Risk, Financial Development and Firm Dynamics

K.7 Risk, Financial Development and Firm Dynamics Morais, Bernardo Please cite paper as: Morais, Bernardo (2015). Risk, Financial Development and Firm Dynamics. International Finance Discussion Papers 1134. http://dx.doi.org/10.17016/IFDP.2015.1134 International Finance Discussion Papers Board of Governors of the Federal Reserve System Number 1134 May 2015

Board of Governors of the Federal Reserve System International Finance Discussion Papers Number 1134 May 2015 Risk, Financial Development and Firm Dynamics Bernardo Morais NOTE: International Finance Discussion Papers are preliminary materials circulated to stimulate discussion and critical comment. References to International Finance Discussion Papers (other than an acknowledgment that the writer has had access to unpublished material) should be cleared with the author or authors. Recent IFDPs are available on the Web at www.federalreserve.gov/pubs/ifdp/. This paper can be downloaded without charge from the Social Science Research Network electronic library at www.ssrn.com.

Risk, Financial Development and Firm Dynamics BernardoMorais (cid:3) FederalReserveBoard May 2015 Abstract I document that the average productivity of firms tends to increase, and its variance to decrease, as they age. These two facts combined suggest that managers learn to reduce their mistakes as they operate. I develop a quantitative framework mimicking these dynamics and find thatyoungfirmshavesubstantiallyhigherfinancingcostsduetolowerandriskierreturns. Inthis scenario, a reduction in the financial development of an economy raises disproportionately the costofcreditofyoung-productivefirmsincreasingtheinputmisallocationwithinthissubgroup. Totestthevalidityofthetheory, Ifindthatthedataconfirmssomenovelpredictionsonaseries offirm-levelmoments. Finally,Ishowthatintroducingthesetwofactsallowsthemodeltobetter explaintherelationbetweenfinancialandeconomicdevelopment. Keywords: productivity, misallocation, financial frictions, learning JEL No. O4, O11 (cid:3)TheviewsexpressedinthispaperdonotreflecttheviewsoftheBoardofGovernorsoftheFederalReserveSystemofits staff. Contact:bernardo.c.morais@frb.gov. IthankDanielDias,SebastianEdwards,PaolaGiuliano,MarkGrinblatt,Hugo Hopenhayn, EdwardLeamer, LoganLewis, LeeOhanian, ClaudiaRuiz, NicoVoigtlander, RomainWacziarg, andMark Wright, fortheircommentsandguidance. IalsothankseminarparticipantsattheAndersonSchoolandtheEconomics department at UCLA, HEC-Montreal, IESE, ITAM, FRB - Boston, FRB - BoG, University of North Carolina Chapel-Hill, UQAM,UniversityofWesternOntario,EconConatPrincetonandSEDinBrugges. FinancialsupportfromtheFundação paraaCiênciaeTecnologiaandFundaçãoGulbenkianaregratefulacknowledged.

1 Introduction The positive relation between firm age and firm productivity has been widely studied.1 Using detailed firm level data for a broad set of countries, I start by examining empirically the dynamics of revenue-productivity (TFPR) of firms as they grow older. I show that after controlling for sample selection: i) there is a positive relation between firm age and TFPR, and that ii) the volatility of the TFPR process of a firm decreases with its age. I create a model incorporating these two characteristics, which I define as Learning, and analyze quantitatively their implications.2 More concretely, I ask: If younger firms have riskier profiles and consequently face more expensive credit, how are theyimpactedbytheleveloffinancialdevelopmentintheireconomy? Areyoung-productivefirms morefinanciallyconstrainedanddotheysufferworsefactormisallocationineconomieswithworse financial development? If so, can this help explain the large cross-country differences in aggregate productivity? To answer these questions I present a quantitative framework based on Midigran and Xu (2013) extended to incorporate the relation between the TFPR process and firm age. I discipline the quantitative analysis by requiring two versions of the model to mimic a series of relevant moments of the UK firm distribution.3 In the first version - defined as No-Learning - I do not require the model to replicate the above described relation between firm age and TFPR, while in the second - denoted as Learning - I do. The comparison between these two sets of simulations allows me to isolate the impact of Learning. Afterwards, and while leaving all estimated parameters unchanged, I vary the financial development of an economy to quantify its impact on firm dynamics and on aggregate income and productivity. The three main results of the paper are that under the Learning scenario: i) the model is better able to replicate a series of relevant firm-level moments not directly targeted by the calibration; ii) financial frictions constrain disproportionately more the capital accumulation of young-productive firms; and iii) financial development can help explain a larger fraction of the 1For empirical evidence on the dynamics of physical productivity (TFPQ) on aviation see Benkard (2000), on shipbuildingseeThorntonandThompson(2001)andoncarmanufacturingseeLevitt,ListandSyverson(2013). Other,more generalstudiesrelatingrevenueproductivityandlaborproductivitywithage,includeBahkandGort(1993)andJensen, McGuckinandStiroh(2001)respectively.Similarly,HuergoandJaumandreu(2004)andWarusawitharana(2012)document that newer establishments have higher rates of revenue-TFP growth. Finally, Foster, Haltiwanger, and Syverson (2008, 2013)decomposefirmproductivityintophysicalandrevenueandreportthatyoungerfirmsaresmallerduetodemand sidefundamentals. 2BydefiningtherelationbetweenfirmageandtheTFPRprocess,asLearning,Iattempttocaptureallfactorsinfluencing the TFPR process of a firm as it grows older. They include changes in supply side fundamentals such as improvementsinphysicalproductivity,aswellasvariationindemand-sidefundamentalssuchastheaccumulationofacustomer base/reputation. 3TheUKisthebenchmarkeconomygiventhequalityofitsdataandthefactthatitisafinanciallydevelopedeconomy. 2

cross-countryvariationinincomepercapitaandaggregateproductivity. Finally,themodeluncovers a series of novel predictions relating financial frictions with a series of firm dynamics - entry/exit, leverageandfirmgrowth-forwhichIdocumentempiricalsupport. The model in Jovanovic and Nyarko (1996) predicts that managers make fewer and less costly mistakesastheylearntorunaproject. Followingthissimpleidea,Istartthispaperbydocumenting empirically that the expected value (variance) of the TFPR of firms increases (decreases) as firms growolder. ThesefactsholdwhenusingdifferentmethodologiestoestimateTFPR,whentestingfor abroadsetofcountries,andcontrollingforsampleselection. Usingwithinfirmvariation,Ifindthat theaverage(variance)TFPRoffirmsrises(falls)atadecreasingrate,duringthefirst20yearsofafirm. Ithencreateaquantitativeframeworkabletoreplicatethesetwofactsalongwithotherrelevantfirmlevelstatistics. TotestthequantitativeimplicationsofthisfindingIuseaone-sectordynamicgeneral equilibriummodel,whereagentshaveanoccupationalchoice-ofmanagingaprojectorworkingfor agivenwage-anddifferintheirentrepreneurialtalentandwealth. Entrepreneurialtalentdevelops stochasticallyanditsunderlyingprocessevolvesasaprojectages. Moreconcretely,asaprojectgrows older entrepreneurs make fewer/less costly mistakes raising expected productivity while lowering volatility. Finally, agents can transfer resources across periods using a one-period defaultable bond, withthecostofcreditreflectingboththeprobabilityofdefaultaswellastherecoveryrate-theproxy forfinancialdevelopment-incasedefaultoccurs.4 Toanalyzetheimpactofintroducingtherelationbetweenfirmageandproductivity-i.e. Learning -Icalibratethemodeltwice. Inthefirstcalibration-No-Learning-thesimulationsdonottargetthis relation, while in the second - Learning - they do. More concretely, in this latter scenario firms can either be young or old and when young they have a probability of receiving a negative transitory shocktotheirproductivityeachperiod. Theresultsofthecalibrationunderthissettingindicatethat eachperiodyoungfirmshavea27percentprobabilityofsufferinga62percentnegativeproductivity shock. AnimportantcharacteristicoftheresultsisthatundertheLearningcalibrationthesimulations can replicate (by construction) the two moments relating firm TFPR with age whereas under No- Learningtheycannot. Theothermaindifferencebetweenthetwocalibrations,isthatthevarianceof productivity shocks is 32 percent lower in the Learning scenario to compensate for the productivity mistakesofyoungfirms. Simulations under both scenarios are fairly successful in replicating important statistics of the 4Both in the model and in the data we use the recovery rate as the main proxy of the financial development of an economy.Therecoveryratemeasurestheshareofdefaultingloanthatisexpectedtoberecoupedbythecreditor. 3

UKeconomynotdirectlytargetedbythecalibrations. Nevertheless,therearethreerelevantstatistics whicharesignificantlybettermatchedunderLearning. Thefirst,istheshareofsalesofyoungfirms. In the Learning scenario young firms have lower average productivity and suffer larger financial constraintslimitingtheirsize. Therefore,theshareofsalesoffirmsyoungerthan6yearsis11percent, asinthedata,whereasintheNo-Learningscenariothisshareis19percent. Thesecondmomentbetter matched under Learning is the leverage ratio of both young and old firms. In this setting - relative to the No-Learning one - young (old) firms are exposed to more (less) intense productivity shocks leading them to support relatively lower (higher) levels of leverage which closely mimic the UK data. Finally, and related with the previous fact, the level of debt to GDP is also better matched in the Learning scenario. Since older firms are subject to smaller productivity shocks and have higher leverageratios,therelativelevelofdebttoGDPishigherinthisscenarioandismorecloselymatched withthedata. As mentioned above, one of the main results of the paper is in establishing that financial constraints have a stronger impact in the resource accumulation of young and young-productive firms. Inthebenchmarkeconomy,Ifindthattheaverageproductofcapital(APK)-aproxyforborrowing constraints - is 50 percent higher for young firms and 95 percent higher for young-productive firms relativetotheaveragefirminthesimulations. Inthedatathesevaluesare15and52percentrespectively. Anadditionalrelatedfindingsuggestingthatfinancialconstraintsarestrongeramongyoung firms,isthatthestandarddeviationofthelog-APK-anindicatorofcapitalmisallocation-ofyoung firmsrelativetotheaveragefirmis15percenthigherinthesimulationsand26percenthigherinthe data. After establishing that both versions of the model replicate well the UK distribution of firms, I analyze the impact that variations in the recovery rate have on a model economy. First, I show that young-productivefirmsaremoreconstrainedineconomieswithworserecoveryrate. Second,Ifind thatundertheLearningscenariothemodeluncoverstwocross-countryfactsonfirmdynamics-leverageandfirmgrowth-thatareverifiedempirically. Third,InotethatundertheLearningframework, the model accounts for a larger fraction of the cross-countries differences in aggregate income and productivity. Thesimulationsunderbothscenariosindicatethatareductionintherecoveryratedoesnothavea significanteffectontheAPKofthemedianfirm. Howevertheyalsoshowthatreducingtherecovery rate increases the APK of young and young-productive firms by around 80 percent. A significant partofthesepredictionsareconfirmedbythedata. 4

The model under Learning provides two new predictions relating the level of recovery rate with firm dynamics: First, selection on productivity is weaker in economies with lower recovery rate. In theseeconomiesolderfirmshaveahigherincentivetocontinueoperating,duetoaloweroutsideoption, even when they become relatively unproductive. Therefore, in these countries there is weaker selection on productivity leading to a lower average TFPR of older firms. This result holds empirically in the cross-country sample of firms.5 I find that while in countries with high recovery rate the average TFPR of firms increases for cohorts of firms up to 20 years-old, in countries with lower recoveryratetheaverageTFPRpresentsaninvertedU-shape.6 Second,firmsstarthighlyleveraged and deleverage as they age. Furthermore, in economies with better recovery rate firms have higher leverage ratios and deleverage more slowly. Given their relative impatience, when agents become entrepreneurs they rely essentially on costly external debt to finance their investment, and decrease their leverage as their projects age.7 The simulations indicate that firms in economies with lower recovery rates deleverage faster due to the higher cost of debt. Empirically, I find that the average leverage ratios are essentially the same for young firms regardless of the recovery rate, and that as firmsagetheydoindeeddeleveragefasterincountrieswithworsefinancialdevelopment. Finally, I analyze the impact that financial frictions have on aggregate productivity. I find that variations in recovery rate can reduce aggregate TFP by 8 percent under the No-Learning scenario and by 20 percent under Learning. As in Buera, Kaboski and Shin (2011) the model provides a clear decomposition of the main margins affected by financial constraints. I find that under Learning oldunproductivefirmsemploytoomuchcapitalandlabor. Thereforeinthisscenario,andunlikeinthe No-Learning one, capital reallocation within age-cohorts only reduces part of the aggregate productivitylossesduetomisallocation. Contribution to the Literature - The paper contributes to the vast literature relating financial with economic development in an attempt to explain the large cross-country differences in aggregate income.8 Nevertheless,themajorityofthisworkdoesnotreplicateimportantfeaturesofthefirm-level 5Usingadifferentsampleofcountries,HsiehandKlenow(2009)alsofindthatolderfirmsinpoorercountries-India andMexico-arerelativelymoreunproductivethanyoungerfirms. 6This result is the consequence of two forces. On the one-hand, firm TFPR increases as firms age and make fewer mistakes. On the other hand, as firms are hit with permanent negative shocks, they have a relatively high incentive to continueoperatinggiventhelowoutsideoption. UnderNo-Learningthereisahigheroutsideoptionasyoungfirmshave thesameproductivityprocessasoldonesleadingagentstoexitwhentheyreceiveapermanentnegativeshock. 7Wefindthatyoungfirms(i.e. age 6)havealeverageratioof0.78whereasolderfirmshavealeverageraioof0.65. (cid:20) HuynhandPetrunia(2013),studyingCanadianfirms,alsofindthatfirmstendtodeleverageastheygrowolder. 8Forexample,JeongandTownsand(2007)attribute70percentofThailand’sgrowthratebetweenthe70sandthe90s to improvements in the financial sector. Buera, Kaboski and Shin (2011) build a two-sector model with fixed costs and showthatadecreaseinfinancialdevelopmentcanresultina40percentdecreaseinaggregateTFP.AmaralandQuintin (2010),ErosaandCabrillana(2010),Castro,ClementiandMacDonald(2009),Greenwood,SanchezandWang(2013),and 5

data such as the volatility of growth rates, leverage dynamics or entry/exit rates.9 This paper attempts to fill this gap by calibrating the parameters using micro moments at the firm level. The paperalsocontributestotheliteraturerelatingfinancialfrictionswithfirmdynamics.10 Inparticular, it is closely related with Hennessy and Whited (2007) and Gomes and Schmid (2010) who also analyzethedynamicsoffirmsinanenvironmentwhereyoungerprojectsareimpactedbyhigherdefault risk.11 In the model, I show empirically and theoretically, that their findings regarding the financial structureoffirmsbecomemoreacuteineconomieswithlowerfinancialdevelopment. The rest of the paper is organized as follows. Section 2 describes the firm-level dataset and documentsthefactsaroundwhichthemodelisbuilt. Section3introducesandcharacterizesthemodel. Section 4, presents the quantitative analysis and counterfactual experiments. Section ?? concludes suggestingfurtheravenuesforfutureresearch. 2 Facts This section describes in detail the datasets used in the paper and documents empirically the two factsrelatingtheproductivityprocessofafirmwithitsage. 2.1 Data Firm-LevelData The cross-country firm level data comes from Analyze Major Database from European Sources (Amadeus). Amadeusisacomprehensivepan-EuropeandatabaseprovidedbyBureauvanDijk. Itis highlyusefulasitnotonlycoversalargefractionofnewandsmallfirmsacrossallindustriesbutit alsoappearstoberepresentativeoftheuniverseoffirmsatnationallevel.12 Thedatabasestartedin 1997,andcollectsstandardizeddatafrom50vendorsacrossEuropewiththelocalsourceofthedata generallytheofficeoftheRegistrarofCompanies. Amadeuspresentsstandardizedannualdata-for up to 10 years - on financial ratios, activities and ownership for approximately 5 million companies MidriganandXu(2013)alsoprovidequantitativeassessmentsoftheimpactoffinancialdevelopmentonaggregateincome andTFP. 9Anexception isMidrigan andXu (2013)who requiretheirmodel tomatch micro-momentssuch asthe volatility of growthratesoffirmsorreturnstocapitalbetweenyoungandoldfirms. 10Itincludesamongothers,Hopenhayn(1992),,ClementiandHopenhayn(2006)andArellano,BaiandZhang(2012). 11HennessyandWhiteduseastructuralmodeltocalculatethecostsofexternalfinancing. Theyfindthatthenatureof theshockshittingmaturefirmsaredifferentfromsmallfirms,witholderfirmshavingalowervarianceintheirshocks. 12Intheappendix,IfollowArellano,BaiandZhang(2012)andprovideacomparisonbetweentheAmadeussampleand theUniverseoffirmspresentintheEurostat. TheEurostatdata-availableathttp://epp.eurostat.ec.europa.eu-usesthe full national business registrars as data sources. Small firms are slightly underrepresented in Amadeus and its fraction variesacrosscountries,butthisvariationisnotcorrelatedneitherwithincomelevelnorwithfinancialdevelopment. 6

per year. It includes accounting data in standardized financial format for balance sheets, income statements, and financial ratios. The accounts are transformed into an universal format to enhance comparisonacrosscountriesthoughthecoverageoftheseitemsvariesacrosscountries. Inadditionto financialinformation, thedatasetprovidesotherinformationsuchastheageofthefirm, ownership and legal status. Amadeus, assigns companies a three-digit NACE code, the European standard of industryclassification,whichcanbeusedtoclassifyfirmsandconstructindustrydummyvariables. Fortheempiricalstudyandforthenumericalexercises,Iusethefinancialinformationpresented for the 2005 to 2009 years. These consecutive years are chosen, as they are they have the most complete coverage. Since all the necessary information for this analysis is unavailable for a series of firms, I need to impose a number of restrictions in order to clean the data.13 For comparison with the vast literature using various Census of Manufacturers, I focus the analysis on firms in the manufacturing sector. Furthermore, I exclude observations of firms incorporated in the same year they were reporting or that do not provide data on assets, liabilities, sales, year of incorporation, or that providenegativesalesorassets. Furthermore,Iexcludecountrieswithlessthan1000firms.14 These criterialeaveuswith3millionobservationsin27countries: Bulgaria,Croatia,CzechRepublic,Denmark,Estonia,Finland,France,Germany,Greece,Hungary,Iceland,Ireland,Italy,Latvia,Lithuania, the Netherlands, Norway, Portugal, Romania, Russian Federation, Serbia, Slovakia, Spain, Sweden, UkraineandtheUnitedKingdom. FinancialDevelopment InadditiontoAmadeus,IusedataprovidedbytheDoingBusinesstoobtainaproxyforfinancial development. Thisdatasetmeasurestheefficiencyoftheinsolvencyprocessforaseriesofcountries, following the methodology developed by Djankov, Hart, McLiesh and Shleifer (2008). Using several outcomes of the insolvency process, such as legal costs, nominal interest rate and length of the procedure, they computed the recovery rate as an indicator of investor protection. It is measured as the present value, net of all costs, recouped by the creditors of a defaulting firm. It summarizes the efficiency of the bankruptcy proceedings and is the main indicator of financial development in thispaper. Anadvantageofusingtherecoveryrate,isthatnotonlyitisacleanmeasureofinvestor protectionbutitalsohasadirectcounterpartinthemodel. In addition to the recovery rate, and to document that the empirical results carry through with othercommonindicatorsoffinancialdevelopment,IpresenttwofurthermeasuresproposedbyKing 13Amoredetailedexplanationonthethecleaningofthedataisdocumentedinsection8.2oftheappendix. 14IexcludefromthisanalysisAustria,Belarus,Cyprus,Liechtenstein,Luxembourg,Macedonia,Moldova,Monacoand Switzerland. 7

andLevine(1993). ThefirstindicatoristheprivatecredittoGDPratio,whereasthesecondistheratio of the liquid liabilities of the financial system to GDP.15 The assumption underlying these measures isthatthelargerthecredittoprivatefirmsthemoreengagedfinancialintermediariesareinresearching those same firms, in providing risk-management services and in facilitating transactions, rather thansimplyfunnelingcredittopublicenterprises. Figure1,presentsthethreemeasuresoffinancial developmentforeachcountry. Allindicatorsarehighlycorrelatedwitheachother.16 2.2 FirmProductivityandFirmAge To document the dynamics of firm productivity, I first present the estimation method. Productivity is not observed directly and must be inferred from sales and inputs used. To calculate revenue TotalFactorProductivity(TFPR)IfollowLevinsohnandPetrin(2003).17 Theydevelopedamethodology for estimating the production function that deals with the endogenous response of inputs to productivity.18 Thisapproachprovidesconsistentestimatesoftheparametersforeachindustryand the details of the estimation procedure are documented in the appendix. After obtaining the input elasticities,IconstructestimatesofthefirmTFPR.19 Expected value - As mentioned in the introduction, I want to understand how the average productivityoffirmsevolvesastheyage. Asafirstpasstothedata,IplottheaverageTFPR-demeaned by average log TFPR - as a function of firm age, using the firm data of the four countries with the largest number of observations.20 The graphs are reported in Figure 2. The solid lines denote the 15ThesealternativemeasuresweretakenfromtheWorldDevelopmentIndicators. 16Thecorrelationbetweenrecoveryrateandthefirstandsecondadditionalindicatorsoffinancialdevelopmentis0.62 and0.93respectively. 17Intheappendix,IshowthattheresultsobtainedinthissectiondonotqualitativelychangeifIestimatetherevenue TFPusingtheinputcostsharesasfactorelasticities. 18Theendogeneityissue,drivenbysimultaneity,arisesfromthefactthatproductivity,whichisdirectlyunobservableby theeconometricianbutknownbythefirm,ispositivelycorrelatedwithinputchoice. ThisimpliesthatanOLSestimation willyieldupwardlybiasedcoefficients.Asecondissue,selectionbias,arisesgiventhenon-randomattritionofthesample. If profitability is positively correlated with firm size, then larger firms will have a lower exit threshold on productivity. Thiswillleadtoanegativebiasinthecapitalcoefficient. ToaddresstheselectionproblemImustgenerateanexitrule. Nevertheless,andasdocumentedinDeLoecker(2007),thisbiasisrelativesmallinpractice,andconsequentlyIwillnot controlforit. 19InordertomaketheestimatedTFPRcomparableacrossindustrieswecreate,forbothmethods,aproductivityindex tfp i;j;t followingAw, Chen, &Roberts(2001). Ineveryindustry, theproductivityindexisobtainedbysubtractingfirm i0spredictedoutputfromitsactualoutput. Thismethodologyinsuresthattheproductivityindexisinsensitivetounitsof measurement. Theindexisobtainedbysubtractingtheproductivityofareferenceplantinabaseyear(inourcase2006) foragivenindustryfromanindividualplant’sproductivitymeasure: tfp =y y (cid:12) l l (cid:12) k k (cid:12) (m m ) i;j;t i;j;t (cid:0) j;t(cid:0) j;l i;j;t (cid:0) j;t (cid:0) j;k i;j;t (cid:0) j;t (cid:0) j;m i;j;t (cid:0) j;t Thebaroveravariableindicatesameanovebral(cid:0)lfirmsinan(cid:1)indbustr(cid:0)yinabasey(cid:1)ear.bThismeasurepresentsthelogarithmic deviationofaplantfromthemeaninagivensectorinabaseyear. 20AsinLevineandWarusawitharana(2013),IpresenttheresultsoftheregressionsfortheUK,ourbenchmarkeconomy, 8

averageTFPRofafirmwhilethedashedlinesindicatetheassociated95thpercentconfidenceinterval. Thefiguresindicatethatinallcountriestheaverageproductivityofthesurvivingfirmstendsto increasewithfirmageatadecreasingrate. Moreconcretely,theaverageTFPRofthecohortoffirms with20yearsofageisonaverage20to85percenthigherthanthatofanentrantcohort. Therateof increaseaverageTFPRforcohortsolderthan20yearsispositivebutsignificantlylowerthanthatof youngercohorts. To study the evolution of firm TFPR more formally, I use regression techniques. First, I analyze the average TFPR of surviving firms without correcting for selection nor using firm fixed-effects. I estimate tfp = (cid:12)logage +Ind Year +" (1) i;t i;t j t i;t (cid:2) j where tfp is the estimated TFPR index of firm i operating in sector j in year t. The coefficient of i;t interest (cid:12) measures the impact of the firm log-age on the TFPR. Finally, a full set of industry-year dummiesInd Year areincluded, offeringaflexiblewaytocontrolforvariationsinproductivity jt (cid:2) inducedbyindustry-widefactorsacrossdifferentyears. Amongthem,areeventslikeweathershocks andbusinesscyclefluctuationswhichtendtohaveasystematicallydifferentimpactacrosssectors. The results of this benchmark regression are reported in the columns (1) of Table 1.21 The coefficient on log-age is positive and significant in all countries. The coefficients in the four countries in thesamplerangefrom0.01to0.25. FocusingontheUK,thecoefficient(0.01)indicatesthatasfirm’s agedoublestheTFPRincreases1percentonaverage. To understand whether firms improve their TFPR as they age, it is more informative to use firm fixed-effects and analyze the within firm variation. The firm fixed-effect accounts for unobserved permanent components that affect firm TFPR. The results are in columns (2) and indicate that it is indeed the case that the within firm variation of TFPR of the surviving firms increases as they age. Forthefourcountriesinthesample,thecoefficientsrangefrom0.05to0.17andareallsignificantat the1percentlevel. However, a concern regarding the above results is the fact that the exit decision may be related with firm characteristics leaving us with sample selection bias.22 To test and correct for it I follow andfortheremainingthreelargesteconomiesforwhichourdatasethasthegoodcoverage(France,ItalyandSpain). We excludeGermany,asitscoverageisparticularlypoorsinceGermanreportingrequirementsarelessstringentthaninother Europeancountries. 21TheAmadeusdatasethasnoinformationonmaterialexpendituresforUKfirms. Therefore,IcalculatetheTFPRfor UKfirmsusingCobb-Douglasspecificationforvalueadded. 22Thisconcernissomewhataddressedbytheinclusionoffirmfixed-effects,astheyimplythatthevariationinthegrowth ratesiswithin-firm. 9

Wooldridge (2002). First, I test the significance of the inverse Mills ratio obtained from a sample selection probit on the benchmark equation 1 and find that indeed there is sample selection bias.23 Tocorrectforit,Iestimatetheequation tfp = (cid:12) logage+(cid:11) +Ind Year +IMR +" i;t 1 i j t t;i i;t (cid:2) where IMR is the inverse Mills ratio. The results of this regression are present in columns (3) and t indicatethatevenaftercontrollingforselection,firmTFPRincreasesasfirmsgrowolder. Variance - In addition to the analyzing the first moment of the firm TFPR, I test Jovanovic and Nyarko’spredictionthatthevarianceoftheproductivityprocessisheteroskedastic. Moreconcretely, Iwanttoknowwhethervariancetendstodecreasewithfirmage. Asafirstpasstothedata,IplotinFigure3,thestandarddeviationofTFPRgrowthforeachagecohort. Inallfourcountriesinthesample, thestandarddeviationisdecreasingandconvexasfirms age. To study the relation between the standard deviation of firm growth and age more formally I follow a two-step procedure based on Castro, Clementi and MacDonald (2009). First I obtain the residual " after regressing the log-change of the productivity process on a series of relevant characteristics. Iestimate b (cid:1)tfp = (cid:12)logage +(cid:11) +Ind Year +" (2) i;t i;t i j t (cid:28);j;t (cid:2) where(cid:1)tfp representsthegrowthrateofthelogTFPRoffirmibetweent 1andt. Iincludelogi;t (cid:0) age as a regressor since firm age is a predictor of productivity growth. Furthermore, and as argued above,Iincludeafirmandsector-yearfixed-effects. Ithenusetheestimatedresiduals"andproceed byestimatingtheequation: b ln"2 = (cid:18)(cid:28) +u i;j;(cid:28) i;t i;j;t (cid:28) where (cid:18)(cid:28) is an indicator variable of the abge-group in which the firm belongs. Letting (cid:18) denote i;t the point estimate of the dummy coefficient, # exp (cid:18) is the estimate of the conditional stan- (cid:28) b (cid:17) r dard deviation for firm TFPR in age group (cid:28).24 The estim (cid:16) a (cid:17) tes # of this regression are reported in b (cid:28) 23Forthefirst-stageprobitequationIregressanexitbinaryindicatoronfirmassets,TFPRandlog-age.Theprobitresults (not shown) indicate that the probability of firm exit is negatively related with firm size, productivity and age in all 4 countries.Duetoalackofqualityintheselectionindicatorsin2006and2009,forthisexerciseIrestrictthesampleto2007 and2008. 24Thisformulationresultsfromtheassumptionofaparticularfunctionalformforthevariance(cid:27)2 j = (cid:27)2exp(cid:18) j :Itisthe 10

columns (1) of Table 2.25 As posited, younger cohorts have a higher variance in their productivity growth-inallfourcountriesinthesample-whichisstatisticallydifferentfromthecoefficientsofthe immediatelyyounger/oldercohorts. Asbefore,aconcernIhaveinthisexerciseregardstheselectioneffect. Itispossiblethatthefirms thatdroppedoutofthesampleatanearlyagehaddifferentcharacteristicsregardingtheproductivity process, such as a larger risk-profile, which may be contaminating the results. To account for this possibility,Iconductatwo-stepHeckmantesttodealwithselectionconcernswhereinthefirststep IincludetheinverseMillsratioin2. Theresultsarereportedincolumns(2)ofTable2,andindicate thattheresultsarerobustwhenusingthisrobustnesstechnique. OnceestablishedthetwofactsrelatingtheTFPRprocessofafirmwithitsage, Icreateageneral equilibriummodelincorporatingthem. 3 Model Toanalyzetherelationbetweenfirmexperiencewithitsproductivityprocess,Icreateaquantitative general equilibrium model based on Buera, Kaboski and Shin (2011) and Midrigan and Xu (2013). The model economy is populated by infinitely lived individuals, heterogeneous in their wealth and intheirabilitytomanageafirm. Theirtalentfollowsastochasticprocesswhichisdependentonthe age of the project. Each period, agents make an occupational choice of working for a given wage or managing a firm. This occupational choice depends not only on the managing ability but also on the agents’ wealth, since the cost of credit to finance project is impacted by the collateral of the entrepreneurandbythequalityoffinancialdevelopmentintheeconomy. 3.1 Outline Environment - The economy is populated by a continuum of infinitely-lived agents of measure 1, whodiscountthefutureatrate(cid:12) andwhosepreferencesarerepresentedby 1 (cid:12)t C i 1 ;(cid:0)t (cid:16) 1 (cid:16) t=0 (cid:0) X whereC isconsumptionofagentiattimetwhile(cid:16) istheintertemporalelasticityofsubstitution. i;t caseofmultiplicativeheterosckedasticitymodelpresentedbyHarvey(1976). 25WehavealsousedtheBreusch-Pagantestforthenullhypothesisthattheconditionalvolatilityisthesameforfirmsin differentage-groups.Itrejectsthishypothesisattheconventionalsignificancelevel. 11

All agents have the ability to run a project. The log of this entrepreneurial talent z follows a i;t continuous-stateMarkovprocesswithatransitiondensitythatdependsontheage(cid:28) oftheproject Pr z i;(cid:28)+1;t = z 0 z i;(cid:28);t = z = (cid:25)(cid:28) z0 z j j (cid:16) (cid:17) (cid:0) (cid:1) Each period, after realizing their current talent level and receiving their periodic income, agents decidehowmuchtoconsume,saveorborrowB,whethertoworkormanageafirm,andinthelatter scenariohowmuchtoinvestonassetsK.26 Whenmakingtheiroccupationalchoiceagentstakeinto accounttheirabilitytorunaproject,theirsavings/debt,thevalueoftheirassets;andtheageoftheir project(cid:28). Ifagentsopttoworktheysupplytheirunitoflaborinelasticallyforawagew. Conversely, iftheychoosetomanageafirmtheydecidehowmuchtoinvestandhowmuchtoborrow. Managers canonlyrunonefirmatatime,andthereisnomarketforentrepreneurialability. Allentrepreneursoperateinaperfectlycompetitiveenvironmentandhaveaccesstoaproduction technologywithdecreasingreturnsgivenby: Y = Z K(cid:11)L (cid:13) i;t i;t i;t i;t 0 (cid:11);(cid:13);(cid:11)+(cid:13) < 1 (cid:20) where Y is revenue, L the amount of labor hired, K the capital stock, and Z the ability of the entrepreneur.27 The capital accumulated by the project depreciates at the rate (cid:14), and is accumulated throughperiodicalinvestmentsI suchthat K = (1 (cid:14))K +I (3) i;t+1 i;t i;t (cid:0) Credit Market - Next, I describe the financial side of the model. I follow Hennessy and Whited (2007)andallowforintertemporaldefaultabledebtcontractsB.28 Allcontractswithagenti;havea one-period maturity where the borrower receives Bi;t units of capital at time t with the promise to Ri;t repayB unitsatt+1. Ifagentssave,theysecurearisk-freerateR determinedexogenously. Ifthey i;t f borrow,theyobtainarisk-adjustedinterestrateR thatdependsontheprobabilityofdefaultandon i 26If B < 0 the agent has savings. As in Hennessy and Whited (2007), I do not allow for simultaneous saving and borrowing. 27Thedecreasingreturnstoscaleassumptioncanbeinterpretedasarisingfromanenvironmentinwhichmonopolistic competitivefirmsfaceaconstantelasticitydemandfunction. 28GomesandSchmidt(2010)andArellano,BaiandZhang(2012)alsomodelthesametypeoffinancialcontract. 12

therecoveryrate(cid:24) incaseofdefault. Moreconcretely,intheeventofdefaultthelenderreceives D (cid:24)min (1 (cid:14))K ;B ; (cid:24) [0;1] i;t i;t i;t (cid:17) f (cid:0) g 2 where(cid:24) representstherecoveryrateoftheeconomy,(1 (cid:14))K theresidualvalueofcapitalandB i;t i;t (cid:0) theoutstandingdebt. Conversely,thedefaultingborrowerappropriates T i;t {max (1 (cid:14))K i;t (cid:24)B i;t ; (1 (cid:24))(1 (cid:14))K i;t (cid:17) f (cid:0) (cid:0) (cid:0) (cid:0) g where{ [0;1]istheshareoftheresidualcollateralthatisappropriatedbytheborrower.29 2 Icannowrepresenttheperiodicbudgetconstraintofagentiattimetas B i;t C +I o (Y wL )+(1 o )w (1 d ) B + +d T (4) i;t i;t i;t i;t i;t i;t i;t i;t 1 i;t i;t (cid:20) (cid:0) (cid:0) (cid:0) (cid:0) (cid:0) R i;t (cid:18) (cid:19) wheretheleft-handsiderepresentstheuseoffunds,whiletheright-handsidetheirorigins. Thevariablesoanddarebinaryindicatorsoftheoccupationalchoiceandofthedefaultdecisionrespectively. Recursive Problem - The timing of the model is as follows: At the beginning of each period, agents make their consumption, investment, financing and occupational choices given their capital K, debt B, ability Z and project age (cid:28). After these decisions, their current ability is realized. If the agent is a worker he supplies his unit of labor, otherwise he makes his hiring decisions. Once productioniscomplete,agentsdecidewhethertodefault. Therefore,thevalueofagentiwithcapital K,debtB,productivityZ andage(cid:28) attimetis V K ;B ;Z ;(cid:28) = max Vc;Vdef (5) i;t i;t i;t i;t n o (cid:0) (cid:1) whichisthemaximumbetweenthevalueofrepayingVc ordefaultingVdef. The value of repaying the full amount of the debt for a worker (o = 0) or for an entrepreneur i;t (o = 1)is i;t Vc(:) = max C i 1 ;(cid:0)t (cid:16) +(cid:12)EV (K ;B ;Z ;(cid:28) :) (6) i;t+1 i;t+1 i;t+1 i;t+1 C;B;I 1 (cid:16) j f g (cid:0) B i;t s:t:C +I o (Y wL )+(1 o )w B + i;t i;t i;t i;t i;t i;t i;t 1 (cid:20) (cid:0) (cid:0) (cid:0) (cid:0) R i;t K = (1 (cid:14))K +I i;t+1 i;t i;t (cid:0) 29When{<1,partoftherenegotiationprocessresultsinlossesofcapitalfortheeconomy. 13

whereasthevalueofdefaultis def C i 1 ;(cid:0)t (cid:16) V (:) = max +(cid:12)EV (I ;B ;Z ;(cid:28) :) t d;B;i;l 1 (cid:16) i;t j;t+1 j;t+1 j;t+1 j f g (cid:0) s:t:C +I o (Y wL )+(1 o )w+T i;t i;t i;t i;t i;t i;t i;t (cid:20) (cid:0) (cid:0) Equilibrium - The framework is designed for the purpose of studying a stationary competitive equilibrium. In this equilibrium, some firms enter and expand, whereas others contract and exit. Despite all these individual changes, aggregate variables remain constant through time. For each debtcontracttheinterestrateR thatallowcreditorstobreakeveninexpectedvalueissuchthat i (1 (cid:14))K i;t R = (1 (cid:25) )R +(cid:25) R (cid:24)min (cid:0) ;1 (7) f i i;t i i;t (cid:0) B i;t (cid:26) (cid:27) where the left-hand side represents the risk-free rate R , while the right-hand side represents the f expected repayment rate of a debtor with default probability (cid:25) and collateral K . Therefore, the i;t i;t riskadjustedinterestrateis R f R = (8) i;t 1 (cid:0) (cid:25) i;t 1 (cid:0) (cid:24)min (1 (cid:0) B (cid:14) i ) ; K t i;t;1 (cid:16) n o(cid:17) Thislendingratedependsontheprobabilityofdefaultandontherecoveryrate.30 A stationary competitive equilibrium for this economy consists of: an invariant distribution of wealth,capitalandtalentG(B;K;Z),policyfunctionsofagentsC(K;B;z;(cid:28)),I(K;B;z;(cid:28)),B(K;B;z;(cid:28)), l(K;B;z;(cid:28))loancontractsofferedbycreditorsR (K;B;z;(cid:28))andpricesw;R suchthat: i;t f 1. Given the schedule of loan contracts offered, the policy and value functions of agents satisfy theiroptimizationproblem 2. Loancontractsreflectthefirm’sdefaultprobabilitiessuchthatallfinancialintermediariesmake zeroprofitsinexpectationonallcontracts 30To provide some intuition to the expression above, assume that (1 (cid:0) (cid:14))Ki;t 1 K;B . In this case, condition 7 Bi;t (cid:21) 8f g becomes R =R (1 (cid:25) )+R (cid:25) (cid:24) f i;t i;t i;t i;t (cid:0) implyingthatinterestrateis R R i;t = 1 (cid:25) f (1 (cid:24)) (9) i;t (cid:0) (cid:0) where it is clear that interest rate is positively related with the probability of default, and negatively related with the recoveryrate. 14

3. Thejointdistributionofwealthandtalentarestationary 4. Labormarketclears 4 Quantitative Analysis Inthissection,Iassessquantitativelytheimplicationsofthemodel. Inascenariowheretheproductivityprocessoffirmsisrelatedwiththeirage,Ianalyzetheimportanceoffinancialdevelopmentin explainingthelargecross-countrydifferencesinfirmdynamicsandaggregateoutcomesinscenarios withandwithoutLearning. 4.1 Calibration A common strategy in quantitative evaluations of the effect of financial frictions is to replicate a relatively undistorted economy, and then use these parameters to evaluate the variations in financial development. Therefore, I calibrate the exogenous parameters so that the benchmark economy matcheskeyaspectsofUKfirmdata. Beforetheparameterization,Imustspecifyafunctionalformfortheproductivityprocessaswell as the mechanism relating firm age with the productivity process. I assume that in the benchmark scenariowherethereisnorelationbetweenthefirmageandTFPR-definedasNo-Learning-thelog productivityofathefirmznl isthesumoftwocomponents: i;t znl = A(cid:28) +" i;t i;t i;t whereAisamorepersistentcomponentwhile"isaperiodicshock. IthinkofAascapturingthefundamentalqualityofanideaforaprojectandassumethatexp(A )isdrawnfromaParetodistribution i Pr(exp(A ) < x) = 1 x (cid:22). TheshapeoftheParetodistributionisgivenby(cid:22)andwillbeestimated. i (cid:0) (cid:0) Althoughthisproductivitycomponentispersistent,eachperiodagentshaveaprobability(cid:20)drawing anewindependentideamakingthemforgotheoldone.31 Thesecondcomponentofproductivity"followsanAR(1)process " = (cid:26)" +(cid:23) i;t i;t 1 i;t (cid:0) 31The inclusion of the somewhat permanent component A, allows the model to replicate the large autocorrelation of salesverifiedinthedata. Theidea/deathshockiscommonintheliteratureandallowsthemodeltoreplicatethefactthat largefirmsexitthemarket. 15

whereshocks(cid:23) aredrawnfromanormaldistribution (cid:23) N 0;(cid:27)2 i;t (cid:23) (cid:24) (cid:0) (cid:1) To capture the relation between firm age and TFPR, I assume that firms have a periodic productivity adjustment s(cid:28) with a probability p(cid:28) which is dependent on firm age/experience (cid:28). The log productivityoftheprojectzl inthisLearningscenariois i;t zl = znl +s(cid:28) i;t i;t s(cid:28) withprob. p(cid:28) s(cid:28) = i;t 8 0 withprob. 1 p(cid:28) < (cid:0) : Furthermore,andtoreducetheage(cid:28) statespace,Iassumethatprojectscanbeeitheryoungormature, and that mature projects have no adjustments sm = 0 . In the period in which workers change i;t their occupation and become entrepreneurs the (cid:16) project (cid:17) is considered young. Thereafter, and if the agentremainsanentrepreneur,projectshaveaperiodicprobability(cid:25)(cid:28) ofbecomingmature. Oncethe projectismature,itremainssoaslongastheagentcontinuesanentrepreneur. Tomeasuretheimpactoftheproductivity/agerelationIcalibrateLearningandNo-Learningeconomies. Fromthe11parameters,fivearesettostandardvaluesintheliterature. AsMidriganandXu(2013), Iset: thecoefficientofrisk-aversion(cid:16) to1, thesubjectivediscountrate(cid:12) to0.92, therisklessrateR f to0.04,theone-yeardepreciationrate(cid:14)to0.06. thecurvatureoftheproductionfunction(cid:11)+(cid:13) to0.85 andtheserialcorrelationparameter(cid:26)to0.35. Iamthusleftwithsix(four)parametersintheLearning(No-Learning)scenariowhicharespecific tothestudy (cid:22);{;(cid:20);(cid:27);s;p andIcalibratethemtomatchsix(four)relevantmomentsoftheUKdata: f g Theemploymentshareofthetopdecileoffirms,thedefaultrateofsmallandmediumfirms,theexit rate of mature firms, the standard deviation of the log value-added of mature (all) firms, and in the Learningscenario,thestandarddeviationoflogvalue-addedofyoungfirms In the model, and for identification purposes, all parameters indirectly impact all moments in a non-linearfashion. Nevertheless,somemomentsareimpactedmorestronglybycertainparameters. Therefore,IcalibratetheshapeoftheParetodistribution(cid:22)tomatchtheemploymentshareofthetop decile of firms. The residual recovery share { allows the simulations to replicate the default rate of smallandmediumfirms. Theprobabilityofreceivinganewidea(cid:20)aimsatreplicatingtheempirical exit rate of mature firms (age > 10). The variance of the productivity process in the Learning (No- 16

Learning)scenarioissetsothatthemodelreplicatesthestandarddeviationofsalesgrowthofmature (all)firms. TheestimatedparametersandtargetmomentsarereportedinTables3and4respectively.32 Overall,bothcalibrationsaresuccessfulinmatchingthetargetmomentsinthedata. Themodelcaptures well the distributions of income and firm size, as well as the standard deviation of output growth ratesandtheproductivityofoldrelativetoyoungfirms. Asexpected,intheLearningcalibrationbothp > 0ands < 0implyingthatyoungfirmsdomake costly productivity mistakes: I also note that the No-Learning calibration cannot capture neither the decrease in the volatility of sales as firms age, nor the increase in average firm productivity of older firms. 4.2 Learning vs. No-Learning Beforecomparingtheempiricalresultswiththemodelpredictionsitisconvenienttounderstandits basicmechanics. InFigure4,IexhibitaseriesofpolicydecisionsinaneconomywiththeUKrecovery rate ((cid:24) = 0:85) in scenarios with and without Learning:33 The first panel, displays the behavior of UK the ratio of investment to assets. Investment rate is similar in the incumbent and in the No-Learning scenarios,andissignificantlylowerforanentrantfirm. Tounderstandifthisisindicativeoffinancial constraints, I follow Midrigan and Xu (2013), and graph in the second panel the average product of capital (APK - ratio of sales to assets). As expected, the APK of an entrant firm is higher than that of an incumbent suggesting stronger constraints. In the fourth panel, I depict the leverage ratio. As expected,entrantfirmshavethelowestleverageratiosduetotheirhighercostofexternalfinance. Figure5,contraststhedynamicsofafirminitsfirsttenyearsinscenariosunderbothLearning/No- Learning.34 The first panel, exhibits the evolution of firm assets. In the No-Learning scenario firm growth is much faster. A firm with four years is almost three times larger in this scenario and is almost twice as large when it reaches ten years. The second panel suggests that the main reason for these size differences is that firm debt can be up to five times larger in the No-Learning scenario. In thethirdpanel,Iexhibittheleverageratiosforthetwoscenarios. Inbothscenarios,firmsstartwith 32ThemajorityofmomentswereobtainedusingtheinformationonUKfirmsdrawnfromAmadeus.Theexceptionsare: top5percentincomeshare,averagefirmemploymentandfirmexitrate. Fortheincomeshareofthetop5percentweuse informationinAlvaredo,Atkinson,PikettyandSaez(2013),whileforthefirmexitrateandaveragefirmemploymentwe usedtheUKindustrialdemographicstatisticsfromtheOECDStructuralStatisticsforIndustryandServicesfor2006. 33Inbothscenarioswedepictthechoicesofatalentedentrepreneurwiththesameamountofdebtforvariouslevelsof assets.Furthermore,weassumethattheentrantentrepreneurismakingnomistakes. 34Underbothscenarios,firmshavethesame(high)productivitythroughoutthe10years. Furthermore,thefirmiskept anentrantintheLearningsimulation. 17

similarandhighleverage. Asfirmsgrowolderthisleverageratiodecreasesinbothscenarios,albeit at a slightly faster pace in the Learning scenario. Nevertheless, this faster pace is reversed around the 10th year (not shown) and the leverage ratios become larger in the Learning scenario due to the lowervarianceinproductivityshocks. Finally,inthefourthpanel,IdepicttheAPK.Asbefore,APK is larger in the Learning scenario throughout the lifecycle of the firm especially for younger firms suggestinghigherfinancialconstraints. To analyze the relative success of both versions of the model in replicating the UK distribution of firms, I report in Table 5 additional empirical and theoretical moments not directly targeted by the calibration. For now, I focus in both Learning/No-Learning scenarios with a RR = 0:85. Both UK Figures 4 and 5, suggested that productivity shocks under Learning increased financial constraints, forcing firms to start smaller. The results reported in Table 5 support this idea, since the share of sales of young firms ( 5 years) is 20 percent in the No-Learning scenario and only 12 percent under (cid:20) Learning, which closely matches the 11 percent in the UK data. Regarding the use of debt, both versionsofthemodelareabletomimicitsgeneraldynamics. Inthedata,theratioofaggregatedebt tototalassetsforyoungfirmsis0.81anddecreasesforolderfirmsto0.72. Bothversionsofthemodel replicate the decrease of leverage as firms age, but the dynamics under Learning are better able to mimic the general pattern, especially for the older cohorts. As discussed above, the cost of debt is strongly impacted by the variance of shocks.35 This implies that while the average leverage ratios yng yng for youngfirms area touch lowerin theLearningscenario (Lev = 0:83vs. Lev = 0:82) for No Lrn Lrn (cid:0) olderfirmsthisresultisreversed(Levold = 0:48vs. Levold = 0:59). No Lrn Lrn (cid:0) 4.3 FinancialDevelopmentandFirmDynamics In this section, and focusing on the Learning scenario, I investigate the impact of recovery rate, the proxy for financial development, on firm dynamics and aggregate outcomes. First, and for illustration purposes, I present a series of exhibits representing the decision rules in scenarios with three different levels of recovery rate.36 In Figure 6, I illustrate the evolution of a series of firm characteristics - assets, debt, leverage and APK - during its first 10 years, while keeping the productivity constant. Asexpected,inthefirsttwopanelsfirmsineconomieswithhigherrecoveryratestartsignificantlylarger,sincetheyareabletoaccesscheaperdebt. Inthethirdpanel,Igraphtheevolutionof 35IntheLearningscenarioyounger(older)firmshaveahigher(lower)varianceintheirproductivityshocksrelativeto theNoLearningversion. 36Thethreerecoveryrateschosenwere 0;0:25;0:85 . Thefirstlevelmimcksaneconomywithoutinvestorprotection, f g thesecondrecoveryrateisthemedianvalueofthelowestquintileandcorrespondstotheRussianrecoveryrate,whilethe finalrecoveryrateisthatoftheUK,thebenchmarkeconomy. 18

firmleverage. Inthescenariowithhighestrecoveryrate(solidline)leverageratiostartsat1,whereas inthescenariowiththeRussianrecoveryrate(dottedline)firmsstartwithaleverageratioof0.5. In allscenarios,firmsdeleverageastheyage. Animportantconsequenceofthedifferencesintheaccess to finance is illustrated in the last panel. Using the APK as a proxy of financial constraints, I note thattheratio APK(cid:24)=0 forentrantfirmsis4anditonlyconvergesto1whenfirmsareolderthan12 APK(cid:24)=0:85 years-old (not shown). Therefore, Figure 6 suggests that financial constraints are dependent on the recoveryrateandhavethestrongesteffectonyoungfirmsunderLearning. VariationsinRecoveryRateunderLearning andNo-Learning AsafirstpassontheimpactofrecoveryratesunderbothLearning/No-Learningscenarios,Iextend Tables4and5toincludemomentsineconomieswithlowfinancialdevelopment(i.e. (cid:24) = 0:25). The four main messages are: i) employment concentration decreases with financial development, especiallyintheLearningscenario. WhileunderhighfinancialdevelopmentinbothLearning/No-Learning scenarios, thetop-10percentoffirmsemployed69percentofworkers, theconcentrationissubstantially higher in the Learning scenario under low financial development (0.86 vs. 0.71). This result is partly driven by the fact that rich-productive entrepreneurs increase their scale due to the lower competition for workers from young firms. For example, under low financial development in the Learningscenario,youngfirmswithlessthanfiveyearsaccountforonlythreepercentoftheshareof sales; ii) there are lower entry/exit rates in the Learning scenario under low financial development. In this scenario young firms (i.e. <10 years) represent less than 10 percent of total firms, whereas in the Learning scenario they represent more than 60 percent. This occurs because workers only become entrepreneurs in those rare instances where they have a highly productive idea. Therefore, in this scenario young firms are scarce but they start relatively large.37 Consequently, the lack of competition from entrants under low financial development in the Learning scenario leads to low exit rates of old firms.38 Rich entrepreneurs, even when they suffer a bad and permanent productivity shock, tend to continue operating due to the low wages prevalent in the economy; iii) the relative productivityofolderfirmsissignificantlylowerintheLearningscenariowithlowfinancialdevelopment. This result is directly related with the previous point. In this scenario, rich but unproductive entrepreneurs are not selected out and continue operating. In fact, and under the Learning scenario, older firms in the economy with high financial development benefit from the reduction in mistakes 37Theaveragesizeofafirmwithlessthanfiveyearsisof33employees.IntheNo-Learningscenariotheaveragenumber ofemployeesis5. 38For example, under low financial development the exit rate is of only 1 percent under Learning, compared with 23 percentunderNo-Learning. 19

and are on average 11 percent more productive. However in the No-Learning scenario, older firms are on average 32 percent less productive, since entrepreneurs hit by bad-permanent shocks do not have the incentives to exit; and finally iv) access to external financing is highly dependent on financialdevelopment. InbothLearning/No-Learningscenarios,ineconomieswithhighrecoveryrateboth interestratesare7p.p. loweronaveragewhiletheleverageratiosareonaveragethreetimeshigher. Thereforevariationsintherecoveryratehavelargerealeffectonthefinancingoffirms. 4.3.1 FirmDynamics To further attest for the relative importance of the relation between firm TFPR and firm age, in this section, I compare the cross-country behavior of other firm dynamics (e.g. TFPR and leverage) in economies with varying levels of recovery rate in scenarios with and without Learning. I start by comparinggraphicallythepredictionsofthesimulatedmodelwiththedata. AfterwardsItestmore formallytheresults,bothintheempiricalandsimulateddata,byestimatingregressionsofthetype x = (cid:12) +(cid:12) logage +(cid:12) logage FD +ind year ctry+v (10) i;t 0 1 i;t 2 i;t c i;t (cid:2) (cid:2) (cid:2) where x is a firm-level variable (e.g. leverage) of firm i, on year t. The explanatory variables are age which indicate the age in years of firm i at time t, while FD corresponds to a measure of i;t c financial development.39 Finally, I also include a series of dummies ind year ctry to control for (cid:2) (cid:2) industry, year and country fixed-effects. The industry fixed-effects, control for all industry features such as capital intensity and tradability that may induce different firm behavior across industries. The year fixed-effects control all year specific characteristics such as the phase of the business cycle in an economy. Finally, the country fixed-effects, control for all country characteristics such as the level of financial development, accounting practices and institutional quality that may impact firm policies. EvolutionofTFPR In this paper, I am interested in the relation between the TFPR of firms and their age. As a first pass into the predictions of the model I display in Figure 7 the relation between firm age and the averagelogTFPRintheeconomieswithhigh/lowrecoveryrate.40 39Inourbenchmarkregressionsweusetherecoveryrateasanindicatoroffinancialdevelopment.Intheappendix,and forrobustnesspurposes,werunallregressionsusingalternativeindicatorsoffinancialdevelopment. 40Due the lack of materials data for a large fraction of firms in lower income countries, the measure of TFPR used in thissectionfollowsSyverson(2004)andusesexpendituressharestocalculatetheelasticitiesofthefactorsofproduction. Detailsofthisapproachareprovidedintheappendix. 20

Inthefirstpanel, IgraphtheaverageTFPRacrossage-cohortsintwosimulatedeconomieswith varying levels of recovery rate (RR = 0:25;0:85 ) under No-Learning. In this framework, the selecf g tion on productivity is similar under the two recovery rates and average TFPR remains relatively unchanged across the age-cohorts. In the second panel, I present the dynamics of firm TFPR under theLearningscenario. Asexpected,giventhestructureofproductivityintroduced,theaverageTFPR is higher for older cohorts in both economies as firms reduce their mistakes with age. Nevertheless, while the increase in average TFPR is concave and increasing in the economy with high RR, for the economy with low RR the average TFPR presents an inverse U-shape. The main reason for this differenceisthelowerselectiononproductivityincountrieswithworseRR.Thisoccurswhenwealthy entrepreneursarehitwithanegativepermanentproductivityshockandprefertocontinueoperating benefiting from the relatively low wages in the economy. In the third panel, I present the empirical comparison. Inthedata, theaverageTFPRincreaseswithfirmageinthehighRRcountriesandthe averageTFPRofthe15-yearcohortis20percenthigherthanthatofanentrantcohort. Interestingly however,thebehavioroftheaverageTFPRincountrieswithlowerRRhasaninverseU-shape. This mimicsverycloselythebehavioroffirmsunderLearning,renderingsupporttotheabovementioned mechanism. In Figure 8, I present the same relation as in Figure 2. However, instead of using four countries with high recovery rate, I follow the evolution of average TFPR in four countries with low levels of recovery rate.41 As predicted from Figure 7, the evolution of TFPR for these countries presentsamuchweakergrowth,andinmanycasesadecrease,thanintherichercountries. AccordingtothemodelunderLearning,thisisdrivenbythepoorerselectiononproductivity. Ithenrunthe moreformalequation10usingtheempiricalandsimulateddataandtheresultsareinTable7. They confirm the analysis from the graphs. The average TFPR increases with firm age with an elasticity around0.05andthiseffectissignificantlystrongerineconomieswithhigherrecoveryrate. DispersionofTFPR The main premise of this paper is that the variance of firm productivity decreases as firms grow older due to Learning. In the model, I introduced this concept by assuming that as firms aged they reducedthefrequencyandmagnitudeoftheirmistakes. InFigure9,Icomparethestandarddeviation of TFPR growth across age-cohorts both in the simulated model and in the data.42 In the first two panels, I display the results of the simulations. Under No-Learning the TFPR dispersion tends to 41Thecountriesare: Ukraine, Romania, CzechRepublicandHungary. Thesecountrieswerechosenastheyhavelow levelsofrecoveryrateandhavealargenumberoffirms(i.e.>1000)acrossall15age-cohorts.Ialsolimitedtheanalysisto theage-cohortsupto15years,totruncatethosefirmscreatedduringthecommunistera. 42ThestandarddeviationofTFPRismeasuredasthestandarddeviationofthelog-TFPRgrowthrateforeachage-cohort. 21

increaseslightly. Conversely,underLearningtheTFPRdispersiondecreasesasfirmsageandreduce theirmistakes. ThereasonforthisdifferenceisthatunderLearningasfirmsageandbecomemature the variance of their shocks decreases. Conversely under the No-Learning scenario, where variance oftheproductivityprocessisindependentofage,thedispersioninTFPRincreasesatouchwithage sincelargerfirmswhicharehitbyaslightlynegativeproductivityshockmaycontinueoperating. In thethirdpanel,IpresenttheempiricalresultsoftheevolutionofthestandarddeviationofTFPRfor all firms in the sample. In both sets of countries - with high and low RR - the standard deviation of theloggrowthoffirmTFPRdecreaseswitholdercohorts. Again,theseresultsaresupportiveofthe productivityprocessintheLearningscenario. Again, the results from the regressions in Table 7 confirm the above analysis. The empirical regression indicates that the elasticity between age and the standard deviation of TFP is negative and between-0.06and-0.25. Leverage Sincethefocusisontheimpactoffinancialconstraintsonfirmbehavior,Iexaminehowleverage - ratio of total liabilities to total assets - varies with different levels of recovery rate. As before, I use themodeltoobtainafirstpassintothepredictionsofthemodel. InthefirsttwopanelsofFigure10 Idisplaytheaverageleverageratioacrossagecohortsintwoeconomieswithvaryingrecoveryrates under both Learning and No-Learning. Both scenarios paint a similar picture. Entrepreneurs start a project with relatively high leverage which decreases as firms age. Furthermore, and even though firms start with similar leverage ratios across economies they tend deleverage faster under low RR due to the higher cost of debt. An important difference between the Learning/No-Learning scenarios, and due to the lower risk of older firms under Learning, leverage tend to decrease more slowly especiallyineconomieswithbetterrecoveryrate. Thethirdpanel,againprovidestheempiricalcomparison. Empirically,thebehavioroftheleverageratioisrelativelysimilartotheLearningsimulation. 43 The results from the regression are reported in Table 7. They confirm that the elasticity of age to leverage is negative and around -0.05. Nevertheless, and unlike the model results, the crossderivativebetweenRRandlog-ageisslightlynegativeimplyingthatfirmsineconomieswithbetter RRdeleveragefaster.44 43These results are in line with existing findings in the literature. Huynh and Petrunia, using a dataset of Canadian manufacturingfirms,findthatthattheleverageratioofthesurvivingfirmsdecreaseswithfirmage. 44InTable9thisresultisreversedwhenweusealternativeindicatorsoffinancialdevelopment. 22

4.4 FinancialDevelopmentandAggregateProductivity Inthissection,Ianalyzetheimpactthatrecoveryratehasonaggregateproductivityandfactormisallocation,underLearningandNo-Learningscenarios. AggregateProductivity Oneofthemostinterestingfeaturesofthemodel,isthatitallowsustoquantifytheimpactLearninghasontherelationshipbetweenfinancialdevelopmentandaggregateproductivity.45 Inthissection,IvarytherecoveryrateandcalculatethecorrespondingaggregateTFPunderbothLearning/No- Learning scenarios. In Figure 11, I report the effect of these financial frictions on aggregate TFP and on that of young and old firms.46 As in the data, the cross-country differences in income per capita (not shown) are mainly driven by variation in aggregate TFP. Variation in RR can reduce aggregate TFP by 8 percent under No-Learning and 16 percent under Learning. Therefore, introducingLearning in a standard model with firm heterogeneity doubles the impact of financial development. Delving deeperintothisresult, inthesecondandthirdpanelsIdividethesampleintooldandyoungfirms. Inthesecondpanel, IpresenttheaggregateTFPoffirmsyoungerthansix-years. IntheNo-Learning scenariotheaggregateTFPofyoungfirmsincreasesslightlywithlevelofrecoveryrate.47 However, under Learning the aggregate TFP actually decreases with recovery rate. This result is driven by the factthatunderLearningandineconomieswithlowrecoveryrate,workers-whohavelowaccessto externalfinancing-onlybecomeentrepreneursinthoserareinstanceswheretheyhaveanextremely productive project. In the third panel, I do the same exercise but for the older firms. In this case the resultunderLearningisreverted,withtherecoveryratebeinghighlyrelatedwiththeaggregateTFP. Inthisscenario,anincreaseinrecoveryrateleadstoariseinaggregateTFPofaround13percent. As noted in the next section, this result is driven by the low exit rates of large-unproductive entrepreneurs. Finally, in the last panel, I graph the relative aggregate TFP of old to young firms. Whereas theratioisconstantandclosetooneunderNo-Learning,itishighlyincreasingwiththerecoveryrate underLearningfortheabovementionedreasons. Furthermore,andgiventhattheratioislowerthan one, it implies that under Learning the aggregate TFP of old firms is significantly lower than that of youngerones. Tocomparethesimulationswiththedata,inFigure12Ireplicatethelastpanelusing the firm level in Amadeus.48 Interestingly, and as predicted by the model in the Learning scenario, 45TocalculateaggregateTFP,IusethedataprovidedbyPWT6.1. TFPisYK(cid:0) 1=3L(cid:0) 2=3,whereY isGDP-PPP,K isthe capitalstock,andListhenumberofworkers. Icalculatethecapitalstockusingtheperpetualinventorymethodata6% depreciationrate. 46Foreachscenario,Iransimulationsfor12differentlevelsofrecoveryrate. 47InthenextsectionsIshowthatthisisdrivenbyvariationsininputmisallocation. 48Fordataconsistency,Iexcludedfirmsthatdidnothaveinformationonassets,materials,employmentandsales. Fur- 23

the aggregate TFP of old relative to young firms is increasing in RR due to the reduction in productivity mistakes. This effect is less prevalent in economies with lower RR due to weaker selection on productivity. FirmTurnover To further analyze the above results, in Figure 13 I exhibit a series of graphs whose main messageis: ineconomieswithlowfinancialdevelopmentundertheLearningscenario,rich-unproductive entrepreneurs do not exit the market, crowding-out younger more productive, albeit poorer, entrepreneurs who find it costly to obtain financing. In the first panel, I exhibit the share of production of young firms with less than six years. In both scenarios, this share increases with RR, however the level is substantially lower under Learning by around 10 percentage points. In fact, when (cid:24) = 0, this share is of only 2.5 percent, compared with 19 percent when (cid:24) = 1. The second and third panels, display the fraction of young firms and the exit rate of old firms respectively. Not surprisingly, both panels paint a similar picture. In economies with low RR under Learning, both the share of youngfirmsandtheexitrateofolderonesareextremelylow. Around5and0.5percentrespectively. Conversely, under No-Learning there is more turnover and it is roughly constant across all levels of financial development, as the exit rate of older firms is roughly determined by the frequency of the permanentshock. CapitalMisallocation As Hsieh and Klenow (2009) noted, dispersion in marginal productivity of capital (MPK) is an importantsourceofmisallocationleadingtolossesofaggregateTFPandincomepercapita.49 Comparing simulations and in the model, I present in Table 6, the theoretical and empirical moments of theAPK. In Panel A of Table 6, I present the APK for a series of simulations in economies with varying levels of financial development. In economies with lower recovery rate, not only is the APK higher onaverageforallfirms,butitisespeciallyhigherforyoungandproductivefirms. Inadditiontothe averageAPK,themodelsimulationsindicatethatfirmsineconomieswithlowerrecoveryratehave ahigherdispersionofaverageproductivityofcapitalconsistentwithhigherinputmisallocation. In Panel B, I present the average APK for firms in the 27 countries in the sample. I divided the countries into two groups depending on their recovery rate. I ended up with 15 High RR and 12 thermore,Iexcludedcountrieswithinformationonfewerthan5000firms. Theseconditionsledtotheexclusionofnine countries,notablytheUKwhichdoesnothavedataonmaterialsused. 49Inouranalysiswefocusontheaverageproductofcapital.Inaone-sectormodelunderCobb-Douglasthetwoconcepts areproportional. 24

LowRRcountries.50 Aspredictedbythemodel,theaverageAPKishigher,17percent,ineconomies with worse financial development and it is also higher for younger firms. Furthermore, and also in accordance with the model, the average APK of younger firms is relatively larger in countries with lower RR, 18 percent vs. 10 percent. Finally, the model simulations are consistent with youngproductive firms being more constrained in economies with low RR. Again, this fact holds in the data. The APK of young-productive firms in countries with low RR is 71 percent higher relative to allfirmsinthosecountries,anditis"only"51percentincountrieswithhighRR.IalsofollowBuera, Kaboski and Shin (2011) in using the standard deviation of the log(APK) of firms as an additional measureoffactormisallocation. Asinthemodel,inthedatathereisahigherdispersionforyounger firms,1.38,thanforolderfirms,1.10. FactorMisallocationandOutputLoss To understand the source of the cross-country differences in aggregate productivity, I follow Buera, Kaboski and Shin (2011) and decompose factor misallocation within and across age-groups under scenarios of Learning and No-Learning.51 First, I estimate the impact of the capital and labor misallocation within age-cohorts by reallocating these factors of production across firms within the samecohortsothatthemarginalproductivityofcapital(MPK)andlabor(MPL)areequalizedacross firms within that cohort. In the second step, I calculate the misallocation of factors of production acrossage-groups,byreallocatingthefactorsofproductiontoequalizebothMPKandMPLacrossall firms. The results of this exercise are reported in Figure 14. In the first (second) panel I report the results under No-Learning (Learning). The main conclusions to draw from this exercise are: i) factor misallocationsaredecreasingwiththeRRandhigherintheLearningscenario. ii)inbothscenariosthe factormisallocationwithinage-cohortsaccountsforallaggregateTFPlossesineconomieswithhigh levelsofRR,whereasineconomieswithlowRRitaccountsforaround75percentofthelosseswith theremaining25percentduetomisallocationacrossage-groups. Thisisequivalenttosaythattransfers of factors of production from old to young firms would improve aggregate TFP by 5 percent in theseeconomies. iii)UndertheNo-Learningscenario,themajorityofthefactormisallocationiswithin age-groups. Under this scenario there would be no improvements in aggregate TFP by transferring factorsofproductionacrossage-groups. 50Wedividedcountriesintohigh(low)recoveryrateiftheirvaluewasabove(below)0.4. Theresultsarenotimpacted qualitativelyifwehadchoseneither0.3or0.5asalternativecutoffvalues. 51Isplitfirmsintoyoungandoldusing8-yearsasathresholdwhichsplitsthefirmsintotwogroupswithroughlythe samenumberofobservations. 25

5 Conclusion I have documented two empirical regularities regarding the dynamics of the TFPR of firms as they age. I found, for a broad set of countries, that the TFPR tends to increase as firms age while the volatility of the productivity process decreases. I created a quantitative framework to analyze the importanceofthismechanism-whichIdefinedasLearning-andnotedthatbyincludingitallowed the model to better match some of its empirical moments such as the dynamics of leverage or the shareofproductionofyoungfirms. UnderstandingthemechanismsdrivingthisLearningprocess,as wellasthefactorsdeterminingitsdynamics,isapossibleavenueforfutureresearch. 6 References References [1] Alvaredo,F.,A.Atkinson,T.PikettyandE.Saez(2013)-"TheTop1PercentinInternationaland HistoricalPerspective",JournalofEconomicPerspectives,27: 3-20. [2] Amaral, P., and E. Quintin (2010) - "Limited Enforcement, Financial Intermediation and EconomicDevelopment: AQuantitativeAssessment",InternationalEconomicReview,51: 785-811. [3] Arellano, C., Y.Bai, andJ.Zhang(2012)-"FirmDynamicsandFinancialDevelopment", Journal ofMonetaryEconomics,59: 533-549. [4] AwB.,X.ChenandM.Roberts(2001)-"Firm-LevelEvidenceonProductivityDifferentialsand TurnoverinTaiwaneseManufacturing",JournalofDevelopmentEconomics,66: 51-86. [5] Bahk,B.andM.Gort(1993)-"DecomposingLearningbyDoinginNewPlants",JournalofPoliticalEconomy,101: 561-583. [6] Benkard L. (2000) - "Learning and Forgetting: The Dynamics of Aircraft Production", American EconomicReview,90: 4034-1054. [7] Buera, F., J. Kaboski and Y. Shin (2011) - "Finance and Development: A Tale of Two Sectors", AmericanEconomicReview,101: 1964-2002. [8] Castro, R., G. Clementi and G. MacDonald (2009) - "Legal Institutions, Sectoral Heterogeneity andEconomicDevelopment",ReviewofEconomicStudies,76: 529-561. 26

[9] Clementi,G.andH.Hopenhayn(2006)-"ATheoryofFinancingConstraintsandFirmDynamics",QuarterlyJournalofEconomics121: 229-265. [10] De Loecker, J. (2007) - "Do Exports Generate Higher Productivity", Journal of International Economics,69-98. [11] Djankov,S.,O.Hart,C.McLieshandA.Shleifer(2008)-"DebtEnforcementaroundtheWorld", JournalofPoliticalEconomy116,6: 1105-1149. [12] Erosa, A., and A. Hidalgo-Cabrillana (2010) - "On Capital Market Imperfections as a Source of LowTFPandEconomicRents",InternationalEconomicReview,49: 437-473. [13] Evans, D., (1987) - "The Relationship Between Firm Growth, Size, and Age: Estimates for 100 ManufacturingIndustries",TheJournalofIndustrialEconomics,35: 566-581. [14] Foster, L., J.HaltiwangerandC. Syverson (2008)-"Reallocation, FirmTurnoverandEfficiency: SelectiononProductivityorProfitability?",AmericanEconomicReview,98: 394-425. [15] Foster, L., J. Haltiwanger and C. Syverson (2013) - "The Slow Growth of New Plants: Learning aboutDemand",mimeo. [16] Gomes,J.andL.Schmid(2010)-"LeveredReturns",TheJournalofFinance,2: 467-494. [17] Greenwood, J., J. Sanchez, and C. Wang (2013) - "Quantifying the Impact of Financial DevelopmentonEconomicDevelopment",AmericanEconomicReview,100: 1875-1891. [18] Hall,R.andCharlesJones(1999)-"WhyDoSomeCountriesProduceSoMuchMoreOutputper WorkerthanOthers?",QuarterlyJournalofEconomics,114,83-116. [19] Harvey, A. (1976) - "Estimating Regression Models with Multiplicative Heteroscedasticity", Econometrica,44: 461-465. [20] Hennessy,C.andT.Whited(2007)-"HowCostlyisExternalFinancing? EvidencefromaStructuralEstimation",JournalofFinance,62: 1705-1745. [21] HopenhaynH.(1992)-"Entry,Exit,andFirmDynamicsinLongRunEquilibrium",Econometrica, 60: 1127-1150. [22] Hsieh, C. and P. Klenow (2009) - "Misallocation and Manufacturing TFP in China and India", QuarterlyJournalofEconomics,124: 1403-1448. 27

[23] Huergo, E. and J. Jaimandreu (2004) - "How Does Probability of Innovation Change with Firm Age",SmallBusinessEconomics,22: 193-207. [24] Hyunh,K.andR.Petrunia(2013)-"AgeEffects,LeverageandFirmGrowth",JournalofEconomic DynamicsandControl,forthcoming. [25] Jensen, J.andR.McGuickinandK.Stiroh(2001)-"TheImpactofVintageandSurvivalonProductivity: EvidencefromCohortsofU.S.ManufacturingPlants",ReviewofEconomicsandStatistics,83: 323-332. [26] Jeong,H.andR.Townsend(2007)-"SourcesofTFPGrowth: OccupationalChoiceandFinancial Deepening",EconomicTheory,32: 197-221. [27] Jovanovic B. and Y. Nyarko (1996) - "Learning by Doing and the Choice of Technology", Econometrica,64: 1299-1310. [28] King, R. and R. Levine (1993) - "Finance, Entrepreneurship and Growth: Schumpeter Might Be Right",QuarterlyJournalofEconomics,108: 717-737. [29] Levine, O. and M. Warusawitharana (2013) - "Finance and Productivity Growth: Firm-Level Evidence",WorkingPaper. [30] Levinsohn, J., and A. Petrin (2003) - "Estimating Production Functions Using Inputs to Control forUnobservables",ReviewofEconomicStudies,70: 317-341. [31] Levinsohn, J. and M. Melitz (2006) - "Productivity in Differentiated Products Market Equilibrium",mimeo [32] Levitt, S., J. List and C. Syverson (2013) - "Toward an Understanding of Learning by Doing: EvidencefromandAutomobilePlant",JournalofPoliticalEconomy,forthcoming [33] OECD(2006)-"StructuralandDemographicBusinessStatistics",EntrepreneurshipIndicatorsProgram [34] Olley,S.andA.Pakes(1996)-"TheDynamicsofProductivityintheTelecommunicationsEquipmentIndustry",Econometrica,64: 1263-1297. [35] Midrigan, V. and D. Xu (2013) - "Finance and Misallocation: Evidence from Plant-Level Data", AmericanEconomicReview,forthcoming 28

[36] Syverson,C.(2004)-"MarketStructureandProductivity: AConcreteExample",JournalofPoliticalEconomy,112: 1181-1222. [37] Thornton, R. and P. Thompson (2001) - "Learning from Experience and Learning from Others: AnExplorationofLearningandSpilloversinWartimeShipbuilding",AmericanEconomicReview, 1350-1368. [38] WarusawitharanaM.(2012)-"ResearchandDevelopment,ProfitsandFirmValue: AStructural Estimation",mimeo. [39] Wooldridge,J.(2002)-"EconometricAnalysisofCrossSectionandPanelData",MITPress. 29

7 Figures and Tables Figure1: IndicatorsofFinancialDevelopment Financial Development per Country Alternative indicators 2 5 .1 1 5 . 0 IceDenNed UK Ire Prt SpnGerSwe Fra Ita Nor Bel Gre Fin EstCro LatHun Bul Lit Czh Slk Pol Srb RusRomUkr Debt/GDP Recovery rate Financial Depth Note:Debt/GDPisthetotalprivatedebttoGDP.FinancialDepthismeasuredastheratioofliquidliabilitiesofthefinancial system to GDP. Recovery rate is the expected amount recouped by creditors per dollar lent to a defaulting debtor. All indicatorsarecountryaveragesforthe2003-2006period. 30

Figure2: AverageTFPRacrossage-cohorts UK 0.2 0 0.2 0.4 0 5 10 15 20 Firm Age RPFT egarevA France 0.2 0 0.2 0.4 0 5 10 15 20 Firm Age RPFT egarevA Italy 0.2 0 0.2 0.4 0 5 10 15 20 Firm Age RPFT egarevA Spain 0.2 0 0.2 0.4 0 5 10 15 20 Firm Age RPFT egarevA Note: ThefigurereportstheaveragedemeanedTFPRforeachage-cohortforthefourcountriesinthesamplewiththe largestnumberoffirms.Thedashedlinesdelimitthe95percentconfidenceinterval. Table1: Regression-FirmAgeandTFPR UK France (1) (2) (3) (1) (2) (3) Benchmark FirmFE Heckit Benchmark FirmFE Heckit 0.01 0.05 0.518 0.10 0.06 0.25 ln(age) (0:004) (0:018) (0:047) (0:017) (0:017) (0:026) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3) Obs. 31,127 31,127 12,325 113,175 113,175 42,097 Italy Spain (1) (2) (3) (1) (2) (3) Benchmark FirmFE Heckit Benchmark FirmFE Heckit 0:19 0:16 0:83 0:25 0:17 0:17 ln(age) (0:001) (0:006) (0:009) (0:002) (0:005) (0:008) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) Obs. 256,995 256,995 124,296 267,819 267,819 108,748 Note: ThistablepresentstheresultsofregressingtheTFPRofafirmonitslog-ageforthe2006-2009period. FirmTFPR wasestimatedusingthemethodproposedbyLevinsohnandPetrin(2003). Thebenchmarkregressiondoesnotinclude firmfixed-effectswhereastheothertwodo. Allregressionshaveindustry-yearfixed-effects. Heckitregressionscontrol forselectionusingtheHeckmantwo-stepprocedure.Duetolowerqualityintheselectionindicator,theHeckitregression usesasamplefrom2007-2008.Theprobitinthefirststepregressestheselectionindicatoronthefirmproductivity,sizeand age. 31

Figure3: Standarddeviationofsalesgrowthandfirmage UK 0.5 0.4 0.3 0.2 0 5 10 15 20 Firm Age RPFT .veD .dtS France 0.5 0.4 0.3 0.2 0 5 10 15 20 Firm Age RPFT .veD .dtS Italy 0.5 0.4 0.3 0.2 0 5 10 15 20 Firm Age RPFT .veD .dtS Spain 0.5 0.4 0.3 0.2 0 5 10 15 20 Firm Age RPFT .veD .dtS Note: ThefigurereportstheaveragestandarddeviationofTFPRgrowthasforeachage-cohortforthefourcountriesin thesamplewiththelargestnumberoffirms. Table2: Regression-FirmAgeandStd. Dev. ofSalesGrowth UK France Italy Spain (1) (2) (1) (2) (1) (2) (1) (2) Age-Group-(cid:28) AllFirms Heckit. AllFirms Heckit. AllFirms Heckit. AllFirms Heckit. age i;t [0;3] 0.121 0.110 0.111 0.117 0.132 0.148 0.156 0.153 2 age i;t [4;7] 0.105 0.096 0.095 0.099 0.119 0.0116 0.138 0.133 2 age i;t [8;12] 0.090 0.087 0.087 0.089 0.108 0.098 0.120 0.115 2 age i;t 13 j 0.076 0.075 0.071 0.073 0.089 0.089 0.100 0.101 (cid:21) Observations 27,727 13,505 88,991 45,827 144,347 82,742 209,344 110,505 Note: This table presents the results of the persistence and variance of the productivity process. These countries were chosenastheyhavethelargestnumberoffirms. 32

Table3: CalibratedParameters Parameter Value AssignedParameters Intertemporalelasticityofsubstitution (cid:16) 1 Risklessinterestrate R 0:04 f Depreciationrate (cid:14) 0:06 Shr. ofoutputtoentrepreneur (cid:11)+(cid:13) 0:85 Capitalshare (cid:11) 0:33 (cid:11)+(cid:13) Autoregressiveparameter (cid:26) 0:35 DiscountRate-Entrepreneurs (cid:12) 0:92 Calibratedparameters No-Learning Learning Permanentproductivity (cid:22) 6:8 6:95 Probabilityofpermanentshock (cid:20) 0:08 0:08 Residualsharekeptbydefaulter { 0:85 0:85 Stochasticshockstandarddeviation (cid:27) 0:175 0:13 " Adjustment s 0 -0:62 Probabilityofadjustment p 0 0:27 Note:Thistablereportstheparametervaluesusedinthesimulations.TheAssignedParameterswereobtained,forcomparisonpurposes, fromaseriesofrelatedquantitativestudies. TheCalibratedParametersminimizedthedistancebetweena seriesofrelevantmomentsofthesimulationsandofthedata.Learning(No-Learning),reportstheestimatedparametersin ascenariowheretheproductivityprocessis(isnot)relatedwithfirmage. Table4: TargetedMoments No-Learning Learning Data-UK Model Model Model Model TargetMoments (cid:24) = 0:85 (cid:24) = 0:85 (cid:24) = 0:25 (cid:24) = 0:85 (cid:24) = 0:25 Top10-pct. empl. share 0.69 0.69 0.71 0.69 0.86 St. dev. sales 0.61 0.61 0.56 0.68 0.63 St. dev. sales<8yrs 0.75 0.64 0.54 0.75 0.93 St. dev. sales 8yrs 0.53 0.57 0.57 0.55 0.58 (cid:21) DefaultRate<250empl 0.06 0.06 0.06 0.06 0.02 TFPRold/TFPRnew 1.12 1.03 1.16 1.11 0.67 Exitrate 10yrs 0.07 0.07 0.08 0.08 0.01 (cid:21) Note: Thistablereportsthetargetmomentsusedinestimatingtheparametersofthemodel. Valuesinitalicdenotemomentsnotuseddirectlyintheestimation.Learning(No-Learning),reportstheestimatedparametersinascenariowherethe productivityprocessis(isnot)relatedwithfirmage. 33

Figure4: Policyfunctions: Learningvs. No-Learning 1 0.9 0.8 0.7 0 20 40 60 80 Assets K etaR tnemtsevnI 0.8 0.6 0.4 No learning 0.2 Entrant Incumbent 0 40 50 60 70 80 Assets K KPA 0.14 0.12 0.1 0.08 0.06 0.04 30 40 50 60 70 80 Assets K etaR tseretnI 1 0.9 0.8 0.7 40 50 60 70 80 Assets K egareveL Note: This figure depicts the decisions of entrepreneurs with median talent and no debt for various levels of assets. It contraststhedecisionsinascenariowherethereisnorelationbetweentheproductivityprocessandfirmage(No-Learning) withonewherethereis.Furthermore,intheLearningscenarioIgraphthedynamicsofbothanentrantfirm-age<6years -andofanolderfirm.MPKisthemarginalproductivityofcapitalwhereasLeverageisdefinedastheratiooftotalliabilities tototalassets. 34

Figure5: FirmdynamicsinscenarioswithandwithoutLearning 2000 1500 1000 500 0 0 2 4 6 8 10 Age stessA 1000 800 600 400 200 No Learning Learning 0 0 2 4 6 8 10 Age tbeD 1.2 1 0.8 0.6 0.4 0.2 0 2 4 6 8 10 Age egareveL 0.9 0.8 0.7 0.6 0.5 2 4 6 8 10 Age KPA Note: ThisfigureillustratestheevolutionofafirmasitgrowsolderinascenariowithandwithoutLearning. MPKisthe marginalproductivityofcapitalwhereasLeverageisdefinedastheratiooftotalliabilitiestototalassets.Inbothscenarios, firmsstartwithoutassetsanddebt,andwiththesameproductivitythroughouttheirlife. 35

Table5: Additionalmoments No-Learning Learning Data-UK Model Model Model Model AdditionalMoments (cid:24) = 0:85 (cid:24) = 0:85 (cid:24) = 0:25 (cid:24) = 0:85 (cid:24) = 0:25 Shareofsales-1-5years 0.11 0.20 0.12 0.12 0.03 Shareofsales-6-10years 0.16 0.19 0.11 0.17 0.08 Fractionoffirms-1-5years 0.31 0.42 0.50 0.42 0.04 Fractionoffirms-6-10years 0.20 0.15 0.12 0.15 0.05 Averagefirmemployment 22 18.9 17.5 19.5 73.2 Averagefirmemployment<5years 11.0 8.0 4.0 4.8 33.3 Averagefirmemployment 5years 35.5 30.4 36.5 32.2 90.2 (cid:21) Leverage(Debt/Assets) 0.73 0.57 0.16 0.63 0.26 Leverage-1-5years 0.81 0.83 0.23 0.82 0.59 Leverage-6 years 0.72 0.48 0.15 0.59 0.26 (cid:20) Averageinterestrate<5years - 0.05 0.13 0.05 0.11 Averageinterestrate 5years - 0.05 0.07 0.04 0.05 (cid:21) Exitrate 0.11 0.14 0.23 0.15 0.01 Top5-pctearningsshare 0.31 0.34 0.40 0.33 0.41 Debt/GDP 2.40 1.25 0.28 1.50 0.54 Note: This table reports a series of relevant moments, empirical and simulated, that were not directly targeted by the calibration.Learning(No-Learning)reportsthemomentsinaneconomywherefirmproductivityis(isnot)relatedwithfirm age.distributionoffirms, 36

Figure6: Firmdynamicsunderdifferentrecoveryrates 1500 1000 500 0 0 2 4 6 8 10 Age stessA 500 400 300 200 RR = 0 100 RR = 0.25 RR = 0.85 0 0 2 4 6 8 10 Age tbeD 1.2 1 0.8 0.6 0.4 0.2 0 2 4 6 8 10 Age egareveL 3 2.5 2 1.5 1 0.5 0 2 4 6 8 10 Age KPA Note: ThisfigureillustratestheevolutionofafirmasitgrowsolderinascenariowithLearningunderdifferentlevelsof recoveryrate.Inallsimulationsproductivityissettoitsmedianvaluethroughoutthelifeofthefirm.MPKisthemarginal productivityofcapitalwhereasLeverageisdefinedastheratiooftotalliabilitiestototalassets. 37

Table6: FinancialDevelopmentandAverageProductionofCapital-DataandModel Model Learning No-Learning PanelA (cid:24) = 0:65 (cid:24) = 0:35 (cid:24) = 0:65 (cid:24) = 0:35 All-firms 0.52 0.53 0.56 0.56 Young 0.75 1.44 0.77 0.69 Mature 0.53 0.49 0.54 0.55 Productive 0.58 0.54 0.60 0.61 Young-productive 1.07 1.86 0.99 1.63 Mature-productive 0.56 0.55 0.56 0.58 St. Dev. 0.65 0.72 0.64 0.75 St. Dev. Young 0.75 0.78 0.70 0.84 Data PanelB Highrecoveryrate Lowrecoveryrate All-firms 1.54 1.82 Young 1.69 2.13 Mature 1.49 1.67 Productive 1.98 2.74 Young-productive 2.34 3.11 Mature-productive 1.91 2.50 St. Dev. 1.09 1.11 St. Dev. Young 1.39 1.36 Note: Thistablereportstheempiricalandsimulatedaverageproductivityofcapital(APK)-ratioofsalestoassets-fora seriesofcross-sectionalgroupsdefinedbytheageofthefirmanditsrelativeproductivity.InPanelB,Ireporttheempirical medianAPKoffirms.Countriesaresplitintotwogroups:Acountryhashigh(low)recoveryrateifitisabove(below)0.4. Forhigh(low)recoveryratecountriestheiraveragerecoveryrateis0.69(0.49). Iclassifyfirmstobeyoung(mature)ifthey areyounger(older)than5yearsold. Furthermore,afirmisconsideredProductiveifitsTFPRisinthetop-10percentile.In PanelB,Iusesalesandnotvalueaddedduetothelackofdataeitheronmaterialsoronvalueaddedforcertaincountries. Tocorrectforoutliers,inallcountries,Idroppedthetopandbottom1percentofAPKobservations. 38

Table7: Regressions-DataandModelSimulations Data TFPR Std. TFPR Leverage 0.057 0.052 -0.061 -0.246 -0.049 -0.052 lnage (0.001)*** (0.004)*** (0.114)*** (0.113)** (0.001)*** (0.001)*** 0.001 0.047 -0.003 FD lnage (cid:2) (0.000)** (0.028)* (0.003)*** Adj:R2 0.52 0.49 0.03 0.03 0.06 0.06 Obs: 1,025,555 552,203 1,454 1,454 3,012,464 1,417,287 Model TFPR Std. TFPR Leverage No-Learn Learn No-Learn Learn No-Learn Learn 0.09 -0.04 0.00 -0.44 -0.29 -0.28 lnage (0.001) (cid:3)(cid:3)(cid:3) (0.001) (cid:3)(cid:3)(cid:3) (0.00) (0.02) (cid:3)(cid:3)(cid:3) (0.001) (cid:3)(cid:3)(cid:3) (0.001) (cid:3)(cid:3)(cid:3) 0.03 0.10 0.00 0.01 0.24 0.22 FD lnage (cid:2) (0.001) (cid:3)(cid:3)(cid:3) (0.001) (cid:3)(cid:3)(cid:3) (0.00) (0.00) (cid:3)(cid:3)(cid:3) (0.001) (cid:3)(cid:3)(cid:3) (0.001) (cid:3)(cid:3)(cid:3) Adj:R2 0.11 0.07 543 674 0.48 0.48 Note:Thistablerepresentsthecoefficientsofaseriesofregressionsusingdataandusingthesimulatedobservations. The three dependent variables are TFPR (revenue-TFP), Growth (log-asset growth rate) and Leverage (ratio of total liabilities to total assets). FD denotes financial development and is measured by the recovery rate in case of default. All regressions have a fixed effect at the country-industry-year level. The standard errors reported in parentheses are robust to heteroskedasticity. 39

Figure7: Firmdynamics-TFPR 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 0.05 0 5 10 15 Age RPFT naeM Firm age and average TFPR No Learning 0.4 High RR (x = 0.65) Low RR (x = 0.35) 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 0.05 0 5 10 15 Age RPFT naeM Firm age and average TFPR Learning 0.4 High RR (x = 0.65) Low RR (x = 0.35) 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 0.05 0 5 10 15 Age RPFT naeM Firm age and average TFPR Data High RR (x >0.4) Low RR (x <0.4) Note:Thisfigurerepresentstheaveragedynamicsassimulatedfirmsgrowolderinscenarioswithhighandlowrecovery rate.TFPRisthelogrevenue-productivity. 40

Figure8: AverageTFPRacrossage-cohorts-CountrieswithlowRecoveryRate 0.4 0.2 0 0.2 0.4 0 5 10 15 Age RPFT naem Ukraine 0.1 0.05 0 0.05 0.1 0.15 0 5 10 15 Age RPFT naem Romania 1 0.5 0 0.5 0 5 10 15 Age RPFT naem Hungary 0.15 0.1 0.05 0 0.05 0.1 0 5 10 15 Age RPFT naem Czech Rep Note: ThefigurereportstheaverageTFPRasafunctionofageforthefourcountriesinthesamplewiththelowRecovery Rate.Thedashedlinesdelimitthe95percentconfidenceinterval. 41

Figure9: Firmdynamics-Std. Dev. TFPR 0.29 0.28 0.27 0.26 0.25 0.24 0.23 0.22 0.21 0.2 0 5 10 15 Age RPFT .dtS Firm age and Std. TFPR No Learning 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0 5 10 15 Age RPFT .dtS Firm age and Std. TFPR Learning 0.65 0.6 0.55 0.5 0.45 0.4 0.35 0.3 0.25 0.2 0 5 10 15 Age RPFT .dtS Firm age and Std. TFPR Data High RR (x = 0.65) High RR (x = 0.65) High RR (x >0.4) Low RR (x = 0.35) Low RR (x = 0.35) Low RR (x <0.4) Note:Thisfigurerepresentstheaveragedynamicsassimulatedfirmsgrowolderinscenarioswithhighandlowrecovery rate. Std.TFPRisthestandarddeviationofthelogrevenue-productivity.growthforfirmsofagivenage-cohort. 42

Figure10: Firmdynamics-Leverage 1 0.8 0.6 0.4 0.2 0 0 5 10 15 Age egareveL Firm age and Leverage No Learning High RR (x = 0.65) Low RR (x = 0.35) 1 0.8 0.6 0.4 0.2 0 0 5 10 15 Age egareveL Firm age and Leverage Learning High RR (x = 0.65) Low RR (x = 0.35) 1 0.8 0.6 0.4 0.2 0 0 5 10 15 Age egareveL Firm age and Leverage Data High RR (x >0.4) Low RR (x <0.4) Note:Thisfigurerepresentstheaveragedynamicsassimulatedfirmsgrowolderinscenarioswithhighandlowrecovery rate.Leverageistheratiooftotaldebttoassets. 43

Figure11: FinancialDevelopmentandAggregateOutcomes 1.05 1 0.95 0.9 0.85 0.8 0 0.2 0.4 0.6 0.8 1 Recovery rate PFT .ggA Aggregate TFP 1.3 1.2 1.1 1 0.9 0 0.2 0.4 0.6 0.8 1 Recovery rate PFT .ggA Aggregate TFP Young Firms No Learning Learning 1 0.95 0.9 0.85 0 0.2 0.4 0.6 0.8 1 Recovery rate PFT .ggA Aggregate TFP Old Firms 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0 0.2 0.4 0.6 0.8 1 Recovery rate PFT .ggA Agg TFP /Agg TFP old young Note: This figure reports relation between the recovery rate and a series of country-level indicators. I graph both the empiricalresultsaswellastheresultsfromthesimulatedmodelwithandwithoutlearning. Thelinesindicatethebest fit under OLS. The solid plot is the line of best fit for the data, whereas the dashed (crossed) line is the best fit for the No-Learning(Learning)model. 44

Figure12: RelativeaggregateTFP-Oldvs. Newfirms HU NO ES FR HRPL PT SE DE SIIT FI SK CZ RUOA EE BG 2. 0 2. 4. 0 20 40 60 80 100 Recovery Rate Relative Aggregate TFP Fitted values Note: This figure graphs the aggregate TFP of new firms with less than eight years-old relative to that of older firms TFP T ol F d P(cid:0) n T e F w Pold . The relation between recovery rate and the relative aggregate TFP is positive and significant both (cid:16)withandwithoutB(cid:17)ulgaria.ThecorrelationbetweenrecoveryrateandtherelativeaggregateTFPis0.46withBulgariaand 0.52withoutit. 45

Figure13: FinancialDevelopmentandAggregateOutcomes 0.25 0.2 0.15 0.1 0.05 0 0 0.2 0.4 0.6 0.8 1 Recovery rate erahS noitcudorP Share of Production of Young Firms (<6 yrs) 0.5 0.4 0.3 0.2 0.1 No Learning Learning 0 0 0.2 0.4 0.6 0.8 1 Recovery rate erahS Fraction of Young Firms 0.08 0.06 0.04 0.02 0 0 0.2 0.4 0.6 0.8 1 Recovery rate etaR tixE Exit Rate of Old Firms (>6 Yrs) 2.6 2.4 2.2 2 1.8 1.6 0 0.2 0.4 0.6 0.8 1 Recovery rate Y/K Capital/GDP Note:Thisfigurereportsrelationbetweentherecoveryrateandaseriesofindicators.Igraphtheresultsfromthesimulated modelwithandwithoutlearning. ThelinesindicatethebestfitunderOLS.Thesolid(dashed)lineisthebestfitforthe No-Learning(Learning)model. 46

Figure14: MisallocationofFactorsofProduction 1 0.95 0.9 0.85 0.8 0.75 0 0.2 0.4 0.6 0.8 1 Recovery rate PFT Impact of TFP from misallocation No Learning 1 0.95 0.9 0.85 0.8 No reallocation Reallocation within age cohort Reallocation across age cohorts 0.75 0 0.2 0.4 0.6 0.8 1 Recovery rate PFT Impact of TFP from misallocation Learning No reallocation Reallocation within age cohorts Reallocation across age cohorts Note: ThisfiguregraphstheimpactthatmisallocationofthefactorsofproductionhasonaggregateTFP.Allvaluesare relativetotheaggregateTFPwithcompletereallocationwhentherecoveryrateis1. 8 Appendix 8.1 ProductivityEstimation 8.1.1 LevinsohnandPetrin Thealgorithmisasfollows. Firmprofitsareafunctionofthefirm’sstatevariablesandfactorprices which are common across all firms. Assuming a Cobb-Douglas production function, the estimating equationforcompanyiinindustryj attimetis y = (cid:12) +(cid:12) l +(cid:12) k +(cid:12) m +! +(cid:17) (11) it 0 l i;j;t k i;j;t m i;j;t i;j;t i;j;t where y;l;k;m are the log value of sales, number of employees, fixed assets and materials. Note f g that production depends, in addition to the inputs, on a firm specific productivity shock ! and on it 47

an unpredictable measurement error unrelated to the input’s choice (cid:17) : Although both ! and (cid:17) are it unobserved, the former is a serially correlated state variable that impacts the firm’s decisions on input demand, whereas the latter is not. To address the simultaneity-bias, I use intermediate inputs’ demandasproxyvariableforproductivity. Undercertainconditions,thedemandofintermediateinputsisstrictlymonotonicintheproductivityofthefirms.52 Invertingm = f (! ;k )providedthat it t it it themonotonicityconditionholds,thedemandforintermediateinputscanbeinvertedinproductivity ! it = f t(cid:0) 1(m it ;k it ) and therefore unobserved productivity can be approximated by the function f 1( ),whichdependsonlyonobservables. Substitutingthisexpressionintheproductionfunction, (cid:0) (cid:1) y = (cid:12) l +(cid:30) (m ;k )+(cid:17) (12) it l it t it it it where(cid:30) t ( ) = (cid:12) k k it +(cid:12) m m it +f t(cid:0) 1(m it ;k it ). InthefirststageoftheroutineofLP,aconsistentestimator (cid:1) forthelaborcoefficientandforthecompositeterm(cid:30) canbeobtainedbytreatingthefunctionf 1( ) it (cid:0) (cid:1) nonparametrically.53 Note that (cid:12) and (cid:12) cannot be identified in this stage since they appear both k m b linearlyandinthenon-parametricterm. Inthesecondstagethecoefficients(cid:12) and(cid:12) areidentified m k byGMMproceduresusingthemomentcondition k it E (cid:24) ((cid:12) ;(cid:12) ) = 0 ijt k m m it 1 (cid:26) (cid:18) (cid:0) (cid:19) (cid:27) Due to the small number of companies in some of the four-digit level companies, I estimate the coefficientsatthetwodigitNACEclassification. 8.1.2 Solowresidual Withthismethod,TFPRiscomputedinthetypicalindexform tfp = y (cid:11) l (cid:11) k (cid:11) m (13) i;j;t i;j;t l i;j;t k i;j;t m i;j;t (cid:0) (cid:0) (cid:0) wherethelowercaseletters tfp;y;l;k;m indicatethelogarithmsoffirm-levelTFPR,output,labor f g employed, capital and materials. These are measured as deflated sales, number of employees, de- 52LevinsohandPetrinrequireacompetitiveenvironmentandthatinvestmentdoesnotrespondtocurrentproductivity. MelitzandLevinsohn(2006)showthatthemonotonicityconditionholdsinnon-competitiveenvironmentaslongasmore productivefirmsdoesnotchargeinordinatelargermarkupsthanlessproductivefirms. 53Inpractice,weapproximatetheunknownfunctionf t(cid:0) 1( (cid:1) )withathird-orderpolynomialincapitalandintermediate inputs. ThispracticeiscommoninOlleyandPakes(1996),LevinsohnandPetrin(2003)andinmostoftheliteraturethat followfromthereafter. 48

flated asset value and deflated total materials of company i in industry j at time t.54’55 To measure inputelasticities(cid:11) ,Iusetheaveragesectoralinputcostshare,oflabor,capitalandmaterials.56 i Table8: Regression-AgeandaverageTFPR UK France (1) (2) (3) (1) (2) (3) Benchmark Weighted Heckit Benchmark Weighted Heckit 0.06 0.06 0.10 0.02 0.02 0.12 ln(age) (0:017) (0:017) (0:056) (0:008) (0:008) (0:028) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) Obs. 31,376 31,376 10,309 113,309 113,309 42,161 Italy Spain (1) (2) (3) (1) (2) (3) Benchmark Weighted Heckit Benchmark Weighted Heckit 0.10 0.11 0.44 0.11 0.12 0.13 ln(age) (0:006) (0:006) (0:017) (0:006) (0:006) (0:016) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) Obs. 257,442 257,442 124,509 268,856 268,856 109,221 Note: ThistablepresentstheresultsofregressingtheTFPRofafirmonitslog-ageforthe2006-2009period. FirmTFPR wasestimatedusingtheSolowresidualmethodproposedbySyverson(2004). Allregressionshavefirmfixed-effectsas wellasindustry-yearfixed-effects. Weightedregressionsusetotalrealsaleshasaweight. Heckitregressionscontrolfor selection using the Heckman two-step procedure. Due to lower quality in the selection indicator, the Heckit regression usesasamplefrom2007-2008.Theprobitregressestheselectionindicatoronthefirmproductivity,sizeandage. 8.2 DataCleaningMethodology In this section I describe in detail de procedure for the cleaning of the data. First, I require firms to have some basic accounting information over the years. Therefore, I drop firms that do not provide positiveassets,sales,laborpaymentsandliabilities,orthatdonotprovideinformationontheageof the firm.57 Next, I delete from the sample, firms that report only consolidated statements, to avoid doublecountingfirmsandsubsidiariesoroperationsabroad. Ialsoexcludecertainindustries. Idrop 54Sales,assetsandmaterialdeflatorswereconstructedusingsectoraldeflators,withsectorsbeingdefinedatthetwo-digit NACEcode. 55Anassumptionof13isconstantreturnstoscale. Ifthescale-elasticityisdifferentfromoneweshouldmultiplythe inputelasticities(cid:11)bythefirm-scaleelasticity. Totestthesensitivityofourresultstothisassumption,wereranourresults assumingbothincreasinganddecreasingreturnstoscale.Testingscaleelasticitiesof0.9and1.1.Thequalitativeresultsare maintained. 56The sectoral cost of capital used was obtained from the BLS. The assumption made was that the cost of capital is thesameacrosseconomies. Totestthesensitivityofourresultstothisassumption,wealsocalculatedacountry-specific sectoralcostofcapitalbymultiplyingthesectoralcostofcapitalprovidedbytheBLSbyacountry-specificcostofcapital. This country-specific cost of capital was obtained using the spread of a country’s sovereign bonds with respect to the Germanonesin2005. 57Lessthan2percentofthefirmsdonotprovideinformationregardingtheirage. 49

Table9: Regressions-RobustnessCheckwithAlternativeDefinitionsofFinancialDevelopment TFPR Growth Leverage Fin. Dev. II Fin. Dev. III FinDev. II FinDev. III Fin. Dev. II Fin. Dev. III -0.04 -0.05 -0.04 -0.09 -0.06 -0.06 lnage (0.002)*** (0:002) (cid:3)(cid:3)(cid:3) (0.002)*** (0:002) (cid:3)(cid:3)(cid:3) (0.001)*** (0:002) (cid:3)(cid:3)(cid:3) 0.08 0.07 0.03 0.04 0.02 0.01 FD lnage (cid:2) (0:003) (cid:3)(cid:3)(cid:3) (0:002) (cid:3)(cid:3)(cid:3) (0:001) (cid:3)(cid:3)(cid:3) (0:002) (cid:3)(cid:3)(cid:3) (0:001) (cid:3)(cid:3)(cid:3) (0:001) (cid:3)(cid:3)(cid:3) Adj:R2 0.30 0.30 Obs: 1,611,723 1,611,723 2,993,504 2,993,504 3,012,464 3,012,464 Note:Thistablereportsthecoefficientsofaseriesofregressionsusingtwoalternativedefinitionsoffinancialdevelopment. Fin.Dev.IIusesasameasureoffinancialdevelopmenttheratioofprivatecredittototalGCPwhereasFin.Dev.IIIusesas ameasureoffinancialdevelopmenttheratioofliquidliabilitiestoGDP.ThethreedependentvariablesareTFPR(revenue- TFP),Growth(log-assetgrowthrate)andLeverage(ratiooftotalliabilitiestototalassets).Allregressionshaveafixedeffect atthecountry-industry-yearlevel.Thestandarderrorsreportedinparenthesesarerobusttoheteroskedasticity. severalprimarysectorswheretheactivityisverycountry-specific. Thesesectorsincludeagriculture (NACE code 1), forestry (NACE code 2), fishing (NACE code 5), and mining (NACE codes 10-14). I alsodropthefinancialservicesindustries(NACEcode65-66)sincethefinancialinformationforthese firms is not-comparable to those of non-financial firms). Finally, I drop the public sector, education, healthandsocialsector,andactivitiesoforganizationsthatcannotbeclassified(NACEcode75-99). 8.3 ComparabilityofCountrySamples ThissectionanalyzesthecoverageoftheAmadeusdatasetacrosscountries. TheEuropeanCommissionReport(ECR)containsinformationontheuniverseoffirmsperbusinesssectorforthemajority ofcountriesinthesample. TheECRreportsthepercentageoffirmsthathave1-49employees,50-249 employees and more than 250 employees. I compare the fraction of firms in each employment categorywiththeonespresentinAmadeus. Unfortunately,andasreportedinArellano,Bai,andZhang (2012), employment information is not reported for all firms in Amadeus. On average, around 70% of the firms in the cleaned sample report the number of employees, with the incidence of missing information being larger for small firms. To deal with this lack of employment information, I follow Arellano et al, and impute employment measures for firms that do not report employment in Amadeus. Irunregressionscountrybycountryoflog(employment)onlog(assets)andlog(sales).58 58Asreferredabove,IsubstituteoperatingrevenueforsalesfortheDanishandNorwegianfirms. 50

NumberofFirmsperAge-Cohort ShareofFirms Age-Cohorts Country RecoveryRate No. Firms <4 [4,7] [8,11] [12,18] 18< Fin 0.883 216,773 0.12 0.18 0.21 0.28 0.21 Ire 0.877 38,204 0.25 0.29 0.17 0.17 0.12 Ned 0.874 45,787 0.10 0.14 0.12 0.18 0.45 Nor 0.944 432,793 0.22 0.25 0.16 0.38 Bel 0.86 974,593 0.18 0.20 0.16 0.26 0.20 UK 0.86 1,139,416 0.34 0.23 0.12 0.13 0.18 Ice 0.82 39,271 0.36 0.27 0.13 0.11 0.12 Spn 0.77 1,807,117 0.22 0.26 0.21 0.20 0.11 Prt 0.73 173,222 0.19 0.22 0.15 0.20 0.24 Swe 0.81 565,565 0.15 0.15 0.16 0.29 0.24 Ita 0.71 1,209,776 0.21 0.22 0.14 0.19 0.24 Den 0.63 165,227 0.30 0.28 0.10 0.17 0.16 Ger 0.56 148,490 0.15 0.19 0.15 0.21 0.30 Lit 0.34 15,998 0.20 0.29 0.36 0.14 Fra 0.46 2,436,583 0.23 0.21 0.17 0.20 0.19 Gre 0.45 71,390 0.19 0.22 0.18 0.22 0.19 Slk 0.40 14,615 0.16 0.31 0.35 0.16 0.02 Hun 0.39 350,581 0.33 0.25 0.26 0.15 Est 0.37 139,259 0.34 0.31 0.24 0.11 0.01 Lat 0.36 13,298 0.17 0.29 0.38 0.16 Bul 0.34 15,475 0.19 0.25 0.27 0.17 0.12 Cro 0.29 62,885 0.04 0.13 0.48 0.30 0.05 Pol 0.26 68,339 0.13 0.24 0.24 0.24 0.15 Rus 0.25 377,692 0.32 0.28 0.32 0.04 0.03 Srb 0.21 49,513 0.06 0.16 0.39 0.33 0.06 Czh 0.15 158,346 0.20 0.31 0.33 0.16 0.01 Ukr 0.08 39,259 0.13 0.37 0.38 0.06 0.06 Rom 0.07 774,217 0.30 0.20 0.33 0.17 Total 0.52 11,543,684 0.21 0.24 0.24 0.19 0.14 Table 10: Recovery rate is the average amount recouped by creditors of an insolvent firm, for each dollar of outstanding credit. The Number of firms is the total number of firm-year observations per country for 2002-2005. The Share of firms represents the fraction of firm-year observations per agecohort. Theshareoffirmsolderthan18years-oldinNorway,Hungary,Latvia,BulgariaandRomania islowerthan0.5percent. 51

CoverageandCross-CountryComparability AmadeusDataset ECData Small Medium Large Small Medium Large Country 1-49 50-250 >250 1-49 50-250 >250 Fin 0.957 0.032 0.953 0.985 0.012 0.003 Ire 0.929 0.070 0.000 Ned 0.610 0.309 0.072 0.981 0.016 0.003 Nor 0.975 0.020 0.004 Bel 0.971 0.023 0.006 0.009 UK 0.907 0.071 0.020 0.978 0.018 0.004 Ice 0.987 0.011 0.002 Spn 0.962 0.031 0.006 0.991 0.008 0.001 Prt 0.929 0.061 0.010 Swe 0.966 0.027 0.006 0.990 0.008 0.002 Ita 0.944 0.047 0.008 0.994 0.005 0.001 Den 0.950 0.041 0.008 Ger 0.691 0.220 0.088 0.972 0.023 0.005 Lit 0.711 0.225 0.039 0.952 0.043 0.005 Fra 0.966 0.027 0.006 0.987 0.010 0.003 Gre 0.894 0.078 0.014 Slk 0.665 0.254 0.081 0.931 0.055 0.014 Est 0.968 0.027 0.004 0.966 0.030 0.004 Hun 0.982 0.015 0.002 0.001 Lat 0.731 0.211 0.053 0.970 0.027 0.003 Bul 0.683 0.217 0.094 0.982 0.016 0.002 Cro 0.905 0.072 0.021 Pol 0.536 0.322 0.109 0.989 0.009 0.002 Rus 0.806 0.139 0.053 Srb 0.856 0.099 0.043 Czh 0.863 0.109 0.028 0.991 0.008 0.001 Ukr 0.325 0.414 0.256 Rom 0.972 0.022 0.005 0.971 0.023 0.006 Table 11: The Amadeus dataset has the information used in the empirical exercises. The European Commission data, is obtained from the National Registrars and includes all formal firms within an economy 52