TOLERANCE FOR FAILURE AND CORPORATE INNOVATION

Transcription

1 TOLERANCE FOR FAILURE AND CORPORATE INNOVATION Xuan Tan Kelley School of Busness Indana Unversty (812) Tracy Yue Wang Carlson School of Management Unversty of Mnnesota (612) Ths verson: July, 2010 Key words: tolerance for falure, nnovaton, patents, venture captal, IPO JEL classfcaton: M14, O31, G24, G34 * We are grateful for comments from Utpal Bhattacharya, Henrk Cronqvst, Douglas Cummng, Nshant Dass, Davd Dens, Wllam Kerr, Gustavo Manso, Robert Marquez, Raghuram Rajan, Amt Seru, Merh Sevlr, Ann Sherman, PK Toh, Andrew Wnton, and Ayako Yasuda. We also thank semnar partcpants at Unversty of Mnnesota, the Ffth Annual Early-Career Women n Fnance Conference, the State of Indana Conference, the Sxth Annual Corporate Fnance Conference at Washngton Unversty n St. Lous, the NBER Entrepreneurshp 2009 Wnter Group Meetng, the SunTrust Bank Sprng Beach Conference at Florda State Unversty, the 2010 Napa Conference on Fnancal Markets Research, the 2010 Fnancal Intermedaton Research Socety Conference, the 2 nd ESSEC Prvate Equty Conference, the Entrepreneural Fnance and Innovaton Conference, the Conference on People & Money at DePaul Unversty, and the 2010 Chna Internatonal Conference n Fnance.

2 TOLERANCE FOR FAILURE AND CORPORATE INNOVATION Abstract We examne whether tolerance for falure spurs corporate nnovaton based on a sample of venture captal (VC) backed IPO frms. We develop a novel measure of VC nvestors falure tolerance by examnng ther tendency to contnue nvestng n a venture condtonal on the venture not meetng mlestones. We fnd that IPO frms backed by more falure-tolerant VC nvestors are sgnfcantly more nnovatve. A rch set of emprcal tests shows that ths result s not drven by the endogenous matchng between falure-tolerant VCs and startups wth hgh exante nnovaton potentals. Further, we fnd that the margnal mpact of VC s falure tolerance on startup nnovaton vares sgnfcantly n the cross secton. Beng fnanced by a falure-tolerant VC s much more mportant for ventures that are subject to hgh falure rsk,.e., ventures born n recessons, ventures at early development stages, and ventures n ndustres n whch nnovaton s dffcult to acheve.

3 1. INTRODUCTION Innovaton s vtal for the long-run comparatve advantage of frms. However, motvatng and nurturng nnovaton remans a challenge for most frms. As Holmstrom (1989) ponts out, nnovaton actvtes nvolve a hgh probablty of falure, and the nnovaton process s unpredctable and dosyncratc wth many future contngences that are mpossble to foresee. Holmstrom thus argues that nnovaton actvty requres exceptonal tolerance for falure and the standard pay-for-performance ncentve scheme s neffectve. Manso (2010) explctly models the nnovaton process and the trade-off between exploraton of new untested actons and explotaton of well known actons. Manso shows that the optmal contracts that motvate exploraton nvolve a combnaton of tolerance for falures n the short-run and reward for success n the long-run. 1 In ths paper we examne whether tolerance for falure ndeed spurs corporate nnovaton. We adopt a novel emprcal approach. We start wth venture captal (hereafter VC) nvestors atttude towards falure and nvestgate how such atttude affects nnovaton n VC-backed startup frms. VC-backed startup frms provde an deal research settng for our study. These frms generally have hgh nnovaton potentals and also hgh falure rsk. Therefore, both tolerance for falure and nnovaton are very relevant for these frms. Further, nnovaton n entrepreneural frms has been an mportant drver of economc growth n the Unted States. Thus t s mportant to understand what factors help to spur nnovaton n startup companes. We beleve that VC nvestors tolerance for falure s crucal for the nnovaton productvty of VC-backed startups. VC nvestors are the prncpal nvestors and mportant decson makers n the startup frms they fnance. They have the fnal decson power on whether to contnue nvestment or to termnate a project. If VC nvestors are not tolerant of falure, then the ventures they fnance are lkely to be lqudated prematurely upon ntal unsatsfactory progress and therefore lose the chance to be nnovatve. Therefore, VC nvestors tolerance for falure can prevent premature lqudaton and allow entrepreneural frms to realze ther nnovaton potentals. We nfer a VC nvestor s falure tolerance by examnng ts tendency to contnue nvestng n a project condtonal on the project not meetng mlestones. A smple model of VC 1 Ederer and Manso (2010) conduct a controlled laboratory experment and provde evdence supportng the mplcatons n Manso (2010). 1

4 project termnaton suggests that a reasonable proxy for a VC s falure tolerance s the VC frm s average nvestment duraton (from the frst nvestment round to the termnaton of follow-on nvestments) n ts past faled projects. The ntuton s that the stagng of captal nfusons n VC nvestments gves VC nvestors the opton to abandon underperformng projects. Such opton s partcularly pertnent n projects that eventually fal because these projects may have faled to meet stage targets even before the lqudaton decsons are made. If a project does not show progress towards stage targets, then the choce between gvng the entrepreneur a second chance by contnung to nfuse captal and wrtng off the project mmedately should to some extent reflect a VC nvestor s atttude towards falure. Other thngs equal, the longer the VC frm on average wats before termnatng fundng n underperformng projects, the more tolerant the VC s for early falures n nvestments. We then lnk a VC nvestor s falure tolerance to IPO frms backed by the VC nvestor. For each IPO frm, the relevant VC falure tolerance s the VC nvestor s falure tolerance at the tme when the VC nvestor makes the frst-round nvestment n the IPO frm. Ths approach s least subject to the reverse causalty problem because the falure tolerance measure captures the nvestng VC nvestor s atttude towards falure before ts very frst nvestment n a startup frm, whch s well before the observed nnovaton actvtes of the startup frm. Our man emprcal fndng s that IPO frms backed by more falure-tolerant VCs are sgnfcantly more nnovatve. They not only produce a larger number of patents but also produce patents wth larger mpact (measured by the number of ctatons each patent receves). The results are robust to alternatve measures of VC falure tolerance and alternatve emprcal and econometrc specfcatons. Although the baselne results are consstent wth the hypothess that VC nvestors falure tolerance leads to hgher ex-post nnovaton productvty n VC-backed ventures, an alternatve nterpretaton of the results could be that falure-tolerant VCs are n equlbrum matched wth projects that have hgh ex-ante nnovaton potentals, and hgh ex-ante potentals lead to hgh expost outcomes. We thus do a rch set of analyss to address ths dentfcaton problem. We frst extend the basc model of VC project termnaton to allow for heterogenety n VC nvestors project selecton abltes that drectly mpact the matchng between VC nvestors and projects. The extended model provdes a clear structure for the causes and solutons to the endogenety problem. Followng the model s gudance, we then show that the effect of VC 2

5 falure tolerance on startup frm nnovaton cannot be explaned away (n fact, the effect s even stronger) by ncludng the lead VC frm fxed effects, whch absorb the tme-nvarant dfferences n project selecton abltes across lead VC nvestors. Further, besdes VC fxed effects we control for the possble tme-varyng component of the VC project selecton abltes by ncludng the VCs past nvestment experences and ndustry expertse. Agan, we fnd that the falure tolerance effect s not only present but even stronger. Fnally, we decompose our falure tolerance measure and remove the VC project selecton component from the measure. We fnd that the varaton of the purer VC falure tolerance measure contnues to powerfully explan the varaton n the IPO frm s nnovaton productvty. Our last set of dentfcaton tests reles on the cross-sectonal heterogenety n the VC falure tolerance effect. If our falure tolerance measure ndeed captures a VC nvestor s atttude towards falure, then the margnal mpact of our measure on nnovaton reflects how valuable a VC s falure tolerance s for startup nnovaton and thus should be stronger n ventures where the falure rsk s hgher. However, f our measure nstead captures the ex-ante nnovaton potentals of ventures as under the alternatve nterpretaton that falure-tolerant VCs are endogenously matched wth hgh-potental ventures, then the margnal mpact of our measure reflects how lkely ex-ante potentals can turn nto successful ex-post outcomes and thus should be stronger n ventures where the falure rsk s lower. We fnd that the effect of VC falure tolerance on startup nnovaton s much stronger when the falure rsk s hgher and thus falure tolerance s more needed and valued. Beng fnanced by a falure-tolerant VC s much more mportant for ventures born n recessons, ventures at early development stages, and ventures n ndustres n whch nnovaton s dffcult to acheve (e.g., the drugs ndustry). These fndngs provde further support for our emprcal proxy of VC falure tolerance and dentfcaton of the falure tolerance effect. Our paper contrbutes to a growng emprcal lterature n corporate fnance on nnovaton. Several recent papers show that the legal system matters for nnovaton. Acharya and Subramanan (2009) fnd that a debtor-frendly corporate bankruptcy code encourages nnovaton. Fan and Whte (2003) and Armour and Cummng (2008) show that forgvng personal bankruptcy laws encourage entrepreneurshp. Acharya, Bagha, and Subramanan (2009) document that strngent labor laws spur nnovaton by provdng frms a commtment devce not to punsh employees for short-run falures. In a smlar sprt, Acharya, Bagha, and Subramanan 3

6 (2010) fnd that wrongful dscharge laws that make t costly for frms to arbtrarly dscharge employees foster nnovaton. These papers show that f the law provdes lenency n the case of ether personal falure or corporate falure, then we observe more entrepreneural actvtes and nnovaton. The forgveness of the law s to some extent related to the noton of falure tolerance. Our paper contrbutes to ths strand of research by documentng a more drect effect of falure tolerance on corporate nnovaton. 2 Our paper also contrbutes to the lterature on VC nvestors role n frm value creaton. Ths lterature has shown that VC nvestors experences, ndustry expertse, market tmng abltes, and network postons can all ncrease the value of VC-backed startup frms (see Gompers 2007 for a survey of ths lterature, the latest studes nclude Hochberg, Ljungqvst, and Lu 2007, Sorensen 2007, Bottazz, Da Rn, and Hellmann 2008, Gompers, Kovner, and Lerner 2009, and Pur and Zarutske 2009). In partcular, Kortum and Lerner (2000) fnd that ncreases n VC actvty n an ndustry lead to sgnfcantly more nnovatons. Our paper shows that the varaton n VCs tolerance for falure can explan the heterogenety n the observed nnovaton productvty of VC-backed frms. The rest of the paper s organzed as follows. Secton 2 dscusses the emprcal measure of VC falure tolerance. Secton 3 descrbes the emprcal specfcaton. Secton 4 dscusses the man results and robustness ssues. Secton 5 addresses dentfcaton ssues. Secton 6 concludes. 2. MEASURING FAILURE TOLERANCE 2.1 VC Frm s Falure Tolerance: A Conceptual Framework Falure n ths study means unsatsfactory progress n the nnovaton process. We wsh to nfer a VC frm s tolerance for falure by examnng ts tendency to contnue nvestng n a project condtonal on the project not meetng stage targets. It s well known that VC nvestments are hghly rsky due to the nature of the busness. Ths s why the stagng of captal nfusons s an essental feature of VC nvestments (Gompers 1995). Stagng allows VC nvestors to gather nformaton and montor the project progress. It also allows VC nvestors to mantan the opton to abandon underperformng projects. If a project does not show progress towards stage targets after the ntal rounds of nvestments, then the choce between contnung to nfuse captal and 2 Other papers have examned the effect of ownershp structure and fnancng on corporate nnovaton (e.g., Atannassov, Nanda, and Seru 2007, Seru 2008, Aghon, Van Reenen, and Zngales 2009, Belenzon and Berkovtz 2010, and Lerner, Sorensen, and Stromberg 2010). 4

7 termnatng fundng mmedately should to some extent reflect a VC s atttude towards falures n nvestments. Put dfferently, a VC s falure tolerance resdes n ts power of termnaton. A challenge for us s how to emprcally capture condtonal on the project not meetng stage targets. Ideally, we would lke to observe the ex-post stage performance of a venture relatve to the ex-ante stage targets specfed n the contract between the VC and the entrepreneur, and the VC s decson after revewng the venture s performance at each nvestment round (.e., contnued nvestment or termnaton of fundng). Unfortunately, the avalable VC nvestment data only provde nformaton about the VC s decson at each nvestment round, but do not provde the condtonng nformaton (e.g., whether the project meets the stage targets). 3 Ths data lmtaton mples that we have to measure a VC s falure tolerance n an ndrect way. Thus we present a smple llustratve model to examne how VC nvestors exercse ther opton to termnate a project upon recevng negatve nformaton, and use the model mplcaton to motvate our emprcal measure of falure tolerance. Suppose that the qualty (or NPV) of a project s, where u. The parameter 0 s a constant and s the average qualty of the projects n the nvestment pool. The parameter u represents the project-specfc qualty. We assume that u s normally dstrbuted wth zero mean and precson h u, and the VC nvestors observe the dstrbuton parameters. When a VC frm starts to nvest n a project, ts pror estmate of the project qualty s smply. As the VC frm nteracts wth the entrepreneur, t learns about the value of u based on a seres of performance sgnals from the nvestment. Let n be the n-th performance sgnal. Specfcally, n u n, where n s ndependent of u and also ndependent of each other. We assume that n s normally dstrbuted wth zero mean and precson h. The VC frm wll stop nvestng n the project when the posteror estmate of the project qualty s below certan threshold. Wthout loss of generalty, we set the threshold to be zero, 3 An alternatve way to capture condtonal on the project not meetng stage targets s to examne a VC s tendency to nvest n a down round,.e., an nvestment round n whch the post-round valuaton s lower than the pror-round valuaton. However, only 16.9% of VC fnancng rounds reported n our VC database (the Thomson Venture Economcs) has nformaton about the post-round valuaton, whch largely reduces the sample sze and may expose our study to serous sample selecton problems. 5

8 consstent wth the postve NPV nvestment rule. Thus the VC wll termnate the project after recevng the n-th sgnal, where n s the smallest nteger that satsfes the followng condton: E u,,..., ) 0 (1) ( 1 2 n Gven the normalty and ndependence assumptons, the expected value of u gven a seres of performance sgnals s as follows: n h nh E( u 1, 2,..., n ) s, (2) h nh h nh u s 1 where s the average of the n sgnals. If a project s eventually abandoned, the average performance sgnal must be negatve. How does falure tolerance affect the VC s project termnaton decson? Note that falure tolerance n ths study s meant to capture the VC s tendency to contnue nvestng n a project condtonal on the project underperformng (.e., condtonal on a negatve performance sgnal). We can vew falure tolerance as the VC s preference that affects how t reacts to ntal negatve performance sgnals. Gven a negatve sgnal, a falure-ntolerant VC tends to adjust the project NPV estmate downward more than a falure-tolerant VC does. As a result, the falure-ntolerant VC tends to termnate a project more quckly upon recevng negatve sgnals. There are dfferent ways to model such preference. We thnk that a smple way s as follows. Let h denote VC- s perceved precson of the sgnal nose. We assume that h h u. (3) Makng proportonal to h means that projects wth hgher uncertanty also tend to have h u u noser sgnals. More mportantly, 0 ntroduces heterogenety among VCs. VCs wth a hgh are falure-ntolerant. They perceve the ntal negatve sgnals as very nformatve and adjust ther estmates of u accordngly. VCs wth a low are falure-tolerant. They vew the ntal negatve sgnals as less nformatve and thus do not adjust ther posteror NPV estmate downward much. Therefore, a lower means a hgher level of falure tolerance. Note that our ntenton s not to argue whch type of VCs s more correct or more ratonal because we assume that nobody observes the true sgnal nose precson. Ths s a reasonable assumpton gven the exceptonally hgh uncertanty n the ntal stages of an nnovaton process, creatng dscreton n nterpretatons and room for judgment. All VC nvestors behave ratonally h 6

9 accordng to ther belefs and preferences. Also, our model s not about dfferent VCs beng dfferentally nformed. It s about dfferent VC belefs and preferences leadng to dfferent reactons to the same nformaton. Pluggng (3) nto equatons (1) and (2), VC- s nvestment duraton n an eventually faled project s the smallest nteger n so that sdes of the nequalty and lettng c log( ), we have ( ) n 1 ). Takng the logarthm on both ( ) ( 1 log( n ) log( ) c. (4) Equaton (4) s the key equaton that provdes the conceptual foundaton for our emprcal measure. It shows that the VC s nvestment duraton n an eventually faled project can serve as a measure of the VC s falure tolerance. The lower the s, the more falure tolerant the VC s, and the longer the nvestment duraton log( n ) s. 2.2 VC Frm s Falure Tolerance: The Emprcal Measure Followng the mplcaton of equaton (4), we construct the measure of a VC frm s falure tolerance based on the average nvestment duraton n the VC s past faled nvestments. Specfcally, VC frm- s falure tolerance n year t s the weghted average nvestment duraton n projects that have eventually faled up to year t (see Fgure 1 for an llustraton). Faled projects are those that are eventually wrtten off by ther nvestng VC nvestors. The nvestment duraton n a project can be descrbed n two ways. One s the tme nterval (n years) between the frst captal nfuson from VC frm- to the termnaton of fundng by VC frm-. The other s the number of fnancng rounds the VC frm nvests before wrtng off an underperformng venture. We use the former as the man proxy and the latter as an alternatve proxy for robustness checks. The weght for a project s VC frm- s nvestment n the project as a fracton of VC frm- s total nvestment up to year t. Usng the average nvestment duraton helps to mtgate the dosyncrases of ndvdual projects. Smlarly, VC frm- s falure tolerance n year t+s s the weghted average nvestment duraton n projects that faled up to year t+s. Snce the number of a VC s faled projects 7

10 accumulates over tme, the falure tolerance measure s tme-varyng, allowng the VC nvestors atttude towards falure to slowly change over tme. 4 Fgure 1: VC Frm s Falure Tolerance VC- has N t faled projects between year 0 and year t. Compute weghted average nvestment duraton n them. VC- s falure tolerance at year t 0 t t+s VC- has N t+s faled projects between year 0 and year t+s. Compute weghted average nvestment duraton n them. VC- s falure tolerance at year t+s We obtan data on round-by-round VC nvestments from the Thomson Venture Economcs database for entrepreneural frms that receved VC fnancng between 1980 and Appendx A pont A dscusses the detals of the data cleanng. To construct the VC falure tolerance measure, we focus on VC frms faled nvestments,.e., entrepreneural frms that were wrtten off by ther nvestng VC nvestors. Venture Economcs provdes detaled nformaton on the date and type of the eventual outcome for each entrepreneural frm (.e., IPO, acquston, or wrte-off). However, the database does not mark all wrtten-down frms as wrteoffs. Therefore, based on the fact that the VC ndustry requres nvestment lqudaton wthn ten years from the ncepton of the fund n the majorty of the cases, n addton to the wrte-offs marked by Venture Economcs, we classfy a frm as a wrtten-off frm f t dd not receve any fnancng wthn a ten-year span after ts very last fnancng round. 4 A subtle but relevant concern s whether our measure s capturng a VC s atttude towards rsk or atttude towards falure. Tolerance for rsk s an nvestor s ex-ante atttude towards uncertantes of nvestment outcomes, whle tolerance for falure reflects how an nvestor ex post reacts to a project s unfavorable outcome. Our measure s more lkely to capture a VC nvestor s tolerance for falure rather than rsk for two reasons. Frst, the venture captal ndustry s known as the hgh-rsk-hgh-return ndustry. Therefore, VC nvestors are relatvely homogenous n ther atttude towards rsk. Otherwse, they wll not nvest n the VC ndustry n the frst place. Second, our VC falure tolerance measure s computed based on the VC nvestor s past faled nvestments. Therefore, how long a VC nvestor wats before wrtng off the project reflects her ex-post reacton to an unsuccessful outcome rather than her ex-ante wllngness to accept hgh uncertanty n the nvestment outcomes. 5 We choose 1980 as the begnnng year of our sample perod because of the regulatory shft n the U.S. Department of Labor s clarfcaton of the Employee Retrement Income Securty Act s prudent man rule n Ths Act allowed penson funds to nvest n venture captal partnershps, leadng to a large nflux of captal to venture captal funds and a sgnfcant change of venture captal nvestment actvtes. 8

11 There are 18,546 eventually faled ventures recevng 67,367 nvestment rounds from 4,910 VC frms n our sample. For each faled venture a VC frm nvested n, we calculate the VC frm s nvestment duraton (n years) from ts frst nvestment round date to ts last partcpaton round date. If the venture contnues to receve addtonal fnancng from other VC nvestors after the VC frm s last partcpaton round, then the duraton s calculated from the VC frm s frst nvestment round date to the next fnancng round date after ts last partcpaton round. Ths s because the decson to contnue or to termnate fundng s generally done at the tme of refnancng (Gorman and Sahlman 1989). We then calculate Falure Tolerance by takng the weghted average of a VC frm s nvestment duraton n ts eventually faled projects up to a gven year. We compute the alternatve falure tolerance measure based on the number of fnancng rounds n a smlar fashon, and call t Falure Tolerance 2. The correlaton between the two measures s Now we lnk the VC s falure tolerance to a future IPO frm fnanced by the VC. Suppose that the VC frm- makes ts frst-round nvestment n a start-up frm-j n year t, and ths frm later goes publc n year t+k. Then the VC falure tolerance relevant to frm-j s VC frm- s falure tolerance n year t (see Fgure 2 for an llustraton). In sum, the relevant VC falure tolerance for an IPO frm s the nvestng VC frm s falure tolerance at the tme when the VC frm makes the frst-round nvestment n the IPO frm. Fgure 2: IPO Frm s Falure Tolerance VC- starts to nvest n Frm-j Frm-j s IPO 0 t t+k VC- has N t faled projects. Compute average nvestment duraton n them. VC- s falure tolerance at year t Frm-j s falure tolerance We obtan the lst of VC-backed IPOs between 1985 and 2006 from the SDC Global New Issues database. 6 We use the standard exclusons and correctons n the IPO lterature (see Appendx A pont B). We then merge the IPO sample wth our VC frm sample. 6 We choose 1985 as the begnnng year of our IPO sample so that we have a long enough tme gap between the begnnng year of our VC sample (.e., 1980) n whch the Falure Tolerance measure s constructed and the 9

12 For each IPO frm n our sample, we observe the dentty of ts nvestng VC frms and the value of each VC frm s falure tolerance measure at ther frst partcpaton round dates. VC nvestments are often syndcated (about 91% of our sample), and the lead VC nvestor usually plays the most mportant role n montorng the venture and decdng f a follow-on fnancng should be made. Ths mples that the lead VC s atttude towards falure should matter the most to a venture s nnovaton. Therefore, we choose the lead VC frm s falure tolerance as the man measure for our IPO frms. Followng the prevous lterature (e.g., Lee and Wahal 2004, Nahata 2008), we defne the lead VC as the one that makes the largest total nvestment across all rounds of fundng n an IPO frm. Alternatvely, snce all VC syndcate members make nvestments n the venture, each VC s atttude towards falure may matter. We thus also construct an alternatve falure tolerance measure by calculatng the weghted average of nvestng VCs falure tolerance f an IPO frm receves fundng from a VC syndcate. The weght s the nvestment by a VC frm as a fracton of the total VC nvestment receved by the IPO frm. Consequently, there are two tme-nvarant VC falure tolerance measures for each IPO frm n our sample: Falure Tolerance s the lead VC s tolerance for falure, and VC Syndcate Falure Tolerance s the weghted average falure tolerance of the nvestng VC syndcate. Table 1 Panel A reports the descrptve statstcs of Falure Tolerance and VC Syndcate Falure Tolerance by IPO frms. The average lead VC s falure tolerance s about two years and seven months and t can be as long as sx years and ten months. The average VC Syndcate Falure Tolerance s about one year and ten months. Ths mples that on average the lead VCs are more falure tolerant than other syndcate members. The dstrbutons of falure tolerance measures are rght skewed. Also, from an economc perspectve there s a large dfference between watng for two years rather than one year before termnatng an nvestment, but probably a smaller dfference between watng for seven years versus sx years. Both the skewness and the lkely nonlnearty n the economc mpact of VC s tolerance for falure suggest that a logarthm transformaton of the falure tolerance measure s approprate. We then use the natural logarthm of Falure Tolerance as the man measure n the rest of the analyss, whch s also consstent wth equaton (4). begnnng year of our IPO sample n whch the Falure Tolerance measure s utlzed. By dong so, we mnmze the possblty that a VC-backed IPO frm has no Falure Tolerance nformaton avalable. 10

13 3. EMPIRICAL SPECIFICATION We use to denote the lead VC frm, and j to denote an IPO frm fnanced by VC-. We use 0 to ndcate the tme when VC- makes the frst-round nvestment n IPO frm-j. Then t ndcates the t-th year after the frst-round nvestment. We generally start to observe nnovaton outcomes n and after the year of frm-j s IPO. To examne how VC falure tolerance affects startup frms nnovaton productvty, we estmate the followng baselne emprcal model: Ln( Innovaton Z Ind Year v (5) j, t ) Ln( FalureTolerance j,0 ) The constructon of Innovaton s dscussed n detal n Secton 3.1. Z s a vector of frm and ndustry characterstcs that may affect a frm s nnovaton productvty. Ind j and Year t capture two-dgt SIC ndustry fxed effects and fscal year fxed effects, respectvely. Snce VC- s falure tolerance s tme-nvarant for IPO frm-j, the panel data regresson as specfed above tends to downwardly bas the estmated effect of falure tolerance. Thus the reported results should be a conservatve estmate of the falure tolerance effect. In robustness checks, we also use both cross-sectonal regressons as well as the Fama-Macbeth regressons. j, t j t j, t 3.1 Proxes for Innovaton The nnovaton varables are constructed from the latest verson of the NBER patent database created ntally by Hall, Jaffe, and Trajtenberg (2001), whch contans updated patent and ctaton nformaton from 1976 to The patent database provdes annual nformaton regardng patent assgnee names, the number of patents, the number of ctatons receved by each patent, the technology class of the patent, the year when a patent applcaton was fled, and the year when the patent was granted. As suggested by the nnovaton lterature (e.g., Grlches, Pakes, and Hall 1987), the applcaton year s more mportant than the grant year snce t s closer to the tme of the actual nnovaton. We therefore construct the nnovaton varables based on the year when the patent applcatons are fled. However, the patents appear n the database only after they are granted. Followng the nnovaton lterature, we correct for the truncaton problems n the NBER patent data (see Appendx A pont C). We construct two measures of nnovatve productvty. The frst measure s the truncaton-adjusted patent count for an IPO frm each year. Specfcally, ths varable counts the number of patent applcatons fled n a year that are eventually granted. However, a smple count of patents may not dstngush breakthrough nnovatons from ncremental technologcal 11

14 dscoveres. Therefore, to capture the mportance of each patent, we construct the second measure by countng the number of ctatons each patent receves n subsequent years. It s true that patentng s a nosy measure of nnovaton productvty because t s only one of several ways frms use to protect returns from nnovatons. However, there s no clear reason to beleve that such nose, whch s n the regresson error term n (5), s systematcally correlated wth the VC falure tolerance measure. Also, we nclude both ndustry fxed effects and VC frm fxed effects (n later specfcatons), whch should effectvely control for the average dfferences n the propensty to patent nnovaton across ndustres and across VC frms. We merge the NBER patent data wth the VC-backed IPO sample. Followng the nnovaton lterature, we set the patent and ctaton count to be zero for IPO frms that have no patent and ctaton nformaton avalable from the NBER dataset. Table 1 Panel B presents the IPO frm-year summary statstcs of the nnovaton varables. On average, an IPO frm has 3.1 granted patents per year and each patent receves 2.5 ctatons. We also report summary statstcs for the subsample of frm-year observatons wth postve patent counts. Ths reduces the sample sze to 5,264 frm-year observatons. The medan patent count per year s 3 and the mean s On average, each patent receves 9.4 ctatons. Snce the dstrbuton of patent counts and that of ctatons per patent are hghly rght skewed, we use the natural logarthms of patent counts and ctatons per patent as the man nnovaton measures n our analyss Control Varables Followng the nnovaton lterature, we control for a vector of frm and ndustry characterstcs (Z) that may affect a frm s nnovaton productvty. In the baselne regressons, Z ncludes frm sze (measured by the logarthm of sales), proftablty (measured by ROA), growth opportuntes (measured by Tobn s Q), nvestments n nnovatve projects (measured by R&D expendtures over total assets), captal expendture, leverage, nsttutonal ownershp, frm age (measured by years snce IPO), asset tangblty (measured by net PPE scaled by total assets), and ndustry concentraton (measured by the sales Herfndahl ndex). Detaled varable defntons are n Appendx B. 7 To avod losng frm-year observatons wth zero patent or patent ctaton n the logarthm transformaton, we add a small number (0.1) to the actual value when calculatng the natural logarthm. 12

15 We extract fnancal nformaton for the IPO frms from Standard & Poor s COMPUSTAT fles, stock prces and shares outstandng data from CRSP, and nsttutonal nvestors ownershp from the Thomson Fnancal 13f nsttutonal holdngs database. In the end, there are 1,848 VC-backed IPO frms n our sample wth non-mssng VC nvestor characterstcs, fnancal and ownershp nformaton. All the fnancal varables n the analyss are wnsorzed at the 1 st and 99 th percentles to mtgate the nfluence of outlers on the results. Table 1 Panel C reports the summary statstcs of IPO frm characterstcs. The average IPO frm has total book assets of $485.5 mllon, sales of $375 mllon, leverage of 34.64%, net PPE rato of 17.36%, and Tobn s Q of FAILURE TOLERANCE AND CORPORATE INNOVATION 4.1 Baselne Results Table 2 reports the baselne results on how VC falure tolerance affects a startup frm s nnovaton productvty. Snce both nnovaton and Falure Tolerance are n the logarthm forms, the regresson coeffcent estmate gves us the elastcty of nnovaton to Falure Tolerance. All regressons nclude year fxed effects and ndustry fxed effects. The Huber-Whte-Sandwch robust standard errors are clustered by IPO frms. Model (1) of Table 2 shows that IPO frms fnanced by more falure-tolerant lead VC nvestors tend to produce more patents. The estmated elastcty of patents to Falure Tolerance s Ths means that a one percent ncrease n Falure Tolerance on average leads to more than a half percent ncrease n the number of patents per year. To be more concrete, consder a VC frm at the 25 th percentle of the falure tolerance dstrbuton. Accordng to Table 1 Panel A, ths VC frm on average nvests for 1.8 years before termnatng a project. If ths VC frm s wllng to nvest for 3.4 years before gvng up a project (the 75 th percentle of the falure tolerance dstrbuton), then holdng everythng else equal the IPO frms backed by ths VC frm tend to have 50.4% ( * ) more patents per year later on. 1.8 In model (2) we repeat the regresson wth the man explanatory varable replaced by VC Syndcate Falure Tolerance. The VC syndcate s falure tolerance also has a postve and sgnfcant mpact on the IPO frm s nnovaton productvty. The estmated elastcty of patents to falure tolerance s Not surprsngly, the margnal mpact of VC syndcate s falure 13

16 tolerance on the IPO frm s nnovaton s much smaller than that of the lead VC s falure tolerance. Ths mples that the lead VC nvestor s atttudes towards falure matters a lot more for the venture s nnovaton. Models (3) and (4) of Table 2 show that frms backed by more falure-tolerant VCs also tend to produce patents wth hgher mpact. Model (3) shows that a one percent ncrease n the lead VC s falure tolerance on average leads to a 0.5 percent ncrease n ctatons per patent. Agan, the effect of falure tolerance contnues to be present when the VC syndcate falure tolerance measure s used n model (4). In un-tabulated regressons, we also exclude selfctatons when computng ctatons per patent. Our results are robust to such modfcaton. 8 We control for a comprehensve set of frm characterstcs that may affect a frm s nnovaton productvty. We fnd that frms that are larger (hgher sales), more proftable (hgher ROA), older, and have more growth potental (hgher Q) and lower exposure to fnancal dstress (lower leverage) are more nnovatve. A larger R&D spendng, whch can be vewed as a larger nnovaton nput, s assocated wth more nnovaton output. Larger nvestment n fxed assets (hgher captal expendtures) s also assocated wth hgher nnovaton productvty. Further, hgher nsttutonal ownershp s assocated wth more nnovaton, whch s consstent wth the fndngs n Aghon, Van Reenen, and Zngales (2009). Fnally, asset tangblty (measured by the net PPE over assets) and ndustry competton (measured by the Herfndahl ndex) do not sgnfcantly mpact a frm s nnovaton productvty. Overall, our baselne results suggest that a VC s tolerance for falure can ncrease a startup frm s nnovaton productvty. These results provde support for the mplcatons of Holmstrom (1989) and Manso (2010) that tolerance for falure s crtcal n spurrng nnovaton. 4.2 Robustness We conduct a set of robustness tests for our baselne results on alternatve econometrc specfcatons. Besdes the pooled OLS specfcaton reported n Table 2, we use the Fama- MacBeth regresson adjustng for auto-correlatons of coeffcent estmates and get an even stronger estmate for the falure tolerance effect. We also use a Tobt model that takes nto consderaton the non-negatve nature of patent data and ctaton data. We run a Posson 8 For example, the coeffcent estmate of Ln(Falure Tolerance) s (p-value<0.001) n model (3) of Table 2 when the natural logarthm of the non-self ctatons per patent s the dependent varable. 14

17 regresson when the dependent varable s the number of patents to take care of the dscrete nature of patent counts. We also control for the IPO year fxed effects nstead of the fscal year fxed effects n order to mtgate the effect of strategc IPO tmng on our results (Lerner 1994). The baselne results are robust n all the above alternatve models, and are thus not reported. The results are also robust to usng the alternatve measure of falure tolerance, Falure Tolerance 2, whch s based on the average number of fnancng rounds the lead VC nvestor made n ts past faled projects. For example, the coeffcent estmate for Ln(Falure Tolerance 2) n model (1) of Table 2 s (p-value = 0.05), and s (p-value = 0.01) n model (3). Focusng on the subsample of frms that has at least one patent n our sample perod yelds smlar results. For example, the coeffcent estmate for Ln(Falure Tolerance) n model (1) of Table 2 s (p-value = 0.05), and s (p-value = 0.01) n model (3). Ths mples that the VC falure tolerance effect s not drven by the large number of frm-year observatons wth zero nnovaton count. The majorty of the IPO sample s backed by lead VC nvestors from Calforna (26%), New York (21%), and Massachusetts (17%). To control for the potental effect of geographc dfferences on our results, we nclude a dummy varable for lead VC nvestors located n each of the three states n the baselne regressons. The estmated falure tolerance effect remans robust. For example, the estmated falure tolerance effect s (p-value < 0.001) n model (1) of Table 2, and s (p-value < 0.001) n model (3). Young VCs may not have a long enough hstory of faled projects and thus the estmate of ther falure tolerance can be nosy. As a robustness check, we exclude IPO frms wth lead VCs less than fve years old from the foundng date (about 21% of the IPO sample). Our man results hold. For example, the estmated falure tolerance effect s (p-value < 0.001) n model (1) of Table 2, and s (p-value < 0.001) n model (3). In Table 2 we control for ndustry fxed effects at the two-dgt SIC level. Alternatvely, we control for ndustry fxed effects usng three-dgt SIC and four-dgt SIC, and the baselne results hold. We also use the 10-ndustry, 18-ndustry, and 354-ndustry specfcatons n the Venture Economcs database for the ndustry fxed effects, and agan the baselne results hold. We also examne whether the effect of falure tolerance on nnovaton s monotonc. Is more falure tolerance always assocated wth hgher nnovaton productvty? In an unreported regresson, we replace Ln(Falure Tolerance) wth Falure Tolerance and ts squared term. We 15

18 fnd that the mpact of Falure Tolerance on patent counts s postve and sgnfcant (coeffcent = 0.589, p-value = 0.03), but the coeffcent estmate of the squared term s negatve and statstcally nsgnfcant (coeffcent = , p-value = 0.19). We fnd smlar results for patent ctatons. The evdence suggests that the effect of falure tolerance on nnovaton productvty s postve and monotonc n our sample. Snce the VC s falure tolerance s tme-nvarant for each IPO frm n our baselne regressons, an alternatve way to analyze the data s to run cross-sectonal regressons. Thus as our last robustness check, we estmate the VC falure tolerance effect n a cross-sectonal regresson and report the results n Table 3. The dependent varables are the total number of granted patents that were fled by each IPO frm wthn the frst fve years after IPO and the average number of ctatons each of these patents receved. We mpose the arbtrary fve-year threshold to facltate comparsons of nnovaton productvty across IPO frms. The ndependent varable s the lead VC s falure tolerance determned at the tme when the VC makes the frstround nvestment n the venture. The values of all control varables are measured as of the venture s IPO year. Unlke Table 2 where the observaton unt s IPO frm-year, the observaton unt n Table 3 s IPO frm. We frst nclude only the lead VC s falure tolerance n Table 3 model (1). The coeffcent estmate of Falure Tolerance s postve and sgnfcant. Also, the cross-sectonal varaton n VC falure tolerance (along wth ndustry and year fxed effects) explans about 39% of the cross-sectonal varaton n startup companes nnovaton productvty n the frst fve years after IPO. In model (2), we nclude all control varables as n Table 2. The coeffcent estmate of Falure Tolerance contnues to be postve and sgnfcant. We repeat the regressons n models (3) and (4) wth ctatons per patent as the dependent varables, and fnd smlar results EMPIRICAL IDENTIFICATION Although our baselne results are consstent wth the hypothess that VC nvestors falure tolerance leads to hgher ex-post nnovaton productvty n VC-backed startup frms, an alternatve nterpretaton of the results could be that falure-tolerant VCs as specfed n our study 9 In untabulated regressons, we replace the lead VC falure tolerance wth VC Syndcate Falure Tolerance, and results contnue to hold. We also replace separate ndustry and year fxed effects wth ndustry-year fxed effects to control for possble ndustry trends n nnovaton, and the results are robust to such modfcaton. 16

19 are n equlbrum matched wth frms that have hgh ex-ante nnovaton potentals, and hgh exante potentals lead to hgh ex-post outcomes. In ths secton we address ths alternatve explanaton as follows. Frst, we extend the smple model n Secton 2.1 to allow for the above endogenety. Then we rely on the model to understand the nature of the endogenety problem and to fnd the approprate soluton. Second, we look for further evdence of dentfcaton n the cross secton. We show that the margnal effect of VC falure tolerance on startup nnovaton s much stronger n ventures n whch the falure rsk s hgher and thus VC s tolerance for falure s more needed and valued. 5.1 What could be the Omtted Varables? In the smple model n Secton 2.1 we assume that VC nvestors are randomly matched wth projects n the nvestment pool wth average project qualty. Now suppose that dfferent VCs have dfferent project selecton preferences or abltes. 10 Such project selecton abltes can be reflected n the average qualty of projects undertaken by the VC. Let be the average qualty (or average NPV) of projects undertaken by VC-. Then the qualty of a project VC- undertakes s u, where u s stll the project-specfc qualty and s ndependent of. Projects undertaken by the same VC are correlated through, but have ndependent u. Ths s the only departure from the basc model n Secton 2.1. The rest of the assumptons are the same as n the basc model. VC- wll stop nvestng n the project when E ( 2 u 1,,..., ) 0. Ths mples that VC- s nvestment duraton n an eventually faled project s the smallest nteger n so that n n 1 ( ). Takng the logarthm on both sdes of the nequalty, we have ( ) 1 log( n ) log( ) log( ). (6) ( ) Equaton (6) provdes gudance for our emprcal dentfcaton. If dfferent VCs have dfferent project selecton abltes, a VC frm s nvestment duraton n an eventually faled 10 Snce we examne equlbrum matchng outcomes, the same analyss apples rrespectve of whether VCs select projects or projects select VCs. Thus for expostonal ease, when we dscuss selecton ablty, we descrbe t as selecton by VC nvestors. 17

20 project s postvely related to not only the VC s tolerance for falure ( 1/ ) but also ts project selecton ablty ( ). Holdng the VC s falure tolerance constant, the better the projects are on average (.e., a hgher ), the longer the VC s wllng to nvest n projects that underperform and eventually fal. The average sgnal s a functon of the project s dosyncratc qualty u. Thus we can wrte our man explanatory varable n the baselne regresson (5) as Ln FalureTolerance ) f (,, u ), ( past where u past represents the dosyncratc qualtes of VC- s past faled projects. The dependent varable n the baselne regresson s the ex-post nnovaton productvty of a future successful project-j undertaken by VC-. Our hypothess s that project-j s ex-post nnovaton outcome depends on VC- s falure tolerance ( 1/ ). But the ex-post nnovaton outcome certanly depends on the ex-ante nnovaton potental of the project as well, whch s related to the qualty of project-j, u. We can nterpret u as the dosyncratc j characterstcs of the future IPO frm-j. Then we can wrte our dependent varable as Ln j ( Innovaton j j ) g(,, u ). Now we can clearly see that can ntroduce an omtted varable problem n our baselne regresson. If VC nvestors ndeed dffer n ther project selecton abltes, then such ablty can postvely affect both the nvestment duraton n ts past faled projects and the nnovaton productvty of ts future successful projects. 11 However, our model mples that nether u nor u ntroduces an omtted varable problem n the baselne regresson. Ths s because projects undertaken by the same VC at dfferent ponts of tme are correlated through, but have ndependent u. The u of a past faled project s uncorrelated wth the u of a future successful project. Thus although past j j u past affects our ndependent varable, t does not affect the dependent varable. Smlarly, although u j affects our 11 One possble concern s that our measure of falure tolerance captures a VC s overconfdence. An overconfdent VC nvestor ncorrectly thnks that ts projects are better-than-average projects, and thus s unwllng to termnate them despte the underperformance. Such overconfdence certanly leads to longer nvestment duraton n eventually faled projects. However, the ex-post nnovaton outcome of a future successful project depends on the true qualty of the project rather than the perceved qualty by the overconfdent VC. Thus f a longer nvestment duraton n a faled project s drven by overconfdence, then there s no omtted varable problem. In other words, we do not expect VC overconfdence to systematcally predct hgh nnovaton outcome n startup frms. 18

21 dependent varable, t does not affect our ndependent varable and thus won t bas our estmate of β n the baselne regresson. Put dfferently, we do not need to worry about the dosyncrases of past faled projects that may affect our VC falure tolerance measure, nor the dosyncrases of the future IPO frms that may affect ther nnovaton productvty. 5.2 Controllng for VC Frm Project Selecton Ablty Extended Emprcal Model How can we effectvely control for the omtted varable problem related to nformaton n VC- s past faled projects can help predct the nnovatveness of ts future successful projects through, then should exhbt predctablty over tme. The smplest case s that? If s constant over tme. That s, the VC frm s project selecton ablty or nvestment preference does not change over tme. In ths case, ncludng VC frm fxed effects can effectvely control for the omtted varable problem. Note that our emprcal measure of VC falure tolerance s tme-varyng. Thus ncludng VC frm fxed effects gves us the estmate of the wthn-vc frm falure tolerance effect. The VC s project selecton ablty may also have a tme-varyng component. That s,. If the tme-varyng component s ndependent over tme, then ncludng VC frm t t t fxed effects s agan suffcent for addressng the omtted varable problem. Ths s because the related to past faled projects cannot predct the related to future projects. More lkely, the tme-varyng component of may exhbt a predctable tme trend. A reasonable conjecture s that the VC frm becomes better at project selecton as t accumulates nvestment experences over tme. Sorensen (2007) shows that more experenced VCs nvest n better projects. Thus we assume that t EXPt, where EXP t s VC- s nvestment experence and expertse at tme t. In ths case, both the tme-nvarant and the tme-varyng EXP t are omtted varables n the baselne regresson. We need to explctly control for both. In sum, to control for the VC s project selecton ablty, we extend our baselne regresson to nclude both the lead VC frm fxed effects and the lead VC frm tme-varyng nvestment experence and expertse and estmate the followng model: Ln ( Innovaton j, t ) Ln( FalureTolerance j,0 ) EXPj,0 controls t v (7) j, t 19

22 Both VC- s falure tolerance and ts experence are measured at the tme when the VC frm makes the frst round nvestment n the IPO frm-j. The parameter effects. The controls are the same as those n the baselne regresson. represents VC frm fxed We measure VC experence from three dfferent angles: past general nvestment experence, past successful experence, and ndustry expertse. For each lead VC frm and each year we compute four VC general nvestment experence measures: a) the total dollar amount the VC frm has nvested snce 1980 (Past Amount Invested); b) the total number of frms the VC frm has nvested n snce 1980 (Past Frms Invested); c) the total dollar amount the VC frm has rased snce 1965 (Past Fund Rased); and d) the age of the VC frm measured as the number of years snce ts date of ncepton (VC Age). These VC experence measures, especally the past funds rased, may also capture the degree of captal constrant the VC frm faces. A VC s project selecton ablty may be best reflected n ts past successes. For each VC frm and each year, we compute Past Successful Ext as the proporton of entrepreneural frms fnanced by the VC frm that have exted successfully through ether gong publc or acquston snce The VC lterature suggests that gong publc s a more desrable outcome than acqustons for both entrepreneurs and VC frms (see, e.g., Sahlman 1990, Brau, Francs, and Kohers 2003). Only frms of the best qualty may access the publc captal markets through an IPO (Bayar and Chemmanur 2008). Therefore, we also calculate Past IPO Ext as the fracton of entrepreneural frms fnanced by the VC frm that has gone publc snce Another mportant dmenson of a VC frm s experence s ts expertse n certan ndustres. We measure such ndustry expertse by examnng the concentraton of a VC s portfolo frms across ndustres. Followng the VC lterature, we construct an nvestment concentraton ndex for each VC frm n each year based on the Venture Economcs ndustry classfcaton (see detals n Appendx B). The measure equals zero f the VC frm s portfolo has exactly the same ndustry composton as the hypothetcal VC market portfolo, and ncreases as the VC s portfolo becomes more concentrated n a few ndustres. Table 1 Panel A shows that the average lead VC frm n a gven year s about 14 years old and has nvested 1.4 bllon dollars n 97 entrepreneural frms. Among all ventures the average lead VC frm has fnanced, 69% had a successful ext but only 24% went publc. The average lead VC s portfolo frms are concentrated n a few ndustres wth the nvestment concentraton ndex of