Reaching for Unique Resources: Structural Holes and Specialization in Scientific Collaboration Networks

Scientists form collaboration ties with others because, among other things, thanks to pooling resources they can jointly benefit from rewards associated with a created outcome, e.g. a scientific publication. We can think of this kind of collaboration as a “co-production” of an outcome. Resources needed to get ahead in science are unequally distributed across the scientific community. For example, scientists at one laboratory specialize in field work and have collected research samples while scientists at some other laboratory might have access to sophisticated equipment that is needed to analyze these samples. Such unequal distribution of resources creates extra incentives to form collaborations and can be linked to decreasing popularity of individual, as opposed to collective, creation of scientific outcomes in contemporary science. An independent mode of work has become less effective for many scientists including those working in disciplines, which were traditionally more individualistic (Moody 2004). The process of formation of collaborative relations involves the matching of resources required for “creating an outcome.” It has many features of a market-like mechanism. To attract desired resources controlled by others, a scientist himself has to offer resources desired by potential collaborators. Simultaneously, scientists face constraints (such as limited time), so there exists some form of competition for access to more desirable resources and more attractive collaborators. We propose a novel approach explaining diversity of resources conveyed in collaboration ties based on complementary mechanisms of structural holes and specialization.

There are a variety of resources relevant for doing science. We can roughly divide them into two categories: (1) resources which are directly engaged in collaboration such as expertise in a particular research topic or a research method, and (2) resources of a more “social” kind, e.g. contacts in academia or prestige. Actors control sets of resources and decide whether to engage them in collaborations according to the demands of collaborators, the desirability of rewards, and time and energy they have. To understand how an actor might engage resources in collaborations let us consider Figure 1 with a simplistic graph of four collaborating scientists. Five types of ties connecting the scientists correspond to different kinds of resources they contribute when collaborating with others. For example, Scientists A and B collaborate and in this collaboration, B conceptualized the research idea while A performed data analysis. At the same time, Scientist A contributes different resources to his other collaborations with D and C. Actors can contribute multiple types of resources in the same dyad, as is the case in pairs A-D or A-C – A contributes a single resource while D and C contribute two types. We may say that Scientist 𝐴 contributes different bundles of resources to B and C. In its entirety, it is a multiplex network or a multigraph (Wasserman and Faust 1994:Ch. 4.6). The interdependencies between different types of ties (here resources) can be quite complex and resulting from, among other things, different availability of, and demands for, different resources among the scientists.

Figure 1:

It often happens that scientists have different responsibilities or “roles”¹ in different scientific projects. Scientist A in Figure 1 is responsible for data analysis in his collaboration with B, but plays a different “role” in his collaborations with C and D by conceptualizing research ideas and providing supervision. From A’s perspective there is a diversity in the sets of resources he contributes to others – his contributions to C and to D are more similar to each other, but both are different from his contribution to B. There is also a diversity in the sets of resources others contribute to A. Contributions of C to A and of D to A are more similar to each other, but different from contributions of B to A. Such defined diversity of contributed resources constitutes an aspect of the above mentioned multiplexity. On the one hand, if an actor is characterized with a low diversity of sets of resources contributed to others, he will have similar types of outgoing ties in all his collaborations. Certain types of ties will co-exist in all the non-empty dyads he is involved in. On the other hand, for an actor characterized with a high diversity, the sets of resources contributed in different dyads will be dissimilar. The general question we ask is:

Assuming that collaboration ties are created purposefully, as a result of goal-directed behavior (Coleman 1994), we may expect actors to seek collaborators that will be “good matches” for the set of resources they possess. A good match here means that it promises a successful collaboration. However, we argue, the formation of collaborative relations is more complex than dyadic resource match-making because, among other things: (1) actors face constraints on the use of resources (e.g. there is a limited number of PhD students one can supervise); (2) the ability to provide certain resources might be related to specific attributes of actors (e.g. only professors can supervise PhD students); and (3) the demand for a particular resource among potential collaborators depends on what other types of resources these collaborators contribute to other collaborations (e.g. a data analyst may need somebody to do a data analysis because he is busy providing similar service to somebody else). In other words, we can expect that mechanisms of diversification (or specialization) of resources contributed to/by others might depend on properties of the actors involved as well as on the patterns of connections among those actors. Therefore, we ask a more specific question:

For example, in the Figure 1, Actor A contributes similar bundles of resources to collaborations with C and D, who also collaborate with one another. At the same time, A contributes a different resource bundle to B, who does not collaborate with any other collaborators of A.

Section 2 starts with providing more background and motivations for our approach. We propose candidate explanations and hypotheses how network structure can influence the diversity of resources contributed to collaborations in Sections 2.1 and 2.2. In particular, we formulate hypotheses built upon concepts of structural holes and specialization. We confront the developed explanations with a small but rich data set based on a qualitative study. Section 3 describes the process of data collection, measurement and data set construction. In Section 4, we introduce the concept of pairwise redundancy and its measure which operationalizes the hypothesized role of structural holes in differentiating the resource bundles contributed to different collaborations. The section also describes the multilevel statistical model we apply to the data. The results are presented in Section 5. The paper is concluded with the discussion in Section 6 and auxiliary details were put into Appendices A1, A2, and A3.

Scientists undertake many kinds of activities, many of which have a collaborative aspect (Boyer 1997). Yet a substantial part of research on scientific collaboration is based on co-authorship data. It brought many insights, such as evidence of different research paths scientists might take corresponding to three basic collaboration network substructures formed by scientists (Moody 2004). Co-authorship data also allow for the analysis of the growth of collaboration structures over time (Wagner 2009), the emergence of new scientific disciplines similar to classical bibliometric studies (Nobre and Tavares 2017; Terekhov 2017), factors improving scientific productiveness (Albarrán, Carrasco, and Ruiz-Castillo 2017), gender inequalities (Hildrun, Alexander, and Johannes 2012), and more. However, scholars indicate that co-authorship data represent only a certain fraction of collaboration activities (Sonnenwald 2007) that are of a very particular kind (Lewis, Ross, and Holden 2012). While co-authorship studies allow for addressing mentioned research questions, and often at a considerable scale, their ability to explain why some scientists collaborate and some other not is limited. The limitation comes, among other things, from the character of bibliographical data which is rather scanty in information relevant for explaining collaboration.

One approach to understand why scientists collaborate is to think about incentives that might lead them to do so. According to Lewis et al. (2012), collaboration in science can be of two types: (1) tangible, concrete, and instrumental, e.g.: designing and conducting a study together, jointly creating specific scientific artifacts such as publications, technological prototypes, production processes, or product designs, and (2) more fluid, relying on discussion, feedback, and commentary. In both cases, the incentives to form a new collaborative relation come from resources scientists possess or control and the interests they might have in resources possessed or controlled by others. For example, an experimentalist might be interested in competencies of a theorist and the resources would correspond to abilities to conduct experiments and develop a theory in a particular research problem. As theorized by Coleman (1994:Ch. 2), actors are interested in resources and pursue these interests through, e.g. engaging in exchanges with other actors. Such exchanges probably take place among scientists too.

Alternatively, collaboration in science might be conceptualized not so much as an exchange, but rather as a process of collaborative creation (Ridley 2011). Scientists’ primary interests are not in the resources themselves, but rather in the outcomes of scientific work that needs these resources as inputs. For example, researchers are interested in publishing an experimental research article. To achieve that goal, it is necessary to design the experiment, conduct it, analyze the data, and write the article. Resources would correspond to abilities to provide or accomplish these smaller tasks efficiently, e.g. specific skills, access to equipment, and so on. To some extent the incentive structure in such a “co-creation” setting seems similar to exchange. By contributing different resources to a common endeavor, actors “exchange” time spent on different tasks as if agreeing to arrangements such as “I’m better at data analysis, so you do the theory.”

Such a resource-based perspective has been elaborated and applied in various settings across the social sciences (e.g. Ekeh 1974; Cook 1987; Lazega and Pattison 1999; Bearman 1997). It has also been applied to the analysis of co-authorship data. However, proper operationalization and measurement of resource contributions become a challenge. For example, Schummer (2004) use departmental affiliation as an indicator of expertise and knowledge bound in scientific disciplines interpreting inter-departmental collaboration as bringing different types of expertise (resources) together. A similar logic was undertaken by, e.g. Bordons et al. (1999) and Qin (1994). It may be debated if the fact of a different departmental affiliation is precise enough as evidence of collaboration through pooling diverse knowledge or methods. Unless additional data on individual responsibilities are available (e.g. Corrêa Jr et al. 2017), co-authorship analysis usually relies on similar assumptions. To overcome such limitations, scientists turn to data of different nature including computational methods such as text mining (e.g. Wang, Notten, and Surpatean 2013; Cheng et al. 2015) and sociological methods, e.g. surveys and interviews (e.g. Youtie and Bozeman 2014; Lazega et al. 2008; Jian and Xiaoli 2013; Laudel 2001). We follow the latter approach as, at the expense of scale, it allows for reconstructing individual collaborations and resources involved in a much greater detail, as we describe later in Section 3.

Convenient abstraction for the approach we advocate in this paper is to represent contributions of different resources between the scientists as a multiplex network. As we have signaled in the Introduction and Figure 1, types of resources contributed are represented with different types of directed ties in a multiplex network of scientists. Study of multiplex social networks is an established area of research (see Kuwabara, Luo, and Sheldon 2010 for a review), especially applied to collective settings such as teams, firms, and other types of organizations. Multiplexity has been used as an explanatory factor. For example, Lazega et al. (2008) investigated how academic success depends on, among other things, position in different types of interpersonal networks. Similar approach was used by Podolny and Baron (1997) to explain intra-organizational mobility (grade advancement).

In contrast to taking multiplexity as given our goal is to explain certain aspects of it. In that sense, our research questions are similar to those of Lazega and Pattison (1999), who analyze how different types of ties (co-working, friendship, and advice) co-exist in a law firm. Their approach was fitting Markov Exponential Random Graph Models (Pattison and Wasserman 1999) to a multivariate data set corresponding to the multiplex network. Coefficients of the estimated models give insights about the interdependencies between different types of ties. For example, they find that it is likely that lawyers seek advice from coworkers of their coworkers, but it is not likely to seek advice of advisors of their coworkers or coworkers of their advisors (Lazega and Pattison 1999:84). Many other equally or more complex interdependencies have been identified. One drawback of their approach is that it is difficult to formulate hypotheses for and interpret models containing many (in the order of dozens) parameters. The other is that it has been since identified that Markovian ERGMs are very often characterized with model degeneracy (Schweinberger 2011) which makes the estimates unreliable.

As will become more evident in the coming sections, we postulate somewhat simpler hypotheses for the dependency structure between ties corresponding to different resources. Implications of our hypotheses can be assessed with a simpler method of looking at, on the one hand, similarity of resource bundles contributed to different collaborations (Section 4.1) and, on the other hand, a measure of relative redundancy of alters in a collaboration network (Section 4.2). When looking at multiplexity from that perspective, there are two relevant concepts of how the structure of collaboration network might influence the diversity in resource bundles contributed to and received from collaborators: Ronald Burt’s concept of structural holes, which we elaborate in Section 2.1 and the concept of specialization which we elaborate in Section 2.2.

2.1

Resources and structural holes

The concept of structural holes introduced by Burt (1995) has become an important approach in social capital research (Crossley et al. 2015). Burt showed how network structure improves access to diverse resources and in return increases individual output like creativity and good ideas. The ideas are built upon Simmels idea of “tertius gaudens” (Simmel 1972). Burt focused on a general setting in which social networks are a source of benefits that can be accessed directly from network peers, but also indirectly – through peers from other members of the network. The theory is well illustrated when we think of the benefits related to resources circulating in the network. In general, it is beneficial to have as many different connections as possible because they provide multiple channels the resources can come to an actor and also the resources can be sent by an actor to others. Access to more groups translates to access to more novel ideas, which can be then adapted in different social circles.

Because maintaining social ties is costly, it becomes necessary to economize and maintain the relations that are efficient and not redundant. Some ties are “redundant to the extent that they lead to the same people, and so provide the same benefits” (Burt 1995:17). Burt distinguishes redundancy by cohesion from the redundancy by equivalence to separate the case in which alters are redundant because they are directly connected to each other from the situation in which alters are redundant because they broke relations to the same set of others (Burt 1995:Ch.1). Figure 2 illustrates this point.

Figure 2:

Actors B and C are redundant to Ego by cohesion because they are directly connected to each other. Simultaneously, actors A, B, and C are redundant by equivalence because they are all connected to the group D-E-F-G. Should an important resource originate from that group, the ego will learn it through any of A, B, or C. Actor Z is a non-redundant contact because he is not related directly or indirectly to other direct contacts of Ego. Following Burt’s argument we can expect that a purposeful collaborator might form non-redundant collaborations because these ties will bring different resources from those that he can acquire elsewhere².

Given the above, we can expect that bridging structural holes by maintaining structurally non-redundant collaborations will be associated with relatively unique sets of resources contributed in those collaborations. Putting it succinctly:

H1. Egos will acquire more similar sets of resources from alters who are redundant by cohesion.

In the ego-network in Figure 2, Actors B and C are redundant, so according to H1 we would expect that their contributions will be more similar to each other than contributions of Actors C and Z to each other as Actors C and Z are not redundant. Do note that the presented argument and hypothesis require a somewhat non-standard operationalization of redundancy. We are not considering redundancy as a property of an alter, but rather as a property of pairs of alters. This concept is further illustrated qualitatively in Section 2.3 and fully developed in Section 4.2.

The structural holes mechanism has been intensively investigated in business environments (e.g. Allen 1977; Katz and Tushman 1981; Burt 2004; Zaheer and Bell 2005; Tiwana and Keil 2007). The proximity of structural holes can also improve access to valuable resources in science. Recently, Bellotti (2012) showed that occupying brokerage position in a scientific community is more important for getting funded than a prestigious position in a scientific field or a recognized affiliation. Lopaciuk-Gonczaryk (2016) indicated that having collaborators who do not collaborate with each other is related to increased publishing productivity. A handful of research shows that sparse networks rich in structural holes result in better scientific output (Abramo, D’Angelo, and Solazzi 2010; Andrade, Los Reyes Lopez, and Martin 2009). Merton and Barber (2006) highlight the role of serendipity in scientific discoveries, which is possible only in sparse, diverse, and “accidental” networks.

2.3

Structural holes and specialization: a rejoinder

Let us bring together and summarize the considerations above. Figure 3 shows two ego-networks of a Scientist 0 in two ideal-case situations corresponding to our hypotheses from Sections 2.1 and 2.2 above. For simplification, but without the loss of generality, we reduced the complexity of the pictures by showing only a single resource type being contributed in a particular direction in every dyad. If the same type of resource is being acquired by Actor 0 in different dyads we expect not necessarily identical resources, but a higher similarity of resource bundles in these dyads.

Figure 3:

Network in Panel A of Figure 3 shows idealized situation in which resources are contributed according to the structural holes argument. In line with our H1, alters who are redundant, here 1, 2, and 3, contribute similar resources to ego 0. Alters who are non-redundant to 0, namely 4 and 5, are more likely to contribute different resources than anybody else.

Network in Panel B corresponds to a situation of perfect within-group specialization. Scientists 0, 1, 2, and 3 form a tightly collaborating group. Should the specialization take place only within such groups, we see, in line with H2, each member of the group to contribute the same type of resources to all other members. In particular, the ego 0 receives different resources from each of the members 1, 2, and 3. Alters 4 and 5 who are not members of that group may contribute still other types of resources.

Table 1 summarizes these hypotheses. According to structural holes argument, we should expect similar bundles of resources contributed by redundant alters (1 and 2) – effect of redundancy on similarity is positive. Specialization argument implies that these contributions will be dissimilar – effect of redundancy on resource similarity is negative.

Table 1:

Directions of effects of redundancy on similarity of acquired resources under H1 and H2.

Hypothesis	Effect of redundancy on similarity of resources acquired by ego
H1: Structural holes	+
H2: In-group specialization	−

We can think of specialization as a process of reducing costs of structural redundancy. Redundant ties may be worth maintaining, but only if they convey resources unique for ego within the group.

Data consists of 40 individual in-depth interviews conducted between April and August 2016 by two interviewers. The interviewees mentioned 333 collaborators in total. The sample consists of 20 female and 20 male scientists from six Polish cities. Respondents represented a broad range of disciplines: natural sciences, social sciences, life sciences, the humanities, engineering, and technology on different levels of career from PhD candidates to professors. The detailed description of the sample can be found in Appendix A2 and in Bojanowski, Czerniawska, and Fenrich (2020). The data set is available online³.

Each interview consisted of four parts. After the initial introduction and a short description of respondent’s professional interests (part one), respondents were asked to name up to 10 most important collaborators during last 5 years (part two). A collaboration might have already ended at the time the interview was conducted, but at least part of it have had to be undertaken in the indicated period. Scientific collaboration was defined broadly as a shared process of creating new knowledge and mutual help in intellectual endeavors to include less standardized collaboration practices. The diversity in scientific collaboration practices is often underlined in the literature (Beaver 2001; Katz and Martin 1997). The definition used during the interviews was built upon definition from Lewis et al. (2012), which includes two types of collaboration: “collaboration” and “Collaboration” (with capital “C”). The first term describes situations in which the relation is fluid, mostly relying on a discussion, feedback, and commentary while the second is more tangible, concrete, and instrumental, including designing and conducting a study together as well as later publications.

Respondents were asked to mention up to 10 “most important” collaborators to reduce the possibility of pointing collaborators not relevant for respondents’ work. The interviewees were asked about collaborators’ gender, scientific degree, and institutional affiliation. If the respondent felt uncomfortable with revealing full names of collaborators, s/he was only asked for unique nicknames. All collaborations were discussed separately (part three). Respondents were asked about the history of collaborations, the merits for collaboration, resources each party engaged in collaboration, and rewards gained from collaboration. Respondents were provided with leading questions about resources which might be engaged and gained from a collaboration such as knowledge and skills, contacts, funding opportunities, equipment, prestige, or joint publications. Respondents were also asked if there were any negotiations of the terms of collaboration. Names of all collaborators were attached to a cork board with respondent in the central point (part four). Respondents were asked to indicate all collaborations among her and her collaborators. Collaborators on the cork board were represented with pins and collaborations were represented with rubber bands (see Figure 4). There were two follow-up questions about mutual dependencies of collaborations and about possible collaborators crucial for any tie, which were not included in the interview.

Figure 4:

Interviews, lasting from 24 to 90 min, were recorded and transcribed. The cork boards with information about collaborations were photographed.

Information about collaboration ties was recovered from photographs of cork boards prepared during the interviews, such as the one presented in Figure 4. Collaborators and collaborations were labeled with unique identifiers and assembled into a two-mode network data set with modes corresponding to persons (pins) and collaborations (rubber bands), respectively.

Information about collaborators was coded based on transcripts. Respondents and collaborators were described with information about gender, scientific degree, scientific discipline, department (if possible), university, city, and country. Some collaborators had more than one affiliation or discipline. The primary discipline and institutional affiliation were chosen based on Polish Science Database. Missing pieces of information were retrieved from the internet.

We have a set of 𝑁 actors. Let us define the collaboration network as an undirected graph X = [x_ij ] _N×N where x_ij = 1 if 𝑖 collaborates with 𝑗 and x_ij = 0 otherwise. No self ties are allowed so x_ij = 0 for all i. Example of such a network is shown on Figure 5.

Figure 5:

Actors can engage various types of resources when collaborating with others. Let us have R types of resources. The engagement of resources in collaborations can be represented as a directed resource flow network Y = [y i j r ] _N×N×R , an array in which y i j r = 1 if resource r is engaged by actor i in her collaboration with actor j. In other words, resource r “flows” from actor i to actor j. As with the collaboration network, no self ties are allowed so y i i r = 0 for all i and r. Let us use ∗ for denoting all elements along the specific dimension, for example y _0∗1 would be a binary vector indicating collaborators of actor 0 with whom she engaged the resource 1. Example of a resource flow network is shown in Figure 6. As the data come from an egocentric study, we impose the convention that ego (the respondent) has an index value of 0.

Figure 6:

To test our hypotheses we need to compare different collaborations in terms of (1) similarity of resources contributed by scientists in the resource flow network; and (2) the extent of structural redundancy in collaboration network. We propose how to measure these concepts below.

The data have a nested structure: for each respondent, the ego, the data contains information about her individual characteristics (Level 2) and information about characteristics of all alter-alter comparisons among his network peers (Level 1). For example, if a particular ego has 8 alters, this is represented with (8×7)/2=28 level-1 observations. We estimate the following specifications:

Estimates and confidence intervals for fixed effects are presented in Figure 7. Detailed model results are available in Appendix A1.

Figure 7:

Figure 8:

Overall, the null model statistics indicate that percentage of variation in dependent variable that can be attributed to between-ego variation is equal to 16.1. AIC values indicate that models are improved by adding level-one and level-two variables (c.f. Table 5).

Let us start with our main variable of interest – pairwise redundancy. Including it as a predictor with a linear effect next to the other explanatory variables significantly improves the fit (Model 2 vs Model 1). The effect is positive and significant, albeit small. This implies that, on average, the more redundant the alters are to ego, the more similar are the resources contributed by those alters. As we have explained in Section 2.3, the structural holes and specialization arguments imply partially conflicting predictions with respect to the effects of redundancy of collaboration ties on similarity of resources contributed. Assuming that, in the structural holes argument, the effect of pairwise redundancy on resource similarity is constant across the redundancy scale, this result is in favor of our H1.

The specialization argument and H2 formulated in Section 2.2 assumed that specialization takes place in densely collaborating groups. This corresponds only to the interval [1; ∞] of pairwise redundancy. As a consequence, a proper test of these hypotheses is included in Model 3 in which pairwise redundancy is represented with a two-segment linear spline with a knot at 1. The effect is positive in the interval [0; 1] and slightly negative in the interval [1; ∞]. This is consistent with H2 implying that in densely collaborating groups alters are more likely to contribute dissimilar resources to ego. It should be noted that the negative effect in the [1; ∞] interval is rather small (although significant).

Next to the specialization in “informal” dense groups we also expected that shared institutional affiliation would foster similarity in sets of resources contributed in collaboration with ego, namely alters from the same scientific institution having access to similar resources and contribute similar resources in collaborations with ego. Alters sharing institutional affiliation are more likely to contribute more similar resources, which is inline with our expectations. The effect persists even when pairwise redundancy is included in the model.

Turning to the career-related effects, the data do not provide enough power to estimate the effects of scientific degree of ego with good precision. The only tendency we can observe shows that if ego is a habilitated PhD or a professor, she is more likely to contribute more similar resources to all her collaborations (as compared to PhDs) irrespective of the properties of the alters and redundancy characteristics. Such trend is not present when analyzing alter contributions. We expected more similar sets of incoming and outgoing resources for alters with the same scientific degree. The coefficients are positive and significant in models for both dependent variables implying that alters with the same scientific degree indeed provide ego with relatively similar sets of resources and vice versa, as we expected.

Collaboration between scientists can be understood as a multiplex network of resource exchanges. We proposed to analyze this multiplexity by looking at the diversity of resource bundles contributed to/by collaborators. We provided two hypotheses referring to two different mechanisms, which regulate the diversity of resources exchanged in collaboration ties between alters and egos: structural holes and specialization. To this end we developed a concept and measurement of “pairwise redundancy” designed to capture relative structural redundancy of alters “vis-à-vis each other.” We also investigated how the diversity of resources depends on institutional co-affiliation and career-related characteristics.

According to the structural holes argument, unique resources should be accessible through collaboration ties bridging structural holes. Our results confirm that collaboration ties with alters who are pairwise redundant are more likely to convey similar sets of resources.

Additional findings indicate that collaboration with scientists sharing institutional affiliation resulted in increased similarity of incoming resources even after controlling for pairwise redundancy. We can conclude that both structural redundancy and institutional factors limit the novelty of sets of resources conveyed in collaboration ties.

The non-linear effect of pairwise redundancy in our models leads to interesting implications for redundancy and brokerage in scientific collaboration networks. The shape of the effect shows that the strongest effect of pairwise redundancy on diversity of resources contributed is observed when comparing a pair of alters who are completely disconnected in ego’s neighborhood to a pair of alters who are directly connected. This corresponds to the [0; 1] interval of our pairwise redundancy measure. Directly connected alters who do not share any other common collaborators with ego constitute a configuration in which the resources contributed by alters to ego are most likely to be similar. Further embedding of the said alter-alter relationship into the collaboration network of ego by adding more shared collaborators stops this tendency or even reverses it – in densely connected groups we find evidence of specialization.

We believe there are several avenues originating from the presented results along which further research can proceed. First, our arguments about the pairwise redundancy and specialization in collaboration networks deserve a more unified theoretical treatment. For example, it is somewhat unclear how these two mechanisms reconcile within, e.g.: (a) institutional research teams that are not very cohesive; and (b) cohesive collaboration groups that span across different institutions.

Second, our approach focuses on a collaboration network as a one mode network. We believe a promising extension, both theoretically and methodologically, would be to approach the problem with the formalism of a two-mode collaboration network. In such a network, scientists (Mode 1) are involved in “projects” (Mode 2). Such an approach should better handle possible multi-party collaborations, which often take place in science. It will also require a different design of data collection.