NETWORK EMBEDDEDNESS, ABSORPTIVE CAPACITY AND
BOARD OF DIRECTORS IN THE CONTEXT OF A FIRM'S
INNOVATION LEVELS
Joris Boer
University of Groningen & Newcastle collaboration (2971143 & 200397506) 12958 Words
07.12.2020
Master Thesis / Dissertation Dual Award Program IB&M Dr. L. Ge
Prof. J. Sapsed
It is in a firm's best interest to increase their understanding of their capabilities and opportunities. Both of these topics relate to being innovative while remaining as effective as possible. A way to comprehend and take advantage of these possibilities in the
environment is to understand the concept of business network embeddedness. This paper proposes the idea that embeddedness can be disentangled into structural and relational embeddedness, where both affect the innovative firm output. Using multiple regression analyses on three alliance networks, I find that both types of business network
2 ACKNOWLEDGEMENTS
In the creation of this master thesis, I received help from various students and professional people. In this section, I would like to first thank the University of Amsterdam and Zoë Kets in particular for helping me with gaining access to the much-needed network data that was one of the most important steps in this study. Also, I would like to thank my supervisors Dr. Ge and Prof. Sapsed for their useful feedback and comments on my draft dissertation and the two meetings for discussion purposes. Lastly, I would like to thank Reddit/r/excel for
3 TABLE OF CONTENTS
INTRODUCTION 4
THEORETICAL BACKGROUND AND HYPOTHESES 10
Innovation and business networks 10
Network embeddedness 13
The role of absorptive capacity 19
The role of the board of directors’ educational level 21
METHODOLOGY 23
Sample & Data collection 23
Network data 23
Company data 25
Measurement of constructs 27
Innovative performance 27
Business network embeddedness 27
Structural embeddedness 27
Relational embeddedness 28
Absorptive capacity 29
Board education level 30
Control variables 30 Analysis 31 RESULTS 31 Network analysis 31 Descriptive statistics 27 Regression results 37 DISCUSSION 41
Findings and contributions 41
Managerial implications 45
Limitations 46
CONCLUSION 47
APPENDIX 48
4 INTRODUCTION
Nowadays, achieving a high level of novelty is something almost every company in any market seems to strive for. Standing out and "being different" can make a company unique in its own way and distinguish itself from competitors (Barney, 1991). For decades now, it has been acknowledged that innovative activities and a successful research and development (R&D) strategy take on a vital role in this process (Cohen & Levinthal, 1989; Bhattacharya & Block, 2004; Mairesse & Mohnen, 2004). However, as more firms recognize the importance of innovations, translating innovative goals into actual output solely through R&D becomes increasingly difficult in this competitive environment (Utterback & Suárez, 1993; Pyka, 2002). Recently, research addressing this issue has been moving towards the impact of the external firm environment. More specifically, business networks have shown signs to directly link to innovation (Narver & Slater, 1990; Afuah, 2000; Rocca & Snehota, 2014). These networks can serve as a valuable source to pursue innovations, depending on the acquirable knowledge gained from external sources in the network (Tsai, 2001). Moore (1993) argues that the linkages with other firms in these networks act as essential facets of a successful innovation strategy since contact with resourceful partners leads to more growth and performance (Roberts, 2001; Buddelmeyer et al., 2009). On top of this, the digital era has caused the scale and scope of business to business communications to significantly increase, which makes expanding the understanding of network effects in the competitive market more important than ever (Bharadway et al., 2013).
5 the importance of firm embeddedness in a business network. He argues that it is crucial to look at this embeddedness from multiple perspectives. More specifically, a distinction of embeddedness is present between structural and relational embeddedness within a business network (Gulati, 1998; Rowley et al., 2000). In these networks, structural embeddedness is determined by analyzing a firm's aggregation of ties, whereas relational embeddedness is about the essence of the relations and involves examining the strength of ties (Burt, 1992; Granovetter, 1985). This division between structural and relational embeddedness in a network is also discussed by Hsueh et al. (2020), who show how innovative performance increased while being more embedded structurally and relationally. At the same time, Uzzi (1997) claims a paradoxical nature of embeddedness: being too embedded can cause firms to experience declining network effects stemming from reduced access to knowledge present at the peripheral of the network. Also, he states that the social pressures from being too
embedded in a network can hinder firms in making functional economic decisions, which causes negative tensions. He even argues that some knowledge ties are permanently blocked off due to this repeated "over embeddedness" (Uzzi, 1996). Lazzarini et al. (2008) confirm this issue and propose that it occurs in a network through uncertainty from engaging with new partners. Firms will be more reluctant to engage in novel relationships because of possible opportunistic behaviors; therefore, they will focus on ties with existing partners. While following this reasoning, there is past evidence to suggest that firms need to balance being structurally and relationally embedded to achieve the most effective strategy that results in the actual innovative output. Currently, academics have not reached a clear consensus on the optimal approach to being embedded both structurally and relationally within a business network.
6 alliance formation. This is because these networks and alliances supposedly involve many different learning processes (Powell et al., 1996). Consequently, firm knowledge absorption capabilities could help clarify differences in structural and relational embeddedness in the earlier mentioned search for innovation. However, few studies have applied these knowledge absorption capabilities in both a relational and structural embeddedness approach separately. Thus, by disentangling the firm embeddedness into structural and relational embeddedness within a business network, I will explore their relation to innovative performance further. This is done through a focus on a company's own ability to absorb external knowledge.
This study proposes two factors that are part of this knowledge absorption capability. First, Cohen and Levinthal (1990) introduce the widely-known absorptive capacity of a firm, which is related to the process of recognizing, assimilating, and exploiting useful knowledge. It is a more general concept and can be regarded as a vital aspect of doing business. Its importance is based upon the idea that the successful transformation of external knowledge into actual innovative ideas is one of the key ways to sustain a certain competitiveness level (Zahra & George, 2002). Tsai (2001) supports this argument and relates it to the previous topic of network embeddedness. He states that even though a firm can be very embedded within a network in general, they must still possess the capability required to tackle the knowledge options within the network.
7 importance; their governing role is crucial in the process of innovation and strategic decisions. The authors point towards the importance of future research on how these boards can be most effective for innovation. Based on these arguments, one can argue that the board of a firm is considered another main factor in knowledge absorption as they act as certain gatekeepers of knowledge (Allen & Cohen, 1969; Kogut & Zander, 1992; Zahra & Filatotchev, 2004).
Ultimately, general absorptive capacity and firms' board educational level are seen as two critical topics that remain largely under-researched while studying network
embeddedness and innovation levels within firms. Therefore, this research will focus on connecting structural and relational embeddedness to innovation while taking into account the absorptive capacity and board educational level of firms. In the end, the discussion depicted above led to the final research question of this study: How does the effect of a firm's structural embeddedness in business networks on innovation compare to the effects of being relationally embedded whilst taking into account the absorptive capacity and educational level in the board of directors?
Consequently, I argue that generally being more embedded within a business network corresponds to higher innovative levels. To specify, when a company possesses a higher structural embeddedness, it will have greater and faster access to a more diverse pool of resources depending on the size of the overarching network. However, in the case of
8 structurally and relationally embedded in the process of innovation. I suggest that since firms with higher levels of absorptive capacity are more effective in transforming external
knowledge into innovative outputs, the structural relational embeddedness effect on innovation will be positively moderated by absorptive capacity.
On the other hand, relational embeddedness effects on innovation would involve a negative moderation from absorptive capacity based on the idea that a higher absorptive capacity is less needed when a firm is more relationally embedded. This is based on the argument that these relationships' strength already provides for ease of knowledge transfer (Uzzi, 1997; Dhanaraj et al., 2004). For the case of the board of directors' educational level, I argue that the educational level involves a certain degree of higher expertise, which is
effective in the area of taking strategic action (Hambrick et al., 2015). Therefore, the moderation effect of the board's educational level on the relationship between structural embeddedness and innovation is also expected to be positive. For the other case of relational embeddedness, the inverse u-shape would experience a negative moderation that weakens the relation to innovation for higher education levels. This is based on the idea that a higher educational level will involve earlier warnings and course of action regarding being over embedded in relationships. On the one hand, this could prevent a relationship from reaping the most amount of benefits, it could also help to prevent the large detrimental effects discussed by Uzzi (1996).
To test the statements above, this study will use quantitative research methods. This is collected from various online databases, which are be analyzed accordingly to test for
9 mapping a whole network accordingly and generating centrality values accordingly.
Moreover, many other scholars have researched the U.S. pharmaceutical market over the recent years, which results in an extensive academic database on topics like competition, alliances, innovation, and knowledge flows (Dong & Yang, 2016; Ball et al., 2018; Roberts, 1999). Next to the fact that this setting will contribute to this research's validity, it will also inhibit a higher chance of explaining any unforeseen results through the other various academic studies on the U.S. pharmaceutical market.
This research contributes to the academic literature on innovation in the following ways. Firstly, it serves as an extension of the already existing studies between business
networks and innovation by applying a different lens of knowledge absorption. On top of that, even though network embeddedness and innovation have been linked before, the area of different types of embeddedness remains mostly undefined in an extensive business context. By including absorptive capacity and board education while studying this relationship, the firm context provides another understanding of network embeddedness in this complex area. Ultimately, this study provides valuable insights for both managers and academics that want to expand their knowledge on the relationship between business networks and innovation. In the following section first, the theoretical background on these concepts is discussed.
10 THEORETICAL BACKGROUND AND HYPOTHESES
Innovation and business networks
The basis for understanding innovation can be derived from an idea proposed by Pierce (1953, p. 117), where he states that "a true innovation must perform some important function". Whilst this seems to narrow the concept down, it is based on only a sole
technological view and seems to lack depth. Baregheh et al. (2009) introduce a study where various definitions of innovation are gathered to generate an overarching multidisciplinary definition of innovation. More specifically, these various definitions were gathered from different research areas such as economics, technology, and science. They conclude that "Innovation is the multi-stage process whereby organizations transform ideas into
new/improved products, service or processes, in order to advance, compete and differentiate themselves successfully in their marketplace". This study will continue this definition by applying it to a business network research setting.
11 value creation (Matinheikki et al., 2017; Holm et al., 1996). Anderson et al. (1994, p.1)
describe such single networks as "sets of two or more connected business relationships, in which each exchange relation is between business firms that are conceptualized as collective actors". Early applications of network theories suggest that they are a source to tap into areas of new possibilities (Walter et al., 2007; Thorelli, 1986). Moreover, they have been directly associated with innovation by their potential to produce new value through their collaborative nature (Dyer & Singh, 1998; Gulati, 1998). The process of open innovation explains this, which was introduced by Chesbrough (2003). Applying the theorem on a firm-level, it can be understood as a process with intended inflows and outflows of knowledge in order to achieve and grow their internal knowledge base (Chesbrough et al., 2006). The process is depicted in figure 1 and shows how external ideas are transferred and internalized within the firm to
generate ideas in current and new markets. The idea here is that a firm's knowledge base does not always contain the most effective and best options, and firms should exploit ideas both externally and internally.
12 In their pioneering article, Powell et al. (1996) propose that another way to look at networks concerning innovation is by seeing them as "networks of learning" (Powell et al., 1996, p. 116). According to them, networks involve collaboration and in return, this collaboration constitutes important aspects in the process of innovating. These include the development of firm capabilities related to managing knowledge, collaborative skills, and awareness of opportunities. In markets where networks are considered the primary source of innovation, these networks involve a path-dependent learning process where all of the above factors also positively influence the connectedness of a firm within that same network. Firms that do not have access to such a network are at an immediate disadvantage when competing with firms connected to the network.
A specific case example of network benefits for the company Toyota is discussed by Dyer and Nobeoka (2000). They state that the Japanese carmaker managed to get significantly ahead of U.S. car manufacturers through an expansive collaboration network in its production area. Through supplier meetings, learning, and problem-solving teams, Toyota was able to create a sizeable knowledge-sharing network where members actively participated. As a result, more valuable knowledge presented itself and the ease of knowledge transfer improved. Through a focus on network identity Toyota was able to create a network
environment that was beneficial for its competitiveness which was predominantly the result of the knowledge diversity and effective processes within the network.
13 Network embeddedness
Next to business networks in general, the embeddedness of a firm in such a network shows signs of playing a significant role in influencing a firm's innovation levels. At first, for understanding structural network embeddedness, the concept of centrality is useful. Centrality in a network concerns measures based on the structure of the nodes and geodesic paths which are two basic components that form every network (Freeman, 1978). In the case of a business network, the firms represent the nodes and the business-to-business communication links represent the geodesic paths. Figure 2 shows an example of what a relatively small business network of 14 companies can look like.
14 Gulati et al. (2000) argue using network structure as a resource, where the most
important notion is that every single firm inside such a network has a unique structural pattern in accessing resources that can create competitive advantages. Knowledge and resource inflows occur for every unit inside a single network differently. Through the communication linkages (nodes) each organization holds different criterium and characteristics that determine how it will learn or exploit the resource and information inflow from other units inside the overarching network (Zack, 1999). Referring this back to the open innovation model of Chesbrough (2003), firms in a business network utilize these communication linkages to exploit external knowledge and internalize it to their advantage. Furthermore, the extent to which this knowledge or other resources are transferable between these businesses is
considered to be mostly dependent upon a single firm's structural embeddedness in a network (Anderson et al., 1994). Generally speaking, when a firm posits a very central structural position in a network, more communication links will be available, resulting in greater access to various resources (Gulati, 1998).
15 Even though an uneven distribution of knowledge might be present in any given network, the firm's structural embeddedness remains of vital importance in accessing these knowledge clusters because of the increased range in the overarching network and their relatively stable position in the dynamic learning environment (Karamanos, 2012). Reagans and McEvily (2003) confirm that this clustering effect, which they explain as "social
cohesion", indeed results in an ease of knowledge transfer. Nonetheless, they also confirm that network range, considered as the degree of communication linkages to other knowledge clusters, plays an important role in this process. The authors indicate that a balance between social cohesion and network range is important, as they warn for the danger of cohesion benefits to be detrimental to the benefits of the range. This reasoning points toward the previous distinction between types of embeddedness in networks and can traces back to Granovetter's arguments (1985). He proposes the idea of social embeddedness influencing behaviors in the economic market where there is a division between the structure of
communication linkages and the extent of interpersonal relations that determine the aggregate formation of a network. On the one hand, structural embeddedness involves "impersonal relationships as a whole" and concerns ideas revolving around the structure of ties, relational embeddedness relates to the specifics of such relationships like trust, size, and scope
(Nahapiet & Ghoashal, 1998, p. 244).
16 expected to increase with network size, but the holes are the key to information benefits". This implies that firms with higher structural embeddedness possess more access options to a greater number of structural holes. These structural holes' potential also links to the relational aspect of embeddedness and can be explained by the strength of weak ties (SWT).
Granovetter (1973) found that men mostly found job opportunities through rather distant contacts in searching for jobs than from close relationships. Consequently, he argued on the strength of weak ties (SWT) stating that a greater number of weaker ties will involve more opportunities. The rationale when applying this to business networks is that stronger ties involve more commonalities between firms and fewer options to actually "learn" or "exploit" new knowledge (Granovetter, 1983). This implies that firms should focus more on the
structural part of being embedded in a network rather than being relationally embedded. Nonetheless, there are other perspectives in the context of relational embeddedness within a network. First, the previous arguments from Reagens and McEvily (2003) issued the idea of cohesion within a network. This is based on certain commonalities between partners that lead to less uncertainty in exchanging resources and knowledge. This argument flows from the fact that knowledge transfer is easier and faster between ties with a stronger connection (Granovetter 1982; Hansen, 1999; Krackhardt et al., 2003). However, Moran (2005) raises an important issue regarding this relational embeddedness by questioning the extent to which these relationships are important. He finds support for the idea that relational embeddedness plays a strong role in the search for innovation outcomes as the facilitation of exchanging complex ideas is more prominent in the presence of stronger ties. These complex ideas and knowledge often are the most important sources for effectively innovating
(Cavusgil et al., 2003). Therefore, one can argue that an emphasis on relational embeddedness must also be important in the search for knowledge through structural holes, as this
17 embeddedness embodies pitfalls because of several reasons. First, when an organization becomes too relationally embedded it may fall at risk because of resource dependency on core partners. As such a core partner fails, negative effects on performance can occur.
Alternatively, being more structurally embedded serves a company with increasingly diverse options that can counter the negative effects of other network partners' failure. Second, institutional shifts could have detrimental effects on firms focusing on certain network ties specifically. For example, sudden regulatory changes may even eliminate certain network benefits and can result in harmful effects on firms that are too relationally embedded . Ultimately, this results in instabilities by being "over-embedded" within a network. Uzzi (1997, p. 58) points out, "redundant ties to the same network partners mean that there are few or no links to outside members who can potentially contribute innovative ideas". On that account, gaining access to structural holes becomes harder as organizations get trapped into a focus of evading network expansion by only building on current relationships. This specific need to balance structural and relational embeddedness has also been confirmed on the individual level by Cattana and Ferrianci (2008). They studied creativity in the Hollywood film industry and found that individuals who are in a position where they have access to both their core and outer section of their social network were more prone to generate creative ideas.
To summarize, in the academic literature, there is an agreement on the importance of network embeddedness concerning being innovative. However, when separating
18 component of accessing novel knowledge sources needed for innovative performance (Burt, 1992). Coincidingly, those firms will have to manage their relational embeddedness as well, both with weak and strong ties. Weak ties can possess valuable knowledge, and being more relationally embedded with those ties eases this important complex knowledge transfer. However, firms are at risk of becoming too relationally embedded with their partners, which can negatively affect performance by being overdependent and trapped within their network (Uzzi, 1997). Thus, there is a need to balance the benefits of relational embeddedness so that they will not reach a phase of decline. The above discussion results in the following two hypotheses:
Hypothesis 1: The structural embeddedness of a firm in a business network is
positively related to its innovation levels
Hypothesis 2: The relational embeddedness of a firm in a business network and its
innovation levels follows an inverted U-shaped relationship
The role of absorptive capacity
When looking back at the structure of a business network shown in figure 2, a
19 innovative performance of a company. Since external inflows of knowledge are crucial to achieving successful innovative goals, a low absorptive capacity poses a great barrier. For example, it could mean that such an organization will experience difficulties in the
recognition, translation, and exploiting processes of important knowledge.
When a firm is in the process of acquiring and using new knowledge for its benefits this study proposes two deciding factors that can be considered to determine its success. First, the assessment of the current level of absorptive capacity is important to correctly analyze the company's current capabilities in relation to its desired goals. In return, when this assessment of the capabilities concerning the goals results in being insufficient for absorbing the
knowledge, firms need to make a strategic decision concerning more investments into their absorptive capacity (Cohen & Levinthal. 1990). This is where the second factor can be introduced: the willingness to invest in the own company to improve the absorptive capacity. Specifically, Cohen and Levinthal (1990) introduce a model (Figure 3) that contributes to the discussion above. They state that the level of R&D activity of firms is very closely associated with their absorptive capacity. At the same time, they argue that the learning incentives to invest in R&D are dependent upon the technological opportunity and appropriability. Where technological opportunity is considered technical knowledge that a firm can use to benefit its knowledge output and appropriability are understood as the level of valuable knowledge that
20 can be effectively learned. As competitors become more technologically advanced, the value of this knowledge might turn out to be lower which involves fewer benefits to invest in R&D to be able to acquire this knowledge.
The above theorem implies that a firm's absorptive capacity inside a business network is dependent on analyzing the environment and its incentives. In return, it influences the level of innovation and as a result of learning to turn valuable knowledge into output. In the case of structural embeddedness, a central network position gives greater access to a variety of resources in a network, where the absorptive capacity of a firm plays a large role in
influencing the number of resources that actually can be exploited in such a position (Helena Chiu & Lee, 2012). For one, it can serve as a barrier; a company not willing to invest in its research capabilities will have greater difficulties recognizing, translating, and exploiting all of the relevant knowledge for successful innovative outcomes. With a more central network position, a company will need a higher absorptive capacity to use the centrality benefits for their own success. This interaction between centrality and absorptive capacity is important as a greater variety of knowledge sources require a higher absorptive capacity (Tsai, 2001).
On a similar note, being more relationally embedded would also involve the ability to exchange a greater amount of complex knowledge (Uzzi, 1997). In this instance, the
21 strengthen the effect of relational embeddedness on innovative outputs (Krackhardt, 1999; McEvily et al., 2003; Tortoriello et al., 2012). Thus, in the first stages of being relationally embedded, the absorptive capacity is expected to take on a similar role as in the case of structural embeddedness, however, this effect will be lower once a firm develops stronger relationships. This results in the following:
H3a: The positive relationship between a firm's network structural network embeddedness
and its innovation levels is moderated by the absorptive capacity of the focal firm: the
relationship will be stronger for firms with a higher absorptive capacity
H3b: The inverted U-shaped relationship between a firm's relational network embeddedness
and its innovation levels is moderated by the absorptive capacity of the focal firm: the
relationship will be weaker for firms with a higher absorptive capacity
The role of the board of directors’ educational level
22 In particular, a board's educational level has been argued to be a critical part of a firm's survival (Brüderl et al., 1992). Education in general has been found to be related to growth and to more social cohesion (Gradstein & Justmen, 2000). In return, this could mean that higher education in boards fosters a common strategic course and decreases communicative barriers and possible conflict. Papadimitri et al. (2020) confirm this effect by finding that credit ratings were significantly better for firms operating under a board with a higher educational level. In another study, Wincent et al. (2010) specifically discuss the relation to innovation and state that board members' education is important in finding solutions for complicated challenges and decisions. They connect the educational level to networks as well and propose that "organizational routines, such as member firms' innovative capabilities are invisible to outsiders and are sticky, the formation of a structural link from a network board by itself may not be sufficient to enhance innovation." (Wincent et al., 2010, p. 267). They propose that firms also need to rely on skills like the educational level to decode and
internalize valuable knowledge stemming from other firms. Ultimately, one can argue that the educational level speeds up the process of effective resource allocation, which in return is a crucial component of an effective innovative strategy (Chen & Huang, 2009).
23 is similarly important. Higher education within the board could indicate a greater quality in awareness towards the actual nature of relationships and foster innovative benefits from those relationships. However, they can also possibly give a word of warning whenever a company becomes too relationally embedded with other firms in a business network. General literature on educational levels also relates higher education to being more risk-averse for samples on individuals (Halek & Eisenhauer, 2001; Jung, 2015). Therefore, it can be proposed that this advice to avoid over embeddedness is more likely to be given as their education improves. Ultimately, this led to the following:
H4a: The positive relationship between a firm's network structural network
embeddedness and its innovation levels is moderated by the educational level within the
board of directors of the focal firm: the relationship will be stronger for firms with a higher
level of education within the board of directors
H4b: The inverted U-shaped relationship between a firm's relational network
embeddedness and its innovation levels is moderated by the educational level within the
board of directors of the focal firm: the relationship will be weaker for firms with a higher
24 Figure 3: Conceptual framework of this study
METHODOLOGY Sample & Data collection
Network data
25 alliance is characterized by technology transfer, research and development, and whether the partners are from the same nation. In this study, three alliance networks are constructed with a 5-year window. This is based on the idea proposed by Kogut (1988), where he states that the average length of an alliance is about five years. Therefore, the SDC data is divided into three separate samples: 2000-2005, 2005-2010, 2010-2015. Where each period begins and ends on the 1st of January every instant. The first sample contains 1492 unique companies, the second 1728, and the third 1102. Before constructing network variables that measure centrality in a network, the data is coded into dyadic relationships, which means that for each alliance where there are more than 2 companies involved, the data can be read as relationships between each company separately. For example, alliance ABC is coded into dyadic form as AB, AC, and BC. This is a necessary step in the process of statistical network data. The program that will construct the alliance networks is called UCINET 6, which is software for the analysis and visualization of network data (Borgatti, Everett & Freeman, 2002). This program requires data to be sent in a matrix format. Where each company is on the first row and column in excel and relationships present or not present are depicted by either a "1" or a "0". By applying queries and visual basic macros within excel, three matrices are constructed and imported into UCINET 6.
Company data
After the calculation and extraction of the network data, this data is matched to data from COMPUSTAT. This process is complex since the SDC database only uses 6 digit-CUSIP identifiers for every company and the COMPUSTAT database can only read digit-CUSIP codes that are 8 or 9 digits. Therefore the CUSIP codes of the initial sample have been coded into 8 digits CUSIPS based on the manual from the Princeton Library found at
26 The company data that is gathered includes total assets, employees, revenue (sales), and R&D expense for the years 2000-2016. When merging this data file to the computed UCINET dataset, the three samples decrease to 262, 234, and 169 companies.
The final data that is collected from BoardEx and concerns information on the board of directors. Specifically, the time on the board of directors, the number of qualifications, and the actual number of board members present throughout the years 2000-2015. Also, patent data is collected that accounts for the measurement of innovative performance. Since the ORBIS database only shows the total number of patent counts and not the patent count throughout the years, this variable is hard to extract and match to the existing dataset as well. The PATSTAT database is one of the largest databases that register patents from a hundred offices around the globe (De Rassenfosse et al., 2014). However, there is a certain complexity surrounding this database as the data has to be extracted using SQL coding, company
identifiers are not used, and downloads are limited under the free trial on 10000 rows (companies). To extract the most amount of data from this database, the minimal number of patents that had to be filed was set on 5. For each of the three alliance networks (2005, 2010, 2015), the number of patents per company from those years and a one-year lag after the alliance network ended were collected. Since company identifiers were not present, matching was done based on the company name. The final merge involved matching the BoardEx and PATSTAT data with the previous SDC/UCINET and COMPUSTAT data and resulted in three final samples of 74, 93, and 98 companies. The table below summarizes this collecting and merging process. After a listwise deletion of observations, the samples included 72, 76, and 77 companies eventually for the regression analyses.
Table 1. Data collection and merging procedure.
Sample 1 (2000-2015) Sample 1.1 (2000-2005) Sample 1.2 (2005-2010) Sample 1.3 (2010-2015)
SDC 5186 alliances 1970 Alliances 1944 Alliances 1271 Alliances
UCINET 6 1492 Companies 1728 Companies 1102 Companies
COMPUSTAT 262 Companies 234 Companies 169 Companies
27 Measurement of constructs
Innovation level
Since a firm's research and development (R&D) expenses is an output variable instead of an input variable. Using this as a measurement for the innovation level might result in misinterpretation or errors (Kleinknecht, 1996; Mork & Yeung, 2001). Therefore, the patent count of a firm is such an output variable that can capture this concept. This has been done before in the context of measuring innovation (Grabwoski, 2002; Branstetter & Sakakibara, 2002). In addition, patents are the direct result of a company's knowledge base and represent the output as the result of this base (Jaffe et al., 1993). Even though measuring the number of patents will give an idea of the innovative performance after the years of participating in a given business network, it may be an incomplete representation of how successful these patents are. Therefore, this variable will be approached with some caution. To ensure the validity of the analysis a 1-year time lag of this variable will be used. This means that the analysis of the alliance of 2005 will involve the patent count of 2006.
Business network embeddedness Structural embeddedness
28 centrality measures are constructed for each observation. This study uses two measures of centrality in the case of structural embeddedness: degree centrality and closeness. Each of these two variables covers the 5-year periods specified above. According to Freeman (1978), degree centrality for a single node concerns the number of nodes it has direct links with. Referring back to figure 2, the degree centrality of company 7 is equal to 4. Closeness
centrality involves a more technical approach and explains the independence of a given node. It measures how close a node is to all other nodes in a network, and shows a relative score based on how close others are in a network. For example, in figure 2 node 11 needs to pass through 4 other nodes to reach node 6, whereas node 2 only needs to pass through 2 other nodes. Node 2 likely has a higher closeness score in this network than node 11 because it can reach all other nodes in the network relatively faster. Closeness centrality therefore can be understood as the sum of distances for a node to all other nodes. The calculations are as follows:
𝐷𝑒𝑔𝑟𝑒𝑒 𝑐𝑒𝑛𝑡𝑟𝑎𝑙𝑖𝑡𝑦(𝑥𝑖) =𝐷𝑒𝑔𝑟𝑒𝑒 (𝑥𝑖)
𝑛−1 𝐶𝑙𝑜𝑠𝑒𝑛𝑒𝑠𝑠 𝑐𝑒𝑛𝑡𝑟𝑎𝑙𝑖𝑡𝑦(𝑥𝑖) =
𝑛−1 ∑𝑗≠𝑖𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒(𝑥𝑖,𝑥𝑗)
Where n is the number of nodes in a network and degree is the sum of all directly connected nodes of node xi. Distance (xi, xj) is the distance in edges from node xi to xj. (Freeman, 1978; Rochat, 2009). These calculations are based on an undirected structure, which means that if company A has an alliance with company B, company B also has an alliance to company A. The scores extracted for both of these centrality measures are normalized within the program of UCINET 6 to account for the size of the network. This results in the ability to compare the three different alliance networks which involved a difference in size.
Relational embeddedness
29 operationalize this concept. However, one could argue that a relationship experienced a large resource transfer only once, which could cause bias in the interpretation of relational
embeddedness. Meuleman et al. (2010) measured relational embeddedness through the number of recurrent deals over a period of 5 years, however, the limited dataset prevents an accurate representation of recurrent deals. More specifically, to measure relational
embeddedness this study will rather focus on the cultural nature of these relationships. For example, two companies from the U.S. are likely to have more shared values than an alliance where one company is from the U.S. and the other from Japan. Dhanaraj et al. (2004) show that trust and shared values can ease tacit and explicit knowledge transfer. Trust can be a difficult concept to operationalize in this research setting. However, capturing shared values can be done by accounting for the number of alliances where partners are from the same nation. More shared values can likely result in a relationship that involves more social embeddedness as a result of a better "cultural fit", this is the result of sharing similar cultural and organizational values (Noorderhaven et al., 2002, p.36). Therefore, the variable of relational embeddedness will be measured as the number of alliances where all partners are
from the U.S. This variable will be constructed for each of the three samples over the three
designated periods of 5 years. Finally, the regression analysis will also include a squared term of this variable, this is needed to test the hypothesized main relationship of h2, where an inverted u-shape relation is predicted.
Absorptive capacity
30 capability involves other factors as well. However, a large amount of literature uses the
measure of R&D intensity as well as this is a relatively accurate indication of the amount a company is willing to invest back into their company for the purpose of R&D (Stock et al., 2001; Tsai, 2001; Schmidt, 2010). Which is argued to be a vital part of a company's learning process. (Cohen & Levinthal, 1990; Tsai, 2001). In this case, this study measures R&D intensity with a 2-year lag on the dependent variable of patent count, this is done to ensure
there is enough time to see the effects of the investment a company makes in R&D on their innovation levels.
Board education level
The level of education in a board of directors will be measured in the average number of qualifications present in a board during the three constructed 5-year windows of the
sample. This includes all of the qualifications at the undergraduate level and above, which has also been used by other studies as indicators for educational levels. Most of these studies referred to this as "educational attainment" where and include bachelor degrees and higher qualifications, which is adjacent to the undergraduate level and above (Headd, 2000; de Jong et al., 2011).
Control variables
31 Analysis
In order to address the above hypotheses, this research design involves a descriptive quantitative approach as it gathers data from databases and analyzes this for possible causal effects of our main relationship studied. This will be done through several multiple regression analyzes and is a deductive process as the analysis is characterized through the application of general theory to more particular effects. The research paradigm concerns the positivism paradigm which corresponds with doing quantitative research and focuses on existing theories in order to predict certain phenomena (Collis & Hussey, 2014: 44). Before computing the different descriptive statistics and correlation matrices for each of the three designated periods, the data undergoes several checks to get a better comprehension of the overall analysis. This is done to gain an overall understanding of the distribution and raise awareness towards possible issues like multicollinearity. While checking the data through histograms and plots for outliers, several observations show signs of having very large values on variables like total assets. I have decided to keep these observations in the sample as the number of observations with extreme values is limited and deleting these companies can jeopardize the representativeness of the analysis.
RESULTS Network analysis
32 alliance network in 2010. The table below shows the number of edges (links) and nodes (companies) within each of the three alliance networks.
Table 2. Alliance networks general measurements
For example, figure 4 shows a visualization of the alliance network present in 2005. Whereas this network contains 1492 companies and there are 4730 ties between companies. These ties are undirected as the matrix that has been imported is binary. This means that when a
relationship is present, this is a two-way relationship. In addition, the average degree
centrality, which is the number of connections per node, is just above 3 for the network at the start of 2005. The analysis also indicates that in the network of 2005, the number of main components is 301. In the figure, they are characterized by a red color.
Figure 4. U.S. Pharmaceutical network in 2005.
Descriptive statistics
The descriptive statistics and correlation matrices can be found below in table 3. There are some important things to note from this table. First, the predicted variable of patent count was transformed into a log function for all three different periods. Since many observations
Alliance network 2005 2010 2015
Number of nodes 1492 1727 1101
Number of ties 4730 4576 2658
Average Degree centrality 3,17 2,65 2,414
Average Distance 5,05 5,05 4,67
Components 301 425 301
33 from the initial sample were not matched, some of the variables possess a highly skewed distribution. Based on visualized histograms, R&D intensity and total assets were also logged in order to generate a more normal distribution. Feng et al. (2014) state that analyzing the results must be done using a cautious approach, as these transformations can involve pitfalls in the interpretation of tests. Other things to take away from the means is that on average, every company has more than 8 directors on their boards each year of the moving window. On top of that, every board member possesses more than 2 qualifications on average and the mean for the number of alliances where both partners are from the U.S. is only above 2 for the alliance network of 2005.
Some of the variables show high correlations which might indicate that
multicollinearity in the models poses an issue. However, after analyzing the VIF scores for every three periods, the means of the VIF's were 2.90, 2.64, and 2.86 respectively. Besides, the individual VIF's did not exceed the threshold of 10 and no individual variable has a VIF that is unusually larger than zero, which constitutes the detection of severe multicollinearity problems that could arise (Alin, 2010; Mansfield & Helms, 1982).
Consequently, the variance of the error term is examined in terms of homoskedasticity. Through the analysis of plotting the independent variables and the residuals, the data is
suspected to be subject to the problem of heteroskedasticity. The plotted residuals show signs of an uneven distribution based on the y-line of zero. The Breusch-Pagan test for
34 Testing on the normality of the residuals of regression models was also done for every three periods using plots and the Shapiro-Wilk test. This confirmed a normal distribution of residuals for regression models in the alliance networks of 2005 and 2015. However, the p-value for the models from 2010 is below 0.01, indicating significant reason to believe a violation of normality for the residuals. This is a serious issue that can cause inaccurate statements about any findings that occur from these regression models (Jargue & Bera, 1980). However, when plotting these residuals there does not seem to be any substantive deviations from normality. Additionally, a kernel density graph to assess normality can be found in the figure below. This plot clearly shows that the distribution is slightly skewed to the left with a few deviations from normality around the tails. Therefore, any interpretations of regression results from models based on the alliance network of 2010 must be treated with caution.
Figure 5. Kernal density graph and normality (model 6)
Lastly, at test was performed for the possible presence of an inverted u-shape relation between number of U.S. alliances and patent count. This was done according to the
Table 3: Descriptive statistics and correlation matrix
Variable Obs Mean Std. Dev. 1. 2. 3. 4. 5. 6. 7. 8. 9. 1. Patents ¹ 72 3.088 1.594 1.000
2. Closeness centrality 72 0.095 0.012 0.387*** 1.000
3. Degree centrality 72 0.005 0.006 0.399*** 0.434*** 1.000
Alliance network 2005 4. R&D intensity ¹ 72 -0.471 2.268 -0.387*** 0.069 -0.069 1.000
5. No. only US alliances 72 2.041 3.548 0.307** 0.334** 0.891*** -0.054 1.000
6. No. only US alliances - squared term 72 16.581 71.473 0.263* 0.204 0.804*** -0.107 0.917*** 1.000
7. Board educational level 72 2.531 0.389 -0.028 0.182 0.139 0.353** 0.144 0.068 1.000
8. Total assets ¹ 72 6.506 2.276 0.730*** 0.288* 0.355** -0.602*** 0.309** 0.297* -0.141 1.000
9. Board size 72 8.746 2.293 0.482*** 0.124 0.374** -0.397*** 0.406*** 0.385*** -0.045 0.681*** 1.000 1. Patents ¹ 76 3.135 1.482 1.000
2. Closeness centrality 76 0.068 0.011 0.160 1.000
3. Degree centrality 76 0.003 0.005 0.285* 0.451*** 1.000
Alliance network 2010 4. R&D intensity ¹ 76 -1.216 1.881 -0.383*** 0.170 0.005 1.000
5. No. only US alliances 76 1.699 2.847 0.250* 0.346** 0.861*** -0.050 1.000
6. No. only US alliances - squared term 76 10.903 48.963 0.205 0.226* 0.808*** -0.048 0.928*** 1.000
7. Board educational level 76 2.521 0.402 -0.011 0.342** 0.155 0.124 0.182 0.075 1.000
8. Total assets ¹ 76 6.974 2.500 0.595*** 0.054 0.423*** -0.631*** 0.388*** 0.315** -0.043 1.000
9. Board size 76 8.959 1.920 0.413*** 0.087 0.408*** -0.297** 0.327** 0.275* 0.048 0.612*** 1.000 1. Patents ¹ 77 2.817 1.389 1.000
2. Closeness centrality 77 0.094 0.010 0.247* 1.000
3. Degree centrality ¹ 77 0.004 0.006 0.402*** 0.470*** 1.000
Alliance network 2015 4. R&D intensity ¹ 77 -0.712 1.946 -0.140 0.118 -0.121 1.000
5. No. only US alliances 77 1.429 2.248 0.400*** 0.477*** 0.887*** -0.099 1.000
6. No. only US alliances - squared term 77 7.041 23.396 0.324** 0.345** 0.901*** -0.099 0.942*** 1.000
7. Board educational level 77 2.541 0.459 0.041 0.217 0.039 0.267* 0.043 0.035 1.000
8. Total assets ¹ 77 6.714 2.507 0.560*** 0.214 0.477*** -0.427*** 0.493*** 0.441*** -0.125 1.000
9. Board size 77 8.593 2.262 0.507*** 0.287* 0.499*** -0.198 0.470*** 0.386*** 0.148 0.743*** 1.000 R&D intensity, research and development expense / total revenue. Correlation matrix follows listwise deletion where n = 72 for 2005, n = 76 for 2010 and n = 77 for 2015. *** p <.01, ** p <.05, * p <.1
Table 4: Regression results on patents using alliance networks 2005, 2010 and 2015 2005 2010 2015 Independent variables 1 2 3 4 5 6 7 8 9 Closeness centrality ᶜ 38.52*** -10.30 -34.31 16.54 170.7 235.3** -7.370 -108.1 -126.5 (12.92) (107.6) (82.36) (17.23) (117.9) (110.2) (18.66) (109.0) (107.0) Degree centrality ᶜ 93.26* -104.5 -179.4 82.20 -733.7 -641.3 92.77** 128.9 177.9 (53.76) (307.8) (290.3) (53.53) (487.6) (462.6) (41.56) (355.2) (323.8)
No. only US alliances ᶜ -0.0533 3.780 2.525 0.0548 -0.338 -2.160** 0.419** 3.200** 3.078**
(0.127) (2.299) (2.062) (0.142) (1.172) (1.003) (0.197) (1.363) (1.373)
No. only US alliances - squared term ᶜ -0.000669 -0.432* -0.286 -0.00455 0.176 0.256** -0.0410** -0.222 -0.197
(0.00330) (0.220) (0.189) (0.00655) (0.127) (0.114) (0.0171) (0.179) (0.172)
R&D intensity ᵃ ᵇ -0.284*** -0.197 0.0800 -0.306*** -0.770 -0.685 -0.0692 0.156 0.594
(0.0772) (0.489) (0.382) (0.0719) (0.558) (0.547) (0.0782) (0.714) (0.730)
Board educational level ᶜ 0.115 -0.694 -1.400 -0.188 3.126 4.100 0.183 -2.326 -2.868
(0.433) (3.171) (2.412) (0.355) (3.121) (2.907) (0.375) (3.569) (3.463)
Interaction effects between R&D intensity and main predictors
Closeness centrality x R&D intensity 0.481 -0.602 9.431 12.43 -1.563 -6.194
(5.957) (5.000) (7.914) (8.120) (7.853) (8.133)
Degree centrality x R&D intensity -99.75* -53.71 7.772 51.44 11.27 23.28
(57.66) (53.49) (52.62) (42.19) (36.91) (34.91)
No. only US alliances x R&D intensity 0.105 0.0862 -0.304 -0.338* -0.241* -0.0958**
(0.104) (0.097) (0.155) (0.148) (0.126) (0.129)
No. only US alliances - squared term x R&D intensity 0.00132 -0.00178 0.0655* 0.0491 0.0369 0.0345
(0.00864) (0.00938) (0.0367) (0.0318) (0.0357) (0.0309)
Interaction effects between board educaitonal level and main predictors
Closeness centrality x Board educational level 17.08 19.91 -58.18 -77.98* 35.49 42.07
(41.30) (32.56) (45.28) (42.13) (40.06) (38.91)
Degree centrality x Board educational level 78.33 98.43 334.6* 267.3 2.870 -16.24
(119.2) (114.4) (169.5) (164.6) (118.6) (106.9)
No. only US alliances x Board educational level -1.488 -0.99 -0.0388 0.590* -1.144** -1.140*
(0.896) (0.809) (0.405) (0.340) (0.488) (0.477)
No. only US alliances - squared term x Board educational level 0.164* 0.107 -0.027 -0.0633** 0.0831 0.0806
(0.0847) (0.0733) (0.0305) (0.0292) (0.0528) (0.0503) Control variables Total assets ᵃ ᵇ 0.429*** 0.378** 0.254*** (0.145) (0.142) (0.0920) Board size ᶜ -0.0167 0.0598 0.0497 (0.112) (0.091) (0.0892) Constant -1.292 1.265 2.222 1.871 -6.789 -12.535 2.415 9.670 9.344 (1.410) (8.221) (5.845) (1.428) (7.972) (7.588) (1.658) (9.500) (9.392) Model Summary N 72 72 72 76 76 76 77 77 77 R2 0.373 0.462 0.615 0.246 0.328 0.341 0.230 0.237 0.458 Adjusted R2 0.315 0.330 0.502 0.181 0.174 0.192 0.164 0.135 0.314 Change adjusted R2 0.015 0.172 -0.007 0.018 -0.029 0.179 F statistic 45.26*** 32.67*** 24.66*** 6.13*** 28.76*** 5.52*** 5.73*** 4.30*** 6.78***
*** p <.01, ** p <.05, * p <.1, Robust standard errors in parentheses ᵃ These variables have a 2 year timelag with the dependent variable patent count ᵇ These variables are logged
37 Regression results
All of the hypotheses were tested in order through multiple regression models that are reported in table 4. First, the results from models 1, 4, and 7 show the regressions for each alliance network with only the independent variables included. Closeness centrality and degree centrality, which capture the concept of structural embeddedness, show positive
coefficients that are both significant for only the alliance network of 2005: 38.52 (p<.01) and 93.26 (p<.05). Another positive coefficient for structural embeddedness is also significant in model 7, however, this involves only the variable degree centrality: 92.77 (p<.05). In the case of the the number of U.S. alliances, which captures relational embeddedness, the variable shows some significant results as well. In model 7 in the alliance network of 2015, there is a significant coefficient for both the normal and the squared term: .419 (p<.05) and -.0410 (p<.05). Finally, R&D intensity as the measurement for absorptive capacity produces significant negative coefficients in models 1 and 4 of .283 and .306 on the level of p<.01.
When adding the interaction effects in models 2, 5, and 8, the amount of variation explained by the predictors in the dependent variable patent count (R2) increases by .015 model 2. However, for models 5 and 8 it decreases by .007 and .029. Also, the interaction term between degree centrality and R&D intensity is significant with a negative interaction term in model 2 (the network of 2005) for p<.1. Another notable significant interaction term before adding the control variables to the models is the .0655 (p<.1) interaction between the squared term of the number of U.S. alliances and R&D intensity in model 5. On top of that, the interaction between the number of U.S. alliances without the squared term and R&D intensity is significant in model 7 with a coefficient of -.241 (p<.1). Moreover, the second
38 the squared term of the number of U.S. alliances in model 2 (.164 for p<.1). In model 7 the moderator exhibits a negative interaction term with the general term of the number of U.S. alliances.
Through the addition of controls in models 3, 6, and 9 the amount of variation explained in the patent count by the regressor variables changes with 17.2, 1.8, and 17.9 percentage respectively for the alliance networks of 2005, 2010, and 2015. In addition, the log of the total assets shows a positive coefficient with patent counts in all of these models for p<.05. Model 6 produces a significant positive coefficient of 235.3 for degree centrality for p<.05. In addition, where a significant negative coefficient is present for the number of U.S. alliances in this model (-2.160 for p<.05) a positive coefficient is found of .256 for the same
squared term of the number of U.S. alliances. Furthermore, the interaction term between the number of U.S. alliances and R&D intensity is also is significant with a negative sign (-.338
for p<.1). Further findings of model 6 include two significant interaction terms on the level of p<.1 for closeness centrality with the board educational level and the number of U.S.
alliances with the board educational level (-77.98 and .059). Additionally, there is also a
significant negative interaction for the squared term of the number of U.S. alliances and board educational level in this model (-.063 on p<.05). For the alliance network of 2015 in model 9,
there is only one main predictor that posits a negative coefficient: the number of U.S. alliances for 3.200 on p<.05. The only two interaction terms that are significant involve the
interaction between the number of U.S. alliances with both R&D intensity and the educational level of the board (-.0958 for p<.05 and -1.140 for p<.1).
Based on these results, hypothesis 1 was partly supported as the measures of structural embeddedness showed significant positive relations in models 1, 6, and 7. In the case of hypothesis 2, the squared term of the number of U.S. alliances indicates some significance in
39 added. Together with the squared term, the general term is also significant in this model. Below, this effect on the dependent variable is plotted in figure 6. The graph shows, that not only there is support to reject hypothesis 2, but the findings also are completely contradictory: the suggested inverted u-shape relationship between relational embeddedness and innovation levels follows a u-shape relationship. At first, the patent count seems to declines as the
number of alliances increases, however, after around 5 U.S. alliances this effect becomes positive and starts to increase.
For the first moderator of absorptive capacity measured as R&D intensity, the positive moderation hypothesized in h3a on structural embeddedness finds no support in any of the most complete models 3, 6, and 9 where control variables were added. In the case of hypothesis 3b, no support indicates a significant interaction with the squared term of the
number of U.S. alliances.
In the case of the second moderator variable of the educational level within the board of directors, there is a significant negative interaction term found in model 6 with the
predictor variable of closeness centrality. This interaction is also plotted below in figure 7 for interpretation purposes. The range of the graph's x-axis is based on the min and max values of the predictor variable closeness centrality. It can be seen that for lower levels 1 standard deviation below the mean of the values for the moderator, the relationship between closeness centrality and the patent count is positive. However, at higher levels for the moderator, this
relationship becomes less positive and crosses over to being negative. Thus, there is no support for hypothesis 4a, instead, the moderation appears significantly weaken the positive effect between the main predictor closeness centrality and the responding variable of the patent count. Lastly, hypothesis 4b argued on a negative moderation of the board of directors'
educational level that would weaken the link between relational embeddedness and patent
40 negative interaction term is found. When graphing this interaction term, it must be noted that hypothesis 2 involved a suggested inverted-u shape, however, the results show a contradictory
finding in this model of a u-shape relation. Therefore, hypothesis 4b in principal can already be rejected. However, there is still a negative moderation for the educational level within the board of directors that weakens this u-shaped relationship, as can be seen from figure 8: for
higher levels of education within the board the relationship between the number of U.S. alliances (relational embeddedness) and patent count (innovation levels) becomes less
pronounced and "flatter", this can be seen from the green line in the graph. Where on the other hand, lower levels of education within the board involves steeper slopes and a more
pronounced u-shape.
Figure 6. Graph of U-shape relation between the number of U.S. alliances and patent count (model 6)
41 DISCUSSION
Findings and contributions
The results described above provide a moderate level of support for a positive effect of structural embeddedness on the patent count. Even though this relationship is only found to be significant on some occasions, it is still important to acknowledge the difference between being relationally embedded as a company in the context of this study. Structural
embeddedness, measured by the number of ties and the connectedness in a network, is likely to play an important role in the process of taking advantage of opportunities and access to knowledge that contribute to a company's innovation levels (Burt, 1992; Moran, 2005; Chiu & Lee, 2012). These findings related to the main effect of structural embeddedness
correspond to Tsai's (2001) ideas where a network position measured based on network structure related to higher innovation levels.
42 (1992) also locks in on this as having more ties serves a firm with the ability to access
information that is characterized by many differences and a high level of uniqueness. Across all alliance networks, the main effects of both degree centrality and closeness centrality show some support for a positive effect on companies' innovative output. Once companies do not directly link with others that can be considered valuable weak ties or access to structural holes, a higher closeness centrality could support the process for a company to reach these key network resources efficiently. This process becomes more difficult once the focal company is separated from others within the network and therefore less connected by the closeness measure.
However, once companies have a higher educational level within the board of directors, the positive effect of closeness centrality starts to decline. Ultimately, this
43 Another explanation for this negative moderation is that higher education levels have also been related to being risk-averse by past scholars (Halek & Eisenhauer, 2001; Jung, 2015). In the end, the findings of this study with relation to board education contain surprising results and have to be approached with some caution. Nevertheless, this shows that it is still a research direction that remains to be important.
On the other hand, relational embeddedness shows different signs. Results from model 9 show a positive relationship between the general term of U.S. relationships and patent count. This relation involves negative interaction terms from both of the moderators, which means that this positive relationship is weakened by both the board of directors' educational level and the R&D intensity. Also, results on relational embeddedness in model 6 seem to play a contradictory role that challenges Uzzi's views (1996). He argues that being more relational embedded includes the danger of being too embedded, which impedes performance and growth, and, therefore, the innovation levels. In fact, the role of relational embeddedness found in this study is more in line with Moran's (2005) general argument, who proposes that relational embeddedness is specifically important in firm performance. Especially under conditions of a higher commonality level in terms of culture, the findings of alliance network 2010 suggest that U.S. firms connecting to partners from the same nation eventually
experience a higher innovative output. However, this effect will first decline as it follows a u-shaped relationship. A higher educational level of a board of directors possibly weakens this relation: the higher the education, the less innovative output is generated by the firm from increasing their U.S. partners. Ultimately, this indicates that it is important to ensure a certain level of cultural commonality between the partners to innovate effectively, but firms have to be aware of their educational board level at the same time.
44 innovative performance is not surprising (Nielsen., 2007). However, trust could play an unobserved role in this process which can explain the initial declining effects of more U.S. alliances, as this concept is often closely related to relational embeddedness (Djanaraj et al., 2004). More familiarity with U.S. partners specifically could be a justification for the change in effects on innovation levels over time, as this influences trust in these partners and
optimizes specific management knowledge for these relationships (Gulati., 1995; Zaheer et al., 1998; Lavie et al., 2012). The negative moderation effect of the educational board level that weakens this u-shaped relation can again be explained through possible more cognitive conflicts and a higher level of risk-aversion that impede the effective decisions and advise the board of directors give on managing these U.S. alliance partners. In the end, this too is an area that must be looked into further.
This study contributes to the past literature on structural and relational embeddedness in networks by applying these concepts on the general level of a firm instead of the individual managerial level of studies like Moran (2005). Moreover, it shows that there might be a reason for firms to analyze network embeddedness in a dual way, where the effects of structural embeddedness in general on innovation are possibly more straightforward than the effects of relational embeddedness. This is because there is reason to account for the board of directors' education level within firms when considering alliance partners in a business
45 centrality score on patent count is negative due to possible overreliance on boards to solve challenges related to a more complex network structure. However, the exact reasons for these effects remain speculative and further studies will have to point out whether this can be attributed to the discussed factors above.
In the end, this study finds that there is no support for any paradox within the concept of relational embeddedness. In fact, as firms become too focused on relations between
partners that share the same cultural values, a higher patent count is actually the result of such an increase. Ultimately, the educational level of the board of directors complexes the
relationships on both the structural and relational side of embeddedness by weakening the positive link between closeness centrality and patent count and at the same time, flattening out the found u-curved relation between relational embeddedness and patent count.
Managerial implications
Research on business networks is continuously evolving and acknowledging this is a first step that can be advised in the process of making strategic considerations. Not only is there a vast amount of evidence to suggest that business networks are important, but there is also much reason to believe that positioning within this network is a crucial process to understand in order to capitalize on the benefits of these same networks (Anderson et al. 2000; Gulati et al., 2000). Increasing direct connections and acquiring a strategic position in terms of network reach can support firms in the search for knowledge that will increase their innovativeness. However, this must be a cautious approach as internal firm characteristics can determine whether actual results will be produced. According to scholars, one of the most prominent characteristics is dependent upon a firm’s willingness to reinvest earnings into their research and developments (Cohen & Levinthal, 1990; Tsai, 2001). This can help firms
46 educational level could hamper progress in the process of increasing both structural and relational embeddedness. Therefore, attention must be pointed to whether firms are actually in a position where it is useful to do this, as it can cause firms to be subject to a possible
information overload, risk aversion, and increasing cognitive conflicts within the
organization. Thus, one of the primary aims of a company should be to recognize that network embeddedness consists of a structural and relational side. Both types of embeddedness posit different effects and firms need to carefully assess the effects of the change to these types of embeddedness on their innovative output.
Limitations
Ultimately, this study suffers from a few limitations that have to be noted. This study's main limitation is that data on all actors within the generated networks of all three years are missing. Therefore, after mapping the network, only a few network scores on centrality are matched to company data. Therefore, the sample suffers from representativity problems and results in the study only being capable of making assumptions based on partly significant findings rather than actual rigorous implications. Besides, the results of this study mainly stem from model 6, which is a model that suffers from some violation of normality in the residuals. Another limitation arises from the operationalization of the relational embeddedness concept. Because the data was limited, alliance characteristics such as trust or strength could not be included in the analysis. These two aspects are argued to be able to explain the extent of a firm’s relational embeddedness (Uzzi & Lancaster, 2003). This study merely captures this variable on shared values and assumes that companies from the same nation are very similar while this conclusion can be inaccurate in some instances.
