The Growth of Interorganizational Collaboration in the Life ...

1,123 views

Published on

0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
1,123
On SlideShare
0
From Embeds
0
Number of Embeds
1
Actions
Shares
0
Downloads
4
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

The Growth of Interorganizational Collaboration in the Life ...

  1. 1. Network Dynamics and Field Evolution: The Growth of Interorganizational Collaboration in the Life Sciences Walter W. Powell Douglas R. White Stanford University UC Irvine and Santa Fe Institute drwhite@uci.edu woodyp@stanford.edu Kenneth W. Koput Jason Owen-Smith University of Arizona University of Michigan kkoput@u.arizona.edu jdos@umich.edu Preprint Forthcoming, American Journal of Sociology Acknowledgements: We want to thank the Santa Fe Institute for providing the venue where these ideas were initially discussed and much of the work was done. We are espe- cially grateful to John Padgett, organizer of the States and Markets group at SFI for his support and insights. We have benefited from comments from the audience at seminars given at Illinois, Michigan, MIT, Oxford, Pennsylvania, SFI, Stanford, UC Irvine, UCLA, and the SSRC’s Economic Sociology Workshop at the Bellagio Center. We received very helpful comments from Peter Bearman, Tim Bresnahan, Paul David, Walter Fontana, Mau- ro Guillen, Heather Haveman, Sanjay Jain, Charles Kadushin, Bruce Kogut, Paul McLean, Jim Moody, Charles Ragin, Spyros Skouras, David Stark, and Brian Uzzi. We are indebted to Duncan Watts for providing access to his unpublished work, and to Mark Newman for sharing software that allowed us to compute the expected size of the largest connected components in our networks. We appreciate very much the excellent research assistance provided by James Bowie, Kjersten Bunker, Kelley Porter, Katya Seryakova, and Laurel Smith-Doerr. We thank Tanya Chamberlain for exceptional help in keeping track of nu- merous successive versions of the paper. The research was supported by grants from the National Science Foundation (#9710729 and SRS 0097970), the Hewlett Foundation, and the EPRIS project at the University of Siena.
  2. 2. Abstract We develop and test four alternative logics of attachment – accumulative advan- tage, homophily, follow-the-trend, and multiconnectivity – to account for both the struc- ture and dynamics of interorganizational collaboration in the field of biotechnology. The commercial field of the life sciences is typified by wide dispersion in the sources of basic knowledge and rapid development of the underlying science, fostering collaboration among a broad range of institutionally diverse actors. We map the network dynamics of the field over the period 1988-99. Using multiple novel methods, including analysis of net- work degree distributions, network visualizations, and multi-probability models to estimate dyadic attachments, we demonstrate how different rules for affiliation shape network evo- lution. Commercialization strategies pursued by early corporate entrants are supplanted by collaborative activities influenced more by universities, research institutes, venture capital, and small firms. As organizations increase both the number of activities on which they col- laborate and the diversity of organizations with whom they are linked, cohesive sub-net- works form that are characterized by multiple, independent pathways. These structural components, in turn, condition the choices and opportunities available to members of a field, thereby reinforcing an attachment logic based on connections to diverse partners that are differently linked. The dual analysis of network and institutional evolution provides an explanation for the decentralized structure of this science-based field.
  3. 3. Table of Contents Abstract Introduction Topology of Large-Scale Networks Field Structuration: Science Meets Commerce Data and Methods Analysis I: Degree Distributions Analysis II: Discrete Time Network Visualization Analysis III: Attachment Bias Conclusion and Implications Appendix I: Network Visualization in Pajek Endnotes Tables 1-7 References Appendix II: Tables A1-A3 Figures List of Tables: Table 1 Top Ten Biotechnology Drugs, 2001 Table 2 Patterns of Entry and Exit into the Network Table 3 Variables in Statistical Tables Table 4 Test of Accumulative Advantage: Odds Ratio from McFadden Mod- el Table 5 Test of Homophily: Odds Ratio from McFadden Model Table 6 Test of Follow-the-Trend: Odds Ratio from McFadden Model Table 7 Test of Multiconnectivity: Odds Ratio from McFadden Model List of Figures: Figure 1 DBF and University Patents, 1976-99 Figure 2 Distribution of Organizational Forms and Activities Figure 3 Degree Distributions by Type of Partner Pajek figures 4 1988 Main Component, All Ties 5 1989 Main Component, New Ties 6 1993 Main Component, All Ties 7 1994 Main Component, New Ties 8 1997 Main Component, All Ties 9 1998 Main Component, New Ties 10 1997 All Ties, Main Component, Cohesion 11 1998 New Ties, Main Component, Cohesion
  4. 4. Introduction the most comprehensive text on network The images of field and network are methods, there is only a paragraph on net- common in both contemporary physical work dynamics in a section on future direc- and social science. In the physical sciences, tions (Wasserman and Faust, 1994). Thus fields are organized by information in the while some progress has been made analyz- form of geometric patterns. The study of ing the dynamics of dyads (e.g., Lincoln et the geometry of fields has attracted consid- al, 1996; Gulati and Gargiulo, 1998; Stuart, erable interest in the statistical mechanics 1998), little attention has been given to the of complex networks. Research by physi- evolution of entire networks. cists interested in networks has ranged There are a number of excellent studies widely from the cellular level, a network of of the structuring of specific organizational chemicals connected by pathways of chem- fields (DiMaggio, 1991; Thornton, 1995; ical reactions, to scientific collaboration Dezalay and Garth, 1996; Ferguson, 1998; networks, linked by coauthorships and co- Scott et al, 2000; Hoffman, 2001; Morrill citations, to the world-wide web, an im- and Owen-Smith, 2002). An organizational mense virtual network of websites connect- field is a community of organizations that ed by hyperlinks (Albert, Jeong, and engage in common activities and are sub- Barabási, 1999; Jeong et al, 2000; New- ject to similar reputational and regulatory man, 2001; Watts and Strogatz, 1998). Al- pressures (DiMaggio and Powell, 1983). bert and Barabási (2002) and Newman Such fields have been defined as “a net- (2003) provide excellent overviews of this work, or a configuration, of relations be- burgeoning literature on the network topol- tween positions” (Bourdieu, 1992), and as ogy of different fields, highlighting key or- “centers of debates in which competing in- ganizing principles that guide interactions terests negotiate over the interpretation of among the component parts. key issues” (Hoffman, 1999:351). Fields In the social sciences, however, analy- emerge when social, technological, or eco- ses of fields and networks have been oddly nomic changes exert pressure on existing disconnected. We say oddly because the relations, and reconfigure models of action study of the macro dynamics of networks and social structures. But despite the rela- should be central to the understanding of tional focus on how different actors and or- how fields evolve. This lack of connection ganizations constitute a recognized arena of is rooted in several features of contempo- social and economic activity, studies of rary research. There is an abundance of re- fields have not analyzed the interactions of search in network analysis on why ties multiple, overlapping networks or the regu- form between two actors and what the con- lated reproduction of network ties through sequences are of having a particular posi- time. This linkage between network dy- tion in a network. Salancik (1995) ob- namics and the evolving structure of fields served, however, that most network re- needs to be made in order to make progress search has taken an individual-level per- in explaining how the behavior of actors or spective, and missed out on the opportunity organizations of one kind or another influ- to illuminate the structure of collective ac- ence the actions of organizations of another tion. McPherson et al (2001) note that there kind. are few studies that employ longitudinal The goal of this paper is to account for data to analyze networks. Burt (2000) has the development and elaboration of the voiced a similar concern that most studies commercial field of biotechnology, show- of network structure are cross-sectional. In ing how the formation, dissolution, and 4
  5. 5. rewiring of network ties over a twelve-year lows us to analyze the nearly 2,800 nodes period, from 1988 to 1999, has shaped the in our sample, and to identify cohesive sub- opportunity structure of the field.1 By map- sets such as multi-connected components ping changing network configurations, we (White and Harary 2001: 12-14). We discern how logics of attachment shift over present a small selection of these network time, and chart multiple influences on the visualizations to highlight both the evolv- varied participants in the field. Our effort is ing topology of the field and the processes part of a more general move in the social by which new ties and organizations are sciences to analyze momentum, sequences, added. (The full ‘movie’, with year-to-year turning points, and path dependencies (see representations of the topology and new Abbott, 2001, for an overview). By linking additions to the network, is available at network topology and field dynamics, we http://www-personal.umich.edu/~jdos/paj_mov.html consider social change not as an invariant for viewing on the web. We then turn to a process affecting all participants equally, statistical examination of network forma- but as reverberations felt in different ways tion and dissolution, and assess the effects depending on an organization’s institution- of alternative mechanisms of attachment. al status and location in the overall network Using McFadden’s discrete choice model as that structure evolves over time. Our aim (McFadden 1973; 1981), a variant of the is to illuminate how patterns of interaction conditional logit model, we test to see if the emerge, take root, and transform, with ram- basis of attraction is accumulative advan- ifications for all of the participants. tage, similarity, follow-the-trend, or diver- We develop arguments concerning how the sity. topology of a network and the rules of at- tachment among its constituents guide the The Topology of Large-Scale Networks choice of partners and shape the trajectory A variety of researchers in physics and of the field. As organizations enter an arena sociology are studying the structure of and relationships deepen and expand, sig- large-scale networks with the intuition that nificant structural changes occur. To ana- complex adaptive systems evince organiz- lyze and understand these emergent net- ing principles that are encoded in their work structures, we use a triangulation of topology. Large-scale networks typically methods. We first analyze the expansion of have characteristic signatures of local the network to see if the process is random structure, such as clustering, and a global or uniform. Prior research suggests that as structure, such as average distance between new organizations join the network, there is nodes. Local and global characteristics of an attachment bias of a higher probability networks help to define network topologies of being linked to an organization that al- such as small worlds, which are large net- ready has ties (de Solla Price, 1965, 1980; works with both local clustering and rela- Barabási, 2002a). We go further and assess tively short global distance. Watts (1999) whether other attachment processes are op- showed that adding only a handful of re- erative as well. We map the development mote links to a large network where the of the field by drawing network configura- level of local clustering is high (e.g., tions to create a framework with which to friends of friends are friends) is sufficient view network dynamics. Pajek (de Nooy, to create a small-world network. Watts and Mrvar, and Batagelj, forthcoming) is our Strogatz (1998) helped to revitalize the ear- software package of choice for the repre- lier line of research introduced by Milgram sentation of network dynamics. Pajek al- (1967) and developed by White (1970). The wide appeal of the small-world idea 5
  6. 6. had been portrayed in the arts, in John power-law tails of degree distributions are Guare’s play Six Degrees of Separation, present in very diverse kinds of networks. and in the popular Kevin Bacon game, In the movie actor network, for example, where virtually every Hollywood actor is new actors tend to start their careers in sup- linked through a few steps. Even a small porting roles accompanying famous actors, proportion of randomly distributed ties can and in science, new publications cite well- knit together diverse clusters of nodes to known papers. Attachment bias in network produce the small world phenomenon. Re- formation bears strong similarity to the searchers have applied the small-world more general phenomenon of accumulative concept to a wide range of activities, in- advantage (Merton, 1973), in which those cluding scientific collaborations (Newman, who experience early success capture the 2001) and corporate board interlocks lion’s share of subsequent rewards. We use (Kogut and Walker, 2001; Davis, Yoo, and early-starter accumulative advantage and Baker, 2003). preferential attachments as baseline argu- Watts and Strogatz’s (1998) formaliza- ments, since they provide potential expla- tion of the small-world problem lacked any nations for growing inequalities in the pro- role for network hubs, which are nodes with cess of network expansion. But not all ear- an unusually large number of ties or edges, ly entrants turn out to be winners, and some in the language of graph theory. This limita- latecomers attain prominence. As the say- tion was also true for early models of ran- ing goes, the early bird may catch the dom networks, in which an equal probability worm, but it is always the second mouse of any given pair of nodes being connected that eats the cheese. Similarly, Albert and generated only a mild tendency for nodes to Barabási’s (2002) formalization of a differ in their number of edges.2 But re- “scale-free” class of networks, where the search on the degree distributions of citation probability that a new entrant will choose networks, however, has shown highly to link to an incumbent node is proportion- skewed distributions, with most nodes hav- al to the number of links it has already, is ing few links, while a handful of nodes have elegantly simple but quite possibly over- an exceedingly large number of ties. Lotka generalized. Other attachment processes, or (1926) and de Solla Price (1965, 1980) a combination of diverse processes, can showed that for the tails of degree distribu- produce power law degree distributions. tions of citation networks, the proportion of Our analyses test for multiple, potential nodes with degree k often varies as a func- probabilistic biases in the processes of net- tion of 1/kα, that is, by the inverse power law work growth. P(k) ~ 1/kα, where alpha is the power coeffi- We enter the discussion of network dy- cient. Barabási and Albert (1999:510) and namics with data from a field where social, Barabási (2002b:70) popularized the term political, economic, and scientific factors ‘scale-free’ for such networks, 3 and con- loom large in shaping patterns of attach- firmed that network growth with preferential ment among the participants. In earlier attachment according to degree4 predicts a work on the biotechnology industry, Pow- scale-free tail of the degree distribution. The ell, Koput, and Smith-Doerr (1996) found a well-connected nodes that newcomers attach liability of disconnectedness, in which old- to become hubs that create short paths be- er, less-linked organizations were the most tween many pairs of nodes in the network. likely to fail. Certainly, early entrants have Preferential attachment to higher degree more time than later arrivals to establish leads to a dynamic of rich-get-richer and connections, but Powell et al (1996) found 6
  7. 7. that how connections were established and To illustrate the questions we are pur- what activities were pursued was critical. suing, consider a contemporary dance club, Biotechnology firms had to both make where revelers compete to get inside and news and be in on the news, that is, they once inside, may dance in groups or with needed to generate novel contributions to only one partner or many partners during the evolving science as well as have the ca- an evening. The mix of available partners pability to evaluate what other organiza- changes as the evening goes on, and di- tions were doing. The pathway to centrality verse styles of music are played in different in the industry network was through re- rooms. While new partners may be chosen, search and development collaborations. the imprint of past choices often lingers. Other routes were either ineffective or Some dancers may be highly sought after much slower in generating centrality. and some music may attract more dancers. Moreover, in a highly competitive world in Or, by way of contrast, consider a much which it is not easy to rest on past accom- more formal setting, such as an early 20th plishments, firms that do not expand or re- century Swedish military ball, where the new their networks lose their central posi- young officers would ride in a horse-drawn tions. At the same time, resource-rich par- carriage around Stockholm in advance of ticipants are more capable of altering their the event with an official dance card, and positions, by reconfiguring their networks. visit the homes of young women to obtain Biotechnology is characterized by a permission to dance with them from their high rate of formation and dissolution of parents. By the time of the dance, the aspir- linkages. Connections are often forged with ing young officers had filled their dance a specific goal in mind, such as taking a card and rehearsed their repertoire of con- company public or selling and distributing versation and dance.5 In such settings, an a new medicine. Once the task is complet- analytically rich set of questions follows: ed, the relationship is ended, and successful who dances which dance with whom and collaborators depart gracefully. There is a when? To address these questions, one good deal of entry and exit into the field, needs information about the cast of partici- with new entrants joining at particular pants, the repertoire of activities per- times when financing is available and novel formed, and the sequences linking partners scientific opportunities can be pursued. The and activities. As the combinations of part- rate at which new nodes appear in the net- ners and dances unfold, collective dynam- work is, in part, determined by the success ics emerge. Individual choices may cumu- that existing nodes have in making late into a cascade, resulting in everyone progress on a technological frontier. More- following similar scripts. Or trends may over, many of the participants in the field cluster and find coherence only in small are ‘multi-vocal’, that is, they are capable densely connected groups. Choices made of performing multiple activities with a va- early may strongly affect subsequent op- riety of constituents (Burt, 1992; Padgett portunities, but path dependence can be and Ansell, 1993; White, 1985; 1992). But offset by a constant flow of new arrivals multi-vocality is not distributed evenly, and departures. The challenge to under- those organizations that are more centrally standing any such highly interwoven sys- located in the industry have access to more tem is to relate the behavior and dynamics sophisticated and diverse collaborators, and of the entire structure with the properties of have developed richer protocols for collab- its constituents and their interactions, and oration (Powell et al, 1996). to discern what types of actors and relation- 7
  8. 8. ships are most critical in shaping the evolu- A preference for diversity, however, tion of the field at particular points in time. suggests a search for novelty, and the incli- We assess different sources of attach- nation to move in different communities ment bias and test to see if these simple and interact with heterogeneous partners. rules guide the process of partner selection, Our ideas concerning multiconnectivity in- and if so, for which participants and at volve several intuitions. A cohesive net- what points in time? We supplement the work, with plural pathways, means partici- idea of accumulative advantage with alter- pants are connected through different link- native mechanisms that sociologists have ages. Thus many nodes must be removed to repeatedly found to be important in the for- disconnect such a structure, meaning such mation of social and economic ties and the groupings are more resilient. The more evolution and replication of social struc- pathways for communication and ex- tures. The first alternative to early advan- change, the more rapidly news percolates tage or rich-get-richer is homophily through the network. In turn, when more (McPherson and Smith-Lovin, 1987), a knowledge is exchanged, participants at- process of social similarity captured best in tend to their network partners more inten- the phrase, ‘birds of a feather flock togeth- sively (Powell, 1990). The enhanced flow er’. A second alternative is based on fol- of ideas and skills then becomes an attrac- lowing the trend, thus the participants ob- tion, making the network more appealing to serve others and attempt to match their ac- join. Rapid transmission and diverse partic- tions to the dominant behavior of the over- ipants enhance both the likelihood of re- all population (White, 1981; DiMaggio and combination and the generation of novelty. Powell, 1983). In this context, action is In the language of organizational learning, triggered by a sense of necessity, by a de- diversity entails a preference for explo- sire to keep pace with others by acting ap- ration over exploitation (March, 1991). Of propriately (March and Olsen, 1989; course, we do not necessarily expect that 160-2). This pattern may also arise from one mechanism dominates at all time peri- participants reacting in similar ways to ods and exerts equal gravitational pull on common exogenous factors. every participant. The very essence of dy- We contrast arguments based on either namic systems is that they change continu- rich-get-richer or processes such as ho- ally over time. The actors may well play by mophily, and a logic of appropriateness different rules at different points in time, with a model based on multiconnectivity depending on the experience of their part- (the multiple linking of partners both di- ners and their position in the social struc- rectly and through chains of intermediaries) ture. Moreover, alternative organizing prin- and a preference for diversity. To pursue ciples may be dominant at different stages the dance imagery, homophily suggests in the formation of the network. Framed that when you select a new partner, he or more formally, the alternative mechanisms she is someone with attributes similar to can be stated as: those with whom you have already danced. H1: Network expansion occurs A rich-get-richer process involves compet- through a process in which the most- ing for the most popular dancer. Following connected nodes receive a dispropor- the trend entails choosing both a partner tionate share of new ties (Accumulative and a dance that are comparable to the Advantage). choices of most other participants. H2: Network expansion follows a process in which new partners are cho- 8
  9. 9. sen on the basis of their similarity to changed accordingly, co-evolving with the previous partners (Homophily). collaborative network. H3: Network expansion entails In the early years of the industry, from herd-like behavior, with participants 1975-87, most DBFs were very small star- matching their choices with the domi- tups, and deeply reliant on external support nant choices of others, either in mutual out of necessity. No DBF in this period had response to common exogenous pres- the necessary skills or resources to bring a sures or through imitative behavior new medicine to market; thus they became (Follow-the-trend). involved in an elaborate lattice-like struc- H4: Network expansion reflects a ture of relationships with universities and choice of partners that connect to one large multinational firms (Powell and another through multiple independent Brantley, 1992). The large multinational paths, which increases reachability and firms, with well established internal career the diversity of actors that are reachable ladders, lacked closeness to the cutting (Multiconnectivity). edge of university science. Lacking a knowledge base in the new field of molecu- Field Structuration: Science Meets Com- lar biology, the large companies were merce drawn to the startups, which had more ca- Our empirical focus is on the commer- pability at basic and translational science cial field of biotechnology, which devel- (Gambardella, 1995; Galambos and Stur- oped scientifically in university labs in the chio, 1996). This asymmetric distribution 1970s, saw the founding of hundreds of of technological, organizational, and finan- small science-based firms in the 1980s, and cial resources was a key factor in driving matured in the 1990s with the release of early collaborative arrangements in the in- dozens of new medicines. This field is no- dustry (Orsenigo, 1989; McKelvey, 1996; table for its scientific and commercial ad- Hagedoorn and Roijakkers, 2002). vances and its diverse cast of organizations, Many commentators at the time argued including universities, public research insti- that these interdependent linkages were tutes, venture capital firms, large multina- fragile and fraught with possibilities of tional pharmaceutical corporations, and “hold-up”, in which one party could oppor- smaller dedicated biotech firms (which we tunistically hinder the other’s prospects for refer to as DBFs). Because the sources of success.6 Some analysts argued that the scientific leadership are widely dispersed field would undergo a “shakeout,” with and rapidly developing, and the relevant large pharmaceutical companies asserting skills and resources needed to produce new dominance and the rate of founding of new medicines are broadly distributed, the par- firms slowing to a trickle (Sharpe, 1991; ticipants in the field have found it neces- Teece, 1986). But as these observers and sary to collaborate with one another. The others came to recognize, a shakeout did evolving structure of these collaborative not occur, nor did cherry-picking of the ties are the focus of our network study. most promising firms by larger companies Concomitant with changes in the network, prove viable.7 Instead, the ensuing period an elaborate system of private governance saw the give-and-take, mutual forbearance has evolved to orchestrate these interorga- of relational contracting (Macneil, 1978) nizational relationships (Powell, 1996), and become institutionalized as a common the internal structure of organizations has practice in this rapidly developing field. By the late 1980s, some of the dedicated 9
  10. 10. biotech firms had become rather large and search institutes at the forefront of basic formidable organizations in their own right, science are necessary (Orsenigo, Pammolli, while many of the big pharmaceuticals cre- and Riccaboni, 2001). The high rate of ated in-house molecular biology research technical renewal is reflected in patent programs (Henderson and Cockburn, 1996; data. Figure 1 shows the brisk rise of life Zucker and Darby, 1997). So even as mutu- science patenting, and highlights the simi- al need declined as a basis for inter-organi- larities in the technological trajectories of zational affiliations, the pattern of dense universities and biotech firms.8 Note the connectivity deepened, suggesting the orig- parallel climb of university and DBF inal motivation of exchanging complemen- patenting in the mid-1980s, a steady ascent tary resources had changed to a broader fo- for the next ten years, and a steep increase cus on utilizing innovation networks to ex- in the late 1990s. Universities start out plore new forms of R&D collaboration and ahead and then are passed by DBFs in product development (Powell et al, 1996). 1997, but the more important point is the This is the period we focus on with data extent to which both become members of a from 1988-99. Table 1, which lists the top- common technological community (Owen- selling biotechnology drugs in 2001, illus- Smith and Powell, 2001a and b). This joint trates the division of innovative labor that membership in a community greatly in- has typified the field. All ten drugs were creases the frequency of interaction be- developed by biotech firms, but only five tween universities and industry. were marketed by biotech firms, and just [FIGURE 1 HERE] four by their originator. In the other five The availability of funding has also in- cases, a large pharmaceutical company creased markedly as biomedicine has be- handled or guided the marketing in return come a major force in modern society. The for a hefty share of the earnings. Compara- total budget of the U.S. National Institutes ble data from the early 1990s indicate that of Health, a key funder of basic research marketing power and control of revenues that allocates approximately 80% of its was dominated overwhelmingly by the budget to external research grants to uni- pharmaceutical giants (Powell and Brant- versities and firms, nearly doubled under ley, 1996). A notable feature of drug devel- the Clinton administration, going from $8.9 opment is that there is no consumer loyalty billion in 1992 to $17.08 billion in 2000. to a company, and limited brand loyalty as The NIH plays a highly significant role in well. Combine these influences with a mar- fostering exploration and variety on the re- ket structure that has many winner-take- search front. Internal R&D expenditures by most features, and the outcome is a biotech and pharmaceutical companies volatile, fast-changing field. have also ramped up, from $6.54 billion in [TABLE 1 HERE] 1988 to $26.03 billion in 2000.9 Venture A number of factors undergird the col- capital disbursements, or seed money to laborative division of labor in the life sci- biotech startups, have flowed into biotech, ences. No single organization has been able but more irregularly as the public equity to internally master and control all the markets have windows of opportunity competencies required to develop a new when particular technologies are in vogue. medicine. The breakneck pace of technical As Lerner, Shane, and Tsui (2003) note, advance has rendered it difficult for any or- unexpected events affecting a single firm – ganization to stay abreast on so many notably the rejection or delay of a drug can- fronts, thus linkages to universities and re- didate by the U.S. Food and Drug Adminis- 10
  11. 11. tration, can have pronounced effects on all ment specializes in R&D, venture capital in firms’ stock prices and ability to raise capi- finance. Some organizations, however, are tal. Consequently, venture funding of able to shift their attention. While enlarging biotech is rather episodic, reaching $395.5 the number of ties, both dedicated biotech million in 1988, declining over the next firms and public research organizations three years, then jumping to $586.4 million broadened their range of activities as well. in 1992, remaining around the half billion The lower figure, which reports the per- level for next four years, then climbing to centage of activity by organizational form, $1.1 billion in 1997, and staying above $1 illustrates the pattern of specialization by billion in 1998 and 1999.10 Biotech financ- government and VCs, and the diversifica- ing by venture capital is also somewhat tion by biotech and pharmaceuticals, while countercyclical, thus when there was great PROs display a trend toward a R&D and li- enthusiasm for internet and telecom star- censing model. Looking at types of activi- tups, interest in biotech waned. In recent ty, VCs come to dominate finance, as phar- years, with the burst of the internet bubble maceuticals master commercial ties, while and a precipitous decline in telecommuni- licensing and R&D are pursued by an array cations, biomedical support has been on the of participants. upswing. Biotech firms that are well posi- [FIGURE 2 HERE] tioned in the network with connections to Finally, as the field gained coherence, basic research funding, industrial R&D and the pattern of reliance on collaboration support, and venture capital financing, are proliferates, institutions emerged to both not only able to obtain money from multi- facilitate and monitor the process. Offices ple sources, they also develop the capabili- were established on university campuses to ty to interact with varied participants. promote university technology transfer These experiences facilitate organizational (Owen-Smith and Powell, 2001a), law learning and expand the scope and depth of firms developed expertise in intellectual an organization’s knowledge base. property issues in the life sciences, and The different members of the field have venture capital firms provided financing, varying catalytic abilities and competen- along with management oversight and re- cies. Some of the participants are quite spe- ferrals to a host of related businesses. As cialized, while others have a hand in multi- these relations thickened and a relational ple activities. Figure 2 provides simple contracting infrastructure grew (Powell, count data to illustrate the relationship be- 1996), the reputation of a participant came tween functional activity and organization- to loom larger in shaping others’ percep- al form, and suggests how the correspon- tions. Robinson and Stuart (2002) argue dence of activity and form has shifted over persuasively that the network structure of time. The top figure emphasizes an overall the field becomes a “platform for the diffu- pattern of expansion in number of ties, and sion of information about the transactional for three of the forms of organization, a integrity” of its participants. Centrality in branching out in terms of activity. Growth the network increases the visibility of a and diversity go hand-in-hand for DBFs, participant’s actions and, they demonstrate, PROs, and large pharmaceutical compa- reduces the need for more overt, contractu- nies. Venture capital growth is notable as al forms of control, such as an equity stake well, but as the percentage figure on the or dominance on the board of directors. We bottom reveals, there is a strong co-occur- turn now to a discussion of the database we rence of some forms and activities: govern- 11
  12. 12. have developed to map the evolving struc- database; 91 (of the 482) exited, due either ture of this field. to failure, departure from the industry, or merger. The database, like the industry, is DATA AND METHODS heavily centered in the U.S., although in re- Our starting point in developing a sam- cent years there has been considerable ex- ple is BioScan, an independent industry di- pansion in Europe. BioScan reports infor- rectory, founded in 1988 and published six mation on a firm’s ownership, financial times a year, which covers a wide range of history, formal contractual linkages to col- organizations in the life sciences field.11 laborators, products, and current research. Our focus is on dedicated biotech firms Firm characteristics reported in BioScan in- (DBFs). We include 482 companies that clude founding data, employment levels, fi- are independently operated, profit-seeking nancial history, and for firms that exit, entities involved in human therapeutic and whether they were acquired or failed. The diagnostic applications of biotechnology. data on interorganizational agreements cov- We omit companies involved in veterinary er the time frame and purpose of the rela- or agricultural biotech, which draw on dif- tionship. Our database draws on BioScan’s ferent scientific capabilities and operate in April issue, in which new information is quite different regulatory climates. The added for each calendar year. Hence the sample of DBFs covers both privately-held firm-level and network data are measured and publicly-traded firms. We include pub- during the first months of each year. We licly held firms that have minority or ma- define a collaborative tie, or alliance, as jority investments in them by other firms, any contractual arrangement to exchange or as long as the company’s stock continues to pool resources between a DBF and one or be independently traded. We exclude orga- more partner organizations. We treat each nizations that might otherwise qualify as agreement as a tie, and code each tie for its DBFs, but are wholly-owned subsidiaries purpose (e.g., licensing, R&D, finance, of pharmaceutical or chemical corpora- commercialization) and duration. Some ties tions. Large pharmaceutical corporations, involve multiple stages of the production health care companies, hospitals, universi- process. All such ties include commercial- ties, or research institutes enter our ization activities, such as manufacturing or database as partners that collaborate with marketing, hence we code complex agree- dedicated biotech firms. Our rationale for ments as commercial ties. We say a con- excluding both small subsidiaries and nection, or link, exists whenever a DBF large, diversified chemical, medical, or and partner have one or more ties between pharmaceutical corporations in the primary them. (DBF) database is that the former do not We seek to explain the processes that make decisions autonomously, while attract two parties to one another, the biotechnology may represent only a minor- evolving patterns of tie formation and dis- ity of the activities of the latter. Their ex- solution, and the overall structure of the clusion from the primary sample of DBFs network. We code the dominant forms of eliminates serious data ambiguities. partner organizations into six categories, The primary sample covers 482 DBFs representing those that populate the field: over the 12-year period, 1988-99. In 1988, public research organizations (including there were 253 firms meeting our sample public and private universities and nonprof- criteria. During the next 12 years, 229 it research institutes and research firms were founded and entered the hospitals); large multinational pharmaceuti- 12
  13. 13. cal corporations (as well as chemical and may involve as little as one or two hundred diversified health care corporations); gov- thousand dollars. The salience of these lim- ernment institutes (such as the National itations recedes as we add more years of Cancer Institute or the Institut Pasteur); fi- observations to the data set. The advan- nancial institutions (principally venture tages of twelve years of fine-grained data capital as well as banks and insurance com- reside in capturing the length of relation- panies); other biomedical companies ships, the dissolution of ties to particular (providers of research tools or laboratory partners and the forging of ties to others, as equipment); and those DBFs that collabo- well as the deepening of some ties. Issues rate with other biotech companies. There of scale are held constant while we exam- are more than 2,300 non-DBFs in the part- ine duration of ties and the extent to which ner database. the parties involved in a relationship share The four types of ties involve different other partners in common at specific points activities, ranging from basic research to fi- in time. This approach allows us to speak nance to licensing intellectual property to to Salancik’s (1995: 348) concern that net- sales and marketing. Thus, the matrix of or- work analysis should show how adding or ganizational forms and activities is 6 x 4, or subtracting a particular interaction in a net- 24 possible combinations of partner forms work changes the coordination among the and functional activities of ties. Some of participants, and either enables or discour- the cells are quite rare, but there are cases ages interactions between parties. in every cell. Some of these activities in- We do not collect data on the ties volve the exchange or transfer of rights, among the non-DBF partner organizations. while others require sustained joint activi- In some cases, such ties would be very ty. The latter obviously entails more inte- sparse or non-existent (e.g., venture capital gration of the two parties to a relationship. funding of universities or pharmaceutical This difference is one reason why we treat companies); in other cases, they are more the four types of activities separately in common (e.g., pharmaceutical support of most analyses. clinical trials conducted at a university Given the differentiation of organiza- medical center).12 The practical problem is tional forms and types of ties, our approach that the data on a network of 2,310 x 2,310 has some limitations. In some, though not disparate organizations would be very diffi- all, of our measures, we treat the type of tie cult to collect. Thus, we analyze the con- or the form of the partner as equally impor- nections that DBFs have to partners, and tant. Obviously, this is not altogether real- the portfolio of DBFs with whom each istic; indeed, if the analyses were based on partner is affiliated. To do this we use k- only a single year of data, this limitation components to identify cohesive subsets of would loom large. The benefit of this as- organizations.13 This measure of multicon- sumption is that it permits comparisons nectivity does not require complete data on across time periods. There is also hetero- relations among non-biotech partners, thus geneity in the nature of participation of dif- allowing us to analyze the total network of ferent organizational forms for a particular nearly 2,800 organizations. Our focus, type of activity. At the extreme, an R&D then, is on cohesion, mediated exclusively partnership between a global pharmaceuti- by ties with DBFs. cal company and a DBF may reach the Network measures such as the analysis scale of $1 billion dollars, while a DBF’s of degree distributions, unlike k-compo- R&D alliance with a university laboratory nents, require complete data. To this end, in 13
  14. 14. the first and third of the analyses that fol- connected. The fat tail contains the hubs low, we separate our database into two of the network with unusually high con- parts: the complete one-mode network (482 nectivity. x 482) of ties among DBFs, where we also Different types of degree distributions have extensive data on the attributes of the can be distinguished when plotted on a 482 biotech firms, and the two-mode net- log-log scale, with log of degree on the x work that consists of complete data on the axis and log of the number of nodes with ties of DBFs to non-DBF organizations. this degree on the y axis. The degree dis- Over the past decade, we have also in- tribution for a network in which the for- terviewed more than 200 scientists and mation of edges is governed by a popular- managers in biotech and pharmaceutical ity bias, where nodes with more connec- firms, as well as university professors who tions have a higher probability of receiv- are actively involved in commercializing ing new attachments, would plot as a basic research. Members of our research straight line on the log-log graph, indica- group have done participant observations in tive of a power law.14 A power-law de- university technology licensing offices, gree distribution is not sufficient by itself, biotech firms, and large pharmaceutical however, to identify the actual mecha- companies. Students working on the nisms that facilitate tie-formation. A pow- project have developed a large data set on er-law degree distribution can reflect not the founders of biotech companies, and an- only preferential attachment by incum- alyzed the careers of scientists who joined bency (degree of attracting node) but biotech companies. In short, while the anal- preferences for attractiveness, legitimacy, yses presented here are based on data de- diversity, or a concatenation of mecha- rived from industry sources, our intuition nisms and still exhibit a power-law degree about the questions to address are grounded distribution. In our study, we test for the in primary data collection. existence of a preferential attachment pro- cess for each year to the next rather than Analysis I: Degree Distributions assume its existence. We use multivariate Research on networks in the graph analysis to try to discern specific mecha- theoretic and statistical mechanics tradi- nisms that govern tie-formation and the tions often utilizes degree distributions as type of degree distributions that different a diagnostic indicator of whether tie-for- substantive processes generate. Before we mation in a network (growth or replace- begin that analysis, however, it is useful ment) is equiprobable (simple random) to examine the degree distributions. for all pairs of nodes or biased propor- In Figure 3, we plot the aggregate de- tional to existing ties of potential partners. gree distributions of DBFs on log-log The degree of each node is measured as scales for the six types of partners for the number of other nodes directly con- biotech firms: other DBFs; pharmaceutical, nected to the focal node. Preferential at- chemical, and diversified health-care cor- tachment to already connected nodes is porations; universities, non-profit research referred to as a popularity bias. Unlike the institutes and hospitals; government agen- tail of a random bell curve whose distri- cies; venture capital firms and other finan- bution thins out exponentially as it de- cial entities such as banks; and biomedical cays, a distribution generated by a popu- companies that supply research tools and larity bias has a “fat” tail for the relatively instruments. For each of the six plots, the greater number of nodes that are highly x-axis reflects log-degree (aggregated over 14
  15. 15. all time periods) and the y-axis the log of other substantive processes are operating. the number of partners of the plotted form These static snapshots of the degree dis- having x degree of attachments to DBFs tributions, however, do indicate the im- (also aggregated over all time periods). A portance of different organizational forms power-law distribution, as noted, would varies with respect to patterns of affilia- plot as a straight line. For a network of suf- tion. To explore the dynamics of the field ficient density, an exponential-decay de- and the changing impact of different orga- gree distribution, mimicking the results of a nizations, we next present a series of simple random process of tie-formation, force-directed network visualizations, fol- would form a convex curve that bows to lowed by a statistical examination of the the upper right away from the origin in the actual attachment processes. log-log plot. An exponential distribution can be statistically rejected from the data Analysis II: Discrete Time Network Vi- presented in Figure 3, along with a simple- sualizations random attachment process that would gen- We utilize Pajek, a freeware package erate degree distributions that decay expo- for the analysis and visualization of net- nentially.15 works, to present a series of discrete-time The least-squares fitted linear slopes for images of the evolution of the biotechnolo- the log-log plot of the degree distributions gy network. Pajek employs two powerful in Figure 3 are in the expected range for minimum energy or ‘spring-embedded’ power-laws, between 1.1 and 2.7, over four network drawing algorithms to represent decades of variation in degree. Only the left network data in two-dimensional Euclidian or low-degree half of one of the distribu- space. These algorithms simulate the net- tions, that for government agencies, has a work of collaborations as a system of inter- slope anywhere near 3, as expected from acting particles, in which organizational pure preferential attachment for large net- nodes repel one another unless network ties works (Barabási and Albert 1999, Bollobás act as springs to draw connected nodes and Riordan 2002). This government de- closer together. Spring-embedded algo- gree distribution is interesting because of rithms iteratively locate a representation of the highly pivotal central role of the Na- the network that minimizes the overall en- tional Institutes of Health, which is a key ergy of the system, by reducing the dis- funder of basic research. The NIH is the tance between connected nodes and maxi- most active partner in the entire network. mizing the distance between unconnected That the slopes of the five other distribu- nodes. (For more detail concerning Pajek, tions are considerably less than 3 may be see the Appendix.) an indication that processes other that pure We generate two sets of images for preferential attachment are operative.16 the time period covered by our database. Although these degree distributions To simplify the presentations, we include are aggregate measures over all time peri- only those organizations in the main com- ods, they give some hint of growth pro- ponent each year, thereby removing the cesses in attachment. The shape of the isolates from this large and expanding net- distributions mimics what would be ex- work.17 The first set of images presents the pected for tie-formation in the biotech full collaboration network, while the sec- network from a process of preferential at- ond set represents the new ties added each tachment to degree, but power-law slopes year. Given space constraints, we select a that are closer to 1 than to 3 suggest that subset of these pictures for inclusion in the 15
  16. 16. paper, but we urge readers to view should reproduce the previous year’s pat- http://www-personal.umich.edu/~jdos/paj_mov.html tern of affiliation. If, for instance, R&D to see the complete set. The visualizations (red) ties dominated in a prior year, neo- afford a multi-faceted view of network evo- phytes would generally enter the network lution, including growth in the number of through this established route and the im- participants, changes in the purpose of ties, ages would be fields of red. A preference and the formation of new ties. for diversity of partners and pathways im- The Pajek figures are designed to visu- plies less visual coherence than the other ally reflect our hypothesized models of at- mechanisms. Multiconnectivity coupled tachment. The graphics cannot adjudicate with heterogeneous activities would show between the various models, which we do clusters in which all four colors would be in statistical analyses below, but they do evident in the springs. The nodes at the provide suggestive evidence, or existence center of the network should be different proofs, if you will. Consider what the full colors, reflecting the array of organization- network and new tie visualizations might al types. Some small nodes should be no- look like if strong versions of our four ticeable at the center of the field, while mechanisms were operating. After a sketch some large nodes should have few new of these stylized images, we then turn to partners. New entrants should be sprinkled the data and examine a series of images. If throughout the network. Empirical reality an accumulative advantage process drives is, of course, more messy than stylized attachment, the spring-embedded images models. We use these thought experiments would display a preponderance of star- to make the detailed visualizations of this shaped structures attached to large nodes. large database more interpretable to readers These stars would be positioned at the cen- not familiar with graphical representations ter of the image and continue to grow of network dynamics. rapidly over time as newcomers (triangles) Figure 4, showing the main component affiliate with the most connected nodes. in 1988, serves as the starting point. The We scale the size of nodes to represent the color of the nodes reflects their organiza- number of ties an organization has. Very tional form, with light blue a dedicated few small nodes will have numerous new biotech firm, yellow a large pharmaceuti- partners, nor will many large nodes be situ- cal, chemical, or diversified medical corpo- ated on the network’s periphery. In con- ration, brown a government institute or trast, if homophily strongly conditions tie agency. In subsequent years gray nodes be- formation, then we would anticipate images come important, reflecting the growing im- differentiated into coherent and loosely in- print of venture capital. The springs are ter-connected single-color clusters. These colored according to the functional activity homophilous clusters should be fairly reflected by a tie, with red a research and dense and dominated by characteristic or- development partnership, magenta a licens- ganizational forms without necessitating a ing agreement, green a financial relation- preponderance of stars tethered to large ship, and dark blue indicating an alliance central nodes. New entrants (pictured as tri- involving one or more stages in the com- angles) would move into the neighbor- mercialization process, ranging from clini- hoods that most closely matched their type cal trials to manufacturing to sales and and profile. marketing. Node size is scaled to standard- Were a follow-the-trend logic domi- ized network degree in the total network, nant, new ties would be overwhelmingly reflecting variation in the extent of degree- uniform in color and the predominant color 16
  17. 17. connectivity among the organizations. To pability are highly sought for collaboration. return to our earlier metaphor of a dance, The large multi-connected nodes in the the representation captures dancers with center of the representation are a small different identities, who may participate in group of established, first-generation different types of dances with one or many DBFs, major multinational firms, and gov- partners. The size of a node, given the pow- ernment institutes (the NIH and the Nation- er-law tendencies in the degree distribu- al Center Institute). In the pullout to the tions, reflects, roughly, a rich-get-richer right of the figure, we identify a handful of process. In the images that follow, particu- the largest nodes: NIH, which will serve as larly the visualizations of new ties, there an orienting node in the full network fig- are several large blue DBF nodes that are ures because of its centrality, and NCI; a clearly the most attractive stars of the net- group of first-wave biotechs founded in the work. In later years, there are nodes that 1970s and early 1980s – Genentech, Cento- have multiple linkages for different activi- cor, Amgen, Genzyme, Biogen, and Chi- ties, reflecting a preference for diversity. ron, which became the largest and most Looking at the overall population rather visible firms by 1988; and three multina- than specific nodes, we observe shifts in tional firms – Kodak, Johnson and John- the dominant activities as well as changes son, and Hoffman La Roche. The close in the composition of the nodes, which il- proximity of the Swiss firm Hoffman La lustrate the overall trends in the field. Roche to Genentech is interesting because [FIGURE 4 HERE] in 1990, the Swiss firm became the majori- Several key features stand out for 1988 ty stockholder of Genentech. Kodak and in Figure 4. The predominant color is blue, J&J reflect the broad interest in biotech by and the most active participants are biotech a range of large firms in different indus- firms, pharmaceutical companies, and sev- tries. Kodak soon drops out of the center of eral government agencies. The strong im- the field, as does J&J, but the latter returns pact of commercialization ties is a clear in- in the late 1990s by acquiring two compa- dication of the dominant strategy of mutual nies, first Centocor and then Alza, both es- need that characterized the industry’s early tablished, well-connected DBFs. Kodak’s years. Biotech firms lacked the capability subsequent loss of centrality and J&J’s pur- to bring novel medicines to market, while chase of two incumbent DBFs illustrates large firms trailed behind in understanding that a first mover, rich-get-richer account new developments in molecular biology does not always hold, even for some of the (Gambardella, 1995; Powell and Brantley, early very resource-rich participants. 1996; Henderson, Orsenigo, and Pisano, In Figure 5, we present the new ties 1999). Finance ties (green) are less preva- added in 1989. To return to the dance lent and only a few venture capital firms metaphor, the music has stopped temporar- (gray) are present, providing further evi- ily, while new partners are being chosen. dence that most DBFs supported them- We add shape to the presentation, with tri- selves by selling their lead product to large angles representing new entrants to the net- corporations, who subsequently marketed work, while circles are the incumbents. the medicine and pocketed the lion’s share Note the very active role of NIH (the of the revenues. Clustered in the center are largest brown node) in forging R&D ties red (R&D) and magenta (licensing) ties, with new entrants, and the appearance of which show that DBFs with significant in- many grey triangles, illustrating the grow- tellectual property and strong research ca- ing involvement of venture capital in fi- 17
  18. 18. nancing DBFs. Node size continues to be have shown elsewhere, the locus of venture scaled for network degree in the full net- capital-financed biotech startups was ini- work, thus graphically representing how tially the Bay Area and Boston, but by the network position in one year may condition end of the 1990s had spread to a number of the addition of new ties in a subsequent key regions in the U.S. and Europe (Pow- year. The size of the node indicates the im- ell, Koput, Bowie, and Smith-Doerr, 2002; portance of rich-get-richer models of at- Owen-Smith et al, 2002). Thus growth in tachment. In the initial years, we see visual the number of new firms, new partners, and affirmation of the positive effect of number new ideas is enhanced by an increase in fi- of previous ties on new tie formation. Note nancial linkages, signaling the important that several large blue DBF nodes are at the role of the public equity markets in foster- center of the new tie network. By contrast ing growth. In contrast, commercialization to the 1988 picture of the overall network, is a more restrictive activity. The ability to the dominance of commercialization manufacture new biomedical products was springs (blue) in 1988 recedes markedly in a relatively scarce skill, as was the ability the 1989 new ties picture. R&D and fi- to market and distribute a new medicine nance are the main avenues generating throughout the world. A relatively small growth in the network. Observe that large number of large firms had these capabili- nodes with just a few ties have little diver- ties, and it took at least a decade before sity in their partners, while some of the DBFs developed these skills. Consequent- smaller nodes adding many new ties have a ly, during the first two decades of the field, wide variety of partners. commercialization ties flowed through a [FIGURE 5 HERE] small set of dominant multinationals and a We fast forward to 1993, but encourage handful of established biotech firms. Thus, readers to follow the annual changes18 in not only were the number of participants the on-line version. Figure 6 portrays a limited, commercialization is a ‘down- large expansion in activity, with green ties stream’ activity; indeed, when it involves (finance) now much more prominent. We the sale of a new medical product, it is the also observe a shift in the composition of ‘last dance’ in the product life cycle. Fi- the most connected participants. Put collo- nance, in contrast, is an ‘upstream’ activity, quially, the music has changed – from which, in turn, fuels R&D, licensing, and commercialization to finance, and accom- commercialization, and thus enrolls more panying the shift in the predominant col- participants in the industry network. laborative activity is both an increase in the The organizational composition of the number of highly connected nodes (there center of the field has shifted as well. Two are roughly three times as many larger research universities, MIT and Harvard, nodes as in 1988) and the march-in of gray along with a handful of leading VC firms (venture capital) and orange (universities) are now at the center. The composition of nodes. This shift in the primary locus of ac- multinational firms shifts from diversified tivity is important on a number of dimen- chemical and medical companies to some sions. Finance, as opposed to commercial- of the giants of the pharmaceutical sector, ization, has a powerful mobilizing effect, e.g., Schering-Plough and Merck. The di- enrolling new types of actors (VCs, and versity of organizational forms at the center subsequently investment banks, pension of the network is notable in that these var- funds, university endowments, etc.) in fi- ied organizations – DBFs, pharmaceuticals, nancing the expansion of the field. As we VCs, research universities, and government 18
  19. 19. institutes – operate in distinct selection en- the 1990s. The picture of all ties in the vironments, subject to very different pres- main component in 1997 (Figure 8) and sures and opportunities. Networks an- 1999 (Figure 10) illustrates a growing chored by diverse organizational forms are number of large nodes, strong expansion of more robust to both failure and attack (Al- collaborative activity and a reaching out to bert, Jeong, and Barabási, 2000). Such di- new entrants, and more varied types of or- versely anchored, multiconnected networks ganizations at the center of the figures. In are less likely to unravel than networks re- 1997 (Figure 8), we see a very cohesive liant on a single form of organization for hub of DBFs, pharmaceuticals, venture their cohesiveness. capital firms, universities, and the NIH [FIGURES 6 AND 7 HERE] complex. The colors of the ties are less dis- The picture of new ties in 1994 (Figure cernable, reflecting the fact that the most 7) reflects growing complexity in the activ- active members of the network are now ei- ity sets of the participants. On the right side ther engaged in multiple activities or con- of the network, finance is very pronounced, nected to DBFs who are. The pharmaceuti- and there are many more gray nodes, which cal sector underwent a period of consolida- are growing in size. But note there are now tion in the mid to late 1990s, as mergers yellow nodes (which are not triangles, so and acquisitions became commonplace. these are not new entrant pharmaceutical Novartis, formed out of the merger of the companies) involved in financing smaller two large Swiss firms Sandoz and Ciba biotech firms. On the left side of the figure, Geigy, and Glaxo Wellcome, the product of we see blue and green ties linked to well- the acquisition of Burroughs Wellcome by connected DBFs. At the center of the net- fellow British giant Glaxo, are in the center work is the NIH, the anchor of R&D activi- of the network. Outside the center there are ty and the largest node, linked to both small several ‘specialist’ DBFs, one with a very as well as large DBF nodes. The picture of active commercialization portfolio and an- new ties in 1994 illustrates the growing other with a licensing cluster. In the new multi-vocality of the industry, with both ties image for 1998 (Figure 9), finance con- well-connected DBFs and pharmaceuticals tinues to be generative in pulling in new developing the capability to finance entrants. For the first time, a handful of younger firms, contribute to basic and clin- pharmaceutical giants, bolstered by a round ical science, and commercialize new of mergers and acquisitions, are central in medicines. The overall picture has shifted the new tie network. While the NIH re- from one in which commercialization and mains in the core, it is no longer clearly the rich-get-richer were the dominant scripts to largest node. Many more DBFs are found one in which finance is generating much in the center and on the edges of the new more diversity in activity and there is more tie network, reflecting the growing pres- heterogeneity in the makeup of the key par- ence of second-generation firms who are ticipants. With greater numbers of new ties active in the field. The consolidation in the added in 1994, the more tree-like structure pharmaceutical sector is producing a rich- of springs in 1989 has now changed to re- get-richer effect among the largest multina- flect multiconnectivity, i.e., a more cohe- tionals, but these ‘survivor’ corporations sive structure even among the new partner- have learned to do more than just commer- ships. cialize the lead compounds of the smaller The density of the field and the number DBFs. The big multinationals have become of participants continue to grow throughout multi-vocal. Meanwhile, a combination of 19
  20. 20. DBFs, universities, and government insti- the diverse types of organizations that are tutes are active in pulling in new partici- driving this connectivity. (There is a fall pants. off in new ties and partner entry in 1999, [FIGURES 8, 9 AND 10 HERE] but this may reflect incomplete reporting in The full network images for 1997 (Fig- 1999 annual reports. As we add more years ure 8) and 1999 (Figure 10) show a notable of data, we will learn whether this is a clustering of financial ties on the right side, downturn or undercount.) The rate of tie with connections to smaller-size DBF dissolution grows, then wanes, then heads nodes. This shift underscores the cyclical up, so there is considerable turnover in in- nature of venture capital, which involves terorganizational relations, reflecting both taking a firm public, ending that relation- successful completion of some projects and ship and moving on to finance new firms. dissolution due to lack of progress on oth- The number of yellow nodes at the center ers. The general picture is one of a continu- has decreased, as consolidation shrinks the ing flow of new entrants into the field, number of big pharmaceutical firms. In alongside the forging of new collabora- turn, the size of each yellow node increas- tions, making for an increasingly dense net- es, as their portfolio of alliances grows and work. To test our hypotheses, as well as the diversifies. Several universities, notably the insights derived from the visualizations, we UC system19 (primarily UCSF and UCSD) turn to a more fully specified analysis in and MIT, are at the center in the 1999 im- which we take organizational variables, age, and the NIH and NCI remain central. network portfolios, and time periods into Recall that the NIH’s budget for R&D account. grew markedly throughout the 1990s, and it [TABLE 2 HERE] remains hugely important as a funder of ba- sic research, and as a participant in licens- Analysis III: Attachment Bias ing the results of intramural NIH research We now turn to a statistical analysis of done. The new tie network in 1998 (Figure attachments between dedicated biotech 9) is the most expansive yet, with more firms (DBFs) and various partner organiza- than 1,100 new ties added in 1998. Multi- tions. For analytic purposes, we distinguish vocality is the dominant pattern here, as between four categories of relationships: nearly every large node is engaged in mul- 1. New 1-mode attachments, in which a tiple types of collaborations. DBF contracts with another DBF as a Pajek visualizations, which we use to partner, and it is the first tie between show the extent of clusterings by type of this dyad. organizations, type of tie, and network de- 2. Repeat 1-mode attachments, in which gree, can be considered to provide a visual a DBF contracts with a DBF partner goodness-of-fit test for our hypotheses. As and it is not the first tie between this a supplement to the graphics, we provide dyad. count data on patterns of entry and exit into 3. New 2-mode attachments, in which a the network. We see in Table 2 that the DBF contracts with a non-DBF part- number of participants, both DBFs and ner and it is the first tie between the partners, grows every year. But the rate of dyad. expansion for total number of ties and new 4. Repeat 2-mode attachments, in which ties outpaces the entry of organizations, a DBF contracts with non-DBF part- suggesting a more connected field. The vi- ner and it is not the first tie between sualizations afford the opportunity to see the dyad. 20
  21. 21. New attachments expand the structure ment is removed from (added to) the risk of the network, whereas repeat ties thicken set for subsequent new (repeat) attach- relations between existing dyads. One- ments. The predictor variables operational- mode attachments create a cooperative ize the attachment mechanisms for each of structure among competing biotech firms, the four hypotheses: accumulative advan- while two-mode attachments engage differ- tage, homophily, follow-the-trend, and ent organizational forms to access re- multiconnectivity. For each observed at- sources and skills. Each relationship in- tachment, partner and dyad measures are volves a focal DBF, a partner to which the computed for all partners in the risk set. attachment occurs, and a set of alternative Accumulative Advantage is reflected partners with whom the DBF might have in the network Degree and Experience of collaborated, but did not. We refer to the both the attaching DBF and partner, which set of partners to which the focal DBF captures both the number of ties and years might link for a particular observed attach- of experience with collaborations. For re- ment, including the partner to which the peat ties, we also include Prior ties and connection occurred, as the risk set for that Prior experience of the DBF-partner dyad. attachment. For new 1-mode attachments, We also use New Partner, an indicator of the risk set is other DBFs, excluding those whether a partner has been in the network to which the focal DBF is currently or has for less than a year, to capture the effect of previously attached. For repeat 1-mode at- being ‘a new kid on the block.’ tachments, the risk set is, conversely, those Homophily is assessed in several ways. current or prior partners of the attaching We measure the Collaborative distance be- DBF. For new 2-mode attachments, the tween the alliance profiles of the DBF and risk set is partners other than DBFs, ex- the partner in the attaching dyad. For 1- cluding those to which the attaching DBF mode attachments, we measure the Age dif- is currently or has previously been linked. ference, Size difference, and Governance For repeat 2-mode attachments, the risk set Similarity between the two DBFs forming is, again conversely, all current or prior the dyad, as well as measuring Co-location. partners (other than DBFs) of the focal For 2-mode attachments, we lack data to DBF. measure partner’s age or size difference, and governance similarity does not apply MEASURES to non-commercial organizations. We did We draw upon multiple measures to measure homophily, however, between the test each hypothesized mechanism across attaching DBF and other DBFs in the part- four classes of ties. We present and define ner’s neighborhood, defined as the set of the variables used in our statistical analy- DBFs to which that partner has direct ties, ses, along with appropriate controls, in Ta- with the variables Collaborative distance, ble 3. Our dependent variables are binary Age difference, Size difference, Gover- indicators of whether an alliance of each of nance similarity, and Co-location, all con- the four types occurs between a DBF and a sidered with respect to a partner’s other af- partner, given that the DBF and partner are filiations. These ‘second order’ measures “at risk” of attaching. A DBF may forge account for the possibility that connections connections to more than one partner in to partner organizations are conditioned by any given window of time, as long as al- the prior experience of those partners with liances can be ordered such that a partner other DBFs. for which a linkage occurs at a specific mo- 21
  22. 22. Follow-the-trend is captured by the discrete-time images as key inflection field-level variable Dominant Trend and points in the pattern of affiliation, along the partner variable Dominant Type. Both with a linear time trend variable, Timeline. measures reflect whether firms are engag- [TABLE 3 HERE] ing in activities that are comparable to those of others in the field. Multiconnec- STATISTICAL METHOD tivity has two facets: cohesion and diversi- Our challenge is to model a set of bina- ty. The former captures the extent to which ry indicator variables, each of which a firm firms are connected by multiple indepen- can have multiple incidences of, and for dent pathways, while the latter reflects each incidence we must update the risk set whether firms are engaged in multiple of alternative partners and the network types of activities. Cohesion is calculated measures. Consequently, our unit of analy- using the maximum level of k-component sis must be the attachment, rather than ei- for each firm and partner, measured sepa- ther the individual firm or the dyad. Our rately as Firm Cohesion and Partner Cohe- choice of a statistical model for analyzing sion, and jointly as Shared Cohesion, attachment bias is based on our unit of which is the maximum level for which both analysis, as well as empirical and theoreti- parties share a common k-component, if cal considerations. Empirically, the design any.20 For diversity, we use Blau’s (1977) of our sample defined the population of heterogeneity measure, as an index of both DBFs and then identified all the partners the range of activities and multiple types of engaged in alliances with them. Theoreti- partners. We measure diversity in four cally, our research question asks what ways: for a DBF (Firm Tie Diversity), mechanisms account for differential (as op- prospectively for a DBF (Prospective Tie posed to simple random) patterns of attach- Diversity, i.e., the diversity of a firm’s col- ment. For these reasons, we use McFad- laborative profile if the attachment were to den’s estimator for multi-probability as- occur), for a partner (Partner Tie Diversity) sessments, which is a variant of a condi- and for the average of a partner’s set of tional logit model that takes each event as a partners (Partner’s Partner Tie Diversity). unit of analysis, which in our case are at- One might think of the last measure as an tachments, and distinguishes between a fo- assessment of the heterogeneity of the cal population and a set of alternatives friends of a friend. (McFadden, 1973; 1981; also see Maddala, The control variables include firm de- 1986; Ben-Akiva and Lerman, 1989).21 We mographics measured at the time of attach- set up the data so that the focal DBFs are ment, including Age (in years), Size (num- the population and the partners in the risk ber of employees), Governance (whether set for an attachment are the alternatives, publicly held), and Location, measured by thus reflecting the one and two mode net- three digit zip code region, or in the case of works in our sample. companies outside the U.S., by the tele- We conducted the analyses in three phone prefix for nation and city. The key stages. First, we perform overall tests of the partner control is the variable Form, re- four hypotheses, on each type of attach- flecting the form of organization, and the ment, by applying McFadden’s estimator key dyad control is Type, reflecting the as follows. For each attachment, the proba- type of activity that is the focus of the col- bility of DBF i attaching to partner j , given laboration. We also include three time Pe- that DBF i attaches at time t to some part- riod variables, which emerged from the 22
  23. 23. ner in the set Ji,t, is specified as a function the logics of attachment may change as the of partner (X) and dyadic (W) variables: overall structure of the network evolves. We estimate conditional models of at- exp( β X j ,t + γ Wij ,t ) tachment, continually updating both the Pr( yij ,t = 1| ∑ yij ,t = 1, X j ,t , Wij ,t ) = J i ,t ∑ exp( β X J i ,t j ,t + γ Wij ,t ) risk set of alternative partners and the net- work measures to lessen problems of ob- servation dependence common in dyad-lev- Some of the partner characteristics, el analyses. Although we define DBFs as such as collaborative diversity, are not de- our focal population and partners as alter- fined for partners without ties prior to the natives, both DBFs and partners are vying attachment. We include these characteris- to connect with one another, and at any tics by interacting them with an indicator point in time a DBF or partner may have a variable for whether the partner has any finite capacity for connections.22 prior ties, and include the main effect of We estimate all models using Stata this indicator variable in the models. The 8/SE. For the 2-mode analyses, two fea- main effect is always negative, showing tures are notable. First, for 2-mode new at- that partners with no ties in the prior year tachments, there are 5,087 events over our are more likely to receive attachments. time period, with up to 1,600 alternative Second, we explore the extent to which partners in any given year. Given the num- the mechanisms of differential attachment ber of variables we tested, the size of the are contingent on various combinations of: data set needed to analyze the full popula- i.) characteristics of the focal, attaching tion of new two-mode attachments became DBF; ii.) the form of a partner; iii.) the type cumbersome. Thus, we obtained a random of activity involved in each case; and iv.) sample of 1,500 2-mode new attachments combinations involving a partner’s part- to form the data on which estimates were ners. We do so by interacting variables rep- obtained, and for each attachment we took resenting each of these categories with a 25% random sample of alternative part- measures of the attachment bias mecha- ners. We then repeated the random sam- nisms. For instance, to explore how the fo- pling process and re-estimated the 2-mode cal DBF’s attributes (Z) may alter the new attachment models 10 times as a test ‘rules’ of attachment, where those rules are of robustness. The results were nearly iden- specified in terms of partner and dyadic tical for every sample and with no sample variables (X,W), we estimate the following did the pattern of significance change, thus specification: we present the initial results. In the 2-mode analyses the comparison of organizations from different sectors (e.g., for-profit corporations, nonprofits, universities, government agencies, and ven- ture capitalists) requires some comment. A Third, we examine how the sources of nonprofit hospital and a venture capital attachment bias shift over time by interact- firm are not substitutes for one another be- ing our measures of the attachment mecha- cause they provide different resources. Yet, nisms with: i.) period effects, as indicated we have included them as alternatives in in Table 3; and ii.) a linear time trend. the risk set for the same attachments. From These measures give us purchase on how a practical standpoint, in this example the inclusion assumes that a focal DBF may 23
  24. 24. have discretion as to whether its next al- For our purposes, social mechanisms are liance will be to obtain further funding or more strongly supported when the bulk of to seek clinical evidence of efficacy. We the effects are in the hypothesized direc- tested the robustness of this assumption on tions and those effects are consistent across our findings in two ways. First, we strati- all four classes of ties. The full table of re- fied the 2-mode risk sets by form, so that sults is both elaborate and complicated; it is only organizations of the same form as the presented in Appendix II. For the sake of partner to which the attachment occurred exposition, we have excerpted the results were included in the risk set. Second, we that are the primary tests of each hypothe- included interactions of our other variables sis into a series of four more manageable with a set of indicator variables for organi- tables, Tables 4-7. To facilitate interpreta- zational form. Interestingly, while there tion, we present the odds ratios, rather than was some variation in magnitude of coeffi- the coefficients. The odds ratios are ob- cients, the pattern of direction and signifi- tained from the coefficients by exponentia- cance of predictor variables was constant tion. That is, across partner forms save for two excep- Odds ratio for X i = exp( β i ) tions: finance ties to venture capitalists and The odds ratio gives the amount by ties of any type to biomedical supply firms. which the odds of an attachment being bi- These exceptions are not surprising, as ven- ased towards a partner are multiplied for ture capitalists are the most specialized and each unit increase in the level of the inde- least multi-vocal partners. (Recall Figure 2, pendent variable, X, for that partner. which illustrates that more than 95% of VC collaborations are financial.) Other Accumulative Advantage. The odds ra- biomedical firms provide equipment and tios for our tests of accumulative advan- research tools, thus they represent bilateral tage, hypothesis 1, are presented in Table 4. exchange relations, and do not participate Note that we employ a narrow reading of in the development of new medical prod- accumulative advantage, focusing on the ucts in the same, mutual manner as do oth- structural feature of popularity bias (partner er participants. Based on these preliminary degree) as well as the prior ties and experi- analyses, we collapsed the 2-mode data ence of partners with DBFs. Odds ratios across all organizational forms and includ- greater than one for the variables partner ed interactions for finance ties to venture degree, partner experience, prior ties and capitalists and ties of any type to biomedi- prior experience offer support for the accu- cal supply firms when exploring contingen- mulative advantage mechanism. For the cies in the mechanisms for attachment bias. variable new partner, hypothesis 1 would predict an odds ratio lower than 1. Overall, RESULTS the evidence does not favor this hypothesis. We organize the presentation of results Only the negative impact of experience is around our four hypothesized rules of at- consistent across all four categories of ties. tachment: accumulative advantage, ho- Whether the relevant experience resides mophily, follow-the-trend, and multicon- with the partner or dyad, this finding indi- nectivity. Since we utilize a number of cates a preference for novelty. This result variables to measure each mechanism, to runs sharply counter to an accumulative ad- assure the relative strength of a logic of at- vantage argument. The pattern of results tachment requires an examination of an ag- for the other measures of accumulative ad- gregate pattern of support or falsification. vantage varies by type of attachment. 24

×