Fnancng Innovaton: Evdence from R&D Grants to Energy Startups Sabrna T. Howell 1 September 1, 2015 s Abstract Governments regularly subsdze new ventures to spur nnovaton. Ths paper conducts the frst large-sample, quas-expermental evaluaton of R&D subsdes. I use data on ranked applcants to the U.S. Department of Energy s SBIR grant program. An award approxmately doubles the probablty that afrmrecevessubsequentventurecaptalandhaslarge,postvempactson patentng and commercalzaton. These effects are stronger for more fnancally constraned frms. Certfcaton, where the award contans nformaton about frm qualty, lkely does not explan the grant effect on fundng. Instead, the grants seem to reduce nvestor uncertanty by fundng technology prototypng. s s 1 NYU Stern. Emal: showell@stern.nyu.edu. I am deeply grateful to Davd Scharfsten, Josh Lerner, Ramana Nanda, Raj Chetty, and Joe Aldy. I also thank Davd Yermack, Ad Sunderam, Jeremy Sten, John Van Reenen, Lug Pstaferr, Arel Pakes, Larry Katz, Adam Jaffe, Sam Hanson, Shane Greensten, Ed Glaeser, Jeff Furman, Lee Flemng, Gary Chamberlan, semnar partcpants, and anonymous referees. I am ndebted to Jame Vernon, Teryn Norrs, Tna Kaarsberg, Carl Hebron, Carla Frsch, Matthew Dunne, Jeff Dowd, and Ken Alston, all currently or formerly at the Department of Energy. Fundng for ths project s from the Harvard Lab for Economc Applcatons and Polcy and a NSF Graduate Research Fellowshp.
1 Introducton Governments regularly subsdze research and development (R&D) n new ventures. 2 One ratonale for such subsdes s that the prvate sector does not nternalze the socal benefts of nnovaton. 3 Another s that fnancal frctons lead to undernvestment n early-stage R&D. 4 Yet crtcs contend that government R&D subsdes are neffectve because they crowd out prvate nvestment or allocate funds neffcently (Lerner 2009). Despte opposng theoretcal arguments, we have lttle emprcal evdence about the effectveness of R&D subsdes. There s also lttle work on whether fnancng constrants are frst-order barrers to nnovatve startups. In the frst quas-expermental, large-sample evaluaton of R&D grants to prvate frms, I show that the grants have statstcally sgnfcant and economcally large effects on measures of nnovatve, fnancal, and commercal success. The study s based on a new, propretary dataset of applcatons to the U.S. Department of Energy s (DOE) Small Busness Innovaton Research (SBIR) program. The data nclude 7,436 small hgh-tech frms and over $884 mllon n awards from 1983 to 2013. The SBIR program has two stages: Phase 1 awards of $150,000, and Phase 2 awards of $1 mllon. A company must wn Phase 1 to apply for Phase 2. DOE offcals rank frms wthn compettons, and I explot these ranks n a sharp regresson dscontnuty desgn that compares frms mmedately around the award cutoff. In a semnal paper, Jaffe (2002) ponts out the challenge that the selecton problem poses a challenge to R&D grant evaluaton, and proposes the regresson dscontnuty approach. The Phase 1 award has powerful effects. In the long term, t ncreases a 2 In addton to the federal SBIR, many U.S. states have smlar programs. Parallels overseas nclude the UK s Innovaton Investment Fund, Chna s Innofund, Israel s Chef Scentst ncubator program, Germany s Mkromezzannfonds and ZIM, Fnland s Tekes, Russa s Skolkovo Foundaton, and Chle s InnovaChle. 3 For evdence that startups contrbute dsproportonately to economc growth, see Akcgt and Kerr (2013), Haltwanger et al. (2013), and Audretsch et al. (2006). 4 Grants mght ncrease nvestment f gven to startups that face excessvely costly external fnance. Frctons that can lead to such costly fnance and thwart prvately proftable nvestment opportuntes nclude nformaton asymmetry, asset ntangblty, and ncomplete contractng (Holmstrom 1989). 1
frm s cte-weghted patents by 2.5 tmes, and almost doubles a frm s chance of recevng venture captal (VC) nvestment from 10% to 19%. A Phase 1 grant also almost doubles the probablty of postve revenue and, condtonal on postve revenue, ncreases t by 30%. It also ncreases the probablty of survval and ext (IPO or acquston). Wthn two years of the grant, the effects on cte-weghted patents and VC are a bt more than half ther long term effect. These results mply that on average the grants do not crowd out prvate captal, and nstead transform some awardees nto prvately proftable nvestment opportuntes. The early stage grants enable new technologes to go forward. The effect does not appear to reflect reallocaton of captal from losers to wnners wthn compettons. Heterogenety across frm types n the grant mpact s consstent wth the grant easng fnancng constrants. Captal ntensve frms should be more fnancally constraned. Consstent wth ths predcton, I fnd that the grant s more useful for hardware than software frms n rasng VC. Young frms, frms wth lttle experence, and frms n emergng sectors should face greater fnancng constrants. Indeed, I fnd that the effects of the grant on VC, survval, and cte-weghted patents declne wth age, and the frst two declne wth prevous cte-weghted patents and wth sector maturty. In contrast to Phase 1, the larger Phase 2 grant has no measurable effect except for a small postve effect on cte-weghted patents. Almost 40% of Phase 1 wnners do not apply to Phase 2. Among frms that get VC wthn two years of Phase 1, 55% opt not to apply. One reason for ths s that frms are nelgble f an outsde nvestor owns more than 50%. Second, whle DOE does not montor how Phase 1 s used, completon of the proposed project s necessary to apply to Phase 2. Frms may have not used the money as orgnally planned. Thrd, the applcaton process s onerous, and some frms who receve external prvate fnance no longer deem t worthwhle to apply. The Phase 1 grant mght ease fnancng constrants through a certfcaton mechansm n whch the government s decson conveys postve nformaton to venture captalsts about the frm s technology. Alternatvely, the money tself may be useful n transformng a project from negatve to postve 2
net present value. A fundng mechansm has two possble channels. Frst, the grant could allow the entrepreneur to retan more equty; n the counterfactual, an nvestor mght requre such a large stake that entrepreneural ncentves could not be mantaned. Second, the startup mght use the grant to prove the vablty of ts technology. Ths prototypng channel could reduce nvestor uncertanty. I test for certfcaton by askng whether applcant ranks are correlated wth outcomes, condtonal on award status. Ratonal nvestors should vew the grant as a postve sgnal only f ranks are relevant to market outcomes. Ths s because a frm s rank wthn a competton, whch the nvestor does not observe, maps drectly to whether the frm wns, whch the nvestor does observe. 5 Condtonal on wn status, the ranks are relatvely unnformatve about outcomes. Intervews wth nvestors as well as addtonal emprcal evdence support the concluson that the grant does not serve as a sgnal. Instead, the evdence s most consstent wth the prototypng channel, where the grant enables proof-of-concept work that the frm cannot otherwse fnance. In a survey of all post-2004 grantees stll n busness, respondents overwhelmngly reported usng the money for ether basc research on a new technology or testng and demonstratng an exstng technology. Ths paper contrbutes to the R&D subsdy program evaluaton lterature. In the U.S., two semnal papers examnng SBIR awardees are Lerner (2000) and Wallsten (2000). Later studes examne non-u.s. R&D programs, such as Jaffe and Le (2015), Lach (2002), and Almus and Czarntzk (2003). Bronzn and Psell (2014) and Bronzn and Iachn (2011) use a regresson dscontnuty desgn to evaluate R&D grants to frms n Northern Italy. Consstent wth my results, they fnd the grants postvely affect patentng for the smaller frms. 6 Ths paper also bulds on the costly external fnance lterature, whch fnds evdence of fnancng constrants but has focused on large publc compa- 5 Offcals that rank frms do not determne the cutoff and are uncertan about the number of awards. 6 Other work ncludes Lnk and Scott (2010), González and Pazo (2008), Hennngsen et al. (2014), and Zhao and Zedons (2013). 3
nes and rarely studed R&D. 7 An excepton s Bond, Harhoff and Van Reenen (2005), who fnd evdence consstent wth fnancal constrants dscouragng nvestment n R&D. Econometrcally, my method s smlar to L (2015) and Jacob and Lefgren (2004), whch examne Natonal Insttutes of Health research grant applcatons. Unverstes and natonal labs must undertake basc R&D. Startups are an mportant mddle ground between these nsttutons and large frms, whch can effcently conduct appled, market-orented R&D (Grlches 1998; Aghon et al. 2008). Early stage grants to small frms releve a crtcal lqudty constrant on R&D nvestment n a captal-ntensve sector. Severe fnancng constrants at the seed stage, however, contrast wth evdence from Phase 2 that later stage projects may not suffer the same frctons. Ths study s man polcy mplcatons, therefore, are that the SBIR program - and potentally smlar programs - could acheve better outcomes through reallocatng money (1) from larger, later stage grants to more numerous small, early-stage grants; and (2) from older frms and regular wnners to younger frms and frst-tme applcants. However, the rght to apply to Phase 2 may have opton value n a bad post-phase 1 state, so reallocatng money to Phase 1 mght ts applcant pool. Optmal program sze and, more generally, whether government should be subsdzng prvate R&D are beyond the scope of ths paper. Ths paper s energy-specfc context relates to an mportant research agenda on nnovaton to mtgate clmate change. Acemoglu et al. (2012, 2014) model the competton between clean and drty technologes n both producton and nnovaton, and ask whether the socal planner should use carbon taxes or research subsdes. They fnd that optmal polcy heavly reles on research subsdes. Ther estmates are senstve to the choce of R&D elastcty. Ths paper provdes emprcal evdence that research subsdes can 7 Fnancng constrants are a central ssue n corporate fnance. A debate begnnng wth Fazzar, Hubbard and Petersen (1988) and Kaplan and Zngales (1997) has for the most part found nvestment to be senstve to cash flow shocks (e.g. Lamont 1997, Rauh 2006, Whted and Wu 2006). However, dentfcaton has been challengng, and there s lttle evdence on small or prvate frms (see Hall 2010). One excepton s Zwck and Mahon (2014), who use ataxpolcychangetoshowthatfnancngconstrantsaremoresevereforsmallerfrms. 4
ncrease clean nnovaton. I fnd that the grant s most effectve n hydropower, carbon capture and storage, buldng and lghtng effcency, and automotve technologes. Small frms n conventonal energy sectors, lke natural gas and coal, do not experence a measurable causal effect of a research subsdy, suggestng that they are not dsadvantaged (.e. are not fnancally constraned). In Secton 2, I explan the DOE SBIR settng and data. Secton 3 descrbes the regresson dscontnuty desgn and establshes ts valdty n my context. Sectons 4 and 5 contan the emprcal estmates of Phase 1 and 2 mpacts, respectvely. Secton 6 examnes how grants mght affect nvestor decsons. Robustness tests are n Secton 7. 2 The Settng: Context & Data Sources In the U.S., grants provde a sgnfcant share of fundng for hgh-tech entrepreneurs. The largest sngle source s the SBIR grant program, whch dsburses around $2.2 bllon each year. Congress frst authorzed the SBIR program n 1982 to strengthen the U.S. hgh technology sector and support small frms. Today, 11 federal agences must allocate 2.7% of ther extramural R&D budgets to the SBIR program; the requred set-asde wll ncrease to 3.2% n 2017. Though mportant n ts own rght, the SBIR program s also representatve of the many targeted subsdy programs for hgh-tech new ventures at the state level and around the world. Akn to staged VC fundng, the SBIR program has two Phases. Phase 1grantsfundproof-of-conceptworkntendedtolastnnemonths. Awardees are gven the $150,000 n a lump sum (the amount has ncreased stepwse from $50,000 n 1983). Phase 2, awarded about two years after Phase 1, funds more extensve, later stage demonstratons. Although the frm proposes to use the grant for R&D n ts applcaton, once the frm receves the lump sum there s no montorng or enforcement of ts use. However, to apply for Phase 2 a frm must (a) demonstrate progress on ther Phase 1 projects to apply for $1 mllon Phase 2 grants; and (b) not be more than 50% owned by a sngle outsde prvate equty nvestor. For both phases, elgblty s also contngent 5
on beng a for-proft, U.S.-based and majorty U.S.-owned frm. There s no requred prvate cost sharng, and the government takes no equty and demands no rghts to IP. The applcaton process s onerous, takng a full tme employee between one and two months. 8 Each year, DOE offcals n technology-specfc programs, such as Solar, develop a seres of compettons. 9 An applcant submts a project proposal to a relevant competton. Program offcals rank applcants wthn each competton based on wrtten expert revews and ther own dscreton, accordng to three crtera: 1) strength of the scentfc/techncal approach; 2) ablty to carry out the project n a cost effectve manner; and 3) commercalzaton mpact (Olver 2012). The ranks and losng applcant denttes are ndefntely non-publc nformaton. 10 The program offcal does not know the award cutoff (the number of grants n a competton) when she conducts the rankng. She submts ordered lsts to a central DOE SBIR offce, whch determnes the cutoff. 11 I use complete data from the two man appled offces at DOE, Fossl Energy (FE) and Energy Effcency & Renewable Energy (EERE). 12 Together, they awarded $884 mllon (2012 dollars) n SBIR grants over the course of my data from 1983 to 2013. The data nclude, for each applcant, the company name and address, funded status, grant amount, and award notce date. My analyss begns n 1995, the frst year for whch I have rankng data. 8 Applcants must descrbe the project and frm n detal, and provde an temzed budget for the proposed work. There are over 100 pages of nstructons on DOE s SBIR Phase 1 applcaton webpage. Intervews wth grantees confrmed the 1-2 month tmeframe. 9 Examples of compettons nclude Solar Powered Water Desalnaton, and Improved Recovery Effectveness In Tar Sands Reservors. 10 Only n my capacty as an unpad DOE employee was I able to use ths data. 11 Ths cutoff vares across compettons and s based on budget constrants. Rankng happens before the SBIR offce determnes how many awards are allocated to each program and competton. Intervews wth a varety of stakeholders at DOE ndcated that the cutoff decson s determned exogenously to the rankng process. Some of the rankng data was provded to me n the form of forwarded emals from program offcals to the SBIR offce, whch also provde evdence of exogenety. Further, observable varables do not predct competton cutoffs. Average award numbers do not vary systematcally by offce or competton sub-sector. 12 Appendx Fgure 1 shows all applcants by offce and award status. 6
Table 1 Panel 1 contans summary statstcs about the applcatons and compettons. Each competton has on average 10.6 applcants (standard devaton s 8). Over 70% of frms appled only once, and 14% appled twce. Panel 2 shows summary statstcs wthn the estmaton sample of varables drawn from sources other than DOE (except for sector and mnorty/womanownershp). I group the competton topcs nto sub-sectors, and create an ndcator for whether the technology s relatvely new, wthout a well-developed supply chan and demand nfrastructure. 13 Prvate fnancng deals are matched to applcant companes by name and state, and then hand-checked for accuracy. There are 838 frms wth at least one prvate fnancng deal. 14 The amount s avalable for 57% of deals; My prmary outcome varable, however, s an ndcator for whether the frm receved VC nvestment after ts frst grant award date (VC post ). I fnd smlar results usng the amount data, but prefer the ndcator because t has better coverage and because the ownershp percentage purchased s almost never avalable. Dollars wthout the equty stake would conflate success at rasng prvate nvestment wth the frm s captal ntensty. Patents and ther ctatons are proxes for nnovaton. Usng comprehensve patent data from 1976 to 2014 from Berkeley s Fung Insttute, I match 1,471 post-1995 applcant frms to at least one non-ressue utlty patent. 15 To frame patents around the award date, I use the patent applcaton date. I do not normalze the patent count by USPTO classfcaton or year because competton fxed effects control for sub-sector and date. Cte-weghted patents are the flow of future ctatons from a frm s future patents. 13 Emergng Sector =1fthesectorssolar, wnd, geothermal, fuel cells, carbon capture and storage, bomass, or hydro/wave/tdal; and zero f the sector s ol, gas, coal, bofuels, or vehcles/motors/engnes. More ambguous sectors are excluded. There are 886 companes assgned to the mmature category, and 481 to the mature. Unclear frms and/or topcs are excluded. 14 I use the ThompsonOne, Preqn, Cleantech Group 3, CrunchBase, and CaptalIQ databases. Summarzed n Appendx Table 1. I categorze angel fnancng as VC, because both nvestor types target hgh-growth startups. The VC Amt post varable omts frms that rase VC after but not before the award for whch there s no deal amount avalable. 15 Subsdary patents are not a problem as the frms are overwhelmngly small, prvate, and sngle-unt. 7
Table 1: Summary Statstcs Panel 1: Applcaton Data from DOE 1983-2013 #Phase1Applcatons 14,522 #UnquePhase1ApplcantFrms 7,419 # Compettons 1,633 1995-2013 #Phase1Applcatons 9,659 #UnquePhase1ApplcantFrms 4,545 # Phase 1 Applcatons wth rankng data used n RD 5,021 # Phase 1 Compettons used n RD ( 1 award) 428 Average # Phase 1 Applcants per Competton 11 (8.3) Average # Phase 1 Awards per Competton 1.7 (1.1) #Phase2ApplcatonsusednRD 919 Panel 2: Varables Used n Analyss from Non-DOE Sources Type Mean Std Dev Medan N Pre-award venture captal (VC) nvestment 0-1.083.27 0 5,021 Pre-award # venture captal deals Count.25 1.3 0 5,021 Pre-award # cte-weghted patents Count 21 122 0 5,021 Pre-award # patents Count 1.9 7.5 0 5,021 Pre-award acquston or IPO 0-1.033.18 0 5,021 Post-award VC VC post 0-1.11.31 0 5,021 Post-award VC (mll real 2012$) Cont. 2.7 26 0 4,964 (VC Amt post ) Post-award # VC deals (VC Deals post ) Count.32 1.4 0 5,021 Post-award # cte-weghted patents Ctes post Count 12 117 0 5,021 Post-award # patents Count 2 11 0 5,021 Post-award Acquston or IPO (Ext post ) 0-1 0.034 0.18 0 5,021 Revenue as of 2016 n $ mllons (Revenue ) Cont. 2.0 6.6 0.20 3,583 Survval as of 2016 (In Bus post ) 0-1 0.67 0.47 1 3,880 Probablty n major metro area (top 6) 0-1 0.30 0.46 0 5,021 Age Cont. 9.5 11 6 3,427 Probablty tech s hardware (Hardware ) 0-1 0.43 0.49 1 2,571 Prob. new sub-sector (Emergng Sector ) 0-1 0.58 0.49 1 2,571 Probablty mnorty owned 0-1 0.077 0.27 0 1,722 Probablty woman owned 0-1 0.084 0.28 0 1,722 # all-gov t SBIR wns Count 10 36 0 5,021 Future patents n modal class Count 9,758 11,809 5,453 1,583 MSA VC nvestment 2011 ($ mll) Cont. 851 1,570 0 4,950 MSA medan per cap. ncome 2011 ($ thou) Cont. 56 14 56 4,603 Note: Panel 1 of ths table summarzes the DOE SBIR data. Panel 2 summarzes applcant characterstcs data. 8
Data was manually collected on frm operatng status and most recent year revenue as of January, 2016 for the 2,994 post-2000 applcants who appear n the prmary sample. 16 Of these, 1,737 were stll n busness (survved) and 1,721 have postve revenue. One thousand frms were determned to be out of busness, and assgned zero revenue. Durng ths research process, 744 companes were determned to produce prmarly hardware, and 971 software (Hardware =0). Companes wth ambguous technologes or that produce both are omtted. Fnally, I use MSA-level data from the Federal Reserve Economc Data research center. 3 Emprcal Strategy Regresson dscontnuty (RD) desgn estmates a local average treatment effect around the cutoff n a ratng varable. Publc agences resst evaluaton by randomzaton, so RD s the best alternatve source of exogenous varaton (Jaffe 2002). Snce the number of applcants and awards vares across compettons, I center the ranks around zero n my prmary specfcaton. The lowest-ranked wnner has centered (normalzed) rank 1, and the hghest-ranked loser has normalzed rank -1. Each competton that I consder has at least ths par. As I expand the bandwdth, [ r, r], I nclude hgher ranked wnners and lower ranked losers. I also use percentle ranks to address composton ssues. I estmate varants of Equaton 1, where Y Post s the outcome. The coeffcent of nterest s on treatment, and f( ) s a polynomal n the frm s rank wthn competton c. 17 Includeafullsetofdummesforeachcompet- 16 Research assstants collected ths data from D&B and company webstes. A company s n busness (survved) f t has an actve webste. I do not use Census data because my data does not nclude frm EINs. Also, the Census Longtudnal Busness Database (LBD) does not contan revenue. Efforts to lnk the LBD to Census Busness Regster tax data, whch provde some ndcaton of revenue, are ongong (see Haltwanger et al. 2016 for dscusson). 17 Standard RD mplementaton pools the data but allows the functon to dffer on ether sde of the cutoff (Imbens and Lemeux 2008). However, I potentally have too few ponts to the rght of the cutoff to estmate a control functon separately on both sdes, so I rely on global polynomals for my prmary specfcaton. I show that my results are robust to allowng the slope coeffcents to dffer. 9
ton c, whch are date-specfc. In some specfcatons I use controls, denoted by X,suchasthethepre-assgnmentoutcomevarable. 18 My estmatons use OLS where possble, but for bnary varables and count data I provde results usng the most approprate model, ncludng logstc and negatve bnomal. 19 Y Post c = + [1 Norm Rank c > 0] + f (Norm Rank c )+ X c + c + (1) " c where r apple Norm Rank c apple r One way ths settng dffers from conventonal RD s that the rankng s ordnal rather than cardnal. However, on average the dfferences n the true dstance between ranks should be the same. That s, errors n dfferences on ether sde of cutoff n any gven competton should average out. An mportant data lmtaton s the ratng varable s dscreteness. Lee and Card (2008) note that dscrete ratng varables can requre greater extrapolaton of the outcome s condtonal expectaton at the cutoff, though the fundamental econometrcs are not dfferent. To determne the approprate polynomal, I employ Lee and Card s (2008) goodness-of-ft test for RD wth dscrete covarates, whch compares unrestrcted and restrcted regressons. 20 by sector-year n the man specfcaton. Iclusterstandarderrors The prmary concern n any RD desgn s whether agents postons (here 18 The RD desgn does not requre condtonng on baselne covarates, but dong so can reduce samplng varablty. Appendx Table 3 projects rank on observable covarates. Prevous non-doe SBIR awards are the strongest predctor of rank. A one standard devaton ncrease n prevous SBIR wns (the mean s 11.4 and the standard devaton s 38) ncreases the rank by nearly one unt. Prevous VC deals also have a small postve mpact. I nclude these two varables n my prmary specfcatons. Lee and Lemeux (2010) advse ncludng the pre-assgnment dependent varable as they are usually correlated. 19 IuseOLSnmyprmaryspecfcatonbecausemanyofthegroupsdefnedbyfxed effects (compettons) have no successes (e.g. no subsequent VC). Logt drops these groups. Also, OLS does as well as logt f not better n estmatng margnal effects (Angrst 2001). 20 The unrestrcted regresson projects the outcome on dummes for each of K ranks. The restrcted regress s a polynomal lke Equaton 1. The goodness-of-ft statstc s: G (ESS Restr. ESS Unrestr. )/(K P ) ESS Unrestr. (N K), where ESS s the error sum of squares from regresson, N s the number of observatons, and P s the number of restrcted parameters. G takes an F-dstrbuton. The null hypothess s that the unrestrcted model does not provde a better ft. If G exceeds ts crtcal value, I reject the null and turn to a hgher order polynomal. Results n Secton 4. 10
ranks) are manpulated around the cutoff for treatment. A vald RD desgn that approxmates a a local randomzed experment must have a cutoff that s exogenous to rank (Lee and Lemeux 2010). In my settng, manpulaton could happen ether f (a) hgher qualty applcant frms tred harder on ther applcatons to make them relatvely more lkely to be accepted; or (b) the program offcal manpulated applcants around the cutoff. Manpulaton s dffcult n ths settng, snce nether the applcant frms nor the program offcers (DOE cvl servants) know the selecton rule (See secton 2). To test for manpulaton, I examne observable varables around the cutoff usng fve types of tests. Frst, Appendx Fgure 2 shows the densty of applcants by normalzed rank. There s no obvous dscontnuty around the cutoff. 21 Second, I demonstrate smoothness n observable baselne covarates (Fgure 1) and pre-assgnment outcome varables (Fgures 2A, 3A). In none of the fgures s there any statstcally sgnfcant dscontnuty around the cutoff. Thrd, I use the baselne covarates to predct the probablty of subsequent VC, and sort the resultng coeffcents by the applcant s rank around the cutoff. 22 There s no dscontnuty, n strkng contrast to the actual outcome n Fgure 1B. Fourth, I conduct a t-test for matched par dfferences of means n baselne covarates mmedately around the cutoff. In no case can I reject the null hypothess that the means are the same for both one- and two-taled tests, except for the one-taled test on prevous ctatons, whch s sgnfcant at the 10% level (Appendx Table 4). 21 Unfortunately, the dscreteness of the runnng varable prevents a McCrary densty test. 22 Iregresstheoutcomeonbaselnecovaratesandcompettondummestoobtana weghted average of the covarates by relevance to the outcome: Y Post c = + X + c + " c. For each applcant I then use the estmated coeffcent vector to predct the probablty of subsequent VC fnancng: Ŷ Post c =ˆ + X ˆ + ˆc. Iaveragetheprobabltesforeachrank and plot them n Appendx Fgure 3. 11
Fgure 1: Observable Covarates Note: Ths fgure demonstrates contnuty n observable covarates. Each varable s measured at the tme of the grant decson. Capped lnes ndcate 95% confdence ntervals. Fnally, I test whether awardees underlyng technologes are more predsposed to growth than those of non-awardees. As a proxy for future nnovaton and growth n a specfc sector, I use future patents n the frm s modal patent sub-class. I demonstrate that future patents do not predct rank or award n Appendx Table 5. Vsually, I confrm ths n Appendx Fgures 4 and 5. The latter fgure s a 3D graph showng that the number of frms wth acertanmodalsubclasssevenlydstrbutedaroundthecutoffforaward.in estmatng the grant effect usng Equaton 1, varaton s wthn competton. It s not possble that rankngs capture unobserved subclass-based predsposton to technologcal growth f there s no varaton n subclasses around the cutoff. AvaldvaldRDdesgnalsorequresthattreatmentnotcauserank. Ths holds here, as the award happens after rankng. I exclude applcants who 12
prevously won n my prmary specfcaton. Also, the functonal form must be correctly specfed, else the estmator wll be based. I perform goodness-of-ft tests and show that rank s unnformatve. Program offcals observe more data than the econometrcan, so t s mpossble to fully test the assumpton of no sortng on observables n the neghborhood of the cutoff. Nonetheless, ths preponderance of evdence suggests the RD desgn s vald. Robustness tests of the man results are provded n Secton 6. 4 The Phase 1 Grant Impact on Frm Outcomes Innovaton I begn wth the best avalable proxy for nnovaton: patentng. Though patents are only one way that frms protect IP, they are postvely assocated wth economc value creaton and stock market returns (Hall et al. 2005). The deal experment observes whether frms nvest exogenous cash n R&D, n whch case costly external fnance must have prevented the frm from explotng exstng proftable nvestment opportuntes. The lterature has establshed that R&D s rarely fnanced wth debt, but t remans unclear whether fnancal constrants cause R&D cash flow senstvty (see Hall 2010). Here I fnd that proftable R&D nvestment would not occur n the absence of a subsdy, suggestng that fnancal constrants nhbt nvestment, as n L (2011) and Faulkender and Petersen (2012). Followng Aghon, Van Reenen and Zngales (2013), I focus on cteweghted patents usng negatve bnomal and OLS specfcatons, wth dependent varables Ctes post and ln 1+Ctes post,respectvely. 23 Fgure 2 shows log cte-weghted patents before and after the grant. DOE clearly does not rank on the bass of prevous patentng, but ranks among hgh-rankng losers may be predctve of future patentng. Ths relatonshp dsappears for wnners, and s not statstcally sgnfcant n regressons. 23 The Pearson goodness-of-ft 2 suggests that the data are excessvely dspersed for the Posson regresson model, so I rely on the negatve bnomal dstrbuton. 13
Fgure 2: Cte-Weghted Patents Before and After Phase 1 Grant by Rank Note: Ths fgure shows log frm patents pror and after the Phase 1 grant award decson. Applcants are bnned by DOE s rank, whch s centered so Rank > 0 ndcates a frm won an award. The date assocated wth a successful patent s the patent applcaton date. Capped lnes ndcate 95% confdence ntervals. Table 2 reports the results. Usng the prmary sample of applcants that have not prevously won wthn my data, I fnd that an award ncreases cte-weghted patents by 2.5 tmes, usng the preferred negatve bnomal specfcaton (Panel 1). 24 Columns I and II use a bandwdth of just one rank (one applcant) on ether sde of the cutoff wthn each competton, and Column III uses all the data and controls for rank lnearly and quadratcally. The OLS specfcaton fnds that the grant ncreases log cte-weghted patents by about 30 percent (Panel 2 columns I-III; the normalzaton and lnearzaton lead to a smaller measured effect). Statstcally, rank s unnformatve. 24 Coeffcents ndcate, for a one unt change n the covarate, the dfference n the logs of expected counts. If s the Posson rate (the number of patents), the model s log ( )= + [1 R c > 0], where covarates other than treatment are omtted. We can R wrte = log ( Rc>0) log ( Rc<0) = log c >0.Exponentatngthecoeffcent gves R c <0 the ncdence rate rato (IRR), how many tmes more patents awardees are expected to have compared to losers. 14
Table 2: Impact of Phase 1 Grant on Cte-Weghted Patents Panel 1: Negatve bnomal model; dependent varable s Ctes post Sample: No prevous wnners All No prev. apps >2 prev. wns Bandwdth: 1 1 All All 1 1 1 I. II. III. IV. V. VI. VII. Award.93***.91***.92***.94***.82*** 2.1***.40*** (.21) (.19) (.33) (.26) (.13) (.34) (.14) Normalzed rank.052 (.074) Normalzed rank 2 -.0072 (.0045) Rank quntles N N N Y N N N Controls N Y N N N N N Sector-year f.e. Y Y Y Y Y Y Y N 1871 1871 5020 5020 2713 972 1477 R 2 0.056 0.084 0.053 0.053 0.034 0.080 0.035 Panel 2: OLS models; dependent varable s ln 1+Ctes post Sample: No prevous wnners All No prev. apps >2 prev. wns Bandwdth: 1 1 All All 1 1 1 I. II. III. IV. V. VI. VII. Award.33**.27**.29***.22**.29***.49**.22* (.15) (.13) (.11) (.087) (.087) (.21) (.13) Normalzed rank -.026 (.02) Normalzed rank 2.00078 (.0011) Rank quntles N N N Y N N N Controls N Y N N N N N Competton f.e. Y Y Y Y Y Y Y N 1872 1872 5021 5021 2714 972 1477 Pseudo-R 2 0.46 0.60 0.53 0.351 0.63 0.71 0.75 Note: Ths table reports regresson estmates of the effect of the Phase 1 grant (1 R > 0) on cte-weghted patents. The specfcatons are varants of Equaton 1. Controls are prevous ctes or log(prevous ctes +1), and prevous SBIR wns. Competton f.e. n the NB model do not permt convergence. Standard errors are clustered by sector-year. *** p<.01. Year 1995 15
Centerng ranks mght obscure nformaton n the raw rank. For example, frms wth centered ranks of two mght have dfferent qualtes n compettons wth two and four awards. I address ths usng percentle (here, quntle) rank controls n Panels 1 and 2 column IV. 25 Condtonal on award status, there s no nformaton n rank vsually or n regressons, regardless of bandwdth. The goodness-of-ft test reveals that once I control for award, no functon s too restrctve. The rght sde of Table 2 splts the sample. Column V consders all applcants, so that frms that wn multple grants are ncluded multple tmes. Column VI lmts the sample to frst tme applcants, and yelds a sgnfcantly larger effect than the standard sample. When only frms wth more than two wns are ncluded (column VII), the effect s smaller than n the man sample. Ths s because the effect declnes n the number of awards. When the award s nteracted wth the number of prevous non-doe awards, the effect declnes sgnfcantly (Appendx Table 7). For each prevous SBIR award, the effect of an award on log cte-weghted patents declnes by 20 percent. I fnd asmlarpatternforotheroutcomevarables,buttsespecallytroublng for patentng. A small subset of applcants wn many awards and may be dependent on grants; whle such frms mght naturally not be seekng VC, f ther R&D s productve t should yeld patents. My results suggest a steeply declnng beneft of addtonal grants to the same frm. Fnance Startups man source of external fundng s venture captal (VC); lesser sources nclude partnerng wth large companes and venture debt. The lterature has establshed that venture captalsts are mportant ntermedares n the U.S. nnovaton system. They select nnovatve frms and brng new technologes to market quckly (Hellmann and Pur 2000, Sørensen 2007). VCs can facltate access to debt fnance, and they provde non-monetary resources lke ntensve montorng, governance, legal servces, and networkng (Hochberg et al. 2014). 25 I fnd smlar results wth the slope controlled for separately on each sde of the cutoff (wth a bandwdth of all specfcaton) and usng quartle ranks. 16
Fgure 3: Probablty of Venture Captal Before and After Grant by Rank Note: Ths fgure shows the fracton of applcants who receved VC pror to (1A) and after (1B) the Phase 1 grant. Capped lnes ndcate 95% confdence ntervals. Not all SBIR applcants are startups, and not all seek external fnance. Ahandfuloffrmsapplymanytmesandapparentlysubsstfordecadeson SBIR grants. However, a majorty of the frms qualty as startups; they are young, small, and have locaton-unconstraned growth potental. 26 The medan frm age s sx years, wth many less than one year old. The lterature has used SBIR wnners as representatve samples of hgh-tech entrepreneural frms (Hsu 2006, Gans and Stern 2003). Also, DOE consders moblzng prvate nvestment an mportant goal. VC s therefore an approprate outcome metrc. Observng VC tests whether the grants moblze or crowd out prvate nvestment, ndcates that the company presents a prvately proftable nvestment opportunty, and s a good early-stage proxy for market success n a context where outcome data are dffcult to collect. Vsual evdence for a grant treatment effect on VC s n Fgure 3. The probablty of subsequent VC jumps from about 10% to 20% around the grant cutoff. Estmates of Equaton 1 n Table 3 Panel 1 vary the bandwdth of ranks around the cutoff. The estmates range from 7 to 14 pp; my preferred estmate 26 For example, among the 23 solar frms that have ever had an IPO, 9 appear n my data; SBIR wnners nclude Sunpower, Frst Solar, and Evergreen Solar. 17
of 10 pp uses the narrowest bandwdth (columns I-II). The average lkelhood of VC after the grant s 11%; among losers t s 9%, and among wnners t s 21%. Quntle ranks yeld stable grant effects, regardless of bandwdth, of 9-10 pp. When I nclude all prvate fnancng events, such as IPOs, acqustons, and debt, I fnd a slghtly larger effect of about 12 pp (Appendx Table 6). 27 I depart from the prmary sample of no prevous wnners n Panel 2. Permttng prevous wnners and excludng prevous applcants have lttle effect on the results (columns I-IV), but when I nclude frms wth more than two prevous wns (columns V-VI) the effect drops and loses sgnfcance. Appendx Table 7 shows that the coeffcent on award remans a sgnfcant 12 pp whle the nteracton between award and prevous wns s -4 pp, sgnfcant at the 10% level. I also estmate the grant effect on the log amount of VC rased subsequently and on the number of deals (Panels 1 and 2 column VII). 28 The results mply that the grant generates over 100% more VC nvestment n dollars and 2.4 addtonal VC deals. A grant mght ncrease a wnner s chance of VC by decreasng the losers chance. To test for such a reallocaton of captal, I frst ask whether the effect vares wth the competton s number of awards. Negatve spllovers should ncrease wth the number of wnners, because compettons are defned by narrow sub-sectors. When the sample s lmted to losers, dummes for the number of awards have no predctve power, suggestng that spllovers do not explan the man effect (Appendx Table 8 columns I-IV). Second, I explot the stylzed fact that VC frms typcally nvest n geographc proxmty to ther offces, and ndeed n frms located n ther cty (Chen et al. 2010, Cummng and Da 2010). Nearby frms are more lkely nvestment from the 27 Dvdng the sample by tme perod reveals that the grants were most effectve between 2009 and 2013 (at 19 pp) and least effectve between 2000 and 2004 (at 5 pp), perhaps because VC frms reduced nvestng when the nternet bubble collapsed. Durng the Stmulus years of 2009-2011, when DOE fundng was unusually hgh, the estmated grant effect s 13 pp. The fundng envronment may explan these temporal dfferences. 28 The amount specfcaton omts frms that rase VC after but not before the award for whch there s no deal amount avalable. I fnd smlar results for the amount usng a zero-nflated model. I use a negatve bnomal specfcaton for deals because the dependent count varable s over-dspersed (varance greater than the mean). 18
same VC frms. If the grant causes reallocaton, ts effect should be larger n compettons where wnners and losers are from the same area. The effect when competng frms are from the same MSA s not sgnfcantly hgher (Appendx Table 8 columns V-VII). I conclude that the measured effect does not reflect wthn-applcant reallocaton, but future research s needed to assess whether captal s reallocated from non-applcant frms. 19
Table 3: Phase 1 Grant Impact on Subsequent Venture Captal Investment Panel 1 Dependent varable: VC post Bandwdth: 1 3 All 1 I. II. III. IV. V. VI. VII. ln 1+VC Amt post Award.098***.10***.12**.094***.072**.10*** 1.4*** (.032) (.033) (.058) (.033) (.033) (.028) (.5) Normalzed rank -.029.0086 (.033) (.0071) Normalzed rank 2.012-7.4e-5 (.0088) (.00043) Rank quntles N N N Y N Y N Controls Y N Y Y Y Y N Competton f.e. Y Y Y Y Y Y Y N 1872 1872 3368 3368 5021 5021 1843 R 2 0.47 0.42 0.35 0.35 0.27 0.27 0.42 Dependent varable: VC post Panel 2 VC Deals post No prev. wns Sample: All applcants No prev. applcants >2 prev. wns Bandwdth: 1 All 2 All 1 All 1 I. II. III. IV. V. VI. VII. Award.11***.075***.11***.10**.080*.074.93*** (.023) (.026) (.037) (.048) (.048) (.053) (.19) Normalzed rank, N Y N Y N Y N normalzed rank 2 Competton f.e. Y Y Y Y Y Y N N 2714 6400 1514 2951 1246 2322 1871 R 2 0.37 0.19 0.50 0.28 0.51 0.35 0.051 (Pseudo-R 2 ) Note: Ths table reports estmates of the effect of the Phase 1 grant (1 R > 0) onsubsequent VC nvestment usng varants of n Eq. 1 (VII s negatve bnomal). Controls: prevous VC, prevous all-gov t SBIR awards. (Not used n Panel 2.) Year-sector. Standard errors are robust and clustered at the sector-year level. *** p<.01. Year 1995 20
Revenue From a program evaluaton perspectve, perhaps the most mportant outcomes are frm revenue and survval. Fgure 5 shows log revenue as of January 2016. 29 There s agan no slope n rank, but a clear jump at the cutoff. Unfortunately, the revenue data are not dated around the award but rather s collected as the most recent year s revenue as of early 2016. Therefore, there s no preaward graph. However, the regresson results nclude tme controls, ether through competton or year fxed effects, so there s no sense n whch recent compettons are advantaged. Iusezero-nflatednegatvebnomal(ZINB),OLS,andTobtmodels. Table 4 Panel I columns I-II use a smple OLS specfcaton and fnd effects of $1.3-$1.7 mllon (relatve to a mean of $2 mllon). When I omt zero revenue observatons and use the log of revenue as the dependent varable (column III), the effect s a 19% ncrease. These results reflect the fact that the grant propels frms from zero revenue to postve revenue. Ths s clear from the ZINB model, whch s my preferred specfcaton; negatve bnomal because the dependent varable s agan over-dspersed, and zero-nflated because the Vuong test statstc ndcates t s preferred to the standard verson. ZINB provdes two estmates of the award effect: frst a logstc porton predctng the lkelhood of a zero, and then a full model predctng revenue. Table 4 Panel 1 columns IV-VI fnd that recevng an award decreases the odds that afrmhaszerorevenuebytwotmes(exponentatngthemostconservatve coeffcent, -0.67). If a frm s not n the certan zeros group, the award ncreases revenue by about 30% (exponentatng the 0.25 coeffcent). Panel 2 columns IV-VI splt the sample. Interestngly, for frms wth more than two prevous wns, the logstc porton becomes slghtly larger and remans hghly sgnfcant; Appendx Table 7 shows usng the OLS specfcaton 29 Unfortunately, there s no pre-applcaton data. The sample conssts of applcants post-2000, and frms found to be out of busness are assgned zero revenue. Frms n busness for whom revenue could not be ascertaned are omtted. 21
that the nteracton between the grant effect and prevous SBIR awards s postve only for revenue. The Tobt specfcaton n Panel 2 columns I-III fnds a treatment effect of $2.3 mllon; the larger effect s because zero values are treated as censored, rather than true zeros. As wth fnancng, I fnd that rank has no predctve power. 30 Fgure 4: Phase 1 Grant Impact on Log Revenue ($, mllons) Note: Ths fgure shows revenue as of January 2016. Capped lnes ndcate 95% confdence ntervals. 30 The G-value from the goodness-of-ft test wth no control for rank s 0.0001, orders of magntude less than the crtcal value of 1.47 wth 5% confdence. 22
Table 4: Phase 1 Grant Impact on Revenue (Mllon $) Panel 1 Model: OLS Zero-nflated negatve bnomal (ZINB) Bandwdth: 1 All 2 1 All All IV. V. VI. I. II. III. DV = LnRev Award 1.7* 1.3**.19*.25**.29**.34*** (.93) (.58) (.11) (.12) (.13) (.096) Award (logstc ZINB) -.67*** -.82*** -.76*** (.16) (.17) (.13) Norm. rank, Norm. N N N N Y N rank 2 Rank quntles N Y N N N Y Controls N N N N Y N Competton f.e. Y Y Y N N N Sector & year f.e. N N N Y Y Y N 1108 3942 1176 1108 3583 3942 R 2 /Dev-based R 2 0.40 0.14 0.16 0.11 0.10 0.06 Panel 2 Model: Tobt Zero-nflated negatve bnomal (ZINB) Bandwdth: 1 All All 1 1 1 Sample: No prevous wnners All apps No prev. apps >2 prev. wns I. II. III. IV. V. VI. Award 2.3*** 2.5*** 2.5***.53***.39**.34 (.72) (.72) (.59) (.10) (.16) (.26) Award (logstc ZINB) -.91*** -1.0*** -1.0*** (.14) (.20) (.33) Norm. rank, Norm. N Y N N N N rank 2 Rank quntles N N Y N N N Sector & year f.e. Y Y Y Y Y Y N 1108 3583 3942 1383 657 329 PseudoR 2 /Dev-based R 2 0.0020 0.0010 0.0010 0.10 0.17 0.18 Note: Ths table reports estmates of the effect of the Phase 1 grant (1 R > 0) onrevenuen mllons of nomnal $ usng varants of the model n Equaton 1. Sample lmted to frms wth revenue > 0. Ths s the coeffcent on the full ZINB model n IV-VI. Controls are prevous VC nvestment and prevous SBIR awards from all gov t agences. Sector and year f.e. used where ML estmator does not converge wth competton f.e. Standard errors are robust and clustered at the sector-year level. *** p<.01. Year 1995 23
Ext & Survval VC nvestors typcally lqudate successful nvestments through an IPO or acquston. 31 Table 5 columns I-II fnd a treatment effect of 3.3-4 pp, ndcatng that the probablty of ext ncreases from about 4% to 7.5%. However, Appendx Fgure 6B reveals that ths result s less consstent across postve ranks than the others. The last outcome s survval; I fnd that a grant ncreases the lkelhood a frm remans n busness as of early 2016 by 15 pp. The majorty of frms reman n busness (the mean of ths ndcator varable s 0.67). For both ext and survval, I fnd that rank has no predctve power. Table 5: Phase 1 Grant Impact on Survval and Ext Dependent varable: Ext post In Bus Post Bandwdth: 1 All 1 All I. II. III. IV. Award.036***.035**.15**.14*** (.013) (.016) (.061) (.047) Normalzed rank -.00018 -.0013 (.0039) (.0096) Normalzed rank 2 -.00011.00033 (.00024) (.00047) Controls Y Y Competton f.e. Y Y Y Y N 4019 7628 1276 3946 R 2 /Dev-based R 2 0.29 0.18 0.47 0.27 Note: Ths table reports estmates of the effect of the Phase 1 grant (1 R > 0) onsurvval (remanng n busness) and ext usng varants of the model n Equaton 1. Controls are prevous VC nvestment and prevous SBIR awards from all gov t agences. Standard errors are robust and clustered at the sector-year level. *** p<.01. Year 1995 31 As n much of the lterature, I am unable to dentfy whether acqustons are hgh return events. However, even an acquston that s unsuccessful from an nvestor s perspectve ndcates that the human captal or IP were valuable. Hochberg et al. (2007) and Pur and Zarutske (2012), among others, employ all M&A events as postve ext outcomes. 24
5 The Phase 2 Grant Impact on Frm Outcomes About nne months after recevng a $150,000 Phase 1 award, a frm may apply for a $1 mllon Phase 2 grant. If successful, the frm receves the frst $500,000 roughly two years after the Phase 1 award, and the second tranche three years after Phase 1. Any Phase 2 effect, therefore, s local to the subset of Phase 1 wnners. I fnd strong postve effects of Phase 2 on cte-weghted patents, but not on other outcomes. Many compettons at the Phase 1 level have zero, one, or two Phase 2 applcants. I therefore use sector and year fxed effects, and I cannot nclude controls for rank. That s, I treat Phase 2 as though treatment were assgned randomly. 32 Table 8 Panel 1 columns I-III show the effects of a Phase 2 award on cte-weghted patents usng the negatve bnomal model and the standard sample of no prevous wnners. The more conservatve estmate of 0.69 means that Phase 2 wnners produced double the cte-weghted patents of Phase 2 losers. Column III ncludes Phase 1 and jontly estmates both stages. The coeffcent on Phase 2 drops to.33, or about 1.4 tmes as many cte-weghted patents. I show OLS results usng log ctatons as the dependent varable n columns VII-IX. These coeffcents mply that a Phase 2 award ncreases cte-weghted patents by about 30%, relatve to a mean of 20 among Phase 2 applcants. Over half of Phase 2 applcants have won multple DOE SBIR awards, and thus are excluded from my prmary sample. In contrast, only 20% of applcants to Phase 1 prevously won an award. Consstent wth the Phase 1 fndngs, the postve Phase 2 effect on patents falls and loses sgnfcance when the sample ncludes prevous wnners. Panel 2 frst shows that there s no measurable Phase 2 effect on VC, regardless of the sample (I-V). Estmates are near-zero and mprecse. The revenue estmates suggest a negatve effect, and the survval and ext mpacts are also mprecse and near-zero. 33 The Phase 2 sample s smaller than ex- 32 If ranks are nformatve - whch the Phase 1 analyss suggested s not the case - Phase 2resultsshouldbebasedupward. 33 In the revenue ZINB model n column VII, the logstc porton coeffcent s nsgnfcant 25
pected because 37% of Phase 1 wnners do not apply for Phase 2. There are three reasons for ths, based on ntervews wth grantees and DOE offcals. Frst, a frm s nelgble to apply f an outsde nvestor owns more than 50%, though there s apparently lttle enforcement of ths rule. Of frms that receve VC wthn two years of the Phase 1 grant, 55% do not apply for Phase 2. Put another way, 19% of non-applers receve VC nvestment wthn two years of ther ntal award, but only 9% of unsuccessful Phase 2 applcants do, and 8% of Phase 2 wnners. 34 Ths supports the possblty that some frms who rase equty after Phase 1 sell too much of the frm to be elgble to apply for Phase 2. The adverse selecton n the Phase 2 sample vs-à-vs nvestment suggests that startups seekng to rase external fnance whose Phase 1 R&D revealed postve nformaton often secured prvate nvestment wthout further government support. Second, frms mght not apply f they changed busness strateges. The only montorng of Phase 1 actvtes occurs when a frm apples for Phase 2; the applcaton requres that the Phase 1 project was completed as descrbed n the orgnal applcaton. Thrd, grantees sad n ntervews that the grant applcaton and reportng processes are so onerous that once they receve external prvate fnance, t s often not worthwhle to apply for addtonal government fundng. 35 Smlarly, Gans and Stern (2003) suggest that prvate fundng s preferred to SBIR fundng. A frm s dscount rate may ncrease once ts ntal R&D s successful, so that applyng to Phase 1 s worthwhle but - despte the larger sum at stake - applyng to Phase 2 s not. Ths would be consstent wth the sharply decreasng rsk premum n Berk, Green and Nak (2004) as an R&D project moves from ntaton to completon. The Phase 2 grant does generate new nventve actvty. But ts effects are much lower per publc dollar. The hgh estmates for the effect of Phase 2 and not reported for brevty. 34 At-testofthedfferenceofmeansstronglyrejectsthehypothessthatnon-applers and applers have the same mean probablty of VC nvestment wthn two years, wth a t-statstc of 5.44. 35 My ntervews suggest that the applcaton cost solely n employee tme s one to two full months, apart from any consultng or legal costs the frm may ncur. 26
are two addtonal cte-weghted patents and - at the most - an ncrease of 6 pp n the probablty of VC. Phase 1 generates about 1.4 addtonal cte-weghted patents and a 10 pp ncrease n the probablty of VC. In 2012 DOE spent $38 mllon on 257 Phase 1 grants and $112 mllon on 111 Phase 2 grants. If all the Phase 2 money were reallocated to Phase 1, DOE could have provded 750 addtonal frms wth Phase 1 grants, ncreasng by a factor of about 2.5 and 3.1 the program s mpact on cte-weghted patents and VC, respectvely. To test the potental of reallocatng funds, DOE could randomly remove Phase 2 from certan compettons, notfyng the publc n advance whch compettons wll be affected. Such expermentaton s requred because there may be opton value n the rght to apply to Phase 2. Elmnatng Phase 2 would lkely alter the applcant pool and affect the Phase 1 mpacts. 27