Engineering serendipity: When does knowledge sharing lead to knowledge production?

2.1 Knowledge production and knowledge sharing

According to the knowledge-based view of the firm, knowledge production is a multicomponent process that involves the transfer, creation and diffusion of knowledge (Grant, 1996; Kogut & Zander, 1992; Nonaka, 1994). Knowledge transfer involves the movement of facts, relationships, and insights from one setting to another, and becomes evident when the experience acquired by an individual, group or organization in one setting is applied in another setting (Argote, 2012; Hansen, 1999; Szulanski, 1996). Knowledge creation refers to the generation of facts, relationships and insights to solve new problems (Nonaka, 1994). The emphasis on new knowledge is what distinguishes the process of knowledge transfer from creation. Because new knowledge is not held by anyone prior to its creation, it cannot be transferred and applied directly (McFadyen & Cannella Jr, 2004). Conversely, knowledge diffusion refers to the process through which transferred or newly created knowledge is disseminated and used by other individuals, groups and organizations (Fleming et al., 2007; Singh, 2005). Knowledge diffusion is a critical component of the knowledge production process because it reduces duplication of effort and promotes efficiency by demarcating what is known from what is yet to be explored on the knowledge frontier (Boudreau & Lakhani, 2015).

Critical to the efficiency of knowledge production is the process of knowledge sharing between individuals possessing different types of knowledge (Grant, 1996; Kogut & Zander, 1992; Nonaka, 1994), and their ability to learn from these interactions (Levinthal & March, 1993; Simon, 1991). Common knowledge between partners is important because it enables individuals to absorb the aspects of their knowledge sets that they do not hold in common, or that are unique to each individual (Carlile, 2004; Cohen & Levinthal, 1990; Maurer & Ebers, 2006). Building on prior work, we propose that there are at least two separate pathways through which people develop their knowledge bases: their field and intellectual specialties. Because of the path-dependent nature of knowledge (Hargadon & Sutton, 1997), the overlap in two individuals' field and intellectual specialties may have different implications on knowledge production. Table 1 provides a summary of our knowledge constructs.

2.2 Two knowledge bases of similarity: Field specialty and intellectual specialty

Field specialties are developed through an individual's educational background, skills and training (Bechky, 2011; Bourdieu, 1984; Haas & Park, 2010). In biomedical research, which is our focal context of study, there are multiple medical field specialties, such as neurology, radiology or oncology that focus on investigations of the biological process and the causes of specific diseases. Organizations often structure their departmental and divisional memberships around field specialties (Biancani et al., 2014). In universities and academia, this means that field specialties are often the primary channels through which resources, such as salaries, funding, tenure, offices and laboratory space, are apportioned to faculty (Biancani, Dahlander, McFarland, & Smith, 2018). For these reasons, field specialties provide professional reference groups (Haas & Park, 2010) for members to identify appropriate behavior, rigor of scholarship, career goals and aspirations (Abbott, 2001). In contrast, intellectual specialties are closely aligned to personal interests and passions that evolve over time. Individuals can have memberships in multiple intellectual specialties depending on their current pursuits and priorities (Caza et al., 2018). For example, in biomedical research, intellectual specialties include research topics such as aging, addictions, Alzheimer's disease, brain injury, sleep, and stem cells; each of these intellectual specialties is associated with memberships in professional reference groups (Haas & Park, 2010) for people motivated by common problems, methods, trends, and processes. These intellectual interests can be specific to a field's research topics, such as Alzheimer's disease, which is typically studied by neurologists, or span different fields, such as stem cell research.

In short, field and intellectual specialties shape people's identities, language, tastes and affiliations through professional memberships (Haas & Park, 2010). Individuals will likely adopt multiple professional identities depending on their field and intellectual memberships: this means that people from different field specialties can share overlapping intellectual interests, while individuals from the same field specialty can have divergent intellectual interests, because field and intellectual interests are two distinct dimensions of individuals' knowledge constructs.

In distinguishing how field and intellectual overlap shape knowledge production, knowledge similarity describes the distance between the field specialties and intellectual specialties of two knowledge sharing partners. Individuals are more likely to recognize and absorb new ideas when they already have some existing expertise and find it more difficult when the ideas are outside their realm of expertise (Carlile, 2004; Cohen & Levinthal, 1990). Put differently, even though distant knowledge tends to be novel and valuable (Jeppesen & Lakhani, 2010; Leahey et al., 2017), people are more likely to experience challenges communicating with one another, and may not recognize the value of external information (Cohen & Levinthal, 1990; Grant, 1996). For example, two neurologists are less likely to experience challenges communicating with each other than a neurologist and oncologist pair because they share more similar educational backgrounds, training and clinical expertise. However, there may be greater opportunities for knowledge production between the neurologist and oncologist pair because their field specialties are more distant. Similarly, although two sleep researchers may communicate and comprehend each other with relative ease because they read and conduct similar research (Dahlander & McFarland, 2013), there is greater potential for knowledge production between an aging and a sleep researcher, who may offer greater utility and value to one another to solve problems related to both sleep and aging processes—provided that they have sufficient common knowledge (Bechky, 2011; Grant, 1996). This suggests that there is an inherent trade-off between the greater ease of conversing and absorbing ideas locally from individuals with many overlapping interests and the greater potential opportunities for novel ideas and knowledge transfer, creation, and diffusion from sharing and assessing the knowledge of those with more distant interests.

2.3 Knowledge transfer and knowledge similarity

Knowledge transfer is a two-sided process because it depends on the efforts of a source to share knowledge with a recipient and the recipient's efforts and capacity to acquire, absorb, and learn it (Argote, 2012). Because of greater common knowledge, local knowledge found within groups of similar individuals tends to be more easily transferred than distant knowledge spanning group boundaries (Carlile, 2004; Kogut & Zander, 1992; Rosenkopf & Almeida, 2003). That said, more distant knowledge may present nonredundant ideas that benefit learning and acquiring new concepts. Consequently, during serendipitous idea exchanges, we would expect that partners from dissimilar field and intellectual specialties offer each other more novel ideas and opportunities to learn (Leahey et al., 2017; Lee, 2019). Divergent interests create a wider pool of knowledge to draw upon, which allows for multiple perspectives and problem-solving approaches that increase the likelihood of new discoveries (Boudreau, Guinan, Lakhani, & Riedl, 2016)—provided that the partners share sufficient common knowledge to make sense of each other's knowledge dissimilarities. Based on these reasons, we expect that pairs with less field or intellectual overlap have a greater potential pool of ideas for knowledge transfer, up until a threshold of dissimilarity, after which they will lack sufficient common interests to benefit from knowledge transfer.

2.4 Knowledge creation and knowledge similarity

There is a growing view among innovation scholars that specialization has created a “knowledge burden” hypothesis that makes collaborative work combining the increasingly narrow niches of specialization imperative to moving the knowledge frontier forward (Jones, 2009; Uzzi et al., 2013). Scientific discovery is a process that combines individually focused tasks—such as reading, experimentation, and writing—with social interactions through joint sense-making with others that can spark new discoveries (Boudreau et al., 2016; Latour & Woolgar, 2013).

That said, collaborations tend to be constrained by social processes, such as preferences for others with shared attributes, as well as search costs that stifle attempts to identify suitable collaborators. First, principles of homophily suggest that people are attracted to others who hold similar values because their interactions are more rewarding and less uncertain (McPherson & Smith-Lovin, 1987), as a result of the benefits of security and mutual attraction (Dahlander & McFarland, 2013). Moreover, joint knowledge creation is a long-term investment in a multiplex relationship (Uzzi, 1997), where partners engage in joint problem-solving and spend time together discussing, reflecting and interacting to achieve mutual benefit (McFadyen & Cannella Jr, 2004). Due to these reasons, collaborations are more likely to form among individuals who share overlapping field and intellectual interests (Biancani et al., 2014; Dahlander & McFarland, 2013).

Second, there is a growing view that search costs and resulting information frictions tend to constrain people's collaboration patterns (Boudreau et al., 2017; Catalini, 2018). Engineered, serendipitous interactions can mitigate some of these frictions by providing people the opportunity to exchange information and develop intellectual links with others they would not otherwise interact with. Such interactions may enable diverse partners to discover the potential complementarities they offer each other, leading to new collaborations. However, to the extent that serendipitous knowledge sharing leads to joint knowledge creation will also likely require that potential partners share some field or intellectual overlap; otherwise their interests may be too disparate to establish synergies in incentives (e.g., publication or tenure requirements) or research interests. Accordingly, we expect that increasing field and intellectual similarity should benefit knowledge creation up to a threshold, over which there are decreasing marginal returns on greater levels of knowledge similarity.

Another pathway that may lead to joint knowledge creation relies on prior collaborations and tie persistence (Hasan & Koning, 2019; Ingram & Morris, 2007; Zhang & Guler, 2020). We examine a specific type of tie persistence, namely the likelihood that an elemental collaboration persists into a complete knowledge product. In science, grant coapplications and copublications represent two essential but opposing ends of knowledge creation, from idea-generation to final output (McFadyen & Cannella Jr, 2004). Both are also essential to scientists' research productivity and often used as evaluative criteria for important decisions, such as promotion, tenure, awards and recognition (Dahlander & McFarland, 2013; Stephan, 1996). The knowledge similarity requirements that improve the likelihood that elemental collaborations persist into copublications may differ from the processes that shape long-term collaborations on peer-reviewed research publications. In addition to being at the open end of research inquiry, grant applications tend to have finite submission deadlines and these can impose resource, attention, coordination and other constraints on potential partners (Dahlander, O'Mahony, & Gann, 2016). Although serendipitous knowledge exchanges can reduce information search costs and expose people to promising new contacts, multiplex relationships require repeated interactions (Ingram & Morris, 2007). In the short-term, potential partners may turn to other individuals with whom they share significant knowledge overlap, due to fewer frictions and relational uncertainties (e.g., priorities, personalities, scheduling constraints) associated with similar partners. Due to these reasons, common knowledge may be even more critical to shaping elemental collaborations. We thereby expect that elemental collaborations are more likely to persist into a final product of knowledge creation when knowledge sharing partners have significant knowledge overlap in terms of their field and intellectual interests.

2.5 Knowledge diffusion and knowledge similarity

Among research scientists, forward citations to others' publications are a primary means for diffusing knowledge. Beyond knowledge diffusion, forward citations constitute a critical means of social recognition for acknowledging the contributions of predecessors (Merton, 1973) and tracing the path of scientific discovery and diffusion (Stephan, 1996). Some scholars argue that the number of citations a publication has received is perhaps the most common way to measure the importance of an individual's contribution to science (Stephan, 1996). Consistent with this view, citations are a critical currency of scientific credit (Latour & Woolgar, 2013), both driving research impact and constituting the basis of reward systems in science, including promotion, status, funding, peer esteem, honors, and awards (Boudreau & Lakhani, 2015).

There are generally two views of how knowledge is diffused in science: openness and secrecy. According to the Mertonian norms of communalism or “openness,” publication enables scientists to establish priority of discovery and allows them to be the first to communicate an advance in knowledge and allow others to freely use it (Merton, 1973). Thus, publication promotes the open diffusion of scientific knowledge, as long as scientists' own internal agents (i.e., other scientists) appropriately recognize and diffuse their work (Boudreau & Lakhani, 2015). On the other hand, social recognition is a discretionary act among scientists, and strong evidence points to the existence of counter-norms promoting secrecy, competition, and information withholding (Haas & Park, 2010; Haeussler, Jiang, Thursby, & Thursby, 2014). Scientists often compete for similar resources, funding, and recognition. Given limited resources, scientists need to be assured that they will be appropriately remunerated for openly diffusing others' knowledge (Hagstrom, 1974; Murray & O'Mahony, 2007; Reschke, Azoulay, & Stuart, 2018).

On balance, the evidence suggests that scientists do not unequivocally diffuse each other's publications. Among scientists with greater knowledge overlap, serendipitous idea sharing may heighten the competition effect, particularly as information exchange can introduce people to novel work they cannot discover on their own via publications or other publicly available sources. These new projects may be similar to their own undertakings, risking duplication of efforts. Because scientists compete for priority (Merton, 1973), recognition (Hagstrom, 1974; Reskin, 1977), and funding (Stephan, 1996), partners with highly overlapping interests may discover a need to differentiate themselves from one another. Social recognition of highly similar peers would not only highlight redundancies but may also detract attention and resources away from one's own work (Campanario & Acedo, 2007). We expect that partners sharing high field or intellectual similarity may refrain from citing each other's research to outperform their peers.

3.1 Setting and research design

We carried out our study in the context of a medical symposium for research on advanced imaging, which is used to detect diseases and other health conditions early, to allow health care practitioners to direct patients to the health care services that they need. We collaborated with the administrators of a large U.S. medical school to layer the medical symposium onto the University's Clinical and Translational Science Center pilot grant program, which provides seed funding in the form of pilot grants to support nascent research efforts that are awarded competitively to faculty within the University.¹ The purpose of the grant opportunity was to solicit proposals to improve methods for using advanced imaging technologies to address unmet clinical needs, and offered $50,000 per award for up to 15 pilot grants. A major challenge in the field of advanced imaging is that furthering the knowledge frontier requires expertise in the latest imaging tools and technologies and a deep understanding of the health problems to which they could be applied. These different types of knowledge are typically held by people from distinct disciplinary backgrounds, which makes advanced imaging an ideal setting for our research.

In November 2011, we invited all life sciences faculty and researchers at the university to a medical research symposium for the unique grant funding opportunity. Our field experiment involved faculty and researchers at the University and its affiliated hospitals and institutions, which are independently owned and managed, with each appearing as a separate entity in hospital rankings and lists of National Institutes of Health (NIH) grant recipients. We communicated to applicants that eligibility to submit a grant application was conditional on attending the symposium. In the first stage, investigators interested in applying for the grants submitted a statement of interest in which they briefly described a specific medical problem that advanced imaging techniques could potentially address. We collected basic biographical information (e.g., degree, institution, department appointment) at this stage.

3.2 Participants and randomization of knowledge-sharing partners

The symposia were held on January 31, February 1, and February 2, 2012, at one of the university's innovation labs. Figure 1 summarizes the key details of the randomization: 402 unique participants were randomly assigned to one of three nights of the symposium, one of four breakout rooms, one of two groups, as well as a poster location to stand next to around the perimeter of the room.² Participants were provided an electronic device, called a “sociometric badge” to automatically record their face-to-face interactions during the symposium (Kim et al., 2012). Each of the three nights featured a 30-min welcome address, followed by two 45-min poster sessions in breakout rooms, with a 15-min social break in between the poster sessions. Within each breakout room, scientists were randomly assigned to a poster location and either the first or second poster group. The scientists presented and exchanged ideas with one another during the poster sessions. The grant administrators prepared the posters to be a standard size and in a standard format, based on each participant's statements of interest. Thus, treatment pairs assigned to the same breakout rooms had more opportunities for knowledge sharing than control pairs.

Shortly after the symposia, all participants received an e-mail invitation to submit applications for the pilot grants or concept awards by the deadline of March 8, 2012. At this time, they also received PDF booklets with the contact information and posters of all participants so that all researchers had identical information apart from the knowledge acquired in the breakout rooms.

3.3 Data collection and variables

We tested our hypotheses using data from a variety of resources from the advanced imaging symposium and the 6 years of publication records from 2013 to 2018 on the attendees. In the analyses, we did not include the year 2012 to remove potential research topics or ideas that were in progress prior to the symposium. We used data from the scientists' registration form for the symposium, which contained information about their institution, department, academic position, self-identity as an imager or clinician, and statements of interest. The sociometric badges automatically recorded their face-to-face encounters when two badges were facing each other with a direct line of sight within a 30° cone of 1 m. We verified that all recorded interactions were within 10 m using Bluetooth proximity data (Kim et al., 2012) and required that two badges be in contact with each other over a span of at least 1 min (Ingram & Morris, 2007). We collected face-to-face interaction data for 306 (74%) scientists who attended the symposium, and the subsequent analyses are based on these scientists.³ After the symposium, we collected information on the coapplicants and the awardees of the advanced imaging grants and used the Scopus database to collect scientists' publication records.

3.4 Estimation approach

We wish to estimate the effects of field and intellectual similarity between knowledge sharing partners on knowledge production outcomes. The unit of analysis is the scientist-pair{i,j}, and pairs are considered to be “at risk” if they attended the same night of the symposium—a total of 15,817 pairs. First, we analyze the effect of being in the same (treatment) versus different (control) breakout rooms on knowledge production outcomes using ordinary least squares (OLS) regression. We then interact Same room with Field similarity, and Same room with Intellectual similarity to examine knowledge similarity effects. Second, we analyze the effect of face-to-face communication on knowledge production outcomes using instrumental variable (IV) regression. We use an IV approach to account for the common endogeneity issue in pairwise interaction data. We exploit exogenous variation in the likelihood of interaction between scientists who are assigned to the same versus different breakout rooms by using Same room as an instrument for F2F communication in the first stage, and the estimates for F2F communication in the second stage. We then interact F2F communication with Field similarity, and F2F communication with Intellectual similarity to examine knowledge similarity effects.

We address the nonindependence, common-person problem of dyadic regressions by estimating robust SEs that are simultaneously clustered on both members of the dyad, using multiway clustering, developed theoretically by Cameron, Gelbach, and Miller (2011) and implemented for Stata in clus_nway.ado (Kleinbaum et al., 2013).

4.1 Knowledge transfer results

We present the OLS and IV regression knowledge transfer results in Table 3. The dependent variable is the percentage of scientist-pair {i,j}'s MeSH postsymposium keywords that were transferred between i and j, with 0% being no transfer, and 100% being complete transfer. Models 1 and 2 present the baseline models with main effects only; Models 3 and 4 add the field similarity interaction terms; Models 5 and 6 add the intellectual similarity interaction terms; and Models 7 and 8 show the full results with both interaction terms.

Hypotheses (H1a) and (H1b) suggest that field and intellectual similarity have an inverted U-shaped effect on the relationship between knowledge sharing and knowledge transfer, respectively. Models 3 and 4 indicate that there is no meaningful association between moderate field similarity and knowledge transfer among same room pairs (0.712, p = .132) and communicating pairs (4.088, p = .197). Models 5 and 6 show that compared to pairs with high intellectual overlap, scientists with moderate intellectual similarity transferred a higher percentage of MeSH keywords for both same room pairs (0.362, p = .042) and communicating pairs (3.348, p = .032), respectively. These results are consistent in Models 7 and 8, which include both interaction terms.

Figure 3a shows the change in knowledge transfer among communicating versus noncommunicating pairs and intellectual similarity from Model 4, with 95% CIs, and illustrates the inverted U-shaped relationship. While communication did not meaningfully impact knowledge transfer among pairs with low field overlap, it led to an increase of 3.581% among pairs with moderate intellectual overlap from 4.085% [3.795%, 4.376%] to 7.666% [4.995%, 10.338%] percent, a nearly twofold increase, compared to a 0.233% [−1.239%, 1.705%] increase for high intellectual overlap pairs. The percentage increase among moderately overlapping pairs corresponds to about six new MeSH keywords per scientist, roughly equivalent to the average number of MeSH keywords in one publication.⁴ This result supports Hypothesis (H1b) by showing cooperative effects of knowledge transfer for pairs with moderate intellectual similarity.

4.2 Knowledge creation

We present the OLS and IV knowledge creation regression results in Table 4. The dependent variable is the number of copublications between scientist-pair {i,j}. Models 1 and 2 present the baseline models with main effects only; Models 3 and 4 add the field similarity interaction terms; Models 5 and 6 add the intellectual similarity interaction terms; Models 7 and 8 show the full results with both interaction terms; Model 9–11 adds Grant coapplicant, followed by the field and intellectual similarity interaction terms, as well as both interaction terms, and controls for Grant awardee.

Hypotheses (H2a) and (H2b) predict that field and intellectual similarity have an inverted U-shaped effect on the relationship between knowledge sharing and knowledge creation, respectively, while Hypotheses (H2c) and (H2d) predict that knowledge sharing partners who start an elemental collaboration are more likely to copublish when they share greater field and intellectual overlap, respectively. Models 3 and 4 show that compared to same field pairs, moderate field similarity has a strong positive relationship with knowledge creation among same room pairs (0.329, p = .047) and communicating pairs (2.159, p = .036). There is also a smaller, positive association between low field similarity and knowledge creation among same room pairs (0.133, p = .092) and communicating pairs (1.040, p = .101). Turning to intellectual similarity, Models 5 and 6 show there is no meaningful association between moderate intellectual similarity and knowledge creation among same room pairs (0.0409, p = .257) and communicating pairs (0.333, p = .234). The results are consistent in Models 7 and 8, which includes both interaction terms.

Figure 3b plots the change in copublications between communicating and noncommunicating pairs and field similarity, with 95% CIs from Model 3, and illustrates the estimated inverted U-shaped relationship. We observe that while communication did not meaningfully benefit pairs with low field overlap, communication led to an increase of 1.174 [−0.207, 2.556] copublications among pairs with moderate field overlap, and −0.984 [−1.731, −0.237] fewer copublications among pairs with high field overlap. The patterns suggest that communication reduced the search costs of identifying synergistic collaborations with scientists sharing some field interests, resulting in a reallocation of resources away from pairs with high field overlap.

We also examined the extent that the knowledge sharing intervention impacted the scientists' overall research portfolios. Turning to the scientists' postsymposium publications, each scientist had an average of M = 32.394 (SD = 35.412) peer-reviewed research articles; the increase of 1.174 publications among pairs with moderate field overlap corresponds to about 7% of a scientist's publications, which is of considerable magnitude from a 90 min intervention. Turning to changes in research direction, we also note that same room pairs were more likely to publish in advanced imaging journals (e.g., Magnetic Resonance in Medicine, NMR in Biomedicine): 21% for same room versus 6% for different room pairs (t = 2.302, p = .0249).⁵ This suggests that the knowledge sharing treatment not only reduced search costs of finding collaborators but also potentially reshaped the pairs' research trajectories.

Turning to Hypothesis (H2c), Model 9 shows that controlling for Grant awardee, grant coapplicants are more likely to copublish when they are from the same field (low: −2.427, p = .189; moderate: −3.303, p = .042). Finally, turning to H2d, Model 10 indicates that intellectual similarity is not a meaningful predictor of elemental tie persistence (low: −0.639, p = .759; moderate: −1.083, p = .465). The results remain consistent in Model 11, which includes both interaction terms. This suggests that although strong field overlap is a catalyzing force for initiating early-stage collaborations, these requirements may dissipate over the long-run as these ties persist into copublications. Thus, our results show support for Hypotheses (H2a) and (H2c).

4.3 Knowledge diffusion

We present the OLS and IV knowledge diffusion regression results in Table 5. The dependent variable is the number of forward citations between scientist-pair {i,j}. Models 1 and 2 present the baseline models with main effects only; Models 3 and 4 add the field similarity interaction terms; Models 5 and 6 add the intellectual similarity interaction terms; Models 7 and 8 show the full results with both interaction terms.

Hypotheses (H3a) and (H3b) predict that knowledge sharing partners are less likely to diffuse knowledge as their field and intellectual similarity increase, respectively. Models 3 and 4 show that increasing field similarity reduced knowledge diffusion among same room pairs (low: 0.366, p = .024; moderate: 1.003, p = .098) and communicating pairs (low: 2.823, p = .033; moderate: 6.603, p = .050). Models 5 and 6 show that there is no evidence that high intellectual similarity lowered knowledge diffusion among same room (low: −0.0487, p = .739; moderate: −0.003, p = .986) or communicating pairs (low: −0.426, p = .712; moderate: −0.091, p = .936). The results are consistent in Models 7 and 8, which include both knowledge similarity interaction terms.

Figure 3c plots the change in knowledge diffusion (forward citations) between communicating and noncommunicating pairs from Model 4, with 95% CIs. We observe that while communication did alter forward citation trends for pairs with low field overlap, and led to a small increase of 3.521 [−1.465, 8.508] citations for pairs with moderate field overlap, we observe a large decrease from 0.874 [0.347, 1.402] to −2.207 [−4.326, −0.088] for same field pairs, corresponding to a reduction of about 3.082 citations per pair.

To generate greater insight into the types of publications being cited among the low and moderate field overlap pairs, we compared the publication titles of cited papers for pairs assigned to the same versus different rooms. We observe that the same room pairs were more likely to cite papers using advanced imaging technologies (e.g., sample paper titles included: “Massively parallel MRI detector arrays”; “An fMRI study of facial emotion processing patients with schizophrenia”). Overall, there was a greater use of advanced imaging technologies in the publication titles of same room pairs: 42.6% for same room pairs versus 31.3% for different room pairs (t = 2.406, p = .017).⁶ This suggests that the knowledge sharing intervention enabled pairs from different fields to learn about “outside” work, which they subsequently integrated into their own future research. In contrast, the knowledge sharing intervention may have led to a crowding out effect among highly similar pairs. In summary, we find evidence for competitive effects of field similarity on knowledge diffusion, which confirms Hypothesis (H3a).