Qinchang Gui, Chengliang Liu, Debin Du
Qinchang Gui is a Ph.D. candidate in the School of Urban and Regional Science, Institute for Global Innovation and Development, East China Normal University, Shanghai 200062, China. His e‐mail address is: 52173902007@stu.ecnu.edu.cn. Chengliang Liu, co‐corresponding author, is a Research Professor in the School of Urban and Regional Science, Institute for Global Innovation and Development, East China Normal University, Shanghai 200062, China. His e‐mail address is: clliu@re.ecnu.edu.cn. Debin Du, co‐corresponding author, is a Professor in the School of Urban and Regional Science, Institute for Global Innovation and Development, East China Normal University, Shanghai 200062, China. His e‐mail address is: dbdu@re.ecnu.edu.cn. The authors would like to acknowledge funding from the National Natural Science Foundation of China (No. 41571123; No.41471108), Science and Technology Commission of Shanghai Municipality (No.17692103600), and Shanghai Pujiang Program (No.17PJC030) and the valuable and constructive comments from the anonymous reviewers. The authors are responsible for the remaining errors. ... Less
Abstract
International knowledge flows have become increasingly common and more frequent and are a key driving force promoting the development of global science. However, little attention has been paid to the determinants of international scientific collaboration. Using co‐publication data from the Web of Science database in the period 2000–2014, this paper illustrates spatial patterns of international knowledge flows and estimates the impact of geographical, technological, social, and cultural proximity on the variation of inter‐country collaboration in science. Our findings demonstrate that the coefficients of the four dimensions of proximity are positive and significant in panel estimates and cross‐section estimates—international knowledge flows are facilitated by geographical, technological, social, and cultural proximity. We also find that the effect of geographical and cultural proximity have waned over time, while the impact of social and technological proximity have strengthened.
Introduction
Since the pioneering contribution of Jaffe, Trajtenberg, and Henderson (1993), knowledge flows and spillovers have attracted increasing attention in economic geography and regional science (Bathelt and Henn 2014; Cassi, Morrison, and Rabellotti 2015; Miguelez and Moreno 2013; Pond et al. 2007; Quatraro and Usai 2017; Rallet and Torre 1999). The core hypothesis is that innovation and new knowledge stem from combining and recombining knowledge elements (Arrow 1962; Schumpeter 1934), and economic development mainly depends on the production, diffusion and application of knowledge, which gradually replace land and labor as increasingly important resources (Laperche 2013; Lucas 1988; OECD 1996). Cross‐pollination and flow of different knowledge elements lead to new knowledge production, advances in science and technology and economic growth. In the light of the importance of knowledge and innovation, there is a growing literature investigating the inter‐regional flows of knowledge in Europe (Barber and Scherngell 2013; Cappelli and Montobbio 2016; Hoekman et al. 2009; Paci and Usai 2009; Quatraro and Usai 2017). Despite the role of country is weakened to a certain extent, in the process of globalization, it remains critical. A systematic investigation of inter‐country knowledge flows could substantially further our understanding of knowledge diffusion.
In the era of the globalizing knowledge economy, the interconnectedness and interdependence of countries have broadened and deepened (Chase‐Dunn and Grimes 1995; Niosi and Bellon 1994). The globalization and networking characteristics of scientific research activities are increasingly prominent, and countries are not only important bodies that take part in global economic competition and cooperation, but are also centers for the integration of human resources and incubators of invention (Guan et al. 2016). To our surprise, empirical evidence of international knowledge flows remains scarce. There are exceptions, such as the global wine industry (Cassi, Morrison, and Rabellotti 2015; Cassi, Morrison, and Ter Wal 2012) and global video game industry (Balland, De Vaan, and Boschma 2013). However, the study of inter‐country knowledge flows is worth pursuing.
In line with this stream of research, the aim of this paper is to assess the role of proximity on inter‐country scientific collaboration. That is to say, we try to investigate the role of multi‐dimensional proximities in international scientific collaboration networks. The paper will answer the following research questions. 1) How does proximity influence international knowledge flows at the country level? 2) Does the impact of different types of proximity evolve over time? More specifically, this study uses bibliographical data from the Web of Science database (WOS) to analyze the evolution over time of international knowledge flows in the period from 2000 to 2014. We employ a gravity model using a negative binomial regression to account for the relationship between knowledge flows and proximity at a global scale.
This paper contributes to the literature on knowledge flows in three broad ways. First, economic geography has emphasized the impact of different geographical scales on the unevenness of knowledge diffusion (Frenken, Hardeman, and Hoekman 2009). As suggested by Balland, Boschma, and Frenken (2015), we not only take into consideration short‐distance relations, but also consider long‐distance relations. Only in this way can we comprehensively understand the determinants of knowledge flows and the roles different types of proximity might have. Second, existing literature takes the proximity framework as a static concept (Balland, Boschma, and Frenken 2015; Broekel 2015). This paper adopts a dynamic approach to analyze the evolution over time of the impact of different forms of proximity on international knowledge flows. Third, our co‐publication data covers all disciplines to provide a wide‐ranging snapshot of global scientific collaboration. Previous studies concentrate on a single industry or specific research field, such as pharmaceutical research (Cantner and Rake 2014; Plotnikova and Rake 2014), biotechnology (Heimeriks and Boschma 2014; Ter Wal 2013), and the navigation satellite system industry (Balland 2012).
The remainder of the paper is organized as follows. The second section provides a short literature review from the perspective of proximity and knowledge flows. The third section introduces the data and methodology used in this paper. The estimation results are reported in the fourth section, and the final section presents the main results, policy implications, and future research.
Proximity and Knowledge Flows
When it comes to investigate the determinants of knowledge flows, proximity is an insightful theoretical framework (Bergé 2017; Kirat and Lung 1999; Torre and Rallet 2005). Boschma (2005) proposes five types of proximity: geographical, cognitive, social, organizational and institutional proximity, which are widely adopted in the literature on the geography of innovation (Balland 2012; Ter Wal 2013; Cassi et al. 2015). This paper explores scientific collaboration at the country‐dyad level. We will focus on geographical, cognitive, social, and cultural proximity.
Geographical proximity refers to the spatial distance between agents (Gilly and Torre 2000), which is widely discussed in economic geography and regional science (Ter Wal 2013). It is well known that knowledge diffusion decays with increasing distance (Audretsch and Feldman 1996). In general, knowledge can be divided into two categories: codified knowledge and tacit knowledge. While the former is easily diffused through books, journals, conferences, and mobility of scientists (Heimeriks and Vasileiadou 2008), the latter is believed to be bound to a certain location since the cost of transmitting tacit knowledge increases with spatial distance (Audretsch and Feldman 1996). In other words, knowledge faces distance‐decay effects (Caragliu and Nijkamp 2016). Communication and learning between actors require face‐to‐face interaction, which facilitates the diffusion of tacit knowledge and increases interpersonal trust, so geographical proximity benefits research collaboration (Boschma and Frenken 2010; Howells 2002; Pond et al. 2007). There is much more consensus that physical distance impedes cross‐region research collaboration and has a negative effect on knowledge spillovers, according to evidence from Europe (Balland 2012; Hoekman, Frenken, and Tijssen 2010), China (Scherngell and Hu 2011), and the globe (Cassi, Morrison, and Rabellotti 2015).
Technological proximity, also called cognitive proximity, refers to the degree of the same knowledge base between two agents (Nooteboom 2000), which can be measured by an uncentered correlation efficient (Cappelli and Montobbio 2016) or Pearson correlation efficient (Morescalchi et al. 2015; Scherngell and Barber 2011) or industry classification (Balland 2012). Based on the absorptive capacity theory, mutual learning need to have a common knowledge base to absorb, interpret, identify, and exploit new knowledge (Boschma 2005; Cohen and Levinthal 1990; Nooteboom 1999). Actors with a technological proximity can access and acquire external knowledge more effectively and efficiently. In addition, the related variety approach also stresses that the importance of technological proximity (Frenken, van Oort, and Verburg 2007). Scherngell and Barber (2011) find that cognitive distance impedes R&D collaboration in the EU, which matters more than other proximities. However, Balland (2012) shows that cognitive proximity does not play a significant role with the example of the European satellite navigation industry.
Social proximity is defined as the socially embedded relations between actors at the micro‐level, and it is regarded as a prerequisite for interaction (Boschma 2005). Relations between agents are established based on friendship, trust and past collaboration, which can help to exchange informal knowledge and thus increase the probability of bilateral collaborations (Boschma and Frenken 2010). From the perspective of social network analysis, knowledge flows are embedded in a social network of collaborations between actors, and social linkages among individuals are the catalyst for knowledge spillover (Granovetter 1985). In network clusters, actors are inclined to connect to partners of partners. This triangle relationship is called closure or triadic closure (Boschma and Frenken 2010; Ter Wal 2013). Dense ties between actors will curb opportunism, enhance trust, and share risk, because friends of friends tend to be friends (Uzzi and Spiro 2005). As Plotnikova and Rake (2014) point out, social proximity is positively and significantly associated with the number of international collaborations. Taking co‐patenting teams in the United Kingdom during the period from 1978 to 2010 for example, Crescenzi, Nathan, and Rodriguez‐Pose (2016) find that social proximity plays a positive role in the invention process.
Cultural proximity indicates the similarity of cultural environments between two actors at the macrolevel, which is associated with institutional proximity (Boschma 2005). Language is regarded as the core of culture, and a common language provides a basis for interacting and communicating to take place. Maskell and Malmberg (1999) argue that knowledge is transferred more effectively with cultural proximity and a common language. In the process of collaboration, speaking the same language makes knowledge spillovers easier, and linguistic or cultural differences may hinder bilateral communication. In European regions, Hoekman, Frenken, and Tijssen (2010) show that research collaborations are more likely to take place when scientists are located in the same linguistic areas. Language barriers significantly hamper industrial and public research collaboration (Scherngell and Barber 2011). On the country level, empirical evidence reveals that language proximity has an insignificant role (Cassi, Morrison, and Rabellotti 2015; Plotnikova and Rake 2014).
The concept of proximity is not static, and the growing number of articles requires a dynamic view to investigate the evolution of proximity over time (Balland, Boschma, and Frenken 2015; Broekel 2015). Balland, Boschma, and Frenken (2015) propose the dynamic of proximity framework, which is conceptualized as a quintuple process of learning, integration, decoupling, institutionalization, and agglomeration. When it comes to how the evolution of proximity changes over time, there is no consensus on this matter. With the development of transportation and Information Communication Technology and globalization, restriction of geographical space is gradually weakening, and some scholars propose the “death of distance” (Cairncross 1997; Castells 1996) and the “world is flat” (Friedman 2005). In the evolution of co‐inventor networks in German biotechnology, Ter Wal (2013) finds that the effect of geographic proximity decreases, and the impact of social proximity increases over time. Using data on patent citations and patent collaborations between 191 European NUST‐2 regions during the period 1981–2000, Cappelli and Montobbio (2016) show that the role of technological proximity, geographical distance, and institutional proximity decrease for patent collaborations, and increase for patent citations. On the contrary, Paci and Usai (2009) provide evidence that the coefficient of geographical distance slightly increases in 1998 compared to 1990, while the opposite effect occurs for national border. Making use of co‐publications between 313 European NUST‐2 regions, Hoekman, Frenken, and Tijssen (2010) show that the importance of geographical proximity has increased for the period 2000–2007, while the territorial borders effect has declined over time. Andersson et al. (2014) demonstrate that the physical distance effect first increases and then declines in the period from 1996 to 2010. Interestingly, Morescalchi et al. (2015) indicate that the effect of country borders and distance first decrease and then grow.
Data and Methods
Data
The empirical data used in this study is retrieved from Thomson Reuters's WoS Core Citation Database, which includes approximately 12,000 authoritative and influential journals and is one of the best source to study international collaboration in science (Cassi, Morrison, and Rabellotti 2015; Hoekman, Frenken, and Tijssen 2010; Plotnikova and Rake 2014). We select 60 countries by considering R&D data availability, because R&D data is regarded as the control variables. If the regression models are short of control variables, the estimation results may be biased (Ter Wal 2013). More specifically, WOS is extracted to retrieve data according to the search term “CU= A and CU= B and PY= year” to acquire the number of co‐publications between country A and country B from 2000 to 2014. In line with Ter Wal (2013), we apply a 5‐year moving window to construct a country‐by‐country collaboration matrix, where the diagonal cell value denotes the number of publications from country A in year t and the other cell denotes the number of co‐publications between country A and country B in year t. That is to say, each yearly network observation includes all co‐publications between year t‐4 and t. The 5‐year periods are consistent with Cappelli and Montobbio (2016), making comparisons possible.
Model and variables
In order to investigate international scientific collaboration, we employ a gravity model, which resembles Newton's law of universal gravitation. Specifically, the gravitational force between two agents depends on mass (economic size or scientific publications) and physical distance. The gravity model is widely used to explore international trade flows (Anderson and Van Wincoop 2003), collaborative knowledge production (Hoekman, Frenken, and van Oort 2009; Scherngell and Hu 2011), collaborative scientific research (Cassi, Morrison, and Rabellotti 2015; Pond et al. 2007), EU Framework programs (Barber and Scherngell 2013; Scherngell and Barber 2011), the patent collaborations (Montobbio and Sterzi 2013).
The first step of the analysis is the panel estimates conducted by pooling the data of three subperiods to explore the role of proximity. The second step is cross‐section estimates using different subperiod data to evaluate the evolution of proximity over time. This gives the following equation:
The dependent variable (Cij) in our model is the number of co‐publications between country i and country j, identified when there is a paper with at least one co‐author from country i and at least one co‐author from country j. It is worth noting that we follow a full counting procedure—if an article has more than one author from the same country the collaboration is measured as only one. Because the author collaboration is meant to measure undirected knowledge flows from one country to another, the final panel data includes 5,310 observations ((60*(60‐1)/2)*3) and 1,770 observations ((60*(60‐1)/2)) for a cross‐section data set. In order to mitigate the effects of annual fluctuations, the dependent variable is the aggregated value in the period 2000–2004, 2005–2009, and 2010–2014. Specifically, the 2004 observations include aggregated data from 2000 to 2004, whereas the 2014 observations include data from 2010 to 2014. The time lags are introduced to minimize the possible problems of endogeneity and reverse causality. The explanatory variables are calculated in the preceding 4 years, that is to say 2000, 2005, and 2010, respectively.
In the empirical model, the explanatory variables contain geographical, technological, social, and cultural proximity between countries, defined as follows.
Geographical proximity (Geoprox) is measured by the inverse of spatial distance between country i and j. Geographical distance is calculated following the great circle formula, which uses the latitude and longitude of the capital cites of the two countries taken from the CEPII data set (Cassi, Morrison, and Rabellotti 2015; Montobbio and Sterzi 2013).
Technological proximity (Techprox) is measured by Jaffe's (1986) index, the uncentered correlation of the two countries' vector of domestic publication across 22 different research areas (Cappelli and Montobbio 2016; Cassi, Morrison, and Rabellotti 2015):
where Pi = ( urn:x-wiley:00174815:media:grow12245:grow12245-math-0003 urn:x-wiley:00174815:media:grow12245:grow12245-math-0004 urn:x-wiley:00174815:media:grow12245:grow12245-math-0005 … urn:x-wiley:00174815:media:grow12245:grow12245-math-0006) indicates the distribution of vectors, and urn:x-wiley:00174815:media:grow12245:grow12245-math-0007 indicates the number of domestic publications published to country i in discipline classification N (N = 1,2…22). The technological proximity takes a value between 0 and 1. When countries i and j have the same research areas, this value equals 1; when their technological vectors are orthogonal, this value is equal to 0.
Social proximity (Socprox) is based on the concept of embeddedness (Boschma 2005; Granovetter 1985). Granovetter (1973) proposes that the degree of overlap of two actors' friendship circles indicates the strength of ties between them, which is consistent with triadic closure in nature and calculated by the Jaccard similarity coefficient (Leydesdorff 2008; Scherngell and Hu 2011). The collaboration matrix is converted into a binary matrix, where 1 represents that a collaborative relationship exists between country i and country j, and 0 otherwise. In this paper, social proximity is defined as.
where qij shows the number of common friends (partners) between country i and j, ri indicates the number of countries that country i collaborates with, and sj is similar to ri.
Cultural proximity (Cultprox) is measured by the common official language. This variable is a dummy variable, which take a value of 1 if two countries share a common official language, and 0 otherwise.
In our model, we also include three control variables. First, the number of papers published in the preceding 4 years (ln(Pubmassit) and ln(Pubmassjt)) is used as a proxy for absorptive capacity. Pubmassit and Pubmassjt respectively indicate the scientific size of country i and j at time t. Second, full‐time equivalent researchers is the proxy used for R&D input, measured by the number of R&D researchers (ln(rdresit) and ln(rdresjt)) in each country (researchers per million people). Third, World Bank income groups are introduced to measure the economic development level (EDLi and EDLj) of a country, that's, Low Income Group(1), Lower Middle Income Group(2), Upper Middle Income Group(3), and High Income Group(4).
The dependent variable, the number of co‐publications between country i and country j, is a count variable. The conditional variance of the dependent variable obviously exceeds its conditional mean, namely, existing over‐dispersion. In line with other studies (Andersson et al. 2014; Plotnikova and Rake 2014; Scherngell and Barber 2011), we employ a negative binomial model instead of Poisson model in our analysis. In addition, a Hausman test rejects the random effects specification. Therefore, we use fixed‐effect negative binomial models. Table 1 lists and describes the dependent and explanatory variables in the econometric model.
Results
Descriptive analysis
Figure 1 sheds some light on the spatial distribution of international scientific collaboration networks. The nodes denote countries, and their sizes are directly proportional to the number of publications that belong to the countries; the links represent the collaborative relationships between two countries i and j, and the thickness of the links is scaled to show absolute strength, i.e., the number of co‐publications. In order to improve the readability of the picture, we prune it—only ties with no fewer than 500 co‐publications are retained, while other are removed.
At first glance, Figure 1A reveals that the international scientific collaboration network is dominated by the science superpowers of Japan, the U.S., and Europe, which form a bipolar world or Atlantic axis. The number of publications of the U.S. is 2,001,000, followed by the U.K. (575,916) and Japan (434,820). Furthermore, over half of the top 10 countries with the most publications are in Europe. From the perspective of co‐publications, of the top 10 inter‐country collaborations, eight involve the U.S. Among them, collaboration between the U.S. and Germany is the largest, reaching 51,508, followed by the U.S. and U.K. (51,101), U.S. and Canada (46,212), U.S. and Japan (36,965), U.S. and France (31,210), U.S. and Italy (25,345), U.S. and China (20,352), and U.S. and Australia (17,813). In addition, the intensity of collaborations between European countries is very high, Germany and U.K. (22,268), Germany and France (17,945), rank seventh and ninth, respectively.
In Figure 1B, there is a surge in terms of the amount of international cooperation. There are 83 links with bilateral co‐publications equal to or greater than 10,000, and seven have more than 50,000. At first look, the backbone of the global scientific collaboration network initially forms a trapezoid‐shaped graph with four vertexes in Western Europe, North America, East Asia and Australia, and the two diagonals are formed by ties from East Asia to North America and Australia to Europe. To our surprise, the bilateral relations between U.S. and China (119351) outnumber all other international pairings, U.S.‐U.K. (115,987) ranks second, followed by U.S.‐Germany (94,352). The rising powers in science, such as China, South Africa, India, and Brazil, develop rapidly, and emerging scientific nations in the Middle East, South east Asia, and North Africa are full of momentum, which points toward an increasingly multipolar scientific world and a reshaping of the global scientific landscape.
Panel estimates of the proximity effect
Table 2 presents descriptive statistics and a correlation matrix. In our paper, the mean of international scientific co‐publications is 1,479.55 and the standard deviation is 5,313.382. The latter is more than twice the former, which means that a negative binomial model is appropriate. The average variance inflation factor is 1.76, and every variance inflation factor value is lower than 3, which suggests that multi‐collinearity is not serious in this study. Table 3 reports the sample estimation results of the fixed‐effect gravity model, with bootstrap‐robust standard errors given in parentheses. Model 1 is the basic model, which only includes the control variables. Model 2 through Model 5 add explanatory variables to Model 1. Model 5 presents the full model, which includes all control variables and explanatory variables.
As expected, the estimates for the mass indicators have a positive sign and are significant, which not only indicates that a higher number of publications in a specific country increases the likelihood of cooperation with other countries, but also corroborates that the global scientific collaboration networks exhibit a “Matthew effect” or “the rich get richer” phenomenon. We conclude that the scientific sizes of countries are positive determinants of international co‐publication, which is in line with other studies (Cassi, Morrison, and Rabellotti 2015; Plotnikova and Rake 2014). In addition, ln(rdres) and EDL are also positive and statistically significant at p < 0.01, which indicates that the number of R&D researchers and the EDL of a country will have an important role in facilitating collaborations.
The geographical proximity coefficient is positive and significant, as expected, which shows that physical proximity facilitates international knowledge flows. This is in line with comparable prior studies that employ co‐publication or co‐inventor as a proxy for scientific collaboration and find evidence that physical distance matters in the establishment of collaboration (Andersson et al. 2014; Hoekman, Frenken, and Tijssen 2010; Scherngell and Hu 2011). On the basis of mergers and acquisitions (M&A) deal data in Italy, Boschma, Marrocu, and Paci (2016) and Usai, Marrocu, and Paci (2017) find evidence that geographical proximity has a positive effect on the probability of inter‐firm knowledge exchanges. As Quatraro and Usai (2017) put it, collaborative relationships concern mainly the flows of tacit knowledge, which are sensitive to physical contiguity.
The coefficients of the technological proximity measures, ranging from 0.1153 to 0.1258, are positive and statistically significant in all model specifications. This implies that international collaboration is more likely to occur between countries with similar technological areas. This result is consistent with various empirical studies for the fifth EU Framework program (Scherngell and Barber 2011), collaborative knowledge production in China (Scherngell and Hu 2011) and R&D collaborations in OECD countries (Morescalchi et al. 2015). As Boschma (2005) points out, scientific collaboration requires a shared cognitive base to identify, absorb, understand, and exploit new knowledge in order to collaborate successfully.
In terms of social proximity, the regression coefficients of model 4 and model 5 in Table 2, are positive and statistically significant at p < 0.01, suggesting that countries are more likely to connect to partners of partners. This result is in line with existing studies showing that social proximity measured through the inverse of geodesic distance or transitivity has a positive role (Lazzeretti and Capone 2016; Ter Wal 2013; Usai, Marrocu, and Paci 2017). Multinational partnerships reduce the network distance between any two countries, that is, indirect ties turn into direct ties, which can enhance the efficiency of knowledge transmission.
Our result shows that the indicator of cultural proximity is significant and positive at p < 0.05, which means that having a common language will promote international collaboration. This result is consistent with the internationalization of inventive activity (Picci 2010) and the globalization of technology (Montobbio and Sterzi 2013). However, Cassi, Morrison, and Rabellotti (2015) and Plotnikova and Rake (2014) point out that a common language does not promote collaboration intensity at the country level. Researchers tend to use their native language to discuss questions. Therefore, international research collaborations are hampered by linguistic difference. It is worth noting that our results rely on the WOS data set, which has geographical and language biases.
Cross‐section estimates for different sub‐periods
The second step is the analysis of the evolution of the different proximities over time. Table 4 presents the cross‐section estimation results of the gravity model, with robust standard errors given in parentheses. In order to account for two separate equations in different years, we use the Chow test (Chow 1960), which is commonly used to determine whether the explanatory variables have significant differences in time series analysis. The Chow test rejects no difference specification (p = 0.000). In addition, using a seemingly unrelated regression approach (SUR) (Greene 2012; Izon et al. 2016), we test the significance of separately explanatory variables, where the coefficients are statistically significant at a 90 percent confidence level.
The two models show that the coefficient of geographical proximity changes from 912.3651 in 2000–2004 to 556.3179 in 2010–2014, which means that the importance of geographical proximity would decrease over time. This result corroborates finding in German biotechnology networks (Ter Wal 2013), patent collaborations between European regions (Cappelli and Montobbio 2016) and inter‐organizational networks of Tuscany (Lazzeretti and Capone 2016). In the context of information and globalization, information communication technology and low‐cost travel facilitate knowledge exchange (Cassi, Morrison, and Rabellotti 2015). On the contrary, some previous studies observe that the effect of geographical distance has slightly increased over time (Cassi, Morrison, and Rabellotti 2015; Hoekman, Frenken, and Tijssen 2010) and there is a “missing globalization puzzle” (Montobbio and Sterzi 2013). In short, we conclude that the role of geography declines over time, but does not die.
In addition, technological proximity shows a gradual increasing trend in coefficients (from 0.1191 to 0.5644). The result of our study contrasts with the existing literature (Cappelli and Montobbio 2016; Lazzeretti and Capone 2016), which find that its importance decreases during the study period. The globalization of science can result in the international division of scientific and technological activities; that is, the increased globalization of knowledge flows would be associated with the process of increased specialization (Cappelli and Montobbio 2016; Durkheim 1997). Niosi and Bellon (1994) find that globalization and specialization are compatible with the openness of countries. In the context of the specialization, international knowledge flows increasingly rely on technological proximity between countries (Montobbio, Primi, and Sterzi 2015).
When it comes to social proximity, the increasing trend is supported by the empirical results (from 1.0828 to 3.7705) in Table 4. The result is in line with Ter Wal (2013), who shows that social proximity is key driver of the evolution of collaboration networks. The globalization of science not only results in specialization, but also brings about networking. All countries are embedded in a global innovation network and social ties can be regarded as a channel through which information and knowledge flow between countries. The process of innovation is full of risk, uncertainty and complexity, and social proximity is likely to curb opportunistic behavior (Boschma 2005; Granovetter 1985).
The last result demonstrates that the role of cultural proximity decreases from 0.9151 in 2000–2004 to 0.6880 in 2010–2014, which indicates that the effect of sharing a common language is finite. Our result is line with Cappelli and Montobbio (2016), who find that language is decreasingly important for tacit knowledge flows and codified knowledge flows. English, the most widely used language and the first choice of second languages, facilitate interactions, and collaborate with different countries. For example, WOS includes 2,072,786 papers in 2014, 96.22 percent of which are authored in English.
Conclusion
International knowledge flows are becoming increasingly common and more frequent. The empirical analysis of knowledge flows is a hot topic in economic geography and regional science. This paper analyzes the dynamics of the international knowledge flows using co‐publication data and investigates the evolution of multi‐dimensional proximity using negative binomial gravity models in the period 2000–2014. More specifically, taking international scientific collaboration networks as an example, we estimate the effects exerted by four different proximity measures (geographical, technological, social, and cultural) on inter‐country knowledge flows through panel estimates and cross‐section estimates.
This paper provides a number of interesting and novel results in the context of the relevant literature. First, we find that all proximity dimensions have a positive and statistically significant effect at a global scale; countries tend to have collaborative ties when they share more proximity. It means that international knowledge flows are facilitated not only by geographical proximity, but also by non‐geographical proximity, for example common language, similar knowledge background, and close triadic closure. Second, the results show that the effects of geographical and cultural proximity have decreased over time, while the opposite effect is true for social and technological proximity. It may be explained by a quadruple process of informatization, specialization, networking, and globalization. The network relationships can act as an information channel and facilitate knowledge exchange between countries. Globalization and informatization result in specialization and networking. International scientific collaboration depends more on technological and social proximity than geographical and cultural proximity.
Several important policymaking implications have been conducted as follows. First, we find that the effect of social and technological proximity has a clearly increasing trend and is more and more important. In order to tap into foreign resources, countries actively take part in international networks, and cooperate with partners of partners. The increased specialization and complexity of scientific activities dictate that countries should collaborate with actors with a common knowledge base. In addition, geographical proximity still plays an important role in knowledge flows, which means that globalization of science is still constrained by geography and spatial barriers. Thus, international organizations and national governments should take measures to minimize barriers to the flow of knowledge, scientists, and researchers, such as border governance. The aim of the European Research Area is to encourage cooperation in science among member countries.
When it comes to future research, several directions are worthy of further study. First, proximity is interdependent, and geographical and nongeographical proximities may be either substitutive or complementary. Second, the study applies the co‐publication indicator to capture knowledge flows is partial and incomplete. We can make use of patent data to build co‐inventor collaboration network. A comparative study of the two networks may provide interesting results. Third, knowledge flows are doubly embedded in collaboration networks and knowledge networks; the former is tacit knowledge flows, and the latter is codified knowledge flows.
https://onlinelibrary.wiley.com/doi/10.1111/grow.12245
