OrganizationScienceVol. 18, No. 2, March–April 2007, pp. 165–180issn 1047-7039 �eissn 1526-5455 �07 �1802 �0165

informs ®

doi 10.1287/orsc.1060.0242©2007 INFORMS

Brokerage, Boundary Spanning, and Leadership inOpen Innovation Communities

Lee FlemingMorgan Hall 485, Harvard Business School, Boston, Massachusetts 02163, lfleming@hbs.edu

David M. WaguespackRobert H. Smith School of Business, 4515 Van Munching Hall, University of Maryland,

College Park, Maryland 20742-1815, dwaguesp@rhsmith.umd.edu

What types of human and social capital identify the emergence of leaders of open innovation communities? Consistentwith the norms of an engineering culture, we find that future leaders must first make strong technical contributions.

Beyond technical contributions, they must then integrate their communities in order to mobilize volunteers and avoid theever-present danger of forking and balkanization. This is enabled by two correlated but distinct social positions: socialbrokerage and boundary spanning between technological areas. An inherent lack of trust associated with brokerage positionscan be overcome through physical interaction. Boundary spanners do not suffer this handicap and are much more likelythan brokers to advance to leadership. The research separates the influence of human and social capital on promotion, andhighlights previously unexamined differences between brokerage- and boundary-spanning positions. Longitudinal analysesof careers within the Internet Engineering Task Force community from 1986–2002 support the arguments.

Key words : brokerage; leadership; open source; careers; social networks

Open innovation communities typically lack finan-cial or corporate backing, forgo personal ownershiprights to their members’ work, rely on volunteers, andeschew formal planning and management structures.Despite these apparent handicaps, they have dramati-cally changed our conceptions of how innovation canand should be managed and have prompted calls fornew theories of innovation (von Hippel and von Krogh2003). Open community development methods appearsuperior to proprietary efforts on some measures (Kogutand Metiu 2001, Mockus et al. 2002). Open sourceoperating systems challenge the world’s most powerfulsoftware firms, and proponents of community innova-tion are extending the model into new contexts suchas gene transfer technology (Broothaerts et al. 2005),medical innovation, crime solving, textbook and ency-clopedia publishing, education, space exploration (Goetz2003), and communities in developing countries, forwhich software customization in local languages remainscost prohibitive (Economist 2003). Open communitieshave spawned some of the most important technologi-cal breakthroughs of our era, including Web browsers,e-mail, and the Web itself. The very protocols thatenable different Internet technologies to work togetheremerged from innovation within a voluntary, nonpropri-etary, and open innovation community.Although most communities take full advantage of

electronic communication media, their members inno-vate—not anonymously and randomly in cyberspace,

but with reference to identity, reputation, technologi-cally derived status, collegial networks, and physicalinteraction (Raymond 2000, Lakhani and von Hippel2003, O’Mahony and Ferraro, forthcoming). Despitetheir bazaarlike, egalitarian, argumentative, unplanned,chaotic appearance, open innovation communities relyheavily on strong leadership to function effectively andto resist splintering, forking, and balkanization (DiBonaet al. 1999, Lerner and Tirole 2001, Kogut and Metiu2001, Lee and Cole 2003, von Hippel and von Krogh2003, Lakhani 2004). Reputation counts: although itremains informal, leadership must be constantly earnedthrough technical acumen and managerial skill. Theimplication is that in order to understand the success ofopen innovation communities, we must understand theemergence of their leaders (von Hippel and von Krogh2003).This understanding can also yield insights into how

human and social capital influence career mobility. Theimportance of social capital is illustrated by Burt (1992),who demonstrates that brokers, individuals who con-nect otherwise disconnected actors, can exploit struc-tural holes to advance more quickly in their careers. Anindividual who works with others who do not otherwiseinteract can control information and shape collegial andmanagerial perceptions. Predating Burt’s structural holetheory, Allen (1977) and Tushman (1977) illustrate awidespread correlation between ability, ties across mul-tiple organizations, and leadership. They describe how

165

Fleming and Waguespack: Brokerage, Boundary Spanning, and Leadership166 Organization Science 18(2), pp. 165–180, © 2007 INFORMS

boundary spanners usually contribute the best engineer-ing; identify, translate, and relay information within andacross engineering firms; and often assume managerialauthority. These two literatures, within sociological andmanagement of technology traditions, have developedwith surprisingly little mutual awareness or interaction,given the similarity of research questions. Furthermore,neither has estimated the relative importance of humanversus social capital, or controlled for an individual’sdesire for promotion.Although the measures of brokerage and boundary

spanning correlate empirically, the concepts remain the-oretically distinct. Brokers can span boundaries, butnot all boundary spanners broker. This unexamined dif-ference leaves a variety of unanswered questions. Forexample, what are the different mechanisms by whichthe occupants of the two positions attain leadership posi-tions? Burt’s early descriptions characterize brokers ascalculating and politically savvy operators, while Allenand Tushman characterize boundary spanners as wellrespected guardians who redirect crucial information,both within and outside the firm. If these characteriza-tions hold any validity, then colleagues of the broker andboundary spanner will surely hold different perceptionsabout individuals in each position. Colleagues will beless likely to trust a broker (Coleman 1988, Burt 2001b),and this lack of trust will be exacerbated within openinnovation communities, which are inherently wary ofbalkanization and cooptation by commercial interests.The most effective strategies and behaviors of aspiringleaders will therefore be different for brokerage- and forboundary-spanning roles.More generally, open innovation communities provide

an opportunity to develop theories of human and socialcapital in a novel context that lacks pecuniary incen-tives, hierarchical authority, and formal structure. Lead-ership in such communities depends more on the trustand mobilization of peers than on approval of superiors.To wit, members cannot be fired or forced to partici-pate in any activity, nor can they be compelled to payattention to any other member. Ascendancy in such rela-tionships relies purely, to borrow a phrase from politics,on “the power to persuade” (Neustadt 1990, p. 11). Inaddition to providing field settings in which to researcha novel social phenomenon, open innovation communi-ties often electronically archive their interactions. Thesepublic records afford social scientists an unprecedentedopportunity to construct longitudinal databases of humanand social capital, social and political processes, and ahost of important outcomes, including the emergence ofleadership.To explain the emergence of leadership within open

innovation communities, we induct theory from inter-views, archival research, and field observations at com-munity events. Estimating rate models for appointment

as a working group leader in the Internet Engineer-ing Task Force (IETF), arguably the world’s first openinnovation community (Bradner 1999), we first demon-strate the importance of technical contribution for futureleaders. Individuals who broker work collaborationsare more likely to assume leadership, but the effectis strongly contingent on physical presence within thecommunity, a consequence of the diminished trust inher-ent in brokered social contexts. Consistent with the argu-ment that they must overcome lack of trust, brokers alsoencounter difficulties when they attempt to span tech-nological boundaries within the community. Boundaryspanners, in contrast, do not suffer from a lack of trustand are more likely than brokers to assume leadershippositions. In summary, future leaders are most likely tobe individuals that make a strong technical contributionfrom a structural position that can bind the communitytogether.

Open Innovation Communities andthe IETFThe institution of science might be considered the firstopen innovation community (Dalle and David 2003),but its technological equivalent began in academic com-puter departments in the 1960s (DiBona et al. 1999).Consistent with commonly espoused norms of science(Merton 1973), department members made their soft-ware readily available to others, forgoing financial com-pensation, and, in return, earned reputation and status.The advent of the Internet and World Wide Web enlargedthe potential scale of these efforts, giving rise to suchextensive communities as Linux, Perl, Apache, Debian,and the IETF. We define an open innovation communityas a group of unpaid volunteers who work informally,attempt to keep their processes of innovation public andavailable to any qualified contributor, and seek to dis-tribute their work at no charge. While we induct the-ory from many open innovation communities, we testour hypotheses with a specific data set culled from thearchives of the IETF, with the dependent variable as timeto appointment as a working group leader within theIETF. We first briefly describe the IETF and process ofappointment.Although strictly speaking the IETF is an open stan-

dards rather than an open source community, it neverthe-less exhibits critical open innovation community featuresin that any individual can volunteer to participate, pro-ceedings remain transparent, and all technology origi-nated therein is made available for free (Bradner 1999).The IETF is also the most long lived of the well-knownopen innovation communities and arguably exerts thegreatest social and economic impact because of its asso-ciation with the Internet. Of its potential as a modelfor open innovation communities Bradner (1999, pp. 47and 52) reports that

Fleming and Waguespack: Brokerage, Boundary Spanning, and LeadershipOrganization Science 18(2), pp. 165–180, © 2007 INFORMS 167

IETF standards are developed in an open, all-inclusiveprocess in which any interested individual can partici-pate. All IETF documents are freely available over theInternet and can be reproduced at will. In fact the IETF’sopen document process is a case study in the potentialof the Open Source movement � � � � The IETF supportedthe concept of open sources long before the Open Sourcemovement was formed.

The IETF emerged in 1986 from an amalgam of ad hocDefense Advanced Research Projects Administration(DARPA) committees and has no official mandate togovern Internet technology. Challenged by more tra-ditional standards-developing organizations and evengovernment bodies, it has nonetheless emerged as thede facto standards-developing organization for the Inter-net (Abbate 1999, Mowery and Simcoe 2002; see alsoHarris 2001; Bradner 1996, 1998; and Hoffman andBradner 2002 for insiders’ descriptions). The IETFdevelops and maintains the core Internet standard,TCP/IP (Transmission Control Protocol/Internet Proto-col), as well as many other standards that are pervasivein modern computing and networking but largely invisi-ble to the average user. Although much of the communi-cation and work of the IETF is conducted via electronicmedia, members meet three times a year, carrying for-ward a tradition begun in January 1986 when 21 IET-Fers met for the first time in San Diego. Membershipremains open to all comers and has continued to grow,with as many as 2,800 individuals attending meetingsand thousands more interacting online. Members par-ticipate in the IETF at least nominally as individuals(Bradner 1999), although they typically work for firms,universities, or governments. There being no dues ormembership lists, in principle any person can join theIETF.The IETF accomplishes most of its work within aptly

named working groups (WGs) organized under largerfunctional areas. Individuals can freely associate, vir-tually or physically, with any of the extant technicalworking groups. Groups and their leaders seek and, ifsuccessful, are granted charters to address specific tech-nical problems within a delimited time and domain.The Secure Shell Group, for example, was charteredto update and standardize the popular SSH protocol(Secure Shell secsh 2003). Working groups have chairsas well as individuals or design teams charged to pro-duce documents that detail proposed standards. Areadirectors (ADs) appoint working group chairs. A nomi-nating committee (NOMCON), randomly selected fromcommunity members who have attended at least twoIETF meetings in the past two years, appoints ADs (gen-erally two for each area).Within the period of study, 344 working groups exist

for varying periods of time, and some 480 individu-als are first appointed to chair one of those groups.Approximately 32% of first-time working group chairs

are appointed to lead existing working groups, eitherreplacing an incumbent chair or working alongside theincumbents in a cochair capacity. The other 68% of first-time appointments are for new working groups. Newworking groups emerge mainly from grassroots interestin a topical area and are typically preceded by a Birds ofa Feather (BOF) meeting convened during a conference.1

For example, the IETF has a series of working groupsrelated to different aspects of the domain-naming sys-tem (DNS). The first stage in group formation entailssoliciting BOF participation via electronic invitationsdisseminated to the IETF mailing list. Organizers whogenerate favorable sign-up response receive physicalmeeting space at the next IETF conference. If a meet-ing is well attended and elicits broad interest, an ADwill charter a group and appoint a chair. The chair isgenerally, but not always, the BOF organizer. Appoint-ment can thus be construed as confirmation of initiallysuccessful leadership and a vote of confidence by com-munity members and current leaders in the individual’sleadership potential.Individuals aspire to working group leadership in the

IETF for a number of reasons: First, it enables some-one to influence the direction of the technology devel-opment. This can happen if an individual inventor seekslegitimacy for her personal inventions (Bradner 2006).Second, it provides an opportunity to gain manage-ment experience without promotion within a privatefirm. Third, it gives an individual visibility in a commu-nity that can provide professional opportunity and careermobility. Finally, firms often support their employees’aspirations for leadership because the firm gains esteemin the larger technological and business community.Such esteem is particularly important for a startup thatwishes to undergo a liquidity event.

Human Capital, Social Capital, andLeadership in Open InnovationCommunitiesOpen innovation communities, like most engineeringcultures, highly value technical contributions, eschew-ing titles and even democracy in favor of proven tech-nology (Wasserman 2003). Despite a strong aversion tonontechnical sources of prestige and authority, commu-nity members readily recognize technical contribution.Successful contributors, according to Raymond (1998,p. ), gain “good reputation among one’s peers, atten-tion and cooperation from others � � � and higher status inthe � � � exchange economy.” Rivlin (2003) illustrates howLinus Torvalds (the original author of LINUX) realizesthat his authority is technically derived, tenuous, andconstantly in need of collective reaffirmation:

His hold over Linux is based more on loyalty than legal-ities. He owns the rights to the name and nothing else.Theoretically, someone could appropriate every last line

Fleming and Waguespack: Brokerage, Boundary Spanning, and Leadership168 Organization Science 18(2), pp. 165–180, © 2007 INFORMS

of his OS [operating system] and rename it Sally. “Ican’t afford to make too many stupid mistakes,” Torvaldssays, “because then people watching will say, hey, maybewe can find someone better. I don’t have any authorityover Linux other than this notion that I know what I’mdoing.” He jokingly refers to himself as “Linux’s hoodornament,” and he’s anything but an autocrat. His poweris based on nothing more than the collective respect ofhis cohorts.

Status accrues to past contributors and translates into ahigher probability of future leadership (Lee and Cole2003). As Hamm describes (2005), “In a world whereeverybody can look at every bit of code that is submitted,only the A+ stuff gets in and only the best program-mers rise to become Torvalds’s top aides.” Individualswho can solve difficult problems gain reputation andesteem in the opinions of their colleagues. Such indi-viduals will become leaders whose opinions are soughtout, respected, and that will influence the community’sfuture.

Hypothesis 1 (H1). Technical contribution will in-crease a member’s likelihood of becoming an open inno-vation community leader.

While technical contribution remains the primaryprerequisite for aspiring leaders, the most importantrole—and challenge—for an open innovation commu-nity leader is to integrate and bind the communitytogether. This occurs because open innovation commu-nities remain voluntary; members can always leave andmobilize a new effort, a process known as forking.(DiBona et al. 1999, Kogut and Metiu 2001, Lerner andTirole 2001). Of Torvalds, Rivlin (2003, p. 157) writes,“More than anything he seeks to avoid taking sides in away that might splinter his followers. ‘I’d much ratherhave 15 people arguing about something than 15 peoplesplitting into two camps, each side convinced it’s rightand not talking to the other,’ he says.”Leaders can forestall this process by creating and

occupying network positions, such as social broker-age and boundary spanning, that integrate and bindthe community together (Perrone et al. 2003). Burt(1992, 1997, 2000, 2001a) defines a broker as the onlysocial connection or bridge among a group of actorsand argues that brokering enhances both career mobilityand promotion. In the current empirical context, bro-kering occurs when coauthors do not collaborate onanother project in the absence of the focal engineer. Thefocal actor who brokers among colleagues thus occupiesand exploits a structural hole. Contrast this with con-straint, which occurs when an actor’s alters know eachother well, and when there exists redundant, dense, andcohesive interaction among the actor’s contacts (Cole-man 1988). Burt’s classic information-control argumentsfor the strategic superiority of brokering are that bro-kers (1) gain first access to information and control

of its diffusion, (2) can present different strategies todifferent groups (because unconnected observers willlack the opportunity to compare the strategies; see alsoPadgett and Ansell 1993), and (3) will be considered foran expanded set of opportunities because they will beknown to a wider set of groups.If open innovation communities were truly open there

would be no opportunity to control information, and thecontext would remain outside the boundary conditions ofBurt’s arguments. Despite the strong functional and nor-mative pressures for openness, it remains impossible forall information within open communities to be shared.Intense working relationships such as collaboration ontechnical publications will involve dyadic or small-groupcommunication that will not be widely shared, nor arepersonal e-mails, or telephone and face-to-face conver-sations likely to be shared. This limitation is recognizedin the working group guidelines for the IETF: “It isoften useful, and perhaps inevitable, for a sub-group ofa working group to develop a proposal to solve a partic-ular problem. Such a sub-group is called a design team.In order for a design team to remain small and agile, it isacceptable to have closed membership and private meet-ings” (Bradner 1998, p. 18; Bradner 2004). For thesereasons, open innovation communities remain within theboundary conditions of Burt’s arguments.Despite the advantages of brokering, the role can also

occasion disadvantages for aspiring leaders. Researchershave long argued that cohesive social structures buildtrust by exerting pressure for consistent norms andreciprocity among individuals within embedded andoverlapping relationships (Granovetter 1992). Cohe-sive networks also increase the likelihood of sanctionsagainst individuals who violate their norms (Coleman1988) and facilitate communication of reputation effects(Reagans and McEvily 2003). Particularly important inthe current context, the potential for information con-trol and political action inherent in brokerage structuresdoes not sit well with cultures obsessed with the open,immediate, transparent flow of information. ConsiderRaymond’s (1998, p. 21) appeal to open innovation com-munities to “be open to the point of promiscuity � � � �When writing gateway software of any kind,” he urges,“take pains to disturb the data stream as little aspossible—and never throw away information unless therecipient forces you to” (p. 44, italics added)! Com-munity members understand and fear the potential forabuse and proclaim strong norms against manipulativebehavior. Raymond (1998, p. 86) again: “Surreptitiouslyfiling someone’s name off a project is, in a culturalcontext, one of the ultimate crimes.” The power of abroker who controls the flow of information to influencethe attribution of credit can undermine the fundamentalmechanics of open innovation communities inasmuch asattribution is a driving incentive and the basis of sta-tus. Most importantly, brokers must consequently allay

Fleming and Waguespack: Brokerage, Boundary Spanning, and LeadershipOrganization Science 18(2), pp. 165–180, © 2007 INFORMS 169

concerns that they might violate norms against hidingand manipulating information. They must also contendwith the inherent and previously recognized challengesof mobilizing across networks (von Krogh et al. 2003).As Coleman (1988) argues and Gould (1991) describeswith reference to the Parisian revolts of 1871, mobiliza-tion will be easier if leaders can call on dense, cohesivenetworks, particularly if they are themselves embeddedin such networks. Brokers face the added difficulty ofbalancing conflicting demands of simultaneous member-ship in multiple groups that might have differing roleexpectations (Podolny and Baron 1997).Because brokerage exerts both positive and negative

influences on leadership, the overall effect of the roleremains an empirical question. We can still test ourmodel of who becomes a leader, however, by developingother potentially observable implications. In particular,if another variable moderates the advantages or disad-vantages of brokerage, we should observe an interactioneffect. For example, if followers’ trust can be gainedthrough assurances that a potential leader will not abusea brokerage position, we should observe a contingentand positive correlation between brokerage and assump-tion of a leadership position. Burt (2001b, p. 2) recog-nizes the problem in contexts beyond open innovationcommunities: “The social capital of brokerage dependson trust—since the value created by brokers by defini-tion involves new, and so incompletely understood, com-binations of previously disconnected ideas—but trust isoften argued to require network closure, precisely thecondition that brokers rise above.”Lerner and Tirole (2001, p. 222) attest to the fun-

damental importance of trust within open innovationcommunities. “The key to a successful leadership,” theyobserve, “is the programmers’ trust in the leadership:that is, they must believe that the leader’s objectivesare sufficiently congruent with theirs and not pollutedby ego-driven commercial or political biases.” Brennanreports that the Debian community developed a webof trust specifically to repel commercial Trojan horses:“The more deep and tightly interlinked the web of trustis, the more difficult it is to defeat” (Brennan 2003,cited in O’Mahony and Ferraro, forthcoming).2 Davies(2003, p. 15) cites an official description of the IETFthat explicitly claims that trust remains intransitive as anetwork relationship:

[IETF processes] are all reliant on personal knowledge ofthe capabilities of other individuals and an understand-ing built on experience of what they can be expected todeliver, given that there are almost no sanctions that canbe applied beyond not asking them to do a similar taskagain � � � � In essence, the IETF is built on a particularkind of one-to-one personal trust relationship. This is avery powerful model but it does not scale well becausethis trust is not transitive. Just because you trust oneperson, it does not mean that you trust (i.e., know the

capabilities of and can rely on) all the people that persontrusts in turn. (italics added)

If absence of trust causes potential followers to doubta leader’s motives, opportunities to observe the leadershould ameliorate that doubt. We suggest that physicalattendance and greater interaction with the communityconstitute such opportunities. Burt (2001b) proposesthat, given sufficient face time and repeated interaction,brokers can overcome potential distrust of the role andturn the position to their advantage. However much elec-tronic interaction predominates in modern open inno-vation communities, trust is still greatly facilitated bypersonal contact. Admission to the Debian community’sweb of trust, for example, involves personal and physicalkey signings, complete with government-issued identifi-cation and a handshake (O’Mahony and Ferraro, forth-coming). A short paper on how open source conven-tions should be run advises planners to shape the socialspace and maximize social interactions through physi-cal layout, after-hours meeting places, message boards,and copious refreshments (Raymond 2000). Face timein the community will facilitate members’ assessmentsof whether an aspiring leader might harbor blatantlypolitical motives or be inclined to abuse the brokerageposition. For these reasons, we argue that trust devel-oped through physical interaction will increase the like-lihood that a broker will advance into leadership. Withlow physical presence, increasing brokerage decreasesthe likelihood of becoming a leader. With high physicalpresence, increasing brokerage increases the likelihoodof becoming a leader.

Hypothesis 2 (H2). The interaction of greater phys-ical presence and social brokerage in working relation-ships will increase a member’s likelihood of becomingan open innovation community leader.

In addition to the integrating role of social broker,individuals can also work across internal communityboundaries and perform the integrating role of bound-ary spanning. The importance of boundary spanners totechnical organizations and the process of innovationwere emphasized in early research on the structure ofengineering firms (Lorsch 1965, Allen 1977, Tushman1977). Individuals who occupy these positions tend tohold advanced technical degrees, make the most impor-tant technical contributions, earn the respect of their col-leagues, communicate with peers in other organizations,and, most important for the current study, advance intomanagement and technical leadership positions. Bound-ary spanners stimulate the innovation process becauseformal organizational boundaries correlate with technicalboundaries (Henderson and Clark 1990). These bound-aries can become barriers to the flow of information dueto the evolution of local dialects (Dougherty 1992) andthe difficulty of transferring complex information acrosssocial boundaries (Sorenson et al. 2006). Technologies

Fleming and Waguespack: Brokerage, Boundary Spanning, and Leadership170 Organization Science 18(2), pp. 165–180, © 2007 INFORMS

can be refined more productively within technologicallyfocused efforts, but at the risk of becoming incremen-tal and obsolete. Boundary spanners reduce this risk ofobsolescence by gathering, interpreting, and disseminat-ing nonredundant information across boundaries (Allen1977, Tushman 1977).The boundaries within open innovation communities,

like those in private firms, usually correspond to theinterfaces between technological subsystems (Hendersonand Clark 1990). Each boundary demarcates a dis-tinct technological area or module. The boundaries aredefined directly by leaders’ architectural decisions andthen implemented by followers’ choices of where tovolunteer their efforts. Torvalds, for example, delegatesresponsibility for different modules to a handful oftrusted lieutenants (Hamm 2005), each of whom reviewssubmitted code, accepts the best, and works with Tor-valds and the other lieutenants to resolve issues thatcross technical boundaries. Bradner (1998, p. 5; 2003b)describes how individuals who aspire to leadership in theIETF must define their technological boundaries relativeto other community efforts as part of their BOF propos-als: “Is there a good understanding of any existing workthat is relevant to the topics that the proposed workinggroup is to pursue � � � and, if so, is adequate liaison inplace?” Although technical and social boundaries corre-late, brokerage and boundary spanning remain distinctroles. Nothing prevents a boundary spanner from beingthe only engineer, in which case she would be a socialbroker. Alternately, she could be one of the many engi-neers who collaborate across a boundary, in which casethe boundary spanner is not a social broker at all.Open innovation communities are well aware of these

issues and value individuals that can span technicalboundaries (Bradner 2004). Boundary spanners shouldbe more creative and able to call on more-diverseresources, as explicitly recognized in the expectationthat leaders promote “cross-pollinating between groups”(Davies 2003, p. 13). In addition to providing the inte-grating advantages of social brokerage, boundary span-ners are in an even stronger position to control the poten-tial for community forking, which generally occurs whenalternative technical solutions attract dedicated coalitionsthat refuse to compromise or resolve incompatibilities.Davies and Hoffman (2004, p. 6) recount recent discus-sions of IETF problems that emphasize the difficulty oftechnological interdependence:

The IETF has effective mechanisms for dealing withwell-defined problems of limited scope. These problemsare well handled in IETF Working Groups, where expertsin a given technology can convene and solve the prob-lems specific to one technology area. However, we aremuch less effective at resolving complex problems thataffect more than one IETF area.

Boundary spanners, being more aware of other effortsacross the larger community, can better negotiate the

boundaries of their own groups’ efforts. Individuals whospan boundaries before attempting to lead are better pre-pared, for a variety of reasons: simple technical aware-ness of interdependent technologies and their inventors;expanded personal creativity; and greater likelihood ofhaving observed or directly participated in conflict res-olution. Potential followers, particularly the most com-petent, will be aware of technical interdependencies andvalue individuals who have worked across their commu-nities’ internal boundaries. Individuals with experienceworking across internal community boundaries are, forthese reasons, more likely to become leaders.

Hypothesis 3 (H3). Internal community boundaryspanning will increase a member’s likelihood of becom-ing an open innovation community leader.

Despite their correlation, social brokerage and bound-ary spanning remain distinct social positions. For exam-ple, a boundary spanner might be the only collaborativelinkage between groups, and thus also be a collaborativebroker; alternately, she could be but one of a numberof linkages across a socially cohesive boundaries. Thesedistinctions provide an additional opportunity to test ourtheory in parallel with Hypothesis 2. Our developmentof Hypothesis 2 held that if a social broker must over-come concerns about trust before becoming a leader,actions and positions that increase trust will exhibit apositive interaction with the social brokerage position.We now propose that the simultaneous pursuit of socialbrokerage and boundary spanning will exhibit a negativeinteraction, because spanning a boundary will diminishpotential followers’ trust.Consider first an individual who spans multiple bound-

aries and simultaneously brokers collaborative rela-tionships. This individual’s collaborators will be lessfamiliar with one another, not having worked togetherpreviously, and, if they contribute in different techni-cal areas, will be less aware of and comfortable withthe technologies, jargon, objectives, and reputations ofcollaborators in the nonlocal areas. In such a situation,concerns about trust will be magnified if an individualinvests serious resources in another group’s efforts, asargued by Podolny and Baron (1997, p. 676): “Eachconstituency grows increasingly suspicious that its needsare receiving less attention from the boundary spannerthan someone else’s needs.” Suspicion will be greater,to the extent that a constituency is locally cohesive andinsulated from other constituencies.Now consider the opposite extreme: an individual who

spans multiple boundaries and simultaneously workswithin cohesive collaborative relationships. This indi-vidual’s collaborators will be quite familiar with oneanother, owing to prior working relationships. Further-more, these cohesive relationships by definition wouldspan technological boundaries within the community.

Fleming and Waguespack: Brokerage, Boundary Spanning, and LeadershipOrganization Science 18(2), pp. 165–180, © 2007 INFORMS 171

Figure 1 Illustration of IETF Ego Publication Network AroundNodes A10267 and H44, December 1999

Notes. The green node, H44, transitions to a first working groupchair leadership appointment in this period. The red nodes rep-resent individuals who do not transition but are at risk of workinggroup chair appointment. The blue nodes represent existing work-ing group chairs and are not in the first appointment risk set (theyare included in network position calculations). The ellipses illustrateworking group memberships. Both H44 and A10267 span workinggroup boundaries, but H44 is in a relatively cohesive ego network.A10267, by contrast, is in a relatively strong brokering position withmultiple structural holes among alters. Consistent with our hazardmodels, A10267 does not advance to leadership in subsequentperiods.

An individual will thus be observed by mutual acquain-tances on both sides and be trusted by both sidesto resolve technological and organizational boundaryconflicts. Such an individual is more likely to satisfyconflicting goals and expectations among collaboratingorganizations. Cohesive boundaries also signify activetechnical areas of innovative search and attempts to re-solve technological interdependencies. Taken together,these arguments imply that leaders will emerge fromamong those who make public contributions at theboundaries of socially cohesive and strongly interactingtechnical organizations within a community.

Hypothesis 4 (H4). The interaction of greater socialbrokerage and internal community boundary spanningwill decrease a member’s likelihood of becoming anopen innovation community leader.

Figure 1 illustrates the hypotheses with two ego net-works taken from December of 1999.

EvidenceWe draw data primarily from the published proceed-ings of the IETF’s triannual meetings from its incep-tion in 1986 until 2002, and from the organization’sRFC (request for comments) publication series over thesame time period.3 Each proceeding represents a snap-shot of the IETF at a given time, complete with a list ofregistered conference attendees (including their e-mail

affiliations) and archived charters for each of the activeworking groups indicating current chairs and area direc-tors. It can thus be determined from the proceedingswhen individuals first attended an IETF conference andif and when they began to lead a working group. RFCpublications indicate, within a designated time frame,which engineers are publishing technical documents andwhether—and with which other engineers and estab-lished working groups—they are collaborating. We sup-plemented these archival data with extensive interviewsof community leaders (Bradner 2003a, b, 2004; Kauf-man 2003) at their places of employment and at the 56thIETF meeting in San Francisco, March 16–21, 2003.Exactly 15,465 individuals attend the IETF meetings

over the course of the study and are observed at four-month intervals. Because the first three years’ of con-ference data were truncated due to leading independentvariable creation, the analysis data span 1989–2002.Subjects enter the risk set on first attending an IETFmeeting or in the first trimester of 1989 if communityinvolvement occurred before 1989. Subjects exit on thefirst working group chair appointment or are treated asright censored.4 For variables relating to the overall stateof the community, such as the number of incumbentworking group chairs at any point in time, we utilizethe full database. Because coauthorship patterns formthe basis of the longitudinal social networks we con-struct, and social brokerage can only occur if an individ-ual writes two or more publications, we restrict analysisof leadership appointment to the 610 individuals whopublish two or more documents prior to working groupchair promotion.5 The logic behind this restriction is thatobservable variation on brokerage or boundary spanningcan only occur when an actor has two or more oppor-tunities on which to interact with other individuals (bydefinition, all coauthors on a single draft are cohesively

Table 1 Descriptive Statistics for Network Risk Set �N = 5�066�

Variable Mean Std. dev. Min Max

Incumbent WG chairs 153�77 51�38 16�00 230�00University affiliation 0�13 0�33 0�00 1�00Government affiliation 0�04 0�19 0�00 1�00Affiliated WG chairs 19�27 35�96 0�00 192�00Executive tie 0�57 0�77 0�00 5�00U.S. patents 0�37 1�38 0�00 17�00Degree 6�97 6�19 0�00 39�00WG contributor 0�82 0�38 0�00 1�00�ln� IETF publications 1�34 0�33 1�10 3�09Presence �uncorrected � 4�53 2�87 0�00 9�00Miles 9�75 0�53 0�00 11�13Social brokeragea 0�00 0�27 −0�66 0�46Spans 1 boundaryb 0�23 0�42 0�00 1�00Spans 2+ boundariesb 0�05 0�22 0�00 1�00

Notes. WG=working group.aMean centered to facilitate interpretation of interaction effects

(Friedrich 1982).bThe omitted reference category is Zero WGs.

Fleming and Waguespack: Brokerage, Boundary Spanning, and Leadership172 Organization Science 18(2), pp. 165–180, © 2007 INFORMS

Table 2 Bivariate Correlation Matrix for Network Risk Set (N = 5�066 Observations)

1 2 3 4 5 6 7 8 9 10 11 12 13 14

1 Incumbent WG chairs 1�002 University affiliation −0�24 1�003 Government affiliation −0�10 −0�08 1�004 Affiliated WG chairs 0�27 −0�12 −0�09 1.005 Executive tie 0�17 −0�05 0�01 0.16 1.006 U.S. patents 0�05 −0�07 −0�03 0.06 0.07 1�007 Degree 0�41 −0�11 −0�02 0.25 0.47 0�08 1.008 WG contributor 0�26 −0�08 0�02 0.15 0.09 −0�03 0.31 1.009 �ln� IETF publications −0�02 0�04 −0�01 0.06 0.23 0�03 0.27 0.06 1.00

10 Presence �uncorrected ) 0�16 −0�04 0�06 0.18 0.05 0�02 0.19 0.24 0.21 1.0011 Miles 0�43 −0�05 −0�08 0.17 0.09 0�01 0.21 0.18 0.07 0.34 1.0012 Social broker (-constraint) 0�27 −0�01 −0�06 0.15 0.23 0�08 0.56 0.19 0.16 0.14 0.17 1.0013 Spans 1 boundary 0�13 0�03 −0�01 0.04 0.06 0�02 0.18 0.26 0.14 0.18 0.11 0.21 1�0014 Spans 2+ boundaries 0�10 0�02 0�03 0.06 0.17 0�00 0.32 0.11 0.39 0.16 0.10 0.20 −0�13 1.00

Note. WG=working group.

tied). We discuss solutions to the biased-sample problemin the variables and results sections.

VariablesTables 1 and 2 describe the 610 subjects who publishedprior to working group chair appointment or censoring,observed over 5,066 trimesters. The dependent variableLeader is a categorical indicator that a subject appears asa working group leader at the end of the current obser-vation period. (ln) IETF publications, which counts totalIETF technical publications authored or coauthored inthe prior three years, is our main measure of individ-ual human capital. These documents also reveal work-ing group affiliation. Spans 1 boundary indicates that anengineer published within two working groups; Spans2+ boundaries indicates that a subject published in threeor more working groups. Although use of a count vari-able for the number of working group membershipsinstead of breakout into mutually exclusive categoriesreturned similar estimates, we prefer the indicators dueto the qualitative differences between no membership,membership in one group, and boundary spanning.We created social network measures based on coau-

thorship by constructing 51 social networks, one for eachtrimester with a three-year window on prior linkages,with valued symmetrical ties between all IETF authors.We calculated the following network position variablesin UCINET version 6 (Borgatti et al. 2001). SocialBrokerage measures the opposite of individual con-straint, where constraint measures the extent to whichego has ties with few alters and those alters have tiesbetween them (Burt 1992). We enter the negative ofconstraint (Equation (1)) as our measure of brokering.In Equation (1), pij is the proportion of i’s relationsdirectly invested in contact j .

∑q piqpqj is the portion of

i’s direct connections invested in contacts q who are inturn invested in contact j . The sum in the parentheses isthe proportion of i’s direct and indirect connections that

are invested in contact j . The sum of squared propor-tions over all contacts j is a measure of individual i’snetwork constraint.

Constraintactor i =∑j

cij � cij =(pij +

∑q

piqpqj

)2�

q �= i� j� (1)

Presence measures the extent of an individual’s phys-ical involvement with the IETF community over theprior three years. A physical presence measure that sim-ply counted a subject’s conference attendance poses aserious endogeneity problem, in that we cannot directlyobserve an individual’s aspirations for leadership fromarchival records. We also cannot observe an employer’sstrategy to coopt the community by influencing the lead-ership processes, and these unobserved aspirations andstrategies would very likely influence conference atten-dance. Endogeneity violates the standard assumption ofno correlation between independent variables and theerror term (Greene 1993, p. 285). Unobserved ambitionalso specifically poses a problem for testing the hypoth-esis that the effect of network position interacts withpresence. To address this problem we employ an instru-mental variable approach (Greene 1993, p. 603). We firstdeveloped a proxy variable that correlates with atten-dance but not with the probability of becoming a leader.Miles accounts for total distance, in miles, between anengineer’s home and all IETF meetings during the priorthree years.6 The instrument implicitly argues that atten-dance will be correlated with both the cost of atten-dance and an individual’s desire to lead. Given that thelocation of each conference is probably set without con-sideration of which members desire to lead, however(and indeed, the conference is moved around intention-ally, both within North America and abroad, to facilitateattendance by a wider variety of members; see Bradner

Fleming and Waguespack: Brokerage, Boundary Spanning, and LeadershipOrganization Science 18(2), pp. 165–180, © 2007 INFORMS 173

2003b), it should be exogenous to the outcome vari-able of leadership. To create the instrument, we assignedeach subject to a latitude and longitude based on thephone number reported in the conference registrationlist, then calculated (using the STATA sphdist ado sub-routine) spherical distance between the subject and eachmeeting’s latitude and longitude (Sorenson and Stuart2001). We then regressed three-year conference atten-dance on Miles and all other independent variables, andrecovered predicted values for Presence. For identifica-tion purposes, Miles was not included when estimatingappointment to leadership. In the first-stage estimation,the predicted variable is purified of the endogenous ele-ment that is correlated with the disturbance term. TheR2 for the regression of these variables on attendance(depending on the final specification) ranged from 22%to 24%. Given the high amount of explained variance,the instrument should not be vulnerable to finite sampleand weak instrument bias (Bound et al. 1995).The variable Nonselection hazard addresses the non-

random sample problem. Because coauthors work, bydefinition, within a cohesive relationship, and becausebrokerage requires at least two different coauthors, wecalculate each of the boundary-spanning and social net-work measures only for cases in which a subject has twoor more publications. Restricting analysis to such a sub-set produces a nonrandom sample and possibly biasesresults by neglecting the factors that led to inclusion inthe sample in the first place. A standard solution to selec-tion bias is to first estimate a selection parameter on thefull data, and then include this parameter in models thatuse the restricted sample (Heckman 1976). Although notas well developed in the hazard model context, prelim-inary work (Boehmke et al. 2006) indicates that sam-ple selection can also bias hazard model estimates. Toaddress this problem, following Heckman we estimatethe probability that any of our 15,465 at-risk engineerspublishes two or more IETF documents. All independentand control variables except those related to boundaryspanning and social network position are included.Finally, the models include a series of control vari-

ables. Entry cohort is a set of 15 dummy variables basedon the year of entry into the IETF community. Incum-bent WG chairs is a count of incumbents (those not atrisk of a first leadership position) holding chair positionsand is intended to gauge the extent to which the com-munity currently fulfills its leadership needs. Universityaffiliation and Government affiliation are dummy vari-ables for employer type. Affiliated WG chairs counts thenumber of extant WG chairs from the same employer,indicating employer commitment to IETF participationor attempts to influence the direction of community evo-lution (Wade 1995). Executive tie measures the numberof collaborative working relationships with a current ADwithin the past three years to control for reliance on per-sonal ties with individuals who have control over the

appointment process. U.S. patents obtained by a sub-ject during the prior three years identifies individualswith an external reputation and propensity to publish.7

Degree measures, on the basis of all drafts published inthe previous three years, the number of individuals withwhom an engineer has published a draft. Finally, theindicator variable WG contributor records subjects thathave participated in an existing working group withinthe prior three years. Such individuals are more likely tolead because they have already demonstrated affinity forsocial interaction and gained greater exposure to organi-zational processes, leadership models, and opportunitiesfor informal leadership.

ModelsWe estimate hazard models of the duration T from firstattendance at an IETF conference until appointment as aworking group leader. Rate models are more appropriatethan choice models because members do not competefor a limited number of positions. Instead, they cre-ate positions based on their understanding of technicalissues and opportunities that confront the Internet com-munity. Rate models also accommodate censored obser-vations. Equation (2) defines the instantaneous hazard ofappointment as a leader. T represents the time between amember’s first attendance at an IETF meeting and (pos-sible) appointment as a working group leader; r(t) isthe instantaneous hazard of making the transition fromindividual member to appointment as a working groupleader. The data change at trimester frequency, basicallyevery four months, to correspond to the frequency of theIETF conferences.

r�t = limt′→t

Pr�t ≤ T < t′ � T ≥ t

t′ → t� (2)

We estimated semiparametric Cox models (Equa-tion (3)) to avoid making parametric assumptions aboutthe form of duration dependence in the underlying haz-ard rate (Cox 1972). Given the relatively few events, weused the Breslow method for handling ties. The model’shazard rate is the product of an unspecified baselinerate, h�t , and an exponential term that includes covari-ates X. The Cox model assumes, however, that the pro-portional hazards remain constant over time. We testedthis assumption both graphically and with a Schoenfeld(1982) test. Neither inspection nor the statistical testindicated a significant relationship between time andmodel residuals (p value < 0�23). Piecewise exponen-tial models returned very similar substantive results. Weestimated all models with STATA version 8.

r�t = h�t e��X � (3)

Fleming and Waguespack: Brokerage, Boundary Spanning, and Leadership174 Organization Science 18(2), pp. 165–180, © 2007 INFORMS

Tab

le3

Cox

Haz

ard

Mo

del

so

fA

pp

oin

tmen

tas

Wo

rkin

gG

rou

pC

hai

rA

mo

ng

IET

FP

arti

cip

ants

,19

89–2

002

(10)

(7)

Impu

ted

(1)

(2)

(3)

(4)

(5)

(6)

w/o

utIV

(8)

(9)

netw

ork

valu

es

Entry

coho

rts

(15,

)(1

5,)

(15,

)(1

5,)

(15,

)(1

5,)

�15�+)

(15,

)�1

5�+)

�15�+)

IncumbentWGchairs

−0�0

39−0

�039

−0�0

37−0

�041

−0�0

42−0

�040

−0�0

43−0

�041

−0�0

39−0

�019

�0�0

10�∗

∗�0�0

09�∗

∗�0�0

09�∗

∗�0�0

09�∗

∗�0�0

09�∗

∗�0�0

09�∗

∗�0�0

09�∗

∗�0�0

09�∗

∗�0�0

09�∗

∗�0�0

03�∗

∗

University

affiliation

−0�1

16−0

�158

−0�1

27−0

�241

−0�2

40−0

�161

−0�0

99−0

�160

−0�1

45−0

�017

�0�2

79�

�0�2

75�

�0�2

77�

�0�2

76�

�0�2

80�

�0�2

83�

�0�2

82�

�0�2

83�

�0�2

84�

�0�1

33�

Governm

entaffiliation

−0�2

93−0

�265

−0�3

09−0

�325

−0�3

27−0

�415

−0�6

83−0

�434

−0�3

03−0

�319

�0�6

06�

�0�5

42�

�0�5

59�

�0�5

41�

�0�5

42�

�0�5

60�

�0�5

31�

�0�5

61�

�0�5

67�

�0�2

41�

AffiliatedWGchairs

−0�0

02−0

�002

−0�0

02−0

�001

−0�0

01−0

�002

−0�0

04−0

�002

−0�0

050�

003

�0�0

04�

�0�0

04�

�0�0

04�

�0�0

04�

�0�0

04�

�0�0

04�

�0�0

04�

�0�0

04�

�0�0

04�

�0�0

02�

Executivetie

−0�3

26−0

�333

−0�3

15−0

�315

−0�3

15−0

�305

−0�2

28−0

�303

−1�4

64−0

�443

�0�1

66�∗

�0�1

52�∗

�0�1

56�∗

�0�1

52�∗

�0�1

51�∗

�0�1

55�∗

�0�1

45�

�0�1

56�

�0�2

98�∗

∗�0�1

55�∗

∗

U.S.p

atents

−0�0

47−0

�050

−0�0

49−0

�035

−0�0

30−0

�024

−0�0

11−0

�020

−0�0

41−0

�012

�0�0

90�

�0�0

89�

�0�0

89�

�0�0

89�

�0�0

87�

�0�0

87�

�0�0

88�

�0�0

87�

�0�0

88�

�0�0

41�

Degree

0�00

3−0

�021

−0�0

45−0

�022

−0�0

23−0

�034

−0�0

20−0

�035

−0�0

360�

040

�0�0

23�

�0�0

25�

�0�0

26�

�0�0

25�

�0�0

27�

�0�0

28�

�0�0

27�

�0�0

28�

�0�0

28�

�0�0

22�

WGcontributor

0�91

81�

120

1�27

90�

773

0�89

10�

997

0�42

20�

919

0�80

40�

628

�0�5

55�

�0�3

71�∗

∗�0�3

91�∗

∗�0�3

62�∗

�0�3

76�∗

�0�3

86�∗

∗�0�3

29�

�0�4

02�∗

�0�3

94�∗

�0�3

14�∗

�ln�IETF

publications

1�33

81�

399

1�30

10�

973

0�97

70�

862

0�58

10�

865

0�95

71�

051

�0�5

62�∗

�0�3

45�∗

∗�0�4

04�∗

∗�0�3

52�∗

∗�0�3

46�∗

∗�0�3

86�∗

�0�2

95�∗

�0�3

86�∗

�0�3

83�∗

�0�1

13�∗

∗

Presence

−0�1

17−0

�158

−0�1

72−0

�156

−0�1

47−0

�129

0�22

0−0

�065

−0�2

28−0

�057

�0�3

20�

�0�1

48�

�0�1

88�

�0�1

63�

�0�1

53�

�0�1

92�

�0�0

45�∗

∗�0�2

13�

�0�2

14�

�0�1

47�

Socialbrokerage

1�11

01�

655

1�26

02�

235

1�07

52�

171

1�97

41�

846

�0�4

04�∗

∗�0�4

43�∗

∗�0�4

43�∗

∗�0�5

24�∗

∗�0�4

46�∗

�0�5

26�∗

∗�0�5

30�∗

∗�0�4

55�∗

∗

Socialbroker

∗Presence

0�79

41�

023

0�09

01�

042

0�91

30�

560

�0�2

59�∗

∗�0�2

95�∗

∗�0�1

30�

�0�3

01�∗

∗�0�3

09�∗

∗�0�2

60�∗

Spans

1boundary

0�73

40�

690

0�72

50�

472

0�79

00�

838

0�85

5�0�2

60�∗

∗�0�2

58�∗

∗�0�2

74�∗

∗�0�2

33�∗

�0�2

82�∗

∗�0�2

82�∗

∗�0�3

27�∗

∗

Spans

1boundary

×Socialbroker

−1�9

70−3

�662

−1�8

93−3

�515

−3�2

42−3

�120

�0�9

15�∗

�1�0

35�∗

∗�0�9

39�∗

�1�0

46�∗

∗�1�0

34�∗

∗�0�9

07�∗

∗

Spans

1boundary

×Presence

−0�1

41−0

�171

−0�1

47�0�1

91�

�0�1

91�

�0�2

02�

Spans

2+boundaries

1�53

41�

701

1�88

31�

341

1�86

71�

848

2�06

1�0�4

08�∗

∗�0�4

85�∗

∗�0�5

05�∗

∗�0�4

48�∗

∗�0�7

31�∗

�0�7

51�∗

�0�6

05�∗

∗

Spans

2+boundaries×Socialbroker

−2�5

03−5

�543

−2�5

19−5

�504

−5�1

35−3

�670

�1�7

93�

�2�0

41�∗

∗�1�8

22�

�2�0

40�∗

∗�2�0

25�∗

�1�8

68�∗

Spans

2+boundaries×Presence

−0�0

53−0

�011

−0�5

22�0�3

28�

�0�3

33�

�0�3

70�

Nonselectionhazard

−1�2

63�0�2

62�∗

∗

Log

likel

ihoo

d−5

79�4

8−5

75�5

0−5

70�7

1−5

71�8

6−5

67�0

3−5

60�8

3−5

53�0

0−5

60�5

6−5

47�1

5−3

�653

�29

Obs

erva

tions

5,05

65,

056

5,05

65,

056

5,05

65,

056

5,05

65,

056

5,05

612

0,19

3In

divi

dual

s61

061

061

061

061

061

061

061

061

015

,465

Even

ts11

811

811

811

811

811

811

811

811

847

2

Notes

.Sta

ndar

der

rors

inpa

rent

hese

s.W

G=

wor

king

grou

p.∗ S

igni

fican

tat5

%;∗

∗ sig

nific

anta

t1%

;+jo

intly

sign

ifica

ntat

5%.

Fleming and Waguespack: Brokerage, Boundary Spanning, and LeadershipOrganization Science 18(2), pp. 165–180, © 2007 INFORMS 175

Figure 2 Publication Spline for Full-Risk Set (15,466 SubjectsOver 243,405 Observation Periods)

Multiplier effect of publication on leadership(+/–95% confidence intervals)

0

20

40

60

80

100

120

140

1 2 3 4 5 6 8 9+

Number of publications in prior three years7

ResultsModels 1 through 5 in Table 3 enter explanatory variablesindividually and then with interactions. We maintainIETF Publications and controls in all specifications bothbecause publications has such a powerful effect andbecause we want to guard against the possibility that dif-ferences in social network position are simply the resultof the number of times (i.e., number of publications)we observe a subject. Models 6 through 10 present fullspecifications. Model 7 illustrates results without use ofan instrumental variable for presence, while Model 8illustrates the lack of effect for a boundary-spanning andpresence interaction. Model 9 illustrates that the effectsremain robust in the presence of a nonselection hazard,and Model 10 establishes that the results hold when esti-mations are run using imputed network values for thefull sample. Before analyzing Table 3, we present anunconditional analysis of the effects of technical contri-bution, as measured by publications for the entire risk setof 15,466 subjects observed over 243,405 time periods.Figure 2 dramatically illustrates the importance of tech-nical contribution on leadership and supports a loggedspecification in the hazard models. The effect sizes arequite large; the first publication increases the likelihoodof becoming a leader by 143%.8 The figure indicatesthat for three publications the point estimate increases1,590% relative to having none. Hence, the results sup-port Hypothesis 1, which argues for the importance oftechnical contribution for leadership.The models also support the hypothesized relation-

ships for both brokerage and boundary spanning. Weargued that the net effect of brokerage remains unpre-dictable due to positive and negative influences ofthe position on the perceptions of potential followers.We proposed in Hypothesis 2, however, that physicalattendance at meetings would ameliorate negative per-ceptions, such that attendance and interaction shoulddemonstrate a positive interaction. All the models sup-port this interaction effect. Model 9 (from which wedraw all interpretations of effect magnitude) indicatesthat an individual who simultaneously increases atten-dance and social brokerage position by one standard

Figure 3 Interaction Between Social Brokerage andCommunity Presence

0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

5.0

–4 –3 –2 –1 0 1 2 3 4

Community presence (inst)

Lea

der

ship

mu

ltip

lier

Low social brokerage (cohesion)Medium social brokerageHigh social brokerage

Note. (Mean+ /− one standard deviation of brokerage fromModel 9.

deviation each will increase the likelihood of becominga leader by 33%. The one standard deviation increasein brokerage alone increases the likelihood of leadershipby 70%, indicating that it has a net positive influence onthe chance of leadership. (Instrumented attendance hasno impact.) Figure 3, which plots the intersection effectfor low, medium, and high values of social brokerage,indicates that most of the positive effect of the brokerageposition accrues to individuals who attend many confer-ences. Conversely, a cohesive ego network structure ispreferred for those who do not attend conferences andlack a strong physical presence. Hence, the models andFigure 3 support Hypothesis 2, that the positive effectsof social brokerage will be enhanced by increased trust,as developed by physical presence in the community.Although social brokerage and its interaction with at-

tendance demonstrate strong positive effects, the bene-fits of boundary spanning are greater. Contributing totwo working groups correlates with a 131% increase inleadership (over and above the 123% benefit of con-tributing within any working group). Contributing tothree or more working groups correlates with a 535%increase. Simultaneous brokerage and membership intwo groups correlates with a negative interaction effectof 59%; in three or more groups, a negative effect of75% (although the overall effect remains positive, dueto the stronger first-order effects). Figure 4 illustratesthe negative consequences of simultaneous brokerageand boundary spanning. With increased brokerage, ben-efits increase for individuals who work within a group,but decrease sharply for those who straddle boundaries.Hence, the models support Hypothesis 3, that bound-ary spanning increases the likelihood of leadership, andHypothesis 4, that simultaneous brokerage decreases thispositive effect. Further supporting our theory, Model 9does not demonstrate an even marginally significantinteraction between boundary spanning and the instru-mented attendance variable. Although prediction and

Fleming and Waguespack: Brokerage, Boundary Spanning, and Leadership176 Organization Science 18(2), pp. 165–180, © 2007 INFORMS

Figure 4 Interaction Between Technical Boundary Spanningand Social Brokerage (Model 9)

0123456789

1011121314151617

–0.28 –0.21 –0.14 –0.07 0 0.07 0.14 0.21 0.28

Brokerage score

Lea

der

ship

mu

ltip

lier

No boundary spanBoundary span (2 WG)Boundary span (3 WG)

Note. WG=working group.

observation of a null result do not test theory (hence, arenot included in our hypotheses), the lack of significanceprovides further evidence of our contention that fellowcommunity members perceive the two roles differently.To directly compare the different structural strategies

and their influence on leadership, imagine two individu-als: first, a broker whose measure is one standard devia-tion above the variable mean and contributes to only oneworking group; and second, a boundary spanner whosemeasure is at the brokerage mean and contributes to twogroups. Both strategies work approximately equally wellif they attend an average number of conferences. Thebroker is then 193% more likely to become a leader(70% for brokerage and 123% for contributing withina working group) compared with a 254% increase forthe boundary spanner (131% for spanning one boundaryand 123% for contributing within a working group). Ifboth individuals increase their attendance at conferencesby one standard deviation, the advantage remains withthe boundary spanner, because the broker gains only33% from the interaction of brokerage and attendance.The advantage shifts decisively in favor of the bound-ary spanner, however, if she spans additional boundaries;assuming all else stays equal, she is 658% more likelyto become a leader (535% for spanning two or moreboundaries and 123% for contributing within a workinggroup). While it appears to be moderately effective tobind a single working group with itself, it appears tobe even more effective to connect three groups withinthe larger community. Moderately ambitious individualsshould therefore look for local integration opportunities;very ambitious individuals should seek out the intersec-tions of communitywide activity. Furthermore, for indi-viduals who have already authored many drafts, it wouldbe easier to span an additional boundary, given the incre-mental impact of a single additional collaboration on adeeply embedded structural position.

With respect to control variables, collaboration withan AD subtracts from the likelihood of leadership, aresult consistent with community-espoused norms. Withthe exception of the incumbent working group popula-tion, none of the controls demonstrate significance in thenetworked risk set. Surprisingly, degree never demon-strates a significant effect in any model; working withadditional coauthors does not correlate with leadership.We attribute this to a dilution effect of the credit thatoccurs when an individual coauthors with too manyothers. Discussion with IETF members lent validity tothis interpretation. Finally, the selection term is negativeand significant, indicating that individuals who are mostlikely not to be selected are also less likely to becomeleaders.Finally, Model 10 in Table 3 takes an alternative tack

at addressing the issue of the small subsample usedfor primary analysis. Recall that very few IETF partic-ipants, just 610 out of 15,465, have two or more pub-lications prior to leadership appointment or censoring.While it is therefore true that many engineers advanceto leadership without observed publications, and hencewithout an observed network position, further investi-gation reveals that 75% of all working group chairseventually publish, and 100% are regular conferenceattendees. There is also anecdotal evidence that most ofthe working group chairs that never publish have, in fact,attempted to do so. Collaborative networks are thus apervasive feature of this organization, but the data arelimited by a nonrandom propensity to observe collabo-rative ties. While Model 9 addresses this problem withthe nonselection hazard term, Model 10 instead utilizesimputed network values for all subjects. Imputed valuesfor Degree, Boundary Spanning, and Social Brokerageare derived from regressions using cumulative confer-ence attendance, time at risk, type of employer, loca-tion outside the United States, cumulative distance fromrecent conferences, publishing by coemployees, work-ing group chair selection among coemployees, patenting,and year dummies. The results in Model 10 comparevery favorable to all other full specifications.

DiscussionWhile the models demonstrate the importance of bothhuman and social capital in the attainment of leadershippositions, it is important to demarcate their theoreticaland empirical limitations. We generalize the theoreticalimplications for brokerage with caution, because thesocial context differs from that of private firms (Burt1992, 1997, 2001a; Podolny and Baron 1997). Thisresearch considers how individuals who participate involuntary communities emerge as leaders, as opposedto how individuals in for-profit firms are promotedto management. Aspiring managers leverage broker-age positions by controlling information, resources, and,

Fleming and Waguespack: Brokerage, Boundary Spanning, and LeadershipOrganization Science 18(2), pp. 165–180, © 2007 INFORMS 177

ultimately, the perceptions of those who do the promot-ing. Such efforts are facilitated by focusing on maxi-mizing uncertainty among a few key decision makers.In a voluntary community, however, leadership requiresmobilizing efforts across a large (and often unknown)population of peers, friends, and colleagues. Memberswho gain reputations as controlling individuals whoactively manage perceptions will probably face infor-mal sanctions (von Hippel and von Krogh 2003). As theresults indicate, aspiring leaders in voluntary communi-ties need to offset these potentially negative perceptionswith additional assurances that they intend to benefit andbind the community as a whole.Caution should also be exercised before generalizing

the results to all open innovation communities, partic-ularly those that lack physical interaction. Our resultsremain conditional on attendance (although our dataindicated that every IETF leader attended at least onemeeting). The issues of trust and forking remain salientfor all open innovation communities and warrant addi-tional research and elaboration. Concern is also meritedwith respect to the monotonically increasing environ-mental munificence during the period of observation—the IETF (and the industries that supported it) expandedduring all but the last few years of the study. Contin-uously expanding resources and leadership opportuni-ties facilitate the dispersion of benefits, enable coalitionbuilding, and probably ease distrust. The IETF’s recentself-examination (Wasserman 2003) admits as much,even as it attempts to grapple with increasingly divi-sive processes and communal fissures. Elaborating thisrelationship between munificence and positional efficacyremains an opportunity for future research within com-munities and firms. Community growth and member-ship change also highlights an opportunity to study thedemographic dynamics of voluntary communities. Forexample, when do current members or leaders attractdemographically extreme recruits? When this occurs, thepossibility of forking or disruptive selection (McPhersonet al. 1992) should greatly increase.Empirical limitations also deserve attention. The de-

pendent variable of appointment does not measure orga-nization of a BOF meeting. Ideally, we would alsoobserve which community members attempted to orga-nize a meeting, and then, conditional on that success-ful organization, which members advance to leadership.Archived BOF meeting documentation remains incon-sistent, however, and more complete for meetings thatbecome working groups. Rather than introduce a non-quantifiable success bias into our models, we developedthe instrumental variable based on attendance, whichcontrols for members’ aspirations and attempts to lead.Without use of the instrument, the influence of techni-cal contribution and structural positions become muchweaker, and mere attendance appears to be a very effec-tive route to leadership. Inferences drawn from this

model would be wrong, however, because it does notcontrol for the self-selection of aspiring leaders to attend.This incomplete specification would therefore downplaythe importance of technical contribution and miss thenuanced influence of brokerage and its interaction withphysical presence. The results should still be regardedwith caution, despite the strength of the instrument,given that the second stage estimates a nonlinear Coxmodel, and the use of instruments with nonlinear mod-els remains controversial (see Bowden and Turkington1981, Box-Steffensmeier and Jones 2004; cf. Hausmanet al. 1995). Limitations notwithstanding, the researchillustrates a method for analyzing social networks witharchival data. Of greater importance to future researchin social networks, the use of an endogeneity instrumentenables control for strategic behavior in the developmentof relationships. To date, network analysis has mainlyignored the problem of separating the causal effects ofposition from the intentions of the individual who mayhave consciously and strategically created the position.Although our analyses used archival data, the more com-mon use of questionnaire data in network analysis couldalso benefit from the explicit design and gathering ofinstrumental variables.In addition to the methodological contributions, the

results contribute to the larger network literature. Giventhat the quantitative research on the benefits of brokeragehas not previously distinguished brokerage from bound-ary spanning, it remains unclear which position confersthe bulk of previously demonstrated benefits. At leastin the IETF, leaders are more likely to emerge fromcohesive boundaries than from isolated brokerage posi-tions. The results present the only analysis (to the bestof our knowledge) that separates the influence of humanand social capital on promotion. They highlight the pre-viously unexamined differences between brokerage andboundary-spanning positions and demonstrate that therole conflict previously associated with both positions(Podolny and Baron 1997) appears to be associated onlywith brokerage.These findings should be elaborated in open and

commercial contexts with dynamic analyses, becauseboundaries change as often as the relationships thatstraddle them. The current research studied boundariesthat emerge from leaders’ architectural and individu-als’ voluntary choices, remain relatively informal, anddissipate with completion of the group’s work. Firmboundaries are more formal, longer lived, and proba-bly support the transformation and hardening of tech-nological boundaries into social boundaries. In privatefirms, brokerage and boundary spanning probably corre-late more strongly, such that individuals who span a cohe-sive boundary should become increasingly rare over time.If this were true, then the combination of boundary span-ning and cohesion might be an even more effective pro-motion strategy in private firms than in open innovation

Fleming and Waguespack: Brokerage, Boundary Spanning, and Leadership178 Organization Science 18(2), pp. 165–180, © 2007 INFORMS

communities. Such a strategy might be conscientiouslydirected by a senior manager, for example, by assigning acapable engineer to a critical boundary-spanning projectand discouraging brokerage of her colleagues. The capa-ble engineer would then be an ideal candidate for promo-tion, in either of the groups or in the larger organization,due to her visible and trusted contributions in managingthe organization’s technological interdependencies.

ConclusionThe IETF archives afford a rare and detailed look atthe social, technical, and political dynamics within anopen innovation community. We used the entire his-tory of 16 years of meetings and work collaborations totrace the emergence of leaders within the organization.After demonstrating the importance of human capital,as measured by technical contributions, the models alsodemonstrated the similarly large importance of socialcapital, as measured by brokerage and boundary span-ning of collaborative relationships. That both brokeringand boundary-spanning roles greatly increase the like-lihood of leadership points to the importance of socialpositions that can unite open innovation communities.We argued that trust does not come easily to communitymembers who fear cooptation by commercial interestsor forking over technical disagreements. Because bro-kers by definition contrive less cohesive and less trustingcontexts, the probability that they will assume leadershiproles remains highly contingent on building trust withcommunity members. We argue that aspiring leaders canbuild trust through physical attendance and, consistentwith this argument, find a positive interaction with phys-ical attendance. Also consistent with our emphasis ontrust in open innovation communities, brokerage andboundary spanning demonstrated a negative interaction,indicating that brokers who span boundaries remain at adisadvantage. While brokerage alone demonstrates posi-tive influence on becoming a leader, boundary spanningdemonstrates a much stronger effect. Finally, we didnot observe a contingent relationship between bound-ary spanning and attendance. Our results emphasize theimportance of intermediary and integrating roles—forbrokers within technological boundaries, and for bound-ary spanners across cohesive technological boundaries.Open innovation communities represent a new and

powerful social context in which to generate knowledgeand advance technology (von Hippel and von Krogh2003). Even if they represent merely a recombinanthybrid between the institutions of science and normsof communities of technological practice, their resultsto date and potential to shape the rate and directionof technological change remain impressive. Increasingtechnological complexity and interdependence put thesecommunities at perpetual risk of forking and balkaniza-tion. Without strong leaders they inevitably splinter and

fail. Leaders, in binding their communities together, fos-ter the integration needed to forestall these eventualities.The contribution of this research is to elucidate howleaders emerge from the interactions of ability, position,and strategy.

AcknowledgmentsThe authors are very appreciative of the help of Scott Brad-ner and many other members of the Internet Engineering TaskForce in completing this study. They would also like to thanktwo anonymous reviewers, Jóhanna Birnir, Ronald Burt, BrentGoldfarb, Michael Hannan, Eric von Hippel, Jerry Kim, KarimLakhani, Siobhan O’Mahony, Renée Richardson, Tim Simcoe,John Simon, Mike Tushman, Steve Wheelwright, and, in par-ticular, Woody Powell and Olav Sorenson for their feedback,and the Harvard Business School Division of Research forsupport. The work has also benefited from seminars givenat Harvard and Stanford Universities and the Universities ofChicago, Maryland, and California, Berkeley.

Endnotes1“One of the principal differences between the IETF and manyother standards organizations is that the IETF is very much abottom-up organization. It is quite rare for the executive lead-ers within the IETF, the IESG (Internet Engineering SteeringGroup) or the IAB (Internet Architecture Board) members, tocreate a working group on their own to work on some prob-lem that is felt to be an important one. Almost all workinggroups are formed when a small group of interested individ-uals get together on their own and then propose a workinggroup to an Area Director” (Bradner 1999, p. 49). It is muchthe same with other open innovation communities. “Today, anopen source software development project is typically initiatedby an individual or a small group with an idea for somethinginteresting they themselves want for an intellectual or personalor business reason” (von Hippel and von Krogh 2003, p. 211).This process for Linux was described thus: “ ‘The lieutenantsget picked—but not by me,’ explains Torvalds. ‘Somebodywho gets things done, and shows good taste—people just startsending them suggestions and patches’ ” (Hamm 2005, p. 66).See also Rosenkopf et al. (2001).2Open innovation communities use the term Trojan horse torefer to the threat of corporate manipulation and malicioussabotage, such as the deliberate planting of insidious bugs(DiBona et al. 1999).3In 1987, 1989, and 1990 the IETF held four meetings eachyear.4We do not estimate repeated events models because priorleadership experience would strongly influence the probabil-ity of an additional appointment. For example, a prior leader,particularly if successful, would probably be more trusted andbenefit less from the publication of standards drafts or advan-tageous network positions.5We investigated other potential measures of relationships, butfound each wanting: e-mail interactions and working groupattendance lists were inconsistently archived, particularly forearlier interactions; e-mail contributions would be skewed infavor of communicative extroverts; interviewees reported thatindividuals often had no knowledge or relationship with affili-ated (that is, same-employer) colleagues. In contrast, standards

Fleming and Waguespack: Brokerage, Boundary Spanning, and LeadershipOrganization Science 18(2), pp. 165–180, © 2007 INFORMS 179

collaborations and the working group publication affiliationmeasure strong, accurately measured, and very importantsocial relationships in this community.6Zero values for miles occur when a member attends her firstconference in her home town.7In order to explore the demographics of IETF members,we linked the IETF data back into the patent databaseand performed some basic regressions on variables fromFleming et al., forthcoming. Relative to the entire popula-tion of inventors of U.S. patents, IETF members with patentshad significantly more patents, collaborators, and number ofassignees (that is, their patents were owned by a greater num-ber of organizations—this correlates most strongly with per-sonnel movement and boundary spanning). IETF membersdemonstrated no significant difference in their brokerage ofother patented inventors, future prior art citations to theirpatents (a widely used indicator of quality), references tothe nonpatent (generally scientific) literature, and proclivity torepeatedly collaborate with the same colleagues. Hence, IETFmembers who are inventors and who patent appear to publishmore, and work with more people and span more organiza-tions, than do non-IETF members who are inventors and whopatent.8As calculated by 100�e�0�887∗1 − 1 = 143.

ReferencesAbbate, J. 1999. Inventing the Internet. MIT Press, Cambridge, MA.

Allen, T. 1977. Managing the Flow of Technology. MIT Press,Cambridge, MA.

Boehmke, F. J., D. S. Morey, M. Shannon. 2006. Selection bias andcontinuous-time duration models: Consequences and a proposedsolution. Amer. J. Political Sci. 50(January) 192–207.

Borgatti, S., M. Everett, L. Freeman. 2001. UCINET 6.0 Version 1.00.Analytic Technologies, Natick, MA.

Bound, J., D. Jaeger, R. Baker. 1995. Problems with instrumentalvariables estimation when the correlation between the instru-ments and the endogenous explanatory variable is weak. J. Amer.Statist. Assoc. 90(430) 443–450.

Bowden, R., D. Turkington. 1981. A comparative study of instru-mental variables estimators for nonlinear simultaneous models.J. Amer. Statist. Assoc. 76(376) 988–995.

Box-Steffensmeier, J. M., Bradford S. Jones. 2004. Event HistoryModeling: A Guide for Social Scientists. Cambridge UniversityPress, New York.

Bradner, S. 1996. The Internet standards process—Rev. 3. RFC 2026.

Bradner, S. 1998. IETF working group guidelines and procedures.RFC 2418.

Bradner, S. 1999. The Internet engineering task force. C. DiBona,S. Ockman, M. Stone, eds. Open Sources: Voices from the OpenSource Revolution. O’Reilly, Sebastapol, CA, 47–52.

Bradner, S. 2003a. IETF structure and Internet standards process. Pre-sentation 57th IETF Conv., Vienna, Austria.

Bradner, S. 2003b. Personal communication. Cambridge, MA(March 24).

Bradner, S. 2004. Personal communication. Cambridge, MA(August 28).

Bradner, S. 2006. Personal communication. Cambridge, MA(April 19).

Brennen, A. 2003. Gnu P G Keysigning Party HOWTO. http://cryptnet.net/fdp/crypto/keysigning-party/en/keysigning_party.html.

Broothaerts, W., H. Mitchell, B. Weir, S. Kalnes, L. Smith, W. Yang,J. Mayer, C. Roa-Rodriguez, R. Jefferson. 2005. Gene transferto plants by diverse species of bacteria. Nature 433 629–633.

Burt, R. 1992. Structural Holes. Harvard University Press, Cam-bridge, MA.

Burt, R. 1997. The contingent value of social capital. Admin. Sci.Quart. 42 339–365.

Burt, R. S. 2000. The network structure of social capital. Res. Organ.Behav. 22 345–423.

Burt, R. 2001a. Structural holes versus network closure as social cap-ital. N. Lin, K. Cook, R. Burt, eds. Social Capital: Theory andResearch. Aldine de Gruyter, New York.

Burt, R. 2001b. Bandwidth and echo: Trust, information, and gos-sip in social networks. A. Casella, J. Rauch, eds. Networks andMarkets: Contributions from Economics and Sociology. RussellSage Foundation, New York.

Coleman, J. S. 1988. Social capital in the creation of human capital.Amer. J. Sociol. 94 S95–S120.

Cox, D. 1972. Regression models and life-tables (with discussion).J. Roy. Statist. Soc. Ser. B 34 187–220.

Dalle, J., P. David. 2003. The allocation of software developmentresources in “open source” production mode. Working paper,Stanford University, Stanford, CA.

Davies, E. 2003. IETF problem statement. http://www.ietf.org/Internet-drafts/draft-ietf-problem-issue-statement-02.txt.

Davies, E., J. Hoffman. 2004. IETF problem resolution process. http://rfc3844.x42.com/.

DiBona, C., S. Ockman, M. Stone. 1999. Introduction. C. DiBona,S. Ockman, M. Stone, eds. Open Sources: Voices from the OpenSource Revolution. O’Reilly, Sebastapol, CA, 1–19.

Dougherty, D. 1992. Interpretive barriers to successful product inno-vation in large firms. Organ. Sci. 3(2) 179–202.

Economist. 2003. Open source’s local heroes. Econom. Tech. Quart.369(8353) 5.

Fleming, L., C. King, A. Juda. Small worlds and regional advantage.Under revision. Organ. Sci. Forthcoming.

Friedrich, R. 1982. In defense of multiplicative terms in multipleregression equations. Amer. J. Political Sci. 26(4) 797–833.

Goetz, T. 2003. Open source everywhere. Wired 11(11).

Gould, R. 1991. Multiple networks and mobilization in the Paris Com-mune, 1871. Amer. Sociol. Rev. 56(6) 716–729.

Granovetter, M. 1992. Problems of explanation in economic soci-ology. N. Nohria, R. Eccles, eds. Networks and Organization.Harvard Business School Press, Boston, MA.

Greene, W. 1993. Econometric Analysis. Prentice Hall, EnglewoodCliffs, NJ.

Hamm, S. 2005. Linux, Inc. Business Week 3918(Jan. 31) 60.

Harris, S. 2001. The tao of IETF—A novice’s guide to the Internetengineering task force. RFC 3160.

Hausman, J., W. Newey, J. Powell. 1995. Nonlinear errors in vari-ables: Estimation of some Engel curves. J. Econom. 65 205–233.

Heckman, James J. 1976. Sample selection bias as a specificationerror. Econometrica 47 153–161.

Henderson, R., K. Clark. 1990. Architectural innovation: The recon-figuration of existing product technologies and failure of estab-lished firms. Admin. Sci. Quart. 35 9–30.

Fleming and Waguespack: Brokerage, Boundary Spanning, and Leadership180 Organization Science 18(2), pp. 165–180, © 2007 INFORMS

Hoffman, P., S. Bradner. 2002. Defining the IETF. RFC 3233.

Kaufman, C. 2003. Personal communication. Westford, MA,(July 28).

Kogut, B., A. Metiu. 2001. Open-source software developmentand distributed innovation. Oxford Rev. Econom. Policy 17(2)248–264.

Lakhani, K. R. 2004. The determinants of developer mobilization inopen source communities. Working paper, Massachusetts Insti-tute of Technology, Cambridge, MA.

Lakhani, K. R., E. von Hippel. 2003. How open source softwareworks: Free user-to-user assistance. Res. Policies 32(6) 923–943.

Lee, G., R. Cole. 2003. From a firm-based to a community-basedmodel of knowledge creation. Organ. Sci. 14(6) 633–649.

Lerner, J., J. Tirole. 2001. Some simple economics of open source.J. Indust. Econom. 2 197–234.

Lorsch, P. 1965. Product Innovation and Organization. McMillan,New York.

McPherson, J., P. Popielarz, S. Drobnic. 1992. Social networks andorganizational dynamics. Amer. Sociol. Rev. 57 153–170.

Merton, R. 1973. The normative structure of science. N. Storer, ed.The Sociology of Science: Theoretical and Empirical Investiga-tions. University of Chicago Press, Chicago, IL.

Mockus, A., R. Fielding, J. Herbsleb. 2002. Two case studies of opensource software development: Apache and Mozilla. ACM Trans.Software Engrg. Methods 11(3) 309–346.

Mowery, D., T. Simcoe. 2002. Is the Internet a U.S. invention? Aneconomic and technological history of computer networking.Res. Policies 31 1369–1387.

Neustadt, R. E. 1990. Presidential Power and the Modern Presidents:The Politics of Leadership from Roosevelt to Reagan. Free Press,New York.

O’Mahony, S., F. Ferraro. Managing the boundary of an “open”project. J. Padgett, W. Powell, eds. Market Emergence andTransformation. MIT Press. Forthcoming.

Padgett, J., C. Ansell. 1993. Robust action and the rise of the Medici,1400–1434. Amer. J. Sociol. 98 1259–1319.

Perrone, V., A. Zaheer, B. McEvily. 2003. Free to be trusted? Orga-nizational constraints on trust in boundary spanners. Organ. Sci.14(4) 422–439.

Podolny, J. M., J. N. Baron. 1997. Resources and relationships: Socialnetworks and mobility in the workplace. Amer. Sociol. Rev. 62(5)673–693.

Raymond, E. 1998. The Cathedral and the Bazaar. O’Reilly andAssociates, Inc., Sebastapol, CA.

Raymond, E. 2000. Conventions at light speed: What hackerscan learn from SF fandom. http://www.catb.org/∼esr/writings/sfshows/.

Reagans, R., B. McEvily. 2003. Network structure and knowledgetransfer: The effects of cohesion and range. Admin. Sci. Quart.48 240–267.

Rivlin, G. 2003. Leader of the free world: How Linus Torvaldsbecame benevolent dictator of Planet Linux, the biggest collab-orative project in history. Wired 11(11).

Rosenkopf, L., A. Metiu, V. George. 2001. From the bottom up?Technical committee activity and alliance formation. Admin. Sci.Quart. 46 748–772.

Schoenfeld, D. 1982. Partial residuals for the proportional hazardsregression model. Biometrika 69 239–241.

Secure Shell secsh. 2003. http://www.ietf.org/html.charters/secsh-charter.html.

Sorenson, O., T. E. Stuart. 2001. Syndication networks and the spatialdistribution of venture capital investments. Amer. J. Sociol. 1061546–1588.

Sorenson, O., J. Rivkin, L. Fleming. 2006. Complexity, networks, andknowledge flow. Res. Policy 35 994–1017.

Tushman, M. 1977. Special boundary roles in the innovation process.Admin. Sci. Quart. 22 587–605.

von Hippel, E., G. von Krogh. 2003. Open source software andthe “private-collective” model: Issues for organization science.Organ. Sci. 14(2) 209–223.

von Krogh, G., S. Spaeth, K. R. Lakhani. 2003. Community, joining,and specialization in open source software innovation: A casestudy. Res. Policies 32(7) 1217–1241.

Wade, J. 1995. Dynamics of organizational communities and tech-nological bandwagons: An empirical investigation of commu-nity evolution in the microprocessor market. Strategic Manage-ment J. 16 111–133.

Wasserman, M. 2003. IETF problem resolution processes. http://www.ietf.org/Internet-drafts/draft-ietf-problem-process-01.txt.