.0( )610.0( )200.0( )200.0( )100.0( )100.0( seY seY seY seY seY seY seY seY seY seY stceffEdexiFyrtsudnI&raeY 640.0 340.0 141.0 721.0 481.0 551.0 472.0 152.0 2RdetsujdA elpmaSsmriFevitavonnI:BlenaP ***500.0- ***320.0- **451.0- ***330.0 ***250.0 serutidnepxED&R )200.0( )400.0( )360.0( )400.0( )300.0( ***010.0- ***830.0- ***041.0- ***360.0 ***270.0 latipaCD&R )200.0( )400.0( )920.0( )400.0( )300.0( seY seY seY seY seY seY seY seY seY seY stceffEdexiFyrtsudnI&raeY 240.0 540.0 721.0 901.0 771.0 151.0 372.0 152.0 2RdetsujdA 40

stiderCxaTD&RetatSnisegnahCfonoitadilaV :3elbaT morfsetamitseretemarapdna,)AlenaP(tidercxatD&Rstisegnahcetatsatahtdoohilekilehtfosnoissergerytilibaborp-raenilfostluserstroperelbatehT foeulavsekattahtymmudasielbairavtnednepedeht,AlenaPnI.)BlenaP(stidercxatD&RetatsnisegnahcnonoitavonninisegnahcfosnoissergerDID ,)1 nmuloC( sesaercni tneuqesbus ro noitcudortni :nrut ni segnahc fo sepyt tnereffid htiw ,tiderc xat D&R sti segnahc etats a nehw raey nevig yna ni eno rehtosedulcnistnanimreteD .)4nmuloC(noitcudortniottneuqesbusstucdna ,)3nmuloC(noitcudortniottneuqesbussesaercni ,)2nmuloC(noitcudortni deretsulcera)sesehtnerapni(seulav-p .snoissergerllanidedulcnieraseimmudetatsdnaraeY .hsacroftnaveleryllaitnetopebnactahtsegnahclevel-etats )seitilitUgnidulcxe(smrfilaicnanfinonSUllafoelpmas-busehtfostsisnochcihw,elpmassmrfievitavonniehtroferaBlenaPnisetamitsE .leveletatsehtta ,)1nmuloC(serutidnepxeD&Rnisegnahcrofdna,doirep1102ot6891ehtrevoraeyenotsaeltaniserutidnepxeD&Rorez-nontropertahttatsupmoCni egnahcehtsitneicfifeocdetroperhcaefonoitaterpretniehT .)4nmuloC(tnempiuqedna,stnalp,ytreporPdna,)3nmuloC(xepaC,)2nmuloC(latipacD&R %8.8 na htiw detaicossa era sesaercni tiderc xat D&R ,2 nmuloC ni ,elpmaxe roF .egnahc tiderc xat D&R na htiw detaicossa elbairav tnedneped eht ni dna,snoitauqeslevelehtnistceffedexfimrfievomerotsecnereffidtsrfinidetamitseerasnoitacfiicepsllA .stessakoobotevitalerlatipacD&Rniesaercni )sesehtnerapni(seulav-p .dedulcnieraseimmud)FF-84(yrtsudni-raeysallewsa)1(noitacfiicepsenilesabehtfoslortnoclevel-mrfiehtnisegnahcdeggal ehT.AxidneppAnierasnoitinfiedelbairaV .ylevitcepser,level%01dna,%5,%1ehttaecnacfiingisgnitoned*dna,**,***htiw,levelmrfiehttaderetsulcera .BxidneppAnisistidercxatD&Retatsnisegnahcfotsildeliated stiderCxaTD&RetatSnisegnahCfostnanimreteD:AlenaP tneuqesbuS tneuqesbuS noitcudortnI -buS&noitcudortnI stuC sesaercnI sesaercnItneuqes )4( )3( )2( )1( snoitidnoCelcyCssenisuBetatS 000.0 100.0 100.0 200.0 htworGPDGdeggaL )100.0( )300.0( )200.0( )300.0( 400.0 400.0 **710.0 **120.0 etaRtnemyolpmenUdeggaL )200.0( )300.0( )600.0( )800.0( snoitidnoClacitiloP 000.0 100.0 100.0 200.0 etadidnaClaitnediserPnacilbupeRfoerahSetoV )000.0( )100.0( )100.0( )100.0( rewoPnoinU 300.0 600.0- 400.0- 010.0pihsrebmeMnoinUdeggaL )300.0( )600.0( )400.0( )700.0( sexaTetatSrehtOnisegnahC 720.0- 300.0- 310.0 010.0 sexaTetaroproCniegnahCdeggaL )710.0( )210.0( )420.0( )620.0( 400.0 310.0- 620.0 310.0 sexaTemocnIlanosrePniegnahCdeggaL )400.0( )110.0( )120.0( )320.0( )553.0=p(41.1 )236.0=p(27.0 )050.0=p( ∗∗92.2 )380.0=p( ∗00.2 0=sffeocllA :tsetdlaW,scitsongaiD ]750.0[seY ]450.0[seY ]470.0[seY ]760.0[seY ]2R[stceffedexfiraeYdnaetatS sexaTD&RetatSnisegnahCfonoitadilaV:BlenaP EPP xepaC latipaCD&R serutidnepxED&R )4( )3( )2( )1( 000.0 100.0 ***880.0 **300.0 )%ni(0=ttaesaercnItiderCxaTD&Rfi1= )200.0( )100.0( )120.0( )200.0( 100.0 100.0 **470.0- **200.0- )%ni(0=ttatuCtiderCxaTD&Rfi1= )300.0( )300.0( )630.0( )100.0( seY seY seY seY stceffedexfiraeY-yrtsudnI,slortnoCmriFenilesaB 41

stiderCxaTD&RetatSnisegnahCfosisylanAenilesaB :sgnidloHhsaCetaroproCnonoitavonnIfotcapmIehT :4elbaT lareves rof stiderc xat D&R etats ni segnahc no stessa koob ot sgnidloh hsac ni segnahc fo snoisserger DID morf setamitse retemarap stroper elbat ehT llafoelpmas-busehtfostsisnochcihw,elpmassmrfievitavonniehtnidetamitseerasnoissergerllA .soitarhsacfosnoitinfieddnasnoitacfiicepstnereffid 785,27( doirep 1102 ot 6891 eht revo raey eno tsael ta ni serutidnepxe D&R orez-non troper taht tatsupmoC ni )seitilitU gnidulcxe( smrfi laicnanfinon SU na htiw detaicossa elbairav tnedneped eht ni egnahc eht si tneicfifeoc detroper hcae fo noitaterpretni ehT .)smrfi euqinu 850,6 rof snoitavresbo raey-mrfi .stessakoobotevitalerhsacniesaercni%1.2ahtiwdetaicossasistidercxatD&Retatsniesaercnina,nmuloctsrfiehtni,elpmaxeroF.egnahctidercxatD&R sgaldnasdaelhtiwnoitacfiicepsevisulcnieromanisallewsa ,)1(noitacfiicepsenilesabehtni ,soitartessakoobothsacrofstlusertroper3ot1snmuloC .ylevitcepser ,elbairavrotacidninafodaetsnidesusitidercxatD&Rehtniegnahcegatnecrepehterehwnoitacfiicepsanidna ,segnahctidercxatD&Rfo ,)5nmuloC(stessafoeulavtekramothsac ,)4nmuloC(stessakoobtenothsacedulcnihcihw ,soitarhsacfosnoitinfiedevitanretlarofera7ot4snmuloC otsecnereffidtsrfinidetamitseerasnoitacfiicepsllA .)7nmuloC(emocnitensulphsacdeggalothsacdna,)6nmuloC(stessakoobotemocnitensunimhsac )FF-84(yrtsudni-raeysallewsa)1(noitacfiicepsenilesabehtfoslortnoclevel-mrfiehtnisegnahcdeggaldna,snoitauqeslevelehtnistceffedexfimrfievomer ,level %01 dna ,%5 ,%1 eht ta ecnacfiingis gnitoned * dna ,** ,*** htiw ,level mrfi eht ta deretsulc era )sesehtnerap ni( seulav-p .dedulcni era seimmud .BxidneppAnisistidercxatD&RetatsnisegnahcfotsildeliatedehT.AxidneppAnierasnoitinfiedelbairaV .ylevitcepser soitaRhsaCfosnoitinfieDevitanretlA foeziS gnimiT enilesaB hsaCgaL(/hsaC /hsaCteN -raM/hsaC teN/hsaC xaTD&R /hsaC /hsaC )emocnIteN+ stessAkooB stessAtek stessAkooB tiderC stessAkooB stessAkooB )7( )6( )5( )4( )3( )2( )1( ***160.0 ***620.0 ***120.0 ***230.0 ***310.0 ***120.0 0=ttaesaercnItiderCxaTD&Rfi1= )710.0( )600.0( )500.0( )600.0( )500.0( )300.0( *900.0 1-=ttaesaercnItiderCxaTD&Rfi1= )500.0( 200.0 1+=ttaesaercnItiderCxaTD&Rfi1= )400.0( ***613.0 0=ttaesaercnItiderCxaTD&R )390.0( **750.0- **610.0- ***610.0- ***920.0- **510.0- ***710.0- 0=ttatuCtiderCxaTD&Rfi1= )720.0( )800.0( )600.0( )700.0( )700.0( )600.0( 110.0- 1-=ttatuCtiderCxaTD&Rfi1= )800.0( 600.0- 1+=ttatuCtiderCxaTD&Rfi1= )900.0( ***912.0- 0=ttatuCtiderCxaTD&R )580.0( seY seY seY seY seY seY seY slortnoCmriFenilesaB seY seY seY seY seY seY seY stceffedexfiraeY-yrtsudnI 191.0 071.0 681.0 931.0 251.0 542.0 861.0 2RdetsujdA 42

stseTnoitacfiislaF :sgnidloHhsaCetaroproCnonoitavonnIfotcapmIehT :5elbaT tnereffidowtrofstidercxatD&RetatsnisegnahcnostessakoobotsgnidlohhsacnisegnahcfosnoissergerDIDmorfsetamitseretemarapstroperelbatehT rof dna )4 ot 1 snmuloC( ssenevitceffe rieht tceffa taht stiderc xat D&R eht fo serutaef lanoitutitsni eht rof seixorp no desab atad eht fo stilps elpmas-bus hcihw,elpmassmrfievitavonniehtnidetamitseerasnoissergerllA .)6ot5snmuloC(stidercxattnemtsevnietatsnisegnahcedulcnioslatahtsnoitacfiiceps revoraeyenotsaeltaniserutidnepxeD&Rorez-nontropertahttatsupmoCni)seitilitUgnidulcxe(smrfilaicnanfinonSUllafoelpmas-busehtfostsisnoc tnednepedehtniegnahcehtsitneicfifeocdetroperhcaefonoitaterpretniehT .)smrfieuqinu850,6rofsnoitavresboraey-mrfi785,27(doirep1102ot6891eht esaercni%2.2ahtiwdetaicossasistidercxatD&Retatsniesaercnina,nmuloctsrfiehtni,elpmaxeroF .tidercxatD&Rniegnahcahtiwdetaicossaelbairav xatetaroprocetats:seixorpetna-xegniwollofehtfoseulav)roirp-raey(naidemwoleb .svevobaneewtebtilpssielpmasehT .stessakoobotevitalerhsacni xat tnemtsevni ni segnahc ehT .)4 dna 3 snmuloC( level tnemtaert-erp rieht ot tcepser htiw serutidnepxe D&R mrfi ni htworg dna ,)2 ot 1 snmuloC( etar .deliatednierehtdna)8002(nosliWdnaoknirihCybdetcelloc-dnahstidercxattnemtsevnietatsnistuc8dna,sesaercni72,stnevenoitcudortni81erastiderc ehtfoslortnoclevel-mrfiehtnisegnahcdeggaldna,snoitauqeslevelehtnistceffedexfimrfievomerotsecnereffidtsrfinidetamitseerasnoitacfiicepsllA *dna,**,***htiw,levelmrfiehttaderetsulcera)sesehtnerapni(seulav-p .dedulcnieraseimmud)FF-84(yrtsudni-raeysallewsa)1(noitacfiicepsenilesab stidercxatD&RetatsnisegnahcfotsildeliatedehT.AxidneppAnierasnoitinfiedelbairaV .ylevitcepser,level%01dna,%5,%1ehttaecnacfiingisgnitoned .BxidneppAnisi nisegnahCotesnopseR sihtworGD&R sietaRxaTetaroproCetatS stiderCxaTtnemtsevnIetatS woL hgiH woL hgiH )6( )5( )4( )3( )2( )1( ***020.0 500.0 ***920.0 400.0 ***220.0 0=ttaesaercnItiderCxaTD&Rfi1= )300.0( )700.0( )600.0( )210.0( )400.0( ***420.0- 800.0- **030.0- 700.0- ***120.0- 0=ttatuCtiderCxaTD&Rfi1= )700.0( )700.0( )510.0( )110.0( )600.0( 010.0- 100.0 0=ttatiderCxaTtnemtsevnIweNfi1= )010.0( )010.0( *610.0- **810.0- 0=ttaesaercnItiderCxaTtnemtsevnIfi1= )900.0( )900.0( **210.0 *900.0 0=ttatuCtiderCxaTtnemtsevnIfi1= )600.0( )500.0( seY seY seY seY seY seY slortnoCmriFenilesaB seY seY seY seY seY seY stceffedexfiraeY-yrtsudnI 161.0 421.0 221.0 091.0 601.0 131.0 2RdetsujdA 43

stseTssentsuboRdnaytidilaVlanretxE :sgnidloHhsaCetaroproCnonoitavonnIfotcapmIehT :6elbaT eerht rof stiderc xat D&R etats ni segnahc no stessa koob ot sgnidloh hsac ni segnahc fo snoisserger DID morf setamitse retemarap stroper elbat ehT ot ssentsubor sredisnoc A lenaP .)C lenaP( srotamitse evitanretla no desab setamitse retemarap dna )D dna ,B ,A slenaP( stset ssentsubor fo stes tnereffid rofgnillortnocotssentsuborsredisnocBlenaP .selbairavlortnoclevel-yrtsudnidna,-etats,-mrfilanoitiddagnidulcnidnasrorredradnatsevitanretlagnisu ytiunitnocsid noisserger cihpargoeg a edulcni hcihw ,srotamitse evitanretla gnisu ot ssentsubor sredisnoc C lenaP .level etats eht ta sdnuofnoc laitnetop )SLS-2( DID-VI na ,)01 woR( )8991( dnoB dna llednulB fo MMG-VI metsys eht ,)9 woR( seitnuoc redrob etats ni smrfi no desab rotamitse )DDR( ngised kcots D&R detnemurtsni htiw noisserger egats dnoces eht fo setamitse( kcots D&R eht rof stnemurtsni sa segnahc tiderc xat D&R sesu taht rotamitse no desab smrfi lortnoc dna detaert sehctam taht ))7991( ddoT dna ,arumihcI ,namkceH( rotamitse DID elpmas-dehctam a dna ,)11 woR ni detroper era gnidulcni fo tsisnoc hcihw ,stnemtaert fo stes tnereffid gniredisnoc yb ytidilav lanretxe senimaxe D lenaP .ecnamrofrep dna ,yrtsudni ,ezis tnemtaert-erp tidercxatD&Retatshcus)elitrauqpot(tsegralehtylnofonoitcudortniehtgniredisnocsallewsa,)51dna31swoR(testnemtaertehtnistnevenoitcudortni smrfi laicnanfinon SU lla fo elpmas-bus eht fo stsisnoc hcihw ,elpmas smrfi evitavonni eht ni detamitse era snoisserger llA .)61 dna 41 swoR( smargorp rofsnoitavresboraey-mrfi785,27(doirep1102ot6891ehtrevoraeyenotsaeltaniserutidnepxeD&Rorez-nontropertahttatsupmoCni)seitilitUgnidulcxe( roF .tidercxatD&RniegnahcahtiwdetaicossaelbairavtnednepedehtniegnahcehtsitneicfifeocdetroperhcaefonoitaterpretniehT .)smrfieuqinu850,6 detamitseerasnoitacfiicepsllA .stessakoobotevitalerhsacniesaercni%2ahtiwdetaicossasistidercxatD&Retatsniesaercnina,wortsrfiehtni,elpmaxe sallewsa)1(noitacfiicepsenilesabehtfoslortnoclevel-mrfiehtnisegnahcdeggaldna,snoitauqeslevelehtnistceffedexfimrfievomerotsecnereffidtsrfini gnitoned*dna,**,***htiw,esiwrehtodetacidnisselnu,levelmrfiehttaderetsulcera)sesehtnerapni(seulav-p .dedulcnieraseimmud)FF-84(yrtsudni-raey ni si stiderc xat D&R etats ni segnahc fo tsil deliated ehT .A xidneppA ni era snoitinfied elbairaV .ylevitcepser ,level %01 dna ,%5 ,%1 eht ta ecnacfiingis .BxidneppA tneicfifeoCdetamitsE tseTssentsuboR tneicfifeoCdetamitsE tseTssentsuboR 0=ttatiderCxaTD&Rfi 0=ttatiderCxaTD&Rfi stuC sesaercnI stuC sesaercnI )2( )1( )2( )1( srotamitsEevitanretlA&spuorGlortnoCreniF:ClenaP slortnoCdnasrorrEdradnatSevitanretlA:AlenaP ∗∗710.0- ∗∗∗910.0 tnecajdagnisuDDRcihpargoeG ]9[ **810.0- ***020.0 roflortnoc,etatsybgniretsulC ]1[ )800.0( )300.0( puorglortnocsaseitnuoc )800.0( )500.0( stceffedexfietats ∗∗∗810.0- ∗∗∗020.0 MMG-VI ]01[ ***510.0- ***910.0 noitairtaperroflortnoC ]2[ )400.0( )500.0( )600.0( )300.0( ∗∗∗860.0 sikcotSD&R,)SLS-2(DID-VI ]11[ ***510.0- ***910.0 etarxatlanigramroflortnoC ]3[ )410.0( stidercxatD&Rgnisudetnemurtsni )600.0( )300.0( **610.0- ***810.0 &yrtsudni,ezis:DIDelpmas-dehctaM ]21[ ***510.0- ***910.0 hcet-hgih&noititepmocroflortnoC ]4[ )700.0( )400.0( ecnamrofreproirp )600.0( )300.0( tiderCweN stuC sesaercnIro ytidilaVlanretxE:DlenaP sdnuofnoClaitnetoProfgnillortnoC:BlenaP ***710.0- ***210.0 llA,SLO ]31[ ***910.0- ***310.0 elcycssenisubetatsroflortnoC ]5[ )600.0( )200.0( )600.0( )300.0( ***710.0- ***120.0 stidercxatwenelitrauqpoT,SLO ]41[ **430.0- ***120.0 snoidnoclacitilopetatsroflortnoC ]6[ )600.0( )300.0( )510.0( )600.0( rewopnoinu& ***820.0- ***810.0 &yrtsudni,ezis:DIDelpmas-dehctaM ]51[ ***610.0- ***120.0 etaroprocetatsroflortnoC ]7[ )700.0( )300.0( CBetats )600.0( )300.0( sexatemocni ***820.0- ***920.0 &yrtsudni,ezis:DIDelpmas-dehctaM ]61[ ***710.0- ***120.0 lanosrepetatsroflortnoC ]8[ )700.0( )400.0( stidercxatwenelitrauqpoT,CBetats )600.0( )300.0( sexatemocni 44

sevitoMyranoituacerP&snoitcirFlaicnaniFfoeloRehtgnissessA :sgnidloHhsaCetaroproCnonoitavonnIfotcapmIehT :7elbaT lareves rof stiderc xat D&R etats ni segnahc no stessa koob ot sgnidloh hsac ni segnahc fo snoisserger DID morf setamitse retemarap stroper elbat ehT foytisnetniehtrofsallewsa,)AlenaP(smrfiybdecafsnoitcirflaicnanfifoytirevesehtrofseixorpetna-xenodesabatadehtfostilpselpmas-bustnereffid SU lla fo elpmas-bus eht fo stsisnoc hcihw ,elpmas smrfi evitavonni eht ni detamitse era snoisserger llA .)B lenaP( hsac dloh ot evitom yranoituacerp eht -mrfi785,27(doirep1102ot6891ehtrevoraeyenotsaeltaniserutidnepxeD&Rorez-nontropertahttatsupmoCni)seitilitUgnidulcxe(smrfilaicnanfinon egnahcahtiwdetaicossaelbairavtnednepedehtniegnahcehtsitneicfifeocdetroperhcaefonoitaterpretniehT .)smrfieuqinu850,6rofsnoitavresboraey .stessakoobotevitalerhsacniesaercni%8.1ahtiwdetaicossasistidercxatD&Retatsniesaercnina ,nmuloctsrfiehtni ,elpmaxeroF .tidercxatD&Rni )7991( selagniZ dna nalpaK ,)1 nmuloC( ezis mrfi :selbairav gniwollof eht fo seulav )tnemtaert-erp( naidem woleb .sv evoba neewteb tilps si elpmas ehT nmuloC(oitarkoob-ot-tekrammrfi,)6nmuloC(egamrfi,)5nmuloC(swoflhsacmrfi,)4nmuloC(xednI-WW)6002(uWdnadetihW,)3nmuloC(xednI-ZK dexfimrfievomerotsecnereffidtsrfinidetamitseerasnoitacfiicepsllA .)2nmuloC(sutatsreyapdnedividybdna,)8nmuloC(ytilitalovwoflhsacmrfi,)7 era seimmud )FF-84(yrtsudni-raey sa llew sa )1( noitacfiiceps enilesab eht fo slortnoc level-mrfi eht ni segnahc deggal dna ,snoitauqe slevel eht ni stceffe elbairaV.ylevitcepser,level%01dna,%5,%1ehttaecnacfiingisgnitoned*dna,**,***htiw,levelmrfiehttaderetsulcera)sesehtnerapni(seulav-p.dedulcni .BxidneppAnisistidercxatD&RetatsnisegnahcfotsildeliatedehT.AxidneppAnierasnoitinfied seixorPevitoMyranoituacerP:BlenaP seixorPsnoitcirFlaicnaniF:AlenaP wolFhsaC B/M mriF hsaC WW ZK dnediviD mriF ytilitaloV oitaR egA swolF xednI xednI reyaP eziS )8( )7( )6( )5( )4( )3( )2( )1( yksiR hgiH gnuoY woL deniartsnoC ∗∗∗910.0 ∗∗220.0 ∗∗030.0 ∗∗120.0 ∗∗120.0 ∗∗∗220.0 ∗∗∗810.0 ∗∗810.0 0=ttaesaercnItiderCxaTD&Rfi1= )400.0( )900.0( )410.0( )110.0( )010.0( )600.0( )124.0( )900.0( ∗∗∗510.0- ∗∗∗910.0- ∗∗∗620.0- ∗∗∗910.0- ∗∗∗910.0- ∗∗∗120.0- ∗∗310.0- ∗∗∗510.0- 0=ttatuCtiderCxaTD&Rfi1= )600.0( )600.0( )600.0( )600.0( )500.0( )700.0( )700.0( )500.0( seY seY seY seY seY seY seY seY slortnoCmriFenilesaB seY seY seY seY seY seY seY seY stceffedexfiraeY-yrtsudnI efaS woL dlO hgiH deniartsnocnU 500.0 010.0 ∗∗800.0 ∗800.0 500.0 ∗800.0 500.0 300.0 0=ttaesaercnItiderCxaTD&Rfi1= )800.0( )600.0( )400.0( )500.0( )400.0( )400.0( )500.0( )400.0( 400.0- 800.0- 700.0- 700.0- 100.0- 300.0- 800.0- 600.0- 0=ttatuCtiderCxaTD&Rfi1= )010.0( )700.0( )500.0( )600.0( )400.0( )700.0( )600.0( )600.0( seY seY seY seY seY seY seY seY slortnoCmriFenilesaB seY seY seY seY seY seY seY seY stceffedexfiraeY-yrtsudnI 45

snoitcirFgnicnaniFtbeDfoeloRehtgnissessA :sgnidloHhsaCetaroproCnonoitavonnIfotcapmIehT :8elbaT ,)A lenaP( gnicnanfi tbed dna erutcurts ytidiuqil etaroproc no stiderc xat D&R etats ni segnahc fo tcapmi eht fo setamitse retemarap stroper elbat ehT -bus tnereffid lareves rof stiderc xat D&R etats ni segnahc no stessa koob ot sgnidloh hsac ni segnahc fo snoisserger DID morf setamitse retemarap dna nidetamitseerasnoissergerllA.)BlenaP(smrfiybdecafsnoitcirfgnicnanfitbedehtfoytirevesehtrofseixorpetna-xenodesabatadehtfostilpselpmas D&R orez-non troper taht tatsupmoC ni )seitilitU gnidulcxe( smrfi laicnanfinon SU lla fo elpmas-bus eht fo stsisnoc hcihw ,elpmas smrfi evitavonni eht ni noitamrofni htiw( smrfi euqinu )897,1( 850,6 rof snoitavresbo raey-mrfi )405,41( 785,27( doirep 1102 ot 6891 eht revo raey eno tsael ta ni serutidnepxe ,elpmaxeroF .tidercxatD&RniegnahcahtiwdetaicossaelbairavtnednepedehtniegnahcehtsitneicfifeocdetroperhcaefonoitaterpretniehT.))QIlatipaC .ytidiuqillatototevitalerytidiuqilknabdesununiesaerced%5.1ahtiwdetaicossasistidercxatD&Retatsniesaercnina,AlenaPfonmulocdnocesehtni ni segnahc fo setamitse DID ,)1 nmuloC( ytidiuqil knab gnivah fo doohilekil eht fo snoisserger tiborp morf setamitse troper ew A lenaP ni ,yllacfiicepS tiborp morf setamitse dna ,)4 nmuloC( egarevel koob ten ni dna ,)3 nmuloC( ecnaussi tbed ten ni ,)2 nmuloC( ytidiuqil latot ot ytidiuqil knab desunu elpmaseht,BlenaPnI .)6nmuloC(tbedderucesnidna)5nmuloC(tbedknabnidezilaicepssitahterutcurtstbedagnivahfodoohilekilehtfosnoisserger dnanainamarbusalaBdna,)2dna1snmuloC(eulavnoitadiuqiltessa)6991( .lateregreBfoseulav)tnemtaert-erp(naidemwoleb .svevobaneewtebtilpssi 5snmuloC(gnitartbedmret-gnolasahmrfiehtrehtehwnodesabsallewsa ,)4dna3snmuloC(ytilibayolpedertessayrtsudnifoxedni)9002(nasadaviS slortnoclevel-mrfiehtnisegnahcdeggaldna,snoitauqeslevelehtnistceffedexfimrfievomerotsecnereffidtsrfinidetamitseerasnoitacfiicepsllA .)6dna ,**,***htiw,levelmrfiehttaderetsulcera)sesehtnerapni(seulav-p .dedulcnieraseimmud)FF-84(yrtsudni-raeysallewsa)1(noitacfiicepsenilesabehtfo xatD&RetatsnisegnahcfotsildeliatedehT.AxidneppAnierasnoitinfiedelbairaV .ylevitcepser,level%01dna,%5,%1ehttaecnacfiingisgnitoned*dna .BxidneppAnisistiderc gnicnaniFtbeDdna,egareveLteN,erutcurtSytidiuqiLnonoitavonnIfotcapmIehT:AlenaP tiborP tiborP kooBteN tbeDteN knaBdesunU tiborP 09deruceS 09knaB egareveL ecnaussI otytidiuqiL knaB ytidiuqiLlatoT ytidiuqiL )6( )5( )4( )3( )2( )1( ∗∗∗820.0- ∗∗120.0- ∗∗∗420.0- ∗∗∗610.0- ∗∗∗510.0- ∗∗∗600.0- 0=ttaesaercnItiderCxaTD&Rfi1= )110.0( )700.0( )600.0( )600.0( )600.0( )200.0( 810.0 910.0 ∗∗220.0 410.0 ∗∗410.0 ∗∗110.0 0=ttatuCtiderCxaTD&Rfi1= )310.0( )910.0( )110.0( )110.0( )700.0( )700.0( seY seY seY seY seY seY slortnoCmriFenilesaB seY seY seY seY seY seY stceffedexfiraeY-yrtsudnI snoitcirFtbeDfoseixorPybsgnidloHhsaCetaroproCnonoitavonnIfotcapmIehT:BlenaP gnitaRtbeD ytilibayolpedeRtessA eulaVelaseR seY oN hgiH woL hgiH woL )6( )5( )4( )3( )2( )1( 900.0 ∗∗∗520.0 ∗210.0 ∗∗∗910.0 400.0 ∗∗810.0- 0=ttaesaercnItiderCxaTD&Rfi1= )600.0( )900.0( )700.0( )400.0( )300.0( )900.0( 300.0- ∗∗∗120.0- 500.0- ∗∗∗710.0- 200.0- ∗∗∗610.0 0=ttatuCtiderCxaTD&Rfi1= )600.0( )400.0( )010.0( )500.0( )400.0( )400.0( seY seY seY seY seY seY slortnoCmriFenilesaB seY seY seY seY seY seY stceffedexfiraeY-yrtsudnI 46

snoitcirFgnicnaniFytiuqEfoeloRehtgnissessA :sgnidloHhsaCetaroproCnonoitavonnIfotcapmIehT :9elbaT DIDmorfsetamitseretemarapdna,)AlenaP(gnicnanfiytiuqenostidercxatD&RetatsnisegnahcfotcapmiehtfosetamitseretemarapstroperelbatehT etna-xenodesabatadehtfostilpselpmas-bustnereffidlarevesrofstidercxatD&Retatsnisegnahcnostessakoobotsgnidlohhsacnisegnahcfosnoisserger stsisnochcihw,elpmassmrfievitavonniehtnidetamitseerasnoissergerllA.)BlenaP(smrfiybdecafsnoitcirfgnicnanfiytiuqeehtfoytirevesehtrofseixorp 6891ehtrevoraeyenotsaeltaniserutidnepxeD&Rorez-nontropertahttatsupmoCni)seitilitUgnidulcxe(smrfilaicnanfinonSUllafoelpmas-busehtfo elbairavtnednepedehtniegnahcehtsitneicfifeocdetroperhcaefonoitaterpretniehT .)smrfieuqinu850,6rofsnoitavresboraey-mrfi785,27(doirep1102ot %6.3 a htiw detaicossa si stiderc xat D&R etats ni esaercni na ,A lenaP fo nmuloc driht eht ni ,elpmaxe roF .tiderc xat D&R ni egnahc a htiw detaicossa koob ot hsac dna )1 nmuloC( ecnaussi ytiuqe ten ni segnahc fo setamitse DID troper ew A lenaP ni ,yllacfiicepS .ytiuqe koob ot evitaler hsac ni esaercni tilpssielpmaseht,BlenaPnI .)2nmuloC()OES(eussiytiuqeyradnocesafodoohilekilehtfosnoissergertiborpmorfsetamitsedna,)3nmuloC(oitarytiuqe dna1snmuloC(raeynevigynaniyrtsudni)FF-84(s’mrfianisOESfoemulovrallodlatotehtfoseulav)tnemtaert-erp(naidemwoleb .svevobaneewteb ehtdna,)4dna3snmuloC(raeynevigynaniyrtsudni)FF-84(s’mrfianistnemecnuonnasOESfo)RAC(nruterlamronbaevitalumuc)1+,0(egarevaeht,)2 mrfievomerotsecnereffidtsrfinidetamitseerasnoitacfiicepsllA .)6dna5snmuloC(SEBImorfstsacerofSPE’stsylanafo)noisrepsid(noitaiveddradnats seimmud)FF-84(yrtsudni-raeysallewsa)1(noitacfiicepsenilesabehtfoslortnoclevel-mrfiehtnisegnahcdeggaldna,snoitauqeslevelehtnistceffedexfi .ylevitcepser ,level %01 dna ,%5 ,%1 eht ta ecnacfiingis gnitoned * dna ,** ,*** htiw ,level mrfi eht ta deretsulc era )sesehtnerap ni( seulav-p .dedulcni era .BxidneppAnisistidercxatD&RetatsnisegnahcfotsildeliatedehT.AxidneppAnierasnoitinfiedelbairaV gnicnaniFytiuqEnonoitavonnIfotcapmIehT:AlenaP ytiuqEkooB/hsaC ecnaussIytiuqEtiborP ecnaussIytiuqEteN )3( )2( )1( ***630.0 ***140.0 ***320.0 0=ttaesaercnItiderCxaTD&Rfi1= )600.0( )700.0( )700.0( **230.0- *720.0- 510.0- 0=ttatuCtiderCxaTD&Rfi1= )610.0( )410.0( )110.0( seY seY seY slortnoCmriFenilesaB seY seY seY stceffedexfiraeY-yrtsudnI snoitcirFytiuqEfoseixorPybsgnidloHhsaCetaroproCnonoitavonnIfotcapmIehT:BlenaP noisrepsiDtsaceroFtsylanA stsoCOES emuloVOES hgiH woL hgiH woL hgiH woL )6( )5( )4( )3( )2( )1( **910.0 900.0 ***020.0 100.0 010.0 ***420.0 0=ttaesaercnItiderCxaTD&Rfi1= )800.0( )600.0( )500.0( )320.0( )610.0( )400.0( **320.0- 200.0- ***920.0- 900.0- 900.0- ***120.0- 0=ttatuCtiderCxaTD&Rfi1= )110.0( )700.0( )900.0( )020.0( )510.0( )300.0( seY seY seY seY seY seY slortnoCmriFenilesaB seY seY seY seY seY seY stceffedexfiraeY-yrtsudnI 47

noititepmoCyrtsudnIdna,noitairtapeR,ycnegAfoeloRehtgnissessA :sgnidloHhsaCetaroproCnonoitavonnIfotcapmIehT :01elbaT lareves rof stiderc xat D&R etats ni segnahc no stessa koob ot sgnidloh hsac ni segnahc fo snoisserger DID morf setamitse retemarap stroper elbat ehT eht fo sa llew sa ,)A lenaP( noititepmoc tekram tcudorp dna sevitnecni noitairtaper rof seixorp etna-xe no desab atad eht fo stilps elpmas-bus tnereffid llafoelpmas-busehtfostsisnochcihw,elpmassmrfievitavonniehtnidetamitseerasnoissergerllA.)BlenaP(mrfiehtybdecafsnoitcirfycnegafoytireves 785,27( doirep 1102 ot 6891 eht revo raey eno tsael ta ni serutidnepxe D&R orez-non troper taht tatsupmoC ni )seitilitU gnidulcxe( smrfi laicnanfinon SU a htiw detaicossa elbairav tnedneped eht ni egnahc eht si tneicfifeoc detroper hcae fo noitaterpretni ehT .)smrfi euqinu 850,6 rof snoitavresbo raey-mrfi ot evitaler hsac ni esaercni %5.1 a htiw detaicossa si stiderc xat D&R etats ni esaercni na ,nmuloc dnoces eht ni ,elpmaxe roF .tiderc xat D&R ni egnahc ,lleztraH ,yeloFfoerusaemnedrubxatngierofehtfoseulav)tnemtaert-erp(naidemwoleb .svevobaneewtebtilpssielpmaseht,AlenaPnI .stessakoob mrfiehtrehtehwnodesabdna,)6dna5snmuloC(selasfo)IHH(xednInamhcsriH-lhadnfireHyrtsudni)FF-84(eht,)2dna1snmuloC()7002(namtiTdna esoht(seirtsudnidetartnecnocylevitalernismrfifoelpmas-busehtylnoredisnocew,BlenaPnI .)4dna3snmuloC(raeyneviganiemocningierofstroper draob ,)2 dna 1 snmuloC( ezis draob fo seulav )tnemtaert-erp( naidem woleb .sv evoba neewteb elpmas eht tilps rehtruf dna )IHH naidem-evoba htiw nidetamitseerasnoitacfiicepsllA .)4dna3snmuloC(srotceridfodraobdefiissalcasahmrfiehtrehtehwnodesabdna,)6dna5snmuloC(ecnednepedni sallewsa)1(noitacfiicepsenilesabehtfoslortnoclevel-mrfiehtnisegnahcdeggaldna,snoitauqeslevelehtnistceffedexfimrfievomerotsecnereffidtsrfi ,%5,%1ehttaecnacfiingisgnitoned*dna,**,***htiw,levelmrfiehttaderetsulcera)sesehtnerapni(seulav-p .dedulcnieraseimmud)FF-84(yrtsudni-raey .BxidneppAnisistidercxatD&RetatsnisegnahcfotsildeliatedehT.AxidneppAnierasnoitinfiedelbairaV .ylevitcepser,level%01dna noititepmoCtekraMtcudorP&noitairtapeRfoseixorPybsgnidloHhsaCetaroproCnonoitavonnIfotcapmIehT:AlenaP IHHyrtsudnI 0>emocnIngieroF nedruBxaTngieroF hgiH woL oN seY woL hgiH )6( )5( )4( )3( )2( )1( 310.0 ***710.0 ***420.0 *700.0 ***510.0 700.0- 0=ttaesaercnItiderCxaTD&Rfi1= )900.0( )300.0( )500.0( )400.0( )500.0( )700.0( 600.0- ***710.0- *320.0- 900.0- *410.0- 500.0- 0=ttatuCtiderCxaTD&Rfi1= )800.0( )600.0( )210.0( )600.0( )800.0( )110.0( seY seY seY seY seY seY slortnoCmriFenilesaB seY seY seY seY seY seY stceffedexfiraeY-yrtsudnI snoitcirFecnanrevoG&ycnegAfoseixorPybsgnidloHhsaCetaroproCnonoitavonnIfotcapmIehT:BlenaP edistuO%(ecnednepednIdraoB nidraoBdefiissalC nieziSdraoB seirtsudnIIHHhgiHni)srotceriD seirtsudnIIHHhgiH seirtsudnIIHHhgiH woL hgiH seY oN hgiH woL )6( )5( )4( )3( )2( )1( ***510.0 **010.0 **810.0 700.0 ***310.0 500.0 0=ttaesaercnItiderCxaTD&Rfi1= )500.0( )400.0( )700.0( )700.0( )500.0( )500.0( *020.0- 700.0- *710.0- 800.0- *710.0- 700.0- 0=ttatuCtiderCxaTD&Rfi1= )210.0( )700.0( )010.0( )800.0( )010.0( )700.0( seY seY seY seY seY seY slortnoCmriFenilesaB seY seY seY seY seY seY stceffedexfiraeY-yrtsudnI 48

Figure1:HistoryofChangesinStateR&DTaxCredits ThisfigureshowsthepatternofchangesinallstateR&Dtaxcredits(subsequenttotheirintroduction)between tax years 1987 and 2011. The left column shows the pattern of cuts in state R&D tax credits, and the right column shows the pattern of increases in state R&D tax credits. We hand-collected this information for all U.S.statesusingavarietyofsources. TheStateTaxHandbook(variousyears)andWilson(2009)providedthe startingpointtogetasummarylistofR&Dtaxcreditsavailableinanygivenstateataparticularpointintime. Wethencross-checkedthisinformationbyverifyingthepassageofstatetaxbillsandobtaininginformationon theyearwhenagiventaxcreditbecomeseffectivebyperformingkeywordsearchesinLexis-Nexis(StateNet) andonlinesearchesofstatecodelegislationandtaxformsonthestates’DepartmentofRevenuewebsites. 49

Figure2:AnnualChangesinCashHoldingsAroundChangesinStateR&DTaxCredits This figure plots average annual within-firm changes in cash to book asset ratios for each year in a five-year window around the year when a state increases (Panel A) or cuts (Panel B) its R&D tax credit (t=0). Plotted changesareinexcessofcontemporaneouscashchangesinthefirm’s48-FFindustry,toremovetheinfluence oftime-varyingchangesinindustryconditions,andofpredictedchangesbasedonpre-treatmentcashlevels, tocontrolforpartialadjustmentofcash. Thesolid,redbarsarefortreatedfirmsandthedotted,bluebarsare forcontrols,andthedifferencebetweenthetwobarsinagivenyear–i.e.,thedifference-in-differenceestimate – is displayed for t=0 (with its respective (two-sided) t-test of significance using standard errors clustered at thefirmlevel,with***,**,and*denotingsignificanceatthe1%,5%,and10%level,respectively). PanelA:IncreasesinStateR&DTaxCredits PanelB:CutsinStateR&DTaxCredits 50