NBER WORKING PAPER SERIES R&D AND THE PATENT PREMIUM. Ashish Arora Marco Ceccagnoli Wesley M. Cohen

Transcription

1 NBER WORKING PAPER SERIES R&D AND THE PATENT PREMIUM Ashsh Arora Marco Ceccagnol Wesley M. Cohen Workng Paper NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cambrdge, MA January 2003 Correspondng author: Marco Ceccagnol. A prevous verson of ths research s part of Marco Ceccagnol s Ph.D. dssertaton (Carnege Mellon Unversty, 2001). We wsh to thank Melssa Appleyard, James Bessen, Tm Bresnahan, Rchard Gold, Shane Greensten, Bronwyn Hall, Thomas Hellmann, Rebecca Henderson, Ulrch Kamecke, Mark Schankerman, Bll Vogt, Rosemare Zedons for helpful comments and suggestons and semnar partcpants at the Amercan Economc Assocaton 2002 meetng, NBER March 2002 productvty workshop, WZB-CEPR Berln 2002 conference on Innovaton Polcy, Harvard Busness School 2002 strategy conference, Academy of Management 2002 conference, Haas School at UC Berkeley, Santa Anna School of Advanced Studes n Psa, INSEAD, Duke Unversty, and H. J. Henz III School of Publc Polcy and Management at Carnege Mellon Unversty. Fnancal support from the Alfred P. Sloan Foundaton and the Natonal Scence Foundaton (Award No. SES ) s gratefully acknowledged. The vews expressed heren are those of the authors and not necessarly those of the Natonal Bureau of Economc Research by Ashsh Arora, Marco Ceccagnol, and Wesley M. Cohen. All rghts reserved. Short sectons of text not to exceed two paragraphs, may be quoted wthout explct permsson provded that full credt ncludng, notce, s gven to the source.

2 R&D and the Patent Premum Ashsh Arora, Marco Ceccagnol, and Wesley M. Cohen NBER Workng Paper No January 2003 JEL No. O31, O32, O34 ABSTRACT We analyze the effect of patentng on R&D wth a model lnkng a frm's R&D effort wth ts decson to patent, recognzng that R&D and patentng affect one another and are both drven by many of the same factors. Usng survey data for the U.S. manufacturng sector, we estmate the ncrement to the value of an nnovaton realzed by patentng t, and then analyze the effect on R&D of changng that premum. Although patent protecton s found to provde a postve premum on average n only a few ndustres, our results also mply that t stmulates R&D across almost all manufacturng ndustres, wth the magntude of that effect varyng substantally. Ashsh Arora Henz School Carnege Mellon Unversty Pttsburgh, PA ashsh@andrew.cmu.edu Marco Ceccagnol Strategy Department INSEAD Fontanebleau France marco.ceccagnol@nsead.edu Wesley M. Cohen Fuqua School of Busness Duke Unversty Box Durham, North Carolna and NBER wcohen@duke.edu

3 1. Introducton Industral R&D s wdely seen as a key drver of productvty and economc growth. In 2000, U.S. frms spent almost 180 bllon dollars on ndustral R&D, n large part because they expected to approprate a substantal part of the return. Many beleve that patent rghts are essental to the protecton of ths return to nnovaton and are consequently a key nducement to R&D. Ths belef n the mportance of patents and ntellectual property protecton has, over the past twenty years, underpnned a trend towards a strengthenng of patent protecton. In 1982, the Court of Appeals for the Federal Crcut was establshed to make patent protecton more unform and, ndrectly, strengthen t. In the early 1980's we have also wtnessed an expanson of what can be patented, when the courts decded that lfe forms and software were both patentable. Patent coverage has been recently extended to busness methods as well. Partly stmulated by the shftng polcy envronment, patents have also become a growng preoccupaton of management (cf. Grndley and Teece [1997]). Indeed, consultants are urgng top management to explot ther patents more aggressvely to the pont of characterzng the untapped knowledge captal of frms as Rembrandts n the attc (Rvette and Klne [2000]). Curously enough, these changes n polcy and manageral practce and percepton have proceeded despte a lmted understandng of the effect of patents--no less stronger patents--on R&D and, n turn, on techncal advance. In ths paper, we begn to address ths gap n the lterature by analyzng the effect of patentng on R&D n a two-step process. We frst estmate what we call the patent premum, defned as the proportonal ncrement to the value of nnovatons realzed by patentng them. We then analyze the effect of changng the premum on R&D. To accomplsh ths, we develop a structural model lnkng a frm s R&D effort wth ts decson to patent, recognzng that R&D and patentng affect one another and are both drven by many of the same factors. Our model accounts for the effect on R&D ncentves of both the drect approprablty ncentve due to patents, and the mpact on R&D productvty of R&D-related nformaton flows orgnatng from other frms patent dsclosures. It also recognzes that stronger patents for a frm means that ts rvals also enjoy stronger patent protecton, to the frm s possble detrment. 1

4 We estmate the model usng unque data drawn from the 1994 Carnege Mellon Survey on Industral R&D n the Unted States. The Carnege Mellon Survey data provde measures of not only R&D and patentng whch tend to be wdely avalable but also on frms evaluatons of the effectveness of patents n protectng the returns to nnovaton, and a measure of the use of patents namely the share of nnovatons that are patented. The avalablty of a measure for frms patent propenstes mples that we can explctly model the determnants of nnovaton separately from the determnants of the decson to patent. In contrast, pror lterature has ether focused on ether the producton of nnovatons (sometmes measured by the number of patents) or the patent decson (more typcally, the patent renewal decson). 1 Thus, n an advance over the lterature, we can emprcally dstngush between how the patent premum affects patentng and R&D, respectvely. Our analyss, however, only consders the mpact of patentng on the R&D of ncumbents. Thus, we do not consder the mpact of patentng on entry and the nnovaton that may be assocated wth t. Indeed, n some ndustres such as drugs, patents may well promote entry by research ntensve frms, whle n others, such as semconductors and telecommuncatons equpment, pervasve cross lcensng of patent portfolos may well deter t (cf. Shapro [2000]). Smlarly, we do not consder the role that patents may play n enhancng ndustry R&D effcency by fosterng the emergence of specalzed technology servce or research frms, as observed, for example, n botechnology, semconductors, scentfc nstruments and chemcals (cf. Arora, Fosfur and Gambardella [2001]). Though dfferent n ts objectve, methods and data from our study, the emprcal lterature that estmates the value of patent rghts usng patent renewal data (e.g., Pakes [1986], Schankerman and Pakes [1986]) partcularly Schankerman s [1998] estmaton of the value of the cash subsdy to R&D conferred by patent protecton n France--provdes a valuable touchstone for that part of our analyss n whch we estmate the patent premum. Asde from our focus on the U.S. rather than 1 Usng data from a 1993 survey on the nnovatve actvtes of Europe's largest ndustral frms, Arundel and Kabla [1998] fnd that frms patent propenstes (the percentage of nnovatons for whch a frm apples for a patent) are postvely related to frm sze and to the degree of patent effectveness. Usng the same data for the French frms, Duguet and Kabla [1998], fnd that the nformaton dsclosed n a patent applcaton lowers the frm s propensty to patent and the number of patent applcatons, whle a desre to acqure a stronger poston n technology negotatons and the avodance of nfrngement suts are assocated wth a hgher number of patent applcatons. However, these studes do not address the queston of the relatonshp between patentng and R&D behavor. 2

5 Europe, our work dffers from ths earler effort n that we develop and test a model that tes together the R&D and patentng decsons. We share wth Schankerman, however, the goal of estmatng a patent premum-cum-subsdy. Our respectve datasets, samples and varables dffer, however, n mportant ways that lead us to expect dfferent estmates. 2 We are able, however, to reconcle our results wth those of Schankerman s [1998], whch s heartenng gven the dfferences n data and approach. The paper s organzed as follows. Followng a background secton, n secton 2 we present a model of R&D and patentng behavor. Secton 3 presents the emprcal specfcaton of the model to be estmated. Secton 4 descrbes the data and measures used for estmaton, whereas secton 5 dscusses a varety of estmaton ssues. Secton 6 contans estmaton results and ther dscusson. A concluson follows n secton 7. Background There are theoretcal as well as emprcal reasons to queston whether patent rghts advance nnovaton n a substantal way n most ndustres. The ratonale for patent protecton s to augment the ncentves to nvent by conferrng the rght to exclude others from makng, usng or sellng the nnovaton n exchange for the dsclosure of the detals of the patented nnovaton. Although the prospect of monopoly rents should nduce nventve effort, the costs of dsclosure can more than offset the prospectve gans from patentng (cf. Horstmann et al. [1985]). In theory, the effect of stronger patents on frms ncentves to nvest n nnovaton are less clear once one recognzes that stronger patents mean that not only any gven frm s patents but also those of ts rvals are stronger. For example, polces that broaden the scope of patents do not unambguously ncrease the expected rents due to nventve actvty when a rval workng n the same technologcal doman may, as a consequence, be able to lmt a frm s ablty to commercalze ts nnovatons (cf. Jaffe 2 For example, whle offerng many advantages, Schankerman s use of patent renewal data mean that hs estmates are condtoned upon frms havng already patented, and therefore cannot consder costs that would tend to affect the ntal decson to patent but not renewal, such as those assocated wth patent dsclosures. In contrast, our sample permts a consderaton of the ntal decson to patent. Also, we allow a frm to fle multple patents for an nnovaton (.e., a new or mproved product or process) as dstnct from the common assumpton of one patent per nventon. Ths dstncton turns out to be mportant for nterpretng our results and reconclng ours wth hs. 3

6 [2000], Galln [2002]). Merges and Nelson [1990] and Scotchmer [1991] further argue that broad patent protecton may slow the rate of techncal change by mpedng subsequent nnovatons where technologes develop cumulatvely. Emprcal work also suggests that the nducement provded by patents for nnovaton s small. The emprcal studes of Scherer et al [1959], Taylor and Slberston [1973], and Mansfeld [1986] suggest that patent protecton may not be an essental stmulus for the generaton of nnovaton n most ndustres. Levn et al. [1987] and, more recently, Cohen et al. [2000] suggest that n most ndustres patents are less featured than other means of protectng nnovatons, such as frst mover advantages or secrecy. Other concerns have been rased. Lerner [1995] suggests that patent ltgaton s especally burdensome for small frms and startups wth less access to fnance, concevably undermnng ther contrbutons to techncal advance. Heller and Esenberg [1998] have clamed that n the doman of genetcs, patentablty has been extended to such fne-graned notons of nventon that the patents and patent owners--coverng any new product nnovaton may now be so numerous that the negotatons necessary to commercalzaton may well break down. Indeed, Cohen et al. [2000] suggest that n ndustres such as electroncs there can be hundreds of patentable elements n one product, wth the consequence that no one frm s lkely to hold all the rghts necessary for a product s commercalzaton. As argued by Cohen et al. [2000] for complex product ndustres generally and Hall and Zedons [2001] for the semconductor ndustry n partcular, such mutual dependence commonly spawns extensve cross-lcensng. Although the knd of breakdown suggested by Heller and Esenberg does not occur n these ndustres, the prospect of extensve cross-lcensng, and the assocated use of patents as barganng chps may stmulate patent portfolo races among ndustry ncumbents that can act as a barrer to entry to frms that possess relatvely few patents. We should not, therefore, assume that patent rghts necessarly nduce nnovaton. Nor, however, should we assume the contrary. Frst, that patents are less featured than other means of protectng nnovatons n the majorty of ndustres does not mply that they yeld lttle return n those ndustres; 4

7 ther effect on R&D ncentves may be consderable. Levn et al. [1987], Mansfeld [1986], and Cohen et al. [2000] also observe that n selected U.S. manufacturng ndustres, such as drugs or medcal equpment, patents are ndeed crtcal to the protecton of nnovatons. Moreover, n contrast to the fndngs for the U.S., Japanese frms report patents to be among the most mportant means of protectng ther nnovatons (Cohen et al. [2002]). 2. A frm level model of R&D and patentng We focus on a typcal product nnovaton that s the output of an R&D project. We also assume that such a product nnovaton may have multple patentable elements. Fgure 1 provdes a schematc representaton of our model of the decson to patent, to nvest n R&D, and the structure of payoffs. If a frm apples for patent protecton t earns x j v - c, where the subscrpt ndexes frms (=1,,n), and j ndexes nnovatons (j=1,,m). The patent premum s defned as the ncremental payoff due to patent protecton as compared to the value of an nnovaton wthout patent protecton, v. 3 A patent premum less than one would actually reflect an expected loss, possbly because nformaton dsclosure costs may be large relatve to benefts. We assume that the patent premum, x j, has a component, ε j, that vares across nnovatons wthn a frm, and s normally dstrbuted wth varance σ 2, and a fxed, frm specfc component, µ. The patent premum, x j =ε j +µ, s thus normally dstrbuted wth mean µ and varance σ 2 (cf. Fgure 2a). The patent premum wll lkely vary across nnovatons wthn a frm. For example, some patents are easer to nvent around than others. Moreover, the premum may vary dependng on how a frm ntends to use a gven patent, ncludng, for example, as a bass for lcensng or perhaps as a barganng chp n a cross-lcensng negotaton. Our specfcaton also assumes, mplctly, that dfferences n the expected probablty that patent protecton wll be obtaned are ncorporated n the patent premum tself. 3 For example, x j =1.2 means that the value from patentng an nnovaton s 20% hgher than the value wthout a patent, gross of the cost of applyng for patent protecton. 5

8 We do not allow for unobserved heterogenety n the value of an nnovaton wthout patent protecton across frms, nor across nnovatons wthn a frm. 4 Also, we assume c to be constant across frms and across nnovatons. The reasons and the mplcatons are dscussed n secton 5.4 below. In lght of data lmtatons and for analytc tractablty, we also do not model strategc nteractons between rvals, and ther possble mpact on the patent premum and the value of an nnovaton wthout patent protecton. However, the model ncorporates the ndrect compettve effects of rvals patentng. To the extent that patents held by others reduce the returns from patentng a partcular nnovaton, the own patent premum, x j, should be lower. Our emprcal specfcaton also controls for the effect of rvals patents by ncludng a measure of rvals patent effectveness among the determnants of the value of an nnovaton when not patented The decson to patent Let y be a bnary varable takng the value of 1 f, gven an nnovaton, a frm apples for patent protecton and zero otherwse. We assume that the frm observes x j, the patent premum specfc to the nnovaton. Gven an nnovaton, y = 1 f the expected net beneft from patentng s greater than the expected net beneft wthout patentng,.e: (1) y = 1 f and only f ( ε j + µ ) v c > v, If π j s the probablty of applyng for patent protecton gven an nnovaton, (1) mples that c c µ (2) 1 c π j = E( y = 1) = 1 F + 1 µ j = Φ = Φ( Z ) v = Pr ε > + 1 µ v v σ σ where Φ s the standard normal cumulatve dstrbuton functon of ε j, σ ts standard devaton, and Z µ 1 c =. Wth data grouped at the frm level, the percentage of nnovatons for whch a frm σ σv apples for patent protecton, π (.e., ts patent propensty), equals: (3) π = Φ( Z ) + η p wth η p representng samplng error. Note that we allow frms to fle for more than one patent per nnovaton. Thus, patent propensty s understood as the probablty of applyng for at least one patent condtonal on an nnovaton. 4 We can allow for both knds of heterogenety n v n a restrcted verson of our model, as dscussed n Secton

9 Though we assume that the premum s normally dstrbuted, the observed dstrbuton of patent prema, x * j, s truncated normal and postvely skewed, as shown n Fg. 2b, where 1+c/v s the cutoff value for applyng for patent protecton, and µ * s the mean of the condtonal dstrbuton. Thus, our specfcaton s consstent wth the fndng reported n the lterature that the dstrbuton of patent values s postvely skewed (e.g., Scherer and Harhoff [2000]). Fgures 2a and 2b also llustrate the pont that even when the average patent premum µ s less than unty, a frm may stll patent a fracton of ts nnovatons. Put dfferently, even f patent protecton s not proftable for most of a frm s nnovatons, ths does not mply that patent protecton s not valuable to the frm. Rather, a frm would tend to apply for patent protecton for a mnorty of ts nnovatons, as descrbed n equaton (2) The producton of nnovatons The nnovaton producton functon s specfed as: β s + η + ηˆ m m (4) m = dr e where m s the number of nnovatons, r s the R&D expendture, d s a constant scale parameter, and s are the factors affectng the average productvty of R&D, such as nformaton flows from other frms, unverstes and government research labs, and β s the elastcty of the number of nnovatons wth respect to R&D. We also assume that other unobserved frm-specfc factors affect the productvty of R&D. In partcular, ηm and ηˆ m are..d. normal errors, wth zero mean and varance σ 2 η m and σ 2ˆm η, respectvely. The former s observed by the frm but not the econometrcan, whereas the latter s unobserved by both the frm and the econometrcan and represents the stochastc component affectng the R&D process The optmal level of R&D The frm maxmzes the expected proft from ts nnovatve actvty, that s the expected payoff per nnovaton, h, multpled by the expected number of nnovatons, E(m ), net of the cost of R&D, measured as the dollars spent on R&D, r. Thus, the frm s objectve s: (5) Max [h E(m ) - r ], r 7

10 wth E ω s ηm ( m ) = dr β e + + and σ η ω = 2ˆm ; h, the expected value per nnovaton, can be expressed as 2 a functon of the value of the nnovaton and the payoff from patentng, weghted by the probabltes of applyng for a patent and not applyng, where the decson to patent s made optmally after observng the patent premum, x j : c v * (6) h = [( εj + µ ) v c] φ( εj ) dεj + v φ( εj ) dεj = Φ( Z )( µ v c) + (1 Φ( Z ) v c v + 1 µ + 1 µ where φ(ε j ) s the standard normal p.d.f. of ε j, and µ * s the mean of the condtonal patent premum dstrbuton: ( Z ) ( ) Z * c φ (7) µ = µ + E( εj εj > + 1 µ ) = µ + σ v Φ Wth further smplfcaton and substtutons we obtan: (8) 1 h = σv φ ( Z ) + Φ( Z ) Z + σ The equlbrum level of R&D for frm s found by solvng (5): 1 ω + s + η = β 1 dhe r m (9) [ ] β The frst and the second order condtons mply 0<β<1, mplyng dmnshng returns to R&D Unobserved varables and emprcal specfcaton We model the nnovaton-specfc random component of the patent premum, ε j, as a latent varable observed by the frm at the tme of patentng, but not the econometrcan. We observe the patent propensty, the total number of patent applcatons and the R&D nvestments of the frm. We do not observe the other frm and nnovaton specfc varables: cost of patentng, value of an nnovaton, the productvty of R&D, the frm specfc average patent premum, and the number of nnovatons. We do have R&D lab, frm and ndustry specfc cross-secton data. Accordngly, we specfy the estmatng equatons as follows. 8

11 3.1. Number of nnovatons We frst transform the nnovaton equaton nto an estmable relatonshp. We thus multply both sdes of the nnovaton producton functon (4) by the frm patent propensty, π, and obtan an equaton explanng the number of patent applcatons, a : (10) a = π k dr β e s + η m + ηˆ m wth k 1 beng the number of patent applcatons per nnovaton. We have measures of both patent propensty and the number of patent applcatons for all frms n our sample, ncludng those who dd not apply for patent protecton, for whom both a and π are smply zero. We do not observe the average number of patent applcatons per nnovaton, k, and thus set k =Kκ, where K represents ndustry dummes and κ s a vector of unknown parameters to be estmated. Thus, k vares only across ndustres The patent premum We do not observe µ, the frm-specfc component of the patent premum. We thus treat t as a frmspecfc constant to be estmated. To do so, we use a self-reported measure of the percentage of a frm s nnovatons for whch patent protecton was rated effectve. Ths measure groups all frms nto one of fve patent effectveness classes. In the emprcal analyss, dscussed n secton 4 below, we assume that frms n a gven patent effectveness class have the same average patent premum, µ. We also allow for possble measurement error and the possblty that our measure of patent effectveness s correlated wth other unobserved factors affectng R&D productvty and estmate a specfcaton where we nstrument for patent effectveness The value of an nnovaton, and the cost of applyng for a patent We do not observe the value of the nnovaton f not patented, v, nor the cost of applyng for patent protecton c. Accordngly we set v =Vα, where V represents vectors of frm and ndustry characterstcs and α a vector of unknown parameters to be estmated. We also set the cost of 5 The f.o.c. for (5) s 1 + s + r β ω η dh e m 1 = 0 β 2 ω + s + ηm β, and the S.O.C. s ( β 1 ) βr dh e < 0 9

12 protecton c=δ, a constant to be estmated. We assume c ncludes the patent applcaton fees, legal fees for draftng and prosecutng patent applcatons and the opportunty cost of the tme of the R&D engneers and scentsts who help draft the patent applcaton. Moreover, as noted above, the frm may apply for more than one patent per nnovaton Other factors affectng R&D productvty R&D productvty s assumed to be a functon of frm and ndustry specfc factors such as the underlyng scentfc and technologcal knowledge base and nformaton flows from other frms and unverstes (see Jaffe [1986] and Cohen [1995], among others). More formally, we set: (11) s = λ 1 S 1 + λ 2 S 2 + λ 3 S 3, where the λ s are parameters to be estmated and S 1, S 2, and S 3 are frm specfc varables: S 1 : S 2: S 3: vector of organzatonal characterstcs condtonng the frm s R&D productvty; measure of nformaton flows from other frms (rvals, supplers, customers, other); measure of nformaton flows from unverstes and government research labs. We also allow for unobserved frm-specfc capabltes to affect both the producton of nnovatons and the knowledge spllovers beneftng the R&D lab. More specfcally, the scentfc and techncal capabltes of the lab s researchers, whch are observed by the frm but not the econometrcan and captured by η m n (4), are lkely to be correlated wth the amount of ncomng nformaton flows from frms and unverstes, S 2 and S 3. In the subsequent model estmaton we nstrument for both types of flows. Moreover, snce patents dsclose nformaton and thus contrbute to the stock of potentally useful technologcal knowledge, we nclude a measure of nformaton flows due to patent dsclosures to nstrument for spllovers The system of equatons to be estmated Takng logs of the R&D and patent equatons, (9) and (10) respectvely, and usng the patent propensty equaton (3), we obtan an estmable system of non-lnear smultaneous equatons 6 : 6 We nclude the non-patentees because they contrbute to estmaton of the patent propensty and the R&D equaton, but not the second equaton n (12). Indeed, for non-patentees both sdes of equaton (10) are null. 10

13 (12) where: v µ 1 c π = Φ + η σ σv p loga logπ = (logk + logd) + s 1 logr = ( γ + logh + s ) + η 1 β r = Vα, c = δ, s γ = log β + log d + ω; η = η a m + ˆ η m + ξ ; 1 η r = η m+ ξ r ; 1 β = Sλ; a + β logr + η a ξ a, ξ r represent measurement error n the product patent applcaton and product R&D measures respectvely and η p represents samplng error. The relatonshps between the endogenous and exogenous varables are summarzed n Fgure 3, whch shows how the decson to patent and to nvest n R&D are jontly determned. Estmaton of the three equatons as a system allows the estmaton of key parameters, such as σ, as well as the separate dentfcaton of the parameters assocated wth the cost of patent protecton and the value of an nnovaton. 7 Snce thes equatons have a number of common parameters, estmatng them together also provdes greater effcency n estmaton. 4. Data and measures We use data from the Carnege Mellon survey (CMS) on ndustral R&D (Cohen, W., Nelson, R., and J. Walsh [2000]). 8 The populaton sampled s that of all R&D labs located n the U.S. conductng R&D n manufacturng ndustres as a part of a manufacturng frm. The sample was randomly drawn from the elgble labs lsted n the Drectory of Amercan Research and Technology (Bowker [1995]) or belongng to frms lsted n Standard and Poor's Compustat, stratfed by 3-dgt SIC ndustry. Vald responses were receved from 1,478 R&D unts, wth a response rate of 54%. 9 The respondents were R&D lab managers who were asked to answer questons wth reference to the "focus ndustry" 7 We dscuss system versus sngle equaton estmaton n the next secton. 8 The survey was admnstered n 1994 by sendng questonnares by mal and conductng follow-ups by telephone. See Cohen, Nelson, and Walsh [2000]. 9 The raw response rate was 46%. A non-respondent survey found, however, that 28% of the non-respondents n the U.S. were not n the target populaton (for example, they dd no manufacturng). After correctng the sample sze accordngly for nelgble cases, the U.S. response rate was adjusted upward to 54%. 11

14 of ther R&D unt, where focus ndustry was defned as the prncpal ndustry for whch the unt was conductng ts R&D. The data refer to the perod. In our theory above, we have taken the frm as our unt of analyss and shall contnue to do so to smplfy exposton. However, the emprcal unt of analyss s the busness unt wthn a parent frm, operatng n the focus ndustry of the respondng R&D lab. In the emprcal analyss, we explctly dstngush between busness unt and frm level measures. Indeed, as dscussed below, we explot the dfferent ndustry sectors to whch the busness unt and the parent frm belong to develop nstruments for reported patent effectveness. For the analyss we restrcted the sample to frms wth busness unts wth 10 or more employees. After droppng observatons wth mssng data for the varables of nterest, we obtan a sample of 737 R&D unts. 10 Ths sample ncludes frms rangng from fewer than 10 to over 700,000 employees, wth annual sales rangng from more than $100,000 to over $130 bllon. The medan frm has 3,000 employees and annual sales of over $500 mllon. The average frm has 21,841 employees and sales of $4.3 bllon. The average frm R&D ntensty (R&D dollars dvded by total sales) s 5.2%. The busness unts range from 10 employees to 448,000, wth annual sales from zero to about $90 bllon. The medan busness unt has 550 employees and $100 mllon n sales. The average busness unt has 6,168 employees and sales of about $1 bllon. Table 1 provdes summary statstcs for the varables used for estmaton Measures of the endogenous varables PRODUCT R&D: Recall that we estmate the model for the case of product nnovatons. To compute the product R&D expendtures we multply company-fnanced R&D unt expendtures n dollars n the most recent fscal year by the percentage of the R&D unt s effort devoted to new or mproved products. The sample average product R&D expendture for a busness unt s about $8 mllon. 10 The sample of 737 observatons also reflects the excluson of 6 R&D unts reportng more than 20 patent applcatons per mllon dollars of R&D, (the 99th percentle value of the dstrbuton). 12

15 PRODUCT PATENT PROPENSITY: R&D managers were asked to state the percentage of R&D unt s product nnovatons n the perod for whch they appled for patent protecton. Patent propenstes n the sample range from zero to 100%, wth a smple, unweghted mean of 32%. PRODUCT PATENT APPLICATIONS: R&D managers were also asked to state the total number of patent applcatons generated by the R&D lab durng To calculate the annual number of product patent applcatons we frst multply the total number of patent applcatons by an adjustment factor based on survey reported measures of the percentage of R&D unt effort devoted to product nnovatons and the reported product and process patent propenstes, as descrbed n the appendx. The resultng number s then dvded by three, yeldng the annual number of product patent applcatons, whose sample average s 6.4, wth actual values rangng from zero to The patent premum EFFECTIVENESS OF PATENT PROTECTION: Respondents were asked to ndcate the percentage of ther product nnovatons for whch patent protecton had been effectve n protectng ther frm's compettve advantage from those nnovatons durng the pror three years. There were fve mutually exclusve response categores. We further assume that all respondents reportng the same level of patent effectveness have a common (unknown and to be estmated) value of µ. In partcular, we can set: (13) µ 1 = τ 1 T 1 + τ 2 T 2 + τ 3 T 3 + τ 4 T 4 + τ 5 T 5 σ wth the τ's beng fve coeffcents to be estmated, and the T s dummy varables defned as: T 1 =1 f patent protecton was rated effectve for 0-10% of the frm s product nnovatons, = 0 otherwse; T 2 =1 f patent protecton was rated effectve for 11-40% of the frm s product nnovatons, = 0 otherwse; T 3 =1 f patent protecton was rated effectve for 41-60% of the frm s product nnovatons, = 0 otherwse; T 4 =1 f patent protecton was rated effectve for 61-90% of the frm s product nnovatons, = 0 otherwse; T 5 =1 f patent protecton was rated effectve for over 90% of the frm s product nnovatons, = 0 otherwse. 13

16 Thus each coeffcent reflects one of fve dscrete levels of the average patent premum: (14) µ = τ σ + ; µ = τ σ + 1; µ = τ σ + 1; µ = τ σ + 1; µ = τ σ Note the mportance of σ (the standard devaton of the dstrbuton of the patent premum wthn frms) for the estmate of the patent premum. We ntally assume σ to be unform across frms and ndustres but later relax ths assumpton by allowng for nter-ndustry dfferences n σ. Snce our analyss hnges upon our measurement of patent effectveness, t s worth consderng the nterpretaton and lmtatons of our survey-based measure. Snce our setup assumes that the patent premum reflects all the ways n whch a frm profts from ts patents, there s some concern about whether the reported effectveness scores accurately reflect ths. As Cohen et al. [2000] fnd, frms patent for reasons that often extend beyond drectly proftng from a patented nnovaton through ts commercalzaton or lcensng. In addton to the preventon of copyng, frms also patent to prevent rvals from patentng related nnovatons (.e., patent blockng ), use patents n negotatons, and to prevent suts. Here, the ssue s whether the respondents scorng of patent effectveness msses some of the latter, conventonally less apprecated, motves for patentng. In a corollary exercse, we estmated an ordered probt model to analyze the relatonshp between frms' reasons to patent and the respondents' patent effectveness scores. We found that the magntude of the coeffcent for conventonal motves for patentng such as lcensng are comparable to those for less conventonal reasons, such as usng patents to nduce rvals to partcpate n crosslcensng negotatons or for buldng patent fences (.e., patentng substtutes) around some core nnovaton. However, one reason for patentng that had no sgnfcant effect on respondents patent effectveness scores was the motve of the preventon of nfrngement suts that s, defensve patentng. Thus, we suggest that wth the possble excepton of defensve patentng, our effectveness measure appears to reflect the broad range of uses of patents observed across the manufacturng sector. It s stll plausble that measurement error, n the form of msclassfcaton across the response scale categores, exsts. The mpact on our results of such measurement error should be, however, mtgated when, as dscussed below, we nstrument for patent effectveness. 14

17 4.3. Value of an nnovaton BUSINESS UNIT AND FIRM SIZE: Busness unt sze, measured by the natural logarthm of the number of busness unt employees, and overall frm sze, measured by the natural logarthm of the total employees of the lab s parent frm, are both ncluded as determnants of v. 11 Frms may proft from an nnovaton by ncorporatng t n ts own output, so that the payoff s ncreasng n busness unt output (Cohen and Klepper [1996]). We also nclude overall frm sze snce large overall sze, especally where t reflects greater dversfcaton, may ncrease the expected value of an nnovaton by provdng economes of scope (cf. Cohen [1995] for a revew of the R&D-sze relatonshp, and Cockburn and Henderson, [2001], showng that scope ncreases the success probablty of drug development R&D). TOTAL NUMBER OF RIVALS AND TECHNOLOGICAL RIVALS: The effect of competton on the expected returns to nventve actvty s not clear a pror (e.g., Needham [1975]). On the one hand, rvals capable of both generatng nnovatons and capturng some of the benefts of ncumbents R&D, what we refer to as technologcal rvals, are expected to dmnsh the value of a frm s nnovaton through mtaton or ntroducton of a substtute product, once the potental postve effect of entry on R&D productvty va ncomng R&D spllovers s held constant. Compettve pressure from the rest of the rvals, on the other hand, has ambguous effects on R&D ncentves. Although average returns to R&D fall wth the ncrease n the number of such rvals, margnal returns to R&D may ncrease (Boone [2000], Ceccagnol [2001]), mplyng that ncreases n the number of compettors may be assocated wth ncreases n R&D. The CMU survey provde measures for both the total number of rvals and technologcal rvals, as categorcal varables n the followng ranges: 0,1-2, 3-5, 6-10, 11-20, or >20 compettors. 12 These responses were recoded to category mdponts. These varables vary across respondents wthn 11 Busness unt employees s reported by R&D managers from the CMU survey, whereas total frm employees were obtaned from sources such as Compustat, Dun and Bradstreet, Moody s, and Ward s. 12 Technologcal rvals are defned n the CMS questonnare as the number of U.S. compettors capable of ntroducng competng nnovatons n tme that can effectvely dmnsh the respondent s profts from an nnovaton, wth reference to the lab s focus ndustry. 15

18 ndustres because they represent each respondent s assessment of hs or her focus ndustry condtons, often reflectng a partcular nche or market segment. RIVALS PATENT EFFECTIVENESS: The effectveness of rvals patents can have a varety of effects on the value of R&D nvestments. The most obvous one s that by dmnshng the "technology space" n whch a frm can work wthout nfrngng rvals patents, ncreases n the patent effectveness of a rval s patents should reduce the expected value of the frm s nnovatons. In terms of our model, ths would decrease R&D nvestments. However, our model s n some sense the reduced form verson of an equlbrum of more complex market nteractons n whch ncreases n rval patent effectveness may spawn offsettng ncentve effects. For example, f one consders the strategc nteractons characterstc of patent races, an ncrease n the effectveness of rvals patents may actually ncrease the margnal payoff to own R&D by ncreasng rval R&D (cf. Renganum [1989]). In the end, we can only hope to estmate a net effect and have no clear pror on the qualtatve mpact of rval patent effectveness. Nonetheless, we clearly need to control for such an effect and accordngly, we nclude, among the determnants of v, the % of frms n an ndustry - excludng the respondent - n each of the fve patent effectveness classes, thus allowng ths measure to vary across respondents n an ndustry. 13 GLOBAL, FOREIGN, PUBLIC: We nclude bnary varables ndcatng whether the frm ownng the lab s GLOBAL (sells products n Japan or Europe), s FOREIGN (the respondent R&D lab s located n the U.S. but the parent frm s located abroad), or t s PUBLIC (publcly traded companes), as controls, to reflect possble dfferences n market opportuntes and cost of captal. INDUSTRY FIXED EFFECTS: We nclude 19 ndustry dummy varables to control for ndustry-level effects of demand and technologcal opportunty n v, constructed usng the SIC code assgned to the focus ndustry of each respondent, where focus ndustry was defned as the prncpal ndustry for whch the unt was conductng ts R&D. The dummes are based on ndustry groupngs descrbed n table A1 n the appendx. 16

19 4.4. Factors affectng R&D productvty INFORMATION FLOWS FROM OTHER FIRMS: We do not drectly measure nformaton flows from rvals, and other frms such as supplers and customers. However, the CMS contans several varables reflectng two key dmensons of the spllover mechansm: a) the frequency wth whch a respondent R&D lab obtans useful techncal nformaton from, respectvely, rvals, customers and supplers n the U.S.; b) the contrbuton of nformaton flows from rvals, customers, and supplers to suggestng or completng R&D projects. We employed factor analyss to develop a sngle factorbased measure of nformaton flows from other frms. The Appendx provdes the detals. In an earler verson of ths paper, we treated nformaton flows from other frms as the dependent varable n an addtonal fourth equaton n our system to hghlght a possble postve mpact on own R&D productvty of the dsclosures assocated wth patents. However, we found that patent dsclosures appeared to have no measurable mpact on nformaton flows from other frms, and therefore no measurable effect on R&D productvty. It s unclear whether patent dsclosures truly have lttle effect on the nformaton flows from others that affect frms R&D productvty, or whether the lack of an observable effect reflects that our measures are too mprecse to dscern t. We do not specfy a fourth equaton but we do contnue to treat nformaton flows from other frms as potentally correlated wth the error terms n both the patent applcaton and R&D equatons. INFORMATION FLOWS FROM UNIVERSITIES: We lack a drect measure here as well. The CMS provdes measures whch reflect two key dmensons of ths varable: a) the frequency wth whch the R&D lab obtans useful techncal nformaton from unverstes or government research labs n the U.S.; b) the contrbuton of nformaton flows from unverstes or government research labs to suggestng or completng R&D projects. We construct a sngle factor-based measure of flows from unverstes, as descrbed n the Appendx. We also nstrument for ths varable, as explaned below. INFORMATION TECHNOLOGY IN ORGANIZATION: We nclude a measure of one feature of the way n whch the R&D process s managed wthn the frm, namely a dummy varable ndcatng 13 Usng alternatve measures, such as the average patent effectveness (computed usng categorcal range 17

20 whether computer network facltes are used wthn the frm to facltate the nteracton between R&D and other functons. Ths varable s ntended to reflect progressve manageral practces more generally and should ncrease s and, n turn, R&D productvty Estmaton We estmate our non-lnear system of equatons (12) wth the method of nonlnear three stage least squares (NL3SLS ) usng the sample of 737 observatons, mposng the cross-equaton restrctons Inter-ndustry versus ntra-ndustry sources of varaton We allow v, the value of an nnovaton, to have an ndustry-specfc fxed effect. Smlarly, the patent applcaton equaton allows for the average number of patents per nnovaton to vary across ndustres. Gven the ncluson of these ndustry fxed effects n these two equatons of our structural model, each of our estmatng equatons has at least one complete set of ndustry dummes. One may nonetheless wonder whether our results are drven prmarly by nter-ndustry varaton n key varables. Cross-ndustry varaton s ndeed mportant, but as table 2 shows, there s very sgnfcant ntra-ndustry varaton n our key varables as well. Indeed, for patent applcatons, R&D, patent propensty and patent effectveness, cross-ndustry varaton represents less than 20% of the total varaton. 16 Thus, our results do not prmarly reflect nter-ndustry dfferences n these key varables. 5.2 Sources of varaton n patent effectveness Another concern for estmaton s that sources of varaton n patent effectveness across respondents wthn an ndustry may be correlated wth unobserved varatons n R&D productvty and spllovers. It s plausble, for example, that managers who manage ther patent holdngs n a mdponts) at the ndustry level (excludng the respondent) does not make a dfference to the results. 14 We expermented wth measures of other characterstcs of the R&D organzaton wthn frms, namely whether the frm rotated ther R&D personnel through other functonal areas n the frm, such as marketng, and whether the frm used project-teams wth cross-functonal partcpaton. 15 NL3SLS s a moments type estmator, where nstrumental varables are used to form the moment equatons (Gallant [1987], p ). We used the exogenous varables ncluded n the equatons, addtonal nstruments to be explaned below, and the squares and cross-products of the contnuous exogenous varables as nstruments. 16 We also estmated the system of equaton (12) wthn the drugs and chemcals ndustres (SIC 28), ncludng botechnology companes, and the computer and electroncs ndustres (SIC 36 electroncs and electrcal equpment, plus SIC 357 computers, and selected frms belongng to 4-dgt electronc nstruments), thus allowng all the parameters to vary across these two samples. The prvately fnanced product R&D performed by these two ndustry clusters amount to more than 60% of the total n our sample, although the smaller number of observatons 18

21 more sophstcated way also manage ther R&D expendtures more effectvely, for example by provdng strong ncentves to generate patentable nnovatons. Smlarly, techncal areas where patent protecton s more effectve may also have more productve R&D because of ther greater proxmty to scence (Arora and Gambardella, 1994). Arguably, ths may also ncrease spllovers from unverstes and frms. There s a related concern. Levn et al. [1987] and Cohen et al. [2000] pont out that frms use approprablty mechansms, such as lead tme and secrecy, n addton to patents. These other mechansms may be substtutes or complements for patentng. 17 Thus systematc dfferences across frms n the effectveness of alternatve approprablty mechansms may also be a source of varaton n reported patented effectveness. Insofar as these alternatve mechansms also condton v, the average value of an nnovaton, ths may bas our estmate of the patent premum. Although we have reported effectveness scores for each of these alternatve mechansms, we do not have any measure of ther actual use--n contrast to patents, where we do observe use n the form of the propensty to patent and numbers of patent applcatons. In a corollary analyss, we estmated a model n whch the effectveness of other appropraton strateges, such as the use of secrecy or lead-tme advantage, were ncluded among the determnants of v. There was no qualtatve change n the results, suggestng that, nsofar as the use of such alternatve strateges s correlated wth ther reported effectveness, any bas due to the omsson of other strateges, s lkely to be small. 18 We also drectly address the possblty that our patent effectveness measure may be correlated wth the errors n the R&D and patent equatons by nstrumentng for patent effectveness. To do so, we explot dfferences n the focus ndustry of the R&D lab (.e., the ndustry sector of the busness unt) and the prmary ndustry of whch the parent frm. We post that factors that condton patent per sample (156 and 184 respectvely) often results n large standard errors. In general, the estmates are smlar to those reported here and consstent wth the dea that our results are not drven by nter-ndustry dfferences. 17 An mplcaton of ths observaton s that our estmate of the patent premum reflects the ncremental payoff to patentng when the frm optmally adjusts ts use of other mechansms. Ths s smlar to estmatng the long run mpact of a change n a gven factor prce on the proft functon. Ths mpact assumes that the frm optmally changes not only the use of the factor whose prce has changed, but also of the other factors nputs. Effectveness measures can, n ths nstance, be analogzed to factor prces. 18 Consstent wth ths result, Cohen et al. [2000] fnd no sgnfcant correlaton between the effectveness of patents and that of any of the other approprablty mechansms, such as secrecy or use of lead tme advantage. 19

22 effectveness and patentng behavor n the prmary ndustry of the parent frm wll reflect the frm s broad approach to ntellectual property management, and thereby affect the perceved effectveness of patents n all the product markets n whch the frm partcpates. We have n mnd notons such as how carefully scentsts and researchers document ther work; how skllfully the n-house lawyers manage patent prosecuton; and how effectvely researchers and n-house lawyers communcate. Smply put, our nstrumentaton strategy s based on the premse that a busness unt whose parent frm operates, for example, n the pharmaceutcal ndustry, where sophstcated management of ntellectual property and a belef n ts value s the norm, wll perceve a hgher effectveness of patents than an otherwse dentcal busness unt whose parent frm s n textles. Although we do not have nformaton about the management of ntellectual property for the parent frm of each R&D lab, roughly half of the respondng busness unts belonged to an SIC dfferent from that of the prmary SIC of the parent frm. We thus use ndustry averages of the patent effectveness and other survey-based dummy varables on the reasons to patent (and not to patent) for the prmary ndustry of the parent frm as nstruments for each respondent patent premum dummy class. 19 We report estmates from the two specfcatons one where we do not nstrument for the patent effectveness varable and another where we do nstrument. Both are qualtatvely smlar. However, arguably, the endogenous patent effectveness specfcaton s theoretcally more defensble and we shall focus on those estmates Endogenety of the spllover measures As noted, we nstrument for nformaton flows from other frms and unverstes. For nformaton flows from other frms, we use two nstruments. The frst measures the technology overlap wth rvals R&D 19 We use as nstruments for the 5 patent effectveness dummes, % of respondents n the ndustry (of the parent frm) that have a postve value for the followng ten ndcator varables, avalable from the CMU survey: 1) the fve patent effectveness ndcator varables (fve nstruments); 2) whether the amount of nformaton dsclosed n a patent applcaton was a reason not to patent for a frm; 3) whether the ease of legally nventng around was a reason not to patent; 4) whether the preventon of other frm's attempts to patent a related nnovaton ("patent blockng") was a reason to patent; 5) whether the earnng of lcensng revenue was a reason to patent; 6) whether the preventon of suts was a reason to patent. The R-squares from the frst stage regresson of the fve patent effectveness dummes on the nstruments are 0.26, 0.17, 0.14, 0.21,and 0.26 respectvely. We also expermented wth usng predcted patent effectveness from an ordered probt regresson of patent effectveness on the above and other exogenous varables as nstruments for the actual patent effectveness scores, wth very smlar results. 20

23 projects, whch should ncrease nformaton flows from other frms. 20 As an addtonal nstrument we construct a survey-based measure of the exogenous stock of patent-related knowledge relevant to the lab, whch reflects nformaton flows due to patent dsclosures. 21 To nstrument for the unverstyrelated nformaton flows, we used the total R&D spendng of doctoral grantng nsttutons by state and feld of scence and engneerng 22, assgned to each respondent accordng to the state n whch t s located and ts reported ratng of the mportance of scence and engneerng feld Unobserved heterogenety n the value of nnovatons across and wthn frms Note that though we allow for unobserved frm heterogenety n R&D productvty (and through that, n the patent applcaton equaton), we do not permt unobserved heterogenety across frms n v. Allowng for frm specfc unobserved heterogenety n v would requre us to move to maxmum lkelhood type estmaton method because t would mply addtvely non-separable error terms, thus rulng out nstrument varable based estmators. Maxmum lkelhood s an unattractve opton because the nonlneartes present n our model, smple as t s, n practce pose convergence problems even for non-lnear least squares estmaton. As noted above, we also do not allow for unobserved heterogenety across nnovatons wthn a frm. Snce we do not observe nnovaton-specfc characterstcs, ths seems lke a sensble way to proceed. In the same sprt, we assume a constant cost of applyng for patent protecton, although t 20 The CMU survey asks a subjectve assessment of the percent of projects started by the R&D unt wth the same techncal goals as an R&D project conducted by at least one of ts compettors. The responses categores are: 1=0%;2=1-25%;3=26-50%;4=51-75%;5=76-100%. Responses were then recoded to category mdponts. 21 Each respondent s assgned the total number of R&D employees multpled by the average patent propensty of the ndustry for whch the feld of scence rated the most mportant contrbuton to R&D s the same as that ndcated by the R&D lab. More formally, the nstrument s constructed as follows: Q =Σ j a j p j r j wth =1,,N, denotng R&D unts; j denotng ndustres; p j s the ndustry average product patent propensty; r j s the sum of R&D employees n ndustry j; a j s a respondent specfc dummy equal to 1 f w j = W j, zero otherwse where: w j s a character varable representng the lab s reported feld of scence and engneerng whose research fndngs contrbuted the most to ts R&D actvty durng the most recent three years (possble felds nclude Bology, Chemstry, Physcs, Computer Scence, Materals Scence, Medcal and Health Scence, Chemcal Engneerng, Electrcal Engneerng, Mechancal Engneerng, Mathematcs); W j s the modal value of w j n ndustry j. All measures avalable from CMS. 22 Unversty R&D expendtures have been taken from 1993 NSF/SRS Survey of Scentfc and Engneerng Expendtures at Unverstes and Colleges. 23 The CMU scence and engneerng felds noted above have been aggregated takng average scores of ther mportance to match the NSF publcaton more aggregated felds (engneerng, physcal scences, math & computer scences, lfe scences). The mportance score assgned to each feld s then used to compute a weghted average of the unversty R&D spendng by state to be assgned to each observaton as an nstrument for the survey based measure of nformaton flows from unverstes. 21