Voluntary technological disclosure as an efficient knowledge management device: an empirical study

Transcription

1 Voluntary technologcal dsclosure as an effcent knowledge management devce: an emprcal study Stéphane LHUILLERY Unversté Pars Nord, UFR d économe, 99, av. J.B. Clément, Vlletaneuse, France lhuller@seg.unv-pars13.fr Revsed draft - January 2003 Abstract Ths paper nvestgates three questons related to endogenous nformaton and knowledge dsclosure by frms: Whch ndustry sectors are more apt to dsclose nformaton and knowledge? Why s such knowledge released? Is knowledge dsclosure an effcent strategy? An emprcal analyss on four French data sets that focus on appropraton, the practces of nnovaton, and the related payoffs suggests answers to these questons. A frm wth hgh R&D ntensty, from a hgh-tech sector, partcpatng n R&D partnershps s found to be more lkely to engage n dsclosure. Frms n the sample were found to "leak" ther knowledge to publc laboratores to a greater degree than to other prvate sector frms. Leakage also was found to be assocated wth mproved nnovaton performance. Ths research helps broaden the lterature on knowledge management practces to nclude not only the pursut of formal ntellectual property rghts such as patents but less formal nter-organzatonal knowledge transmsson mechansms. Keywords: nnovaton, endogenous spllovers, cooperaton, appropraton JEL classfcaton: O32, L21, C81 I thank B. Hall, F. Sachwald, C. Le Bas, L. Mott, E. Pfster, an anonymous referee, and the partcpants n the 2001 semnar of the doctoral school on Economcs and Management of Innovaton and Knowledge (Measurements and econometrcs of nnovaton) n Strasbourg (BETA) for helpful comments and suggestons. The Mnstry of Industry (SESSI) and the mnstry of research and hgher educaton (DPD C3) are also acknowledged for the avalablty of data. The opnons and remanng errors are the author s alone. 1

2 I. Introducton Snce Arrow (1962), emprcal and theoretcal studes of the nnovaton process have underlned the mportance of knowledge spllovers. Many emprcal works have been made to assess the exstence and the mportance of the external avalable stocks of knowledge on R&D decsons and productvty (See Klette et al. (1999) for a survey). Some authors have developed a theoretcal olgopoly framework where spllovers may deter R&D nvestments (see De Bondt (1996)) for a revew) or even the mpact of spllovers on economc growth. Ths lterature consders that spllover levels are exogenous. A frst step towards the concept of endogenous spllover s to consder that frms develop costly strateges n order to obtan external knowledge. Whle the usual models assume that frms absorb knowledge wthout sgnfcant costs, several emprcal studes (Cohen & Levnthal (1989), Henderson & Cockburn (1996), Adams (2000)) show that the capacty to absorb external knowledge from other frms knowledge depends on ther own R&D efforts. Theoretcal models are also avalable to ground the emprcal fndngs (Cohen & Levnthal (1989), Kamen and Zang (2000)). A second step towards ths concept s to consder that frms are also able to desgn appropraton mechansms n order to prevent spllovers to compettors (See Levn et al. (1985)). Nevertheless, there s an mportant feature of spllovers that these prevous theoretcal and emprcal models have faled to take nto account. Emprcal evdence suggests that frms let ther knowledge flow ether at the level of the nnovaton process or at the nnovatve product or process level. Ths purposeful knowledge sharng behavor s now qute well documented n case studes and ndustrybased emprcal studes, whch provde useful nsghts nto the dfferent means of dsclosure and the characterstcs of the frms engagng n transmsson. A varety of research works suggest that sharng knowledge cannot be restrcted to a research jont venture framework or techncal consorta. Ths paper s an attempt to brdge between the theoretcal treatment of knowledge spllovers and busness practce through an emprcal nvestgaton nvolvng a large number of frms across multple ndustres. The emprcal nqury focuses on three questons: What motvates frms to dsclose nformaton and knowledge? Whch ndustres are more apt to engage n leakage? Does dsclosure enhance a frm's nnovaton productvty? Usng data compled from four dfferent data sets on nnovaton, appropraton, and nnovaton success, ths research explores answers to these questons. The surveys nclude nformaton on ether knowledge sharng mechansms or dsclosure practces mplemented by nnovatve frms. We examne the mpact of ndustry sectors, frm sze, R&D ntensty, and R&D partnerng on four dchotomous proxes of knowledge dsclosure. An nnovaton producton functon s also consdered to nvestgate the mpact of endogenous leakage on nnovaton performance. Fnally, the relatonshp between a frm's dsclosure strategy and ts pursut of formal ntellectual property rghts (IPRS) s examned. 2

3 The paper s organzed as follows. In Secton II, we survey the theoretcal lterature that examnes the ratonale behnd dsclosure strateges as well as the prevous emprcal work done n ths area. In Secton III, we develop a framework of analyss and measurement that ncorporates both the strategy of the transmttng frm to leak and approprate ts knowledge and the ablty of the recevng frm to absorb t. Secton IV presents the survey data and the emprcal methods employed. Secton V dscusses the results, and the fnal secton concludes. II. Knowledge dsclosure: theory and evdence A. Theoretcal models Several recent theoretcal works suggest that endogenous leakage occurs not only n a cooperatve settng but n a compettve settng as well. Frms can establsh knowledge management systems to retan nformaton and knowledge but also to boost the release of knowledge. In demonstratng these outcomes many authors use a two-step olgopoly model wth cost-reducng R&D to lnk the R&D nvestment and subsequent proft level to an exogenous externalty rate (for a survey, see De Bondt (1996)). When comparng non-cooperatve and cooperatve R&D nvestments, a frst natural hyptothess s that dfferent spllover levels occur for cooperatng groups and compettve groups. The models sometmes assume that spllovers are endogenous when cooperaton occurs (Katz (1986); Kamen et al. (1992); Poyago-Theotoky (1995), Beath et al. (1998)). These cooperatve models suggest that complete knowledge dsclosure s requred to maxmze profts when R&D expendtures are coordnated. 1 In fact, an RJV wll obtan any gven R&D output more cheaply than n a non-cooperatve equlbrum. Indeed, compared to the non-cooperatve equlbrum t may acheve greater R&D output wth a smaller total expendture on R&D. An RJV may acheve more R&D output than the noncooperatve equlbrum, especally when the products are dfferentated, ndependent or even complementary. Here the dsclosure rate s subordnate to R&D cooperaton. Katsoulakos & Ulph s paper (1998) presents a model n whch the R&D and nformaton sharng are endogenously and ndependently set even for a non-cooperatve equlbrum. They fnd that t s possble that noncooperatve equlbrum acheves maxmal nformaton sharng, especally when frms conduct complementary research and operate n dfferent ndustres. In an RJV, nformaton sharng s more lkely to occur. Two general conclusons seem to be found n the olgopoly lterature dealng wth R&D nvestment. Frst, nformaton or knowledge dsclosure s more lkely to occur n R&D cooperatve projects. A more precse predcton s that the more the products are dfferentated between R&D partners, the hgher the level of R&D nvestments and the hgher the leakage rate. Secondly, prmarly when products are dfferentated, frms may be able to acheve full research desgn and nformaton 3

4 sharng, and thus they are more lkely to ncrease ther proftablty. More precsely, these theoretcal results suggest the two followng hypotheses : H1 The probablty to dsclose s hgher when a frm mplements R&D cooperatve agreement(s); H2 The probablty to dsclose s hgher for R&D partners as clents or customers (strategc complements), than for cooperaton wth publc labs (consdered here to be nether strategc complements nor substutes) and much more than partnershps wth compettors (strategc substtute). The prevous RJV models are smplstc snce the sharng s found to be an all or nothng proposton. However ncomplete sharng occurs for dfferent reasons. Imperfect nformaton sharng may come up aganst an upper lmt. Such a celng can exst for techncal or economc reasons. In the case of the former, complete codfcaton of the knowledge for transmsson to another frm s smply consdered mpossble to acheve because the nature of the technology (complexty, tactness); De Bondt and Wu (1997) suggest also that knowledge sharng may be mperfect because of technologcal or cultural dstance between frms. In ths case, ncreasng codfcaton costs or transmsson costs would lead to an ntermedate level of nformaton sharng (as suggested by Davd and Foray (1995)). Of course, frms can also hde some of ther research results from ther partners. Adverse selecton n knowledge sharng and moral hazard problems can also be consdered (See d Aspremont et al. (1998), for a survey of the lterature on R&D teams). The volumnous lterature, however, requres data on codfcaton costs and asymmetrc postons n RJV that are very dffcult to obtan for a large representatve sample. Ths topc s out of the scope of the present paper. The concept of dsclosure strategy s not restrcted to the RJV lterature, and voluntary leakage can also be mplemented n a pure R&D compettve framework wth races for nventon. Usng such a framework, Lchtman et al. (2001) show that an effcent strategy for a laggard frm may be to narrow the scope of or fully preempt a rval s patent. In ths case, the laggard dscloses any research results t had n common wth the rval and can compete n a duopoly sub-market nstead of a monopoly market. From the leader s perspectve, the frst stage of nformaton dsclosure s a means of drvng laggards from the race. However a leader wll not adopt a dsclosure strategy at the nnovaton output level (second stage). Laggards and leaders do not necessary dsclose the same knd of nformaton. Ths knd of model seems thus very hard to test because t detaled descrptons of the nformaton and knowledge flows as well as the strategc poston of the compettors. De Fraja (1993) obtans dfferent and more tractable results employng a smlar modelng approach: frst, frms tend to let ther knowledge leak when they have to share the market wth mtators, when the loss of the laggard s small compared to the 1 Kamen et al. (1992) then refer to these arrangements as Research Jont Ventures (RJVs). 4

5 gan to the leader f he wns, and when both frms choose to shorten the nventon race (See also Von Hppel (1987), Haroff et al. (2002)). Second, De Fraja shows that the level of R&D efforts experences a smlar outcome: a hgh requred R&D ntensty tends to nduce dsclosure n order to shorten the length of the nventon race. A hgher dsclosure rate s thus assocated wth a hgher R&D ntensty. De Fraja s results are consstent wth the conclusons from the ltterature on R&D cooperaton. These theoretcal models yeld the followng hypothess: H 3 Informaton dsclosure may be complementary to appropraton strateges rather than a substtute. De Fraja s framework, combned wth olgopoly models wth R&D cooperaton 2 between frms wth dfferentated product suggest an addtonal hypothess: H 4 Informaton dsclosure s hgher for R&D ntensve frms. B. Emprcal evdence At the emprcal level, a number of studes provde drect evdence of ntentonal knowledge sharng by frms and the means mplemented to dsclose nformaton and knowledge. A number of dfferent works suggest that the dsclosure strategy s much more frequent and mportant than usually consdered among scholars; that t occurs usng a varety of spllover channels and that the phenomenon can be nformally or formally organzed. Even f the emprcal evdence s more extensve than the theoretcal lterature, the emprcal fndngs are consstent wth the theoretcal predctons. A frst step to understandng the knowledge dsclosure possbltes s to recognze that a market for knowledge exsts. Arora et al. (2000) suggest an ncreased roll for market medated technology transactons coverng lcensng and or contract R&D. The rse s sometmes presented as a general movement of codfcaton of knowledge that facltates knowledge transacton and absorpton at lower costs. The emergng nformaton and communcaton technologes make t easer to codfy, transport and trade knowledge (Antonell (1992)). Voluntary knowledge dsclosure transactons are made possble thanks to the mplementaton of property rghts. Market transactons nclude labor moblty as an mportant source of knowledge transmsson: n ths settng, a frm may refuse to bd up an employee s wage n order to retan her. Ths result s found at the local level (See Saxenan (1994)) even f moblty may also found to be assocated wth knowledge flows regardless of geographc proxmty (Rosenkopf & Almeda (2001)). Even f a departng employee s requred to sgn a nondsclosure statement that 2 R&D ntense frms are frms wth a hgh R&D level compared to ther turnover. The relaton of R&D ntensty wth the R&D dsclosure rate s however not smple n R&D cooperatve models. In D Aspremont & Jacquemn (1989) framework, t s possble to show that R&D ntensty s hgher for a cooperatve R&D frm wth complete knowledge sharng compared to a compettve frm wth a lower R&D externalty rate. 5

6 prevents her from revealng company secrets, the possble cross-fertlzaton through moblty seems to weaken ts enforcement (Saxenan (1994); Glson (1999)). At the same tme, many case studes document that knowledge s revealed outsde the market n a more nformal way. Allen (1983) provded hstorcal evdence, based on the Brtsh blast furnace ndustry n the nneteenth century, that nnovatve frms release of technologcal nformaton and performance to ther compettors. From the mn-mll frms of the US steel ndustry, Von Hppel (1987) and Schrader (1991) show that many employees frequently gve techncal nformaton to colleagues from compettors (and from other ndustres as well) through nformal networks. 3 Contrastng Slcon Vallye, Calforna wth Route 128 n Massachusetts, Saxenan (1994) shows that such dsclosures by Slcon Valley engneers can even become socal norms. Powell et al. (1996) remnds us that nformal nformaton and knowledge tradng often occurs durng the course of a formal knowledge exchange agreement (See also Arora (1995) on lcenses). The last references underlne the mportance of spontaneous and even nvoluntary face-to-face communcatons. Lynn et al. (1993) or Appleyard (1996) faled to fnd a statstcally sgnfcant dfference n knowledge-sharng usng personal contacts between U.S. and Japanese engneers. Several recent papers provde evdence that the dssemnaton of nformaton and knowledge by frms s also made through publcatons: corporate researchers publshed about 8 percent of U.S. scentfc papers n 1994 (Hcks (1995). Henderson & Cockburn (1996)) underlne that researchers n the pharmaceutcal ndustry are often aware of ther rval s research thanks to a hgh rate of open scentfc publcatons. Recent studes pont to the strkng number of dsclosures among compettors. Freely revealed nformaton between users s also common especally when network effects occur (See Harhoff et al. (2000), for the case of IBM or Brand s ntervew (2001), on the DoD and nternet)). Furthermore, knowledge transfers and communcatons are often encouraged among afflates wthn a group. The knowledge sharng s expected to be especaly hgh when a specalsed R&D afflate takes the place of nternal R&D department. Knowledge sharng may be dfferent for afflates of foregn multnatonal enterprses. Foregn frms may encounter dffcultes n cooperatng or enterng local networks n host countres. To overcome such dffcultes, knowledge dsclosure may be an effcent strategy (See Blomström & Kokko (1998)). Taken together, the arguments leads to the followng hypotheses : H 5 Informaton dsclosure s hgher for afflates who are encouraged to share knowledge nsde ther group; H 6 Informaton dsclosure s hgher for foregn afflates relatve to host country afflates. 3 See also Harhoff et al. (2002) for examples. 6

7 Many studes suggest that dsclosure s often embedded n a more formal framework such as offcal R&D cooperatve agreements. Followng the theoretcal lterature, several econometrc studes gve clues to the exstence of endogenous spllovers n blateral cooperaton agreements or larger R&D consorta (See Irwn & Klenow (1996); Sebert (1996)). Endogenous spllovers are however rarely drectly measured. The exstence of endogenous leakage s evoked as a means to nterpret n the dfferent emprcal results. Branstetter & Sakakbara (1998) take a more drect approach and consder than the degree of R&D spllovers realzed wthn a consortum may be affected by the organzatonal structure of the consortum (p 4). Usng a sample of 226 Japanese frms, observed from 1983 to 1989, the authors fnd that frms partcpatng n RJVs have hgher R&D nvestment and productvty, related to an ncrease of potental endogenous spllovers. The potental spllovers are defned as a postve functon of the technologcal proxmty of partcpants or the degree of basc technologcal orentaton of the research. The paper makes an nterestng dstncton between knowledge sharng at the nput level due to the characterstcs of the knowledge producton process from that at the output level of the nnovaton process. The paper concentrates on the frst category to explan the potental knowledge sharng n Japanese consorta, but t downplays the subsequent outcome of the phenomenon that frms may also decde to dsclose nformaton and knowledge concernng ther nnovaton output. Cassman and Veugelers (2002) dstngush ncomng spllovers and outgong spllovers. Usng a sample of 411 Belgan frms, the authors fnd that frms wth hgher ncomng spllovers are more lkely to be actvely engaged n cooperatve R&D agreements. The author's results suggest that the level of knowledge leakage and absorpton s not exogenous to the frm, whch manages ther absorpton and appropraton capabltes durng the nnovaton process. In ths framework, they fnd however that frms whch are more effectve n appropratng the results from ther nnovaton process are also more lkely to cooperate n R&D. The result seems at odds wth theoretcal models where nformaton sharng and R&D cooperaton are lnked. More generally, consorta are lkely to organze the access to nformaton and knowledge n ways that make t acceptable to the members. Three dfferent formal collectve structures can be found n the lterature: patent pools (Merges (1999)), prvate offcal and unoffcal standardzaton consorta (e.g., ATVEF on televson, IETF on Internet), and research consorta (such as SEMATECH or MCC). Consderng dfferent types of structures, Rosenkopf and Metu (2001) suggest that the knd of techncal commttee actvty facltates the entry of less-establshed frms nto allance networks where knowledge sharng occurs. Intermedaton can also be mplemented by ndependent frms as jont venture captalst frms (See Aok (1999)) or contract research unts (See Arora & Merges (2001)). Case studes also show that the transmtter s far from nave. Aok (1999) asserts that frms n Slcon Valley do not dsclose specfc nformaton crucal for ther own compettve nches but nstead only 7

8 nformaton and knowledge relevant to dmnshng the turbulence of the evolvng technologcal frame. Usng data on ndustry-fnanced research, Cohen et al. (2000) fnd that organzed leakage by frms through academc publcatons s, n more than half the cases, less complete or slower than n the usual academc settng. In addton, several case studes menton the refusal to allocate the best researchers to RJVs (e.g. Bhattacharya et al. (1992)). Opportunsm seems however not to be a compulsory trat of nformaton and knowledge sharng. Innovaton through an nformal, entrely open and decentralzed epstemc communty such as Lnux s a well-documented workable model (See Von Hppel (2002)). Three aspects reman qute hazy n the lterature surveyed. Frst, there s lttle evdence of the effect of sze on the dsclosure phenomenon. If the dsclosure process can be seen as negotated by two monopolsts, one could conclude that large frms have stronger barganng power than ther small counterparts and are thus more lkely to receve than gve. However, one can argue that large frms are more lkely to protect ther monopoly rents (through credble threats or hgher ltgaton capactes as n Lerner (1995) and that the leakage rate s thus postvely nfluenced by the sze of the frms. Another argument would be that large frms are more lkely to nduce standardzaton effects than small ones and are thus more lkely to freely dsclose knowledge to ther users. These arguments leads to the followng hypothess: H 7 Informaton dsclosure s hgher for large frms whch are more able to manage ther envronment and to enforce networkng effects. Second, the degree of novelty of knowledge s not drectly nterpreted as a determnant of knowledge dsclosure. Even f the phenomenon s documented n low-tech ndustres (as n Von Hppel (1987) or even Merges (1999)), most of the case studes deal wth hgh tech sectors, such as sem-conductors, telecommuncaton or pharmaceutcals (See Saxenan (1994); Aok (1999); Cockburn & Henderson (1996)). Many studes are then largely focused on manufacturng ndustres wth a rapd pace of technologcal change, suggestng that ndustres wth a slow pace of technologcal change are less lkely to share knowledge, yet some counter arguments can be found (See Appleyard (1996)). Ths observaton underpns the followng hypothess: H 8 Informaton dsclosure s hgher for Hgh Tech sectors, especally n ICT ndustres where network effects and or techncal comty exst. Fnally few general results are avalable on the effects of knowledge dsclosure on frm performance. Some work based on specfc ndustres suggest thats nformal tradng knowledge networks have a postve effect on frm performance (Schrader (1991)); that a provson such as the publcaton of academc papers s an effcent long-term ncentve for researchers (Cockburn et al. (1999)); that RJV frms partcpatng n RJVs have hgher R&D productvty (Branstetter & Sakakbara (1998)). 8

9 III. Knowledge dsclosure: further buldng blocks Unlke much of the past lterature, ths study underscores the potental mportance of knowledge leakage that goes beyond formal ntellectual property rghts. Few theoretcal and even fewer emprcal studes provde ratonale for dsclosure behavor broadly defned or deal wth measurement ssues. In order to develop a coherent and workable framework to analyze and explan knowledge sharng, three addtonal buldng blocks must be frst put n place: the place of voluntary knowledge dsclosure n a theory of endogenous spllovers, the ncentves to dsclose, and the ssue of measurng the dsclosures. A. Integratng dsclosure strategy Whle establshng the endogenous nature of absorpton strateges, the economcs lterature has tended to assume that leakage s "unntentonal" (Mshan (1971)) or reles on the nature of technology. In ths vew, a frm can control and endogenze the level of leakng knowledge only through ts R&D nvestment decson. A frst step n an analyss of endogenous outgong spllovers s to consder that frms make costly nvestments n order to lmt the level or the usefulness of leakng knowledge. The dfferent appropraton strateges are now well documented (Levn et al. (1985); Cohen et al. (2001)). A complementary vew s to assume that the share of basc research n R&D s a means to control outgong knowledge (Vonortas (1994), Yn & Zuscovtch (1994)). The addton of appropraton strateges to absorptve strateges gve a more complete and factual vew of endogenous externaltes (Cassman & Veugelers (2002); Gersbach & Schmutzler (2003)). Purposeful knowledge sharng remans however outsde the scope of these papers gvng only a partal vew of the leakage assocated wth knowledge externaltes. Two mportant features of an endogenous theory of knowledge flows are mssng: Frst the leakage rate should be dvded nto an appropraton strategy and a dsclosure strategy that both requre dfferent costly nvestments. Second, the lnk between the two strateges needs to be clarfed. On one sde, snce Arrow (1962), the two strateges are consdered to be antagonstc (Cassman & Veugelers (2001)). On the other sde, emprcal nvestgatons of the electroncs and pharmaceutcal ndustres suggest that nformal knowledge dssemnaton occurs even when frms follow an aggressve patent applcaton strategy and/or are compettors (Appleyard (1996), Hall and Zedons (2001), Henderson and Cockburn (1996)). Usually assumed as at odds wth the appropraton mechansms, the case studes on patent pools and research or standardzaton consorta suggest also that dsclosed knowledge can be a complementary and not a strategc substtute wth patent strateges. It also suggests that the knd of knowledge freely dssemnated by frms s not the same as knowledge that s protected or sold (See Von Hppel (1987), Merges (1999), Aok (1999)). Formally, the technologcal externalty rate χ j between two frms and j can be defned as the product of a leakage rate λ of the transmtter and an absorptve capacty α j of the receptors j. Often the 9

10 externalty rate between frms s consdered exogenous. In ths case, a gven externalty rate s consstent wth an nfnte combnaton of dsclosure and absorptve rates (See the so-externalty curve for χ j = 0.1 n Fgure 1). The understandng of a gven externalty rate n an economy thus reles on the dentfcaton of the dsclosure capacty as well as the absorptve capacty. [Fgure 1 about here] Past research has emphassed ths dual requrement both at a theoretcal and emprcal level. Wth respect to ncomng spllovers, the endogenety of χ j often reles on the endogenety of α (e.g. Cohen and Levnthal (1989); Adams (2000)). Fgure 1 shows that χ rses wth α ( α 1 < α 2 < α3 ) for a gven exogenous leakage rate λ 4. α depends on nternal absorptve capactes. We propose n ths paper to focus on the characterstcs and endogenety of the λ parameter consderng that the absorptve capacty of frms s ndependent of dsclosure strateges. Followng Merges (1999) or Aok (1999), we consder that knowledge that s ntended for retenton (RK) may dffer from knowledge that s meant for dsclosure (DK). Thus the knowledge potentally useful to receptor frm j from frm s a weghted sum of the two publc knowledges: j K j [( ρ ) RK + δ DK ] α, j = 1, j where ρ, δ are respectvely the rate of retenton and dsclosure of frm, and j α the rate of absorpton of frm j. 5.Ths paper attempts to clarfy the determnants of the dsclosure coeffcent and ts lnk wth the appropraton coeffcent consderng that the organsatonal practces dedcated to dsclosure and retenton actvtes are at the core of the two parameters. For tractablty, n order to test emprcally the hypotheses from the large body of theoretcal and emprcal lterature, t s necessary to neglect ths mportant aspect of appropraton. B. The mpact of knowledge sharng on performance The decson to dsclose knowledge or not depends on whether the expected benefts from the leakage outwegh the costs resultng from dsclosure. The costs may nclude the declne of market power attrbutable to the knowledge plus the transacton costs of dsclosure (See Appleyard (1996)). The latter costs can be broken down nto the codfcaton costs assocated wth the dffuson of knowledge and the communcaton costs. Of course, lke the retenton strategy, the dsclosure actvty s mperfect. The degree to whch the dsclosure capacty s really the choce of the frm depends also on the locaton and 4 In ths case, the dsclosure rate becomes the lmtng component of χ. 5 All the three parameters can be consdered as bounded between 0 and 1. A celng value other than unty can be ntroduced to nclude the ncreasng dffcultes of the three dfferent strateges. 10

11 nature of the technology. Tact knowledge can therefore be emboded at the ndvdual level as tact human sklls or may be at the organzatonal level, embedded n collectve routnes (Wnter (1987)). The benefts expected by a "leakng" frm span at least three categores. Frst, a frm wth a lengthy nventon race may dsclose knowledge to shorten the development process (De Fraja (1993)) especally when there s a small compettve advantage to be the leader. In ths context, there s a drect effect on costs and a possble ndrect effect nduced by the compettve frm whch also shares ts knowledge. The case studes also suggest that sharng knowledge s done to secure the nnovaton process. Thus, even wthout any formal R&D agreement, sharng knowledge can be consdered as a wde long-run precompettve cooperaton among frms. In ths context, Haroff et al. (2002) suggest that the nnovator may create or dscharge through knowledge sharng generalzed recprocty oblgatons combned wth reputaton effects. Second, frms may, beneft from ntentonally leakng the knowledge through network effects (wth or wthout externaltes) resultng from the pervasve adopton of technology sometmes nfluenced by standards (Katz & Shapro (1994)). Thrd, followng Katz (1986), many models have also shown that R&D partners can mprove ther profts f they maxmze ther (symmetrc) dsclosure rate. The dsclosure strategy has therefore several possble mpacts n terms of frm performance. Frst, sharng knowledge can lower nnovaton costs, wth or wthout formal RJV agreements. Second, free dsclosure of nformaton and knowledge wth recprocty may ncrease the qualty of nnovaton process or product. Thrd, t may protect the exstence of a market for nnovaton n a olgopoly market where a monopoly poston by others would be damagng. Fnally, knowledge dsclosure may wden the market for new products or processes. Ths rough summary, combned wth theoretcal results and emprcal evdence, suggests three addtonal hypotheses: H 9 Innovaton productvty s hgher for frms sharng knowledge. The assumpton s thus not restrcted to the R&D cooperatve framework. However, the olgopoly models on R&D cooperaton yeld to the two followng hypotheses: H 10 Innovaton productvty s hgher for frms sharng knowledge through R&D partnershp(s), because of nternalzed spllovers. H 11 Innovaton productvty s hgher for frms belongng to a group where the R&D costs and knowledge are more lkely to be spread among the afflates actng as a natural consortum. Here nnovaton productvty can be measured by the share of nnovaton sales or, f avalable, the nnovaton sales dvded by nnovaton costs. C. Measurng knowledge dsclosure 11

12 Several surveys deal wth the dsclosure of knowledge by frms. However only a few aspects of the dsclosure process are covered by questonnares pror to those consdered here. For example, the R&D survey harmonzaton resultng to the Frascat manual allows us to obtan the level of annual external R&D expendtures and recepts and thus to dentfy the level of R&D market transactons. One dffculty wth these data s n determnng the R&D fnancal flows earmarked for R&D subcontracts versus R&D cooperaton. RJV actvtes are also hard to dentfy snce, even f the R&D frm s surveyed, ther recepts from partners are not well defned. Furthermore, fnancal flows may not accurately reflect the knowledge flows for groups wth vared transfer prcng. Fnally, R&D recepts are hardly precse when the knds of partners are consdered. The data may dstngush between prvate ntegrated frms, publc labs and specalzed R&D frms, but rarely are there dentfers that ndcate whether the organzatons are compettors or customers or supplers.. A more general shortcomng s that the quanttatve R&D data do not capture the dfferent knds of knowledge flows occurrng as a consequence of an R&D contract or an nformal cooperatve relatonshp. Examnaton of patents s also useful to dentfy the level of nformaton released. The value of knowledge codfed n a patent s hard to measure even f patent ctatons can mprove the measure of the qualty of knowledge contaned n the patent. Unfortunately, many patent systems do not nclude meanngful nformaton on qualty, so t s dffcult to nduct a systematc assessment of the qualty of ndvdual patents and the degree of access to the knowledge. From a methodologcal pont of vew, the measurement of voluntary dsclosure requres great care. The concept of leakage mxes voluntary and nvoluntary dsclosure strateges. The dfferent agents (the frm, the group, and the employee) nvolved n dsclosure need to be taken n to account. The dstncton between formal and nformal nstances of dsclosure cannot be made solely based on channel of dsclosure (lunch, phone call ). A challenge arses when attemptng to classfy exchanges because multple dsclosure categores may apply to sngle exchange (e.g. a paper publshed by a member of a research consortum). The pror research surveyed here deals wth several aspects of the dsclosure phenomenon (market, consorta, nformal transfer) but hardly exhausts the possble relatonshps between two partes engaged n knowledge sharng. In Fgure 2, we depct the possble relatonshps and dsclosure decsons (See also Appleyard (1996)). [Fgure 2 about here] A horzontal readng makes us aware of the number of useful dstnctons that need to be drawn between nstances of knowledge dsclosure. The modes at the rght sde of the fgure are numerous and smlar to each other. Thus, even f the bulk of dsclosed knowledge s covered by the dfferent sources or surveys, an ndcator across the dfferent channels may reflect mperfectly the knowledge sharng strateges of a frm. An ndcator can also overlap another one and measure the same amount of knowledge sharng. 12

13 Takng note of these possble shortcomngs, ths paper seeks to test the dfferent hypotheses outlned above. The followng secton presents the varables collected as well as the econometrc strateges mplemented to make operatonal the eleven research hypotheses. An addtonal secton presents the results of the dfferent estmatons. IV. Varables and econometrc models A. The explaned (dependent) varables The data set n ths paper s constructed from fve separate surveys: the frst and second French communty nnovaton survey (CIS1 and CIS2) survey, the frst appropraton survey (PACE), the survey on nnovaton competences (COMPET), and the annual R&D survey. The CIS1, PACE, and COMPET surveys deal wth knowledge acquston and knowledge dsclosure. We focus n our study on the last aspect of endogenous externaltes that s rarely consdered n the emprcal lterature (see however Cassman & Veugelers (2001)). These three surveys are combned wth the annual mandatory R&D survey (R&D survey) as well as the non-nnovatve frms from the second CIS sample. 6 Of the surveys used to collect the data n ths study, the frst CIS survey s partcularly well adapted to pn down the means of knowledge transmsson. A queston s asked concernng the dfferent means of transferrng knowledge: Dd your enterprse transfer to a thrd party (ncludng a frm wthn the enterprse group) new technologes by the followng means? 1) R&D carred out for a thrd party ; 2) the rghts to use a thrd-party nventon (among them patent or lcense to sell) (LICEN) ; 3) consultng for other frms (CONSU) ; 4) equpment sales (EQUIPT) ; 5) the departure of qualfed employees (LEAV) ; 6) communcaton wth other frms (COMM) ; 7) the sale of a part of the frm (SELLING) ; and 8) strategc jont venture allances. The survey then asks the dentfcaton of the type(s) of receptor: wthn country, wthn the EU, n the USA, Japan, or Other). As mentoned, the lstng of the dfferent means of transfer proposed shows a bas toward market-medated trade of knowledge and nformal or temporary transfer s not ncluded as an opton. An alternatve way characterze knowledge transmsson s to obtan nformaton on the knowledge management practces at the frm level. Here two dfferent ndcators are avalable. An ntal step s to focus on the organzatonal rules surroundng nformaton dsclosure: Are your engneers allowed to dffuse non-confdental aspects of your technologes (publshng artcles, conferences)? asked n the 1994 PACE questonnare, or Do you montor communcatons on your strategc knowledge? asked n the 1997 survey on Competences, offer two examples contaned n ths research. The last queston may 6 See the appendx on data sources for more nformaton on the varous surveys used n ths paper. 13

14 be broader than n the PACE formulaton snce strategc knowledge can nclude non-confdental as well as confdental nformaton or knowledge. Usng the CIS1, PACE and Competences surveys, four proxes of knowledge dsclosure are computed and used as varables to be explaned. From CIS1, we compute a dchotomous varable LEAK that s one when at least one means of transfer among the proposed seven tems s used durng the perod. 7 Only 26% of nnovatve frms 8 note such a technology transfer. As mentoned, the ndcators are based toward market-medated dsclosure. To nvestgate the degree of bas, we also ncluded a smpler measure of dsclosure based on COMM ( Communcaton wth other frms ), whch s used by 12% of the nnovatve frms from CIS1. [Table 2 about here] The 1994 PACE survey s questons on dffuson are used drectly to buld a dchotomous varable called AUTHO, equal to one when engneers are allowed to dffuse non-confdental aspects of technologes, otherwse zero. Of 983 nnovatve frms from PACE wth more than 49 employees forty-four percent allow such communcaton (weghted usng populaton weghts). Smlarly, the 1997 Competences survey s used to compute an ndcator PERMI (standng for permsson ), equal to one when an nnovatve frm allows ther employee to communcate strategc knowledge to outsde partes. Despte the bas assocated wth the queston, the results are very close to the results for the PACE survey: 45% of the nnovatve frms gve such permsson to ther employees. The share of product nnovaton turnover (IT ) n total sales s also avalable 9. Dfferences exst across the two dfferent CIS questonnares n terms of how nnovaton turnover s measured. In CIS1, the part of product nnovatve sales of the total 1992 sales (IT CIS1 thereafter) s classfed and ordered nto four classes ( 0 to 10%, 10 to 30%, 30 to 70%, 70% and above). In CIS2, the level of the 1996 nnovatve turnover s avalable (to get IT CIS2 the fgure s dvded by the 1996 total sales). The PACE and CIS2 surveys contaned a dchotomous varable PATENT. The ndcator s one, when a patent applcaton was 7 The dfferent tems are those lsted above. LEAK = Max (LICEN, CONSU, EQUIPT, LEAV, COMM, SELLING). Thus the frst alternatve ( R&D carred out for a thrd party ) s not ncluded because nformaton dsclosure s supposed to be ncluded n the transacton. 8 An nnovaton, as defned n the CIS and Competences surveys, s a new or sgnfcantly mproved product (good or servce) ntroduced to the market or the ntroducton wthn your enterprse of a new or sgnfcantly mproved process. The nnovaton s based on the results of new technologcal developments, new combnatons of exstng technology or utlzaton of other knowledge acqured by your enterprse. The R&D manager, the general manager, or the frm drector s supposed to answer to the queston. 9 In CIS2, nnovatve turnover s computed from the answers to the followng queston: Please estmate, n percentage, how your turnover n 1996 was dstrbuted between new or sgnfcantly mproved products (goods or servces) ntroduced durng the perod and unchanged or only margnally modfed products (goods or servces) durng the perod Innovatve turnover s the sum of turnover attrebuted to new or sgnfcantly mproved products. The defnton was smlar n CIS1. 14

15 submtted on the perod (null otherwse) whch s the case n 52% n PACE and 49% n CIS2 (or 38% for all szes as shown n Table 2). B. The ndependent varables Four sets of ndependent varables are ncluded as explanatory varables: sze, ndustry, group membershp, and R&D ntensty. In the CIS1 and Competences survey, frm sze s a categorcal varable based on the number of employees. Sze s taken as the reference group snce the PACE survey concerns only frms wth more than 49 employees. Thus, when the dependent varable s AUTHO, only fve sze dummes are ntroduced. As sze ncreases, the coeffcents of these dummy varables are expected to ncrease, wth more dsclosure for larger frms. At the same tme, the relatve effort of R&D, measured by the average R&D ntensty (the logarthm of the rato of R&D to sales, measured between 1988 and 1992 for CIS1 and PACE and between 1992 and 1996 for Competences), s also expected to be postve and sgnfcantly correlated to the probablty of dsclosure. However, sze s expected to be ndependent of the share of nnovatve sales as typcally stated n the lterature on the Schumpeteran conjecture on sze (See Cohen & Klepper (1996)). Wthn-ndustry determnants of knowledge dsclosure are controlled for by fxed effects defned at the 2-dgt level of the NACE. Takng the Clothng, leather, and footware ndustry as a reference, a set of 13 dummes denoted IND m (m = 1,..,13) s ncluded. Followng Gerosk s (1998) argument, such sector dummes capture dfferences n technologcal opportuntes facng frms located n dfferent sectors. We expect that greater technologcal opportuntes are lnked to technologcal uncertanty as well as to network effects. The expected sgn s thus expected to be postve and the mpact on dsclosure hgher n the hgh tech sectors. 10 The ndustry dummes are also expected to greatly mprove the explantory power of the regresson. The group dmenson s avalable only for the Competences questonnares through CIS2. FRGROUP and FORGROUP are dummes for belongng to French and foregn groups respectvely. In our sample, 49% of nnovatve frms belong to a group whereas 21% belong to a foregn group. The expected sgns of FRGROUP and FORGROUP dummes are postve and the two parameters are expected to be jontly sgnfcant. Followng the H6 hypothess, we also expect that the FORGROUP parameter s postve and hgher than the FRGROUP parameter when explanng dsclosure behavor. In terms of nnovatve productvty, the two parameters are expected to be postvely lnked wth IT (H11). However, the relatve rankng of ther values s unclear n ths case. 10 The hgh-tech, low-tech dchotomy s here s based on the OECD classfcatons whch take the R&D-ntensty of dfferent ndustres as a measure of ther knowledge-ntensty, strategc mportance and future potental. Pharmaceutcals, electrc & electronc supples, electrc & electronc components, and to a lesser extent aeronautcs, tran, shps and motor vehcles ndustres are thus consdered to be hgh tech ndustres. 15

16 In order to test hypothess 3, we assume that the determnants of patentng (PATENT) are the same as those for PERMI. Usng sze, R&D ntensty and ndustry dummes as explanatory varables, the model s very smlar to the usual specfcaton of a patent propensty model (e.g. Brouwer & Klenknecht (1998)). The expected sgns of the second (patent) equaton are predcted to be smlar to the frst despte prevous contradctory results on patent propensty (e.g. on the effect of sze, see Duguet & Kabla (1997), Brouwer & Klenknecht (1998)). Furthermore, the dfferent types of R&D partners are avalable through CIS1 and CIS2. 69% of the nnovatve frms clam to be engage n R&D partnershps (COOP = 1, null otherwse), and the rato falls to 39% n the CIS2 survey. We aggregated the dfferent types of R&D partners to three ndcators: VERTICAL, whch s one when an nnovatve frm declares at least one partner among ts customers, materal or equpment supplers (null otherwse); GRPE, whch s one when at least one R&D lnk s used wth the group laboratory of another sster frm (null otherwse); and lastly, PUBORG., whch s one when partners such as unverstes, publc laboratores or techncal centers are encountered (null otherwse). Dummes for partner frms desgnated as consultng partners (CONSULT=1, null otherwse) as well as compettors (HORIZONTAL=1, null otherwse) are ncluded wthout any modfcaton. The ncluson of the R&D partner varables s possble only for the estmates based on CIS1 and CIS2 (merged wth Competences) samples, that s for the LEAK, COMM and PERMI endogenous varables. As stated n H2, a herarchy of the R&D partners dummes s expected: the Vertcal coeffcent s expected to be sgnfcantly hgher than the PUBORG and HORIZONTAL coeffcents. Overall, engagng n R&D partnershps s expected to be postve for dsclosure (hypothess H1). We can summarze the multple varables and the expected sgns of parameters n two smple equatons. The frst one explans the dsclosure strateges of frms: DISC = l= 5 l l= 1 > 0 τ RDPART l + γ INTENS + ϕ FRGROUP + ϕ FORGROUP + 1 > 0 1 > 0 2 > 0 + γ 2 < 0 NORD m= 21 m= 9 γ m + k = 8 k k = 3 > 0 IND m γ SIZE k + γ ONE + µ 22 (1) In ths case, the varous endogenous varables DISC are LEAK, COMM, AUTHOR or PERMI. Moreover, R&D partners are not avalable for the PACE questonnare, and FRGROUP and FORGROUP are ncluded only for the PERMI varable. Dfferent hypotheses (H1, H2, H4, H5, H6, H7 and H8) can be tested usng these specfcatons. The second equaton s an nnovaton producton functon, mplemented to test the H9 to H11 hypothess: 16

17 IT = δ DISC + > 0 l= 5 l l= 1 > 0 τ ( RDPART DISC ) + γ INTENS + γ + ϕ FRGROUP + ϕ FORGROUP + 1 > 0 2 > 0 l m= 21 m= 9 1 > 0 γ m IND k 2 < 0 NORD + γ ONE + µ 22 + k = 8 γ k k = 3 = 0 SIZE k (2) CIS1 where DISC takes the same four values as above and IT (share of nnovatve sales) s ether IT or IT CIS2. As above, the Group dmenson s only avalable usng CIS2. As found prevously by sectoral studes the effect of dsclosure DISC on IT s expected to be postve. R&D cooperaton should encourage frms to dsclose nformaton.. Here the broad defnton of cooperaton s helpful: the mplementaton of the R&D relatonshp can be formal as well as nformal. Introducng dfferent cross effects between the dsclosure varable and the R&D partnershp dummes, we expect to fnd that a frm dsclosng nformaton and cooperatng s more lkely to be productve n terms of nnovaton. Even f we do not have a dsclosure ndcator for each type of R&D partnershp (the leakage may be authorzed when partners are publc labs and not wth compettors), we ntroduce the fve cross effects to gve a crude estmate of the hypotheszed cross effects. C. Econometrc ssues The frst model explans dsclosure usng a latent varable formulaton. In ths case, the latent varable DISC* = γz + ν where z s a vector of explanatory varables and DISC = 1 f DISC* > 0, DISC = 0 f DISC* 0, and DISC represents LEAK or COMM or AUTHO or PERMI. The four equatons are estmated usng maxmum lkelhood estmaton. As a frst step, model (1), whch ncludes the R&D partner varables s estmated wthout the cross effects, whch are then entered one by one. Based on the theoretcal models of RJVs, a postve cross effect between R&D ntensty and VERTICAL or PUBORG s expected whle the cross effect wth HORIZONTAL s expected to be smaller. Thus fve separate Wald tests, for the fve dfferent types of R&D partners, are conducted to test the nullty of the nteracton term, along wth a Wald test of the jont hypotheses that all the nteracton terms are smultaneously non dfferent from zero. In order to test the H3 hypothess, we estmate the frst equaton, together wth an equaton explanng the probablty of flng at least one patent (avalable n CIS2, PATENT=1, null otherwse). Consderng the observed varables as ndcators of latent varables measurng the IPR appropraton strategy, we use a smple bvarate probt model n whch the frst equaton s the same as the prevous: PERMI* = γ 1 z 1 + ξ, where PERMI = 1 f PERMI* > 0, and PERMI = 0 f PERMI* 0. For patentng there s a smultaneous second equaton: 17

18 PATENT* = γ 2 z 2 + ζ, where PATENT = 1 f PATENT* > 0, PATENT = 0 f PATENT* One way to test the nterdependent hypothess s by assumng that the error terms of both decsons are correlated n the form Cov[ξ, ζ ] = ρ (See Goureroux (2000)). In ths case, f ρ>0, the complementarty between the endogenous varables s found as expected. Independence s rejected f the ρ varable s sgnfcantly dfferent from 0. A thrd case may be that IPR and dsclosure are substtutes (ρ<0), as usually consdered among scholars. A fnal model s consdered to test the two hypotheses on the mpact of knowledge dsclosure on nnovaton productvty (H9, H10, H11). One problem s that IT and R&D partners are observable only for nnovatve frms. In order to mtgate ths selectvty bas n the estmaton of model (2), we ntroduce a correcton for selectvty n the estmaton procedure. The selecton mechansm s modeled usng a probt equaton: INNO* = αx 1 + u CIS. where INNO = 1 f INNO* > 0, and INNO = 0 f INNO* 0 and INNO s the observable declaraton n CIS1 and CIS2 of the latent varable INNO* CIS.. We use smlar selecton equatons for the two CIS samples. The group dmenson s however avalable only for CIS2. Thus the selecton equaton can be summarzed as the followng: 14 * = α 1 SIZE + ϕ1 FRGROUP + ϕ 2 FORGROUP + > 0 > 0 > 0 k= 2 INNO α IND + α ONE + e k k 15 We expect the lkelhood of beng an nnovator to be a functon of the condtons of technologcal opportunty: t s easer to nnovate n hgh tech sectors than n others. We also ntroduce a sze effect (SIZE) wth a postve expected sgn (See Cohen & Klepper (1996)). Because the share of nnovaton sales IT s reported by ndcator varables (brackets) n CIS1, n the frst selecton model, ths varable s modeled by a latent varable IT* where IT s the observed counterpart of the varable only f INNO = 1. Snce the observed varable s an ordered varable, the frst selecton model s an ordered logt model where IT* = β X 2 + e, wth 0% IT * 100%. The four IT brackets are reduced to three due to the lmted number of frms n the last bracket. Thus, IT = 0 f IT * 10%, = 1 f 10% IT * 30%, = 2 f 30% IT * 70% and = 3 f IT * 70%. 11 A smultaneous equatons model wth two bnary dependent varables was not consdered n ths frst stage of our nvestgaton. Even f ndependence between the two error terms s rejected, the ntroducton of one endogenous varable n the set of ndependent varables s hard to justfy because the causalty between PATENT and PERMI s not well defned n the lterature. The postve effect of PATENT on PERMI may be due to the partcular case of standardzaton consorta: a patent s often presented as a prerequste to bargan and dsclose nformaton and knowledge. 18

19 A generalzed Tobt model s appled n a second selecton model used to explan the contnuous endogenous varable IT CIS2 (See Mohnen & Dagenas (2000)). The ordered Logt wth sample selecton as well as the generalzed Tobt model are estmated by a two-step procedure whose last step uses the full nformaton maxmum lkelhood method. Compared to X l, X 2 ncludes the R&D ntensty, the dfferent leakage proxes (LEAK, COMM) and the knd of R&D partners. The effcency of an nnovatve frm s lkely to be hgher when dsclosure s mplemented. To test the H10 hypothess, a Wald test s conducted to test the joned hypotheses that all the nteracton terms between dsclosure and R&D partnershps are smultaneously non dfferent from zero. V. Results The estmates of model (1), shown n Table 3, confrm that on average, the R&D ntensty s postvely and sgnfcantly lnked to the probablty of transferrng knowledge or of allowng people to release t. The results are found for the four dfferent measures of knowledge dsclosure. The probablty of dsclosng knowledge s postvely assocated wth the sze of the frms. The estmated parameters suggest that nonlneartes occur: sze dummes are found to be postve and sgnfcant only for frms wth more than 250 employees (usng the LEAK, COMM and AUTHOR varables). The effect of sze seems weaker when PERMI s consdered. [Table 3 about here] The coeffcents of the sector dummes confrm that there are ndeed substantal dfferences between sectors. Manly hgh tech sectors are found to be more open to knowledge dsclosure. Textles, and wood and paper are not sgnfcantly more lkely to dsclose nformaton and knowledge than the clothng sector. In contrast, pharmaceutcal ndustres, electrc and electronc equpment, shps and aeronautcs as well as stone and glass are more systematcally open than frms from the clothng ndustres. The results do confrm that frms nsde hgh tech sectors (as electrc and electronc supples ndustry) are lkely to release nformaton or knowledge. 12 The result for the pharmaceutcal ndustry s much less pronounced than n Cockburn and Henderson s work (1996). In terms of low tech ndustres, as shown by Von Hppel (1987) and Schrader (1991), a sector such as basc metal and metal products seems to engage n some communcaton wth other frms, at least for the CIS1 sample. Compared to the avalable lterature based on sector studes, the results hghlght the mportance of knowledge dsclosure n car and aeronautcal ndustres. Lastly, to belong to a group s sgnfcantly lnked to the probablty of releasng knowledge. A Wald test ndcates that the two coeffcents are dfferent from 12 The French E3 set (electrc and electroncs) n the French NACE classfcaton ncludes computers (CITI 300) and telecommuncaton products (CITI 322) 19