International research collaboration

The topic of research collaboration as a subject of inquiry in science policy is wide-ranging and well-developed. The term “invisible college” captures the early roots of scientific collaboration, referring to a group of seventeenth-century scientists who organized activities informally through the collaborative exchange of letters, which later formed the Royal Society of London. Derek de Solla Price [20] documented a steady acceleration of coauthored papers in the early 20^th century, arguing that science had moved from “little science” in the seventeenth century to a contemporary form of “big science” involving national expenditures, economic impacts, computers, rockets, and the Manhattan Project. One of the most astonishing examples of contemporary international research collaboration includes the Atlas Experiment in particle physics, which produced a series of publications with thousands of authors from dozens of countries [21]. In the contemporary setting, the “new invisible college” refers to a sprawling global network of authors connected by papers rather than institutions [8,22,23]. The growth and density of the international collaboration network continue to increase [24]. However, relatively few scholars have questioned whether and how the overarching governance structures of nations influence self-organization among scientists in the international system.

Numerous factors have been implicated in the dramatic evolution of research collaboration as a mode of scientific inquiry over the past century. The costs of cooperation have declined substantially, particularly given the rise of the internet, email, and video conferencing [25,26]. The steadily increasing specialization of scientific disciplines [27], as well as the evolution of multidisciplinary research projects [28], have also contributed. The complexity, scope, and magnitude of scientific programs have increased, requiring participation by numerous researchers to execute single experiments, particularly those involving complex and expensive equipment [29]. Further, the institutional norms of higher education institutions and governmental organizations, have adjusted to accommodate and promote collaborative teams [30]. Meanwhile, scholars have adapted to the new norm, realizing gains in “scientific human capital” and productivity through collaboration domestically and across national borders [31,32]. The journey toward promotion and tenure through the traditional route of “publish or perish” may be less perilous with collaborative partners shouldering a part of the burden of knowledge creation. Lastly, collaboration facilitates friendship, career advancement, learning [33], and the epistemic goals of researchers [27].

Abramo et al. [34] showed that the association between research collaboration and the scientific productivity of universities varies substantially across disciplines, between basic and applied science, and in the context of domestic versus international collaboration. Recently, Fox and Nikivincze [35] showed that collaboration is an important predictor of prolific research productivity. International research collaboration, as distinct from domestic collaboration, may have even greater returns. As Glanzel and de Lange [36] showed, internationally coauthored papers tend to have a higher citation impact than domestically published papers. Reversing the question, Abramo et al. [34] test whether the productivity of the scientist predicts a higher level of international research collaboration. Their findings indicate that more productive scientists tend to be internationally collaborative. Thus, there appears to be an endogeneity/chicken-egg problem at work, where collaboration begets performance begets collaboration. However, the literature has repeatedly shown that international research collaboration results in higher productivity and performance [37]. Wagner et al. [23] questioned whether international collaboration tended to produce conventional, novel, or atypical research, finding, contrary to expectation, that conventional research tended to be the most typical outcome. Using a different measure of internationalization in research, Sugimoto et al. [38] find that mobile scientists with many country affiliations in their publication record tend to build up international collaboration networks as they travel within the international system and also tend to have the highest citation impact [39]. Similarly, Robinson-Garcia et al. [40] found that mobile scientists tend to have higher publication output and citation impact.

Fewer studies have examined the effects of national-level characteristics. Glanzel and Schubert [41] suggest that larger countries, such as India, Japan, the United States, and Turkey tend to produce a higher proportion of domestic publications than smaller countries. Conversely, smaller or more remote countries, such as Switzerland, Austria, and South Korea, tend to have higher international publication rates. Leydesdorff et al. [42] recently examined whether the influence of government funding or international research collaboration affects a nation’s citation impact finding that the balance favors international research collaboration.

Finally, Subramanyam [43] suggested that coauthorship as an indicator of collaboration has its advantages as a measure, being non-reactive, invariant, quantifiable, and inexpensive. However, Katz and Martin [29] cautioned against reading too much into the term ‘collaboration.’ As Davidson and Carpenter [44] suggest, there are also ambiguities that should be appreciated in international collaboration, such as single authorship with multiple affiliations, but also suggested such errors are small relative to the size of bibliometric databases. Thus, international co-authorship may not always reflect true collaboration. Rather, the approach of this study is that the term “international research collaboration” refers minimally to instances in which two or more affiliations referring to distinct countries co-occur on a single research paper.

Governance effects on innovation in science and technology

The rise of China in the international system of science increases the salience of questions regarding democratic governance and research capacity. These are traditionally assumed to be complementary if not critically compatible. But few large-scale empirical studies have examined the general effect of governance on scientific processes and outcomes or on how the relationship might influence international research collaboration. Scholars suggest that social structure influences organizations and individuals, who reproduce and sometimes transform the structures within which they are embedded [9,45]. However, while a large body of research has focused on the individual, the discipline, and the institution, the effects of political governance structures on the self-organization of scientific praxis are not empirically well established.

The conjecture that democracy is the optimal political environment for the advancement of science has existed since both popular sovereignty and natural philosophy have co-occurred under the same jurisdiction of governance. In the classic two-volume work, The Open Society and its Enemies, the philosopher of science Karl Popper [46] defends liberal democracy against both fascism and communism by disarticulating the ideological commitments to historicism found in the latter two systems. The notion that history follows a pre-determined trajectory toward some final, perfected, and inevitable political end-state is antithetical to both the notion of political liberalism and the necessity for a free and open society. Parsons [47] suggested that a society’s social structures shape its scientific development, while Barber [48] suggested that the values of democracy and science are compatible with each other and incompatible with autocratic rule. Merton [49,50] suggested that a scientific enterprise concerned with the pursuit of impersonal truths should be unfettered by ideological dogma, is consonant with democratic governance, and clashes with totalitarianism. Work in this conceptual vein suggests a fundamental compatibility between the normative elements of democracy and science.

As Taylor [51] suggested, “[i]nnovation is inherently political in that it involves highly contested decisions over the allocation of resources, institution and policy design, and the formation and maintenance of domestic and international political economic networks” (p.150). Wang et al. [52] conducted a test of Popper’s hypothesis regarding the positive effects of democracy on innovation using patents and trademarks. They show a significant positive effect of democracy on innovation across a variety of empirical tests, with the greatest returns occurring for countries with lower levels of economic development. These results are contradicted by Gao et al. [53], who test the hypothesis on a subset of patents. Thus, the results of this literature are somewhat mixed. While these observations refer primarily to technology, it is reasonable to suggest that political governance also has significant effects on scientific innovation. In one empirical article specifically on the topic, Whetsell et al. [11] recently provided evidence in support of the democracy-science hypothesis, showing significant positive effects of democratic governance on scientific performance. They suggested that democratic governance might enhance the internal and external complexity of the country, such that science is provided wider latitude for self-organization on intellectual fitness landscapes than in autocratically governed countries [54,55]. International collaboration as a type of external complexity may produce a greater variety of conventional, novel, and atypical ideas, c.f. [56,57]. Whetsell et al. [11] also showed a negative interaction effect between democracy and international collaboration intensity, where the effect of democracy on scientific performance diminished for higher levels of international collaboration intensity. This counter-intuitive finding incites further research into the effects of democracy on international research collaboration.

Democracy and academic freedom

To more clearly articulate a potential causal pathway from political governance to collaborative self-organization, a newly developing empirical literature on democracy and academic freedom is briefly reviewed. Coppedge et al. [1,2] suggest liberalism is based on a negative concept of liberty that is realized when the political power of the executive is constrained by effective checks and balances, the rule of law, constitutionally protected civil liberties, and an independent judiciary. At the same, liberalism must be combined with a structural view of electoral democracy, under which suffrage must be extensive, there must be freedom of association and expression, and there are clean elections for public officials. This structural element is based on Dahl’s [58,59] concept of polyarchy. However, liberal democracy is not the system of government for the majority of the world. Roughly 70% of the global population lives under closed or electoral autocracies [2,4]. As Lührmann and Lindberg [60] suggested, the world is experiencing a “third wave” of autocratization. Further, as several observers have noted, the latest wave of autocratization is less blunt and more sophisticated. In other words, autocratization is now less likely to occur as a result of an outright military coup and more likely to involve legalistic machinations, misinformation/disinformation campaigns, toxic political polarization, and sham elections [4]. As Hellmeier et al. [61] point out, autocratizing countries appear to follow a predictable pattern which typically begins by repressing academic freedom, media, and civil society. For instance, Enyedi [62] details the methods by which academic freedoms have been eroded in recent years in Hungary coinciding with its descent into electoral autocracy. According to theory, the autocratization of nations should hinder the self-organization of scientists within and across national borders.

In summary, theory suggests that liberal democracy provides the optimal conditions for the self-organization of science as a whole through the development of academic freedom, where members of the academic community retain individual and collective rights to develop and communicate knowledge, are autonomous from the state and other political forces, and are free to cooperate with others across boundaries and borders [63]. Berggren and Bjørnskov [64] show that democratic political institutions are a significant predictor of academic freedom. Thus, there is a clear connection between democratic governance and the free conduct of science. In liberal democracies, where scientists are permitted and encouraged to freely associate with others across ideological, political, and national borders, we should expect greater levels of international research collaboration. In more autocratic systems, where academics are censored, monitored, and their movements are restricted, we should expect lower levels of international research collaboration. As such, the following hypothesis is warranted.

1. Hypothesis 1: Democratic governance is positively associated with international research collaboration a) tie formation and b) tie strength in the global research collaboration network.

Assortative mixing effects of governance structure

Assortative mixing refers to a process within networks in which actors tend to form connections with actors of a similar (homophily) or dissimilar (heterophily) characteristic [65]. When actors in a network form social ties based on similarities, such as race and ethnicity, religion, group membership, or any other construct real or imagined, this is often referred to as homophily in the social network literature. A common metaphor used to describe this process is “birds of a feather flock together” [66]. In contrast to homophily, when actors in a social network form connections based on dissimilarity on a given attribute, this is referred to as heterophily, and includes examples such as heterosexual mate selection, mentees seeking out older “wiser” mentors, or constructing skill complementarity in team structures. In the context of the present study, the question may be formulated as follows. Are countries with similar governance structures more likely to collaborate with each other?

The question of democratic collaboration is perhaps best situated within a stream of international relations theory loosely referred to as liberal institutionalism. In contrast to realist and neorealist theories of international relations, which deprioritize the internal characterization of a nation’s governance structures and instead prioritize its military capabilities in balance with those of respective competitors [67,68], liberal institutionalism stresses the importance of reducing the transaction costs of cooperation through the construction of international regimes [69]. As the logic goes, conflict can be reduced by erecting international institutions, and democracies tend to engage in institution building. Therefore, democracies tend to cooperate more with each other and have less conflict. This vein of theory is more broadly known as the “democratic peace”, the “liberal peace”, or the “Kantian peace”. Within the framework of liberal internationalism, Wagner’s [22] “new invisible college”, as a global network of science, might be considered a type of international institution that promotes cooperation and generally reduces the prospect of conflict between nations. These broader effects, however, are beyond the scope of this article. As Wagner et al. [24] suggested, the global network of science should be conceptualized as a new type of institution. It is reasonable then to question whether democracies may be more engaged in homophilous collaboration in this new institution.

The broad question posed in this article, how do the domestic characteristics of a country affect its tendencies to engage in international research collaboration, is quite like a question that political scientists have been interested in for decades. Lai and Reiter [70] argued, “the central question of international relations in the past decade has been, What are the connections between domestic and international politics”(p. 204). Leeds [71] argued that democratic leaders, by virtue of their accountability, are more likely to make and uphold credible foreign policy commitments. According to this theory, democracies might be viewed as more reliable partners for international cooperation and may also be less likely to form weaker agreements with other non-democratic countries. As Gallop [15] suggests, “it seems both plausible and empirically justified that states should gain somewhat more utility when they cooperate with like regimes” (p.315). The logic of international cooperation based on regime similarity appears to be isomorphic with the research question of international research collaboration based on governance structure. However, an important difference here is that the interactions between countries discussed above tend to be formal agreements between governments, while international collaboration is the outcome of the self-organization of scientific activity.

As Mansfield et al. [72,73] show, countries with democratic governance are more likely to cooperate on trade policy. In contrast, Lai and Reiter [70] provided evidence against the hypothesis that democracies tended to engage in more formal alliances with each other. However, they did show that democracies were more likely to form defense pacts after 1945 and that, more broadly, countries with similar regime types are more likely to cooperate with one another in the international system. Using inferential network analysis techniques on a data set of international agreements between WWII and 1980, Kinne [14] shows that international cooperation is significantly driven by endogenous network effects, such as preferential attachment and transitivity. He also shows significant positive effects of democracy, which appear to overwhelm the effects of military capabilities on military agreement formation. However, interestingly he shows a negative direct effect on formal science agreements during this time period and a non-significant and positive homophily effect. In a subsequent study, Kinne [74] shows positive significant direct effects and homophily effects of democracy on diplomatic tie formation. In a similar vein, Warren [75] also uses inferential network analysis to show that democracies tend to form alliance ties with one another.

1. Hypothesis 2: Countries with similar political governance are likely to a) form and b) have stronger international research collaboration ties in the global research collaboration network.

Modeling approach

This article employs inferential social network analysis techniques in the family of models referred to as exponential random graph models (ERGM) [76–78]. Exponential random graph models are appropriate to account for dyadic dependencies found in relational international collaboration data. While there are numerous studies on research collaboration, fewer have employed ERGMs. However, a number of ERGM studies of collaboration have recently emerged [79–84]. In a nutshell, the ERGM provides estimates on parameters of interest by simulating a probability distribution of thousands of networks and then comparing elements of the observed network to the simulated distribution. In addition to including exogenous parameters of interest, ERG models also include endogenous terms that account for complex dependencies in the data, such as density, preferential attachment, and transitivity, which are known to affect tie formation in networks. Two specific types of ERG models are used to address the hypotheses: bootstrapped temporal exponential random graph models (TERGM) are used to assess tie formation, while valued exponential random graph models (VERGM) and used to assess tie strength.

First, to test hypotheses 1a and 2a, bootstrapped TERGMs are used to account for temporal dependence in longitudinal network data [85,86]. The TERGM uses information from prior network time points to account for subsequent network configurations, incorporating special endogenous temporal terms that account for linear time-trends and dyadic stability over time [86]. Significance tests for the parameters of interest are generated through a confidence interval approach using bootstrapped pseudolikelihood [87,88]. The data in the current study have discrete cross-sections for the international research collaboration network between 2008 and 2017. The TERGM approach may be preferable to the stochastic actor-oriented (SAOM) model in this specific case because co-authorship ties are non-durational instances measured temporally at the year in which the paper was published. Thus, there is no actor-oriented decision to dissolve the tie in the following year, which the SAOM model assumes [89].

Second, valued ERG models (VERGMs) are used to test H1b and H2b. There currently exist limited options available to analyze edge(tie) weights in valued networks. Because the ties in the international research collaboration network are collapsed on countries as nodes, ties are stacked according to the number of instances where countries have authors collaborating with each other each year. These frequencies are referred to as edge(tie) weights or valued ties in a weighted/valued network. Edge weights are important to understand in international research collaboration because much of the variability in the data is on the weights between countries rather than the presence or absence of ties. Valued networks remain less well understood and developed and also pose greater computational challenges than binary ERGMs [90]. Often, such networks are simply binarized and analyzed using standard ERG models [87]. One method to binarize valued networks based on tie weights is a disparity approach to estimating the statistical significance of ties using the R package ‘backbone’ [91,92]. More recently, VERGMs have been advanced and implemented in the R package ‘ergm.count’ to analyze valued networks [93,94], which is part of the broader R package ‘statnet’[95]. Rather than focusing on the probability of a tie forming between two nodes, VERGMs produce estimates based on the strength of ties. Thus, VERGMs are suitable for addressing the hypotheses regarding the strength of research collaboration ties in the international system. As Huang and Butts [90] suggested, there is an “acute” need for analysis of the strength of social connections between nations. To test for effects over time, VERGMs are replicated for each yearly network cross-section from 2008 to 2017. More details on the implementation of the models are described below in the results section.

Data and variables

This research combines data from numerous sources. First, international research collaboration metadata were gathered from Indiana University’s CADRE portal containing the Web of Science database on all articles and reviews for the years 2008 to 2017 [96]. The metadata includes author-country affiliations for individual papers. These data are aggregated at the country level, where papers represent network ties between two or more affiliated countries. Under this construction, the network is a symmetrical adjacency matrix, M, where Mij equals the number of times a paper published each year includes an author from country i and an author from country j. Aggregating at the country level produces a valued network where all ties between the countries are aggregated as frequency weights on the network ties. There are at least three reasons for aggregating at the country level, as opposed to using author-author network data. First, ERGMs tend to be computationally infeasible with extremely large networks, i.e., the global co-author network for all research publications comprises hundreds of thousands, if not millions of nodes. Although, work on this problem appears to be advancing [97]. Second, disambiguation of names remains a persistent problem in studies of co-author networks, where names are entered differently across publications or multiple individuals have identical names. Treating countries as nodes avoids the name disambiguation problem because a network tie is created between two countries when a paper includes more than one country affiliation, independent of how the author lists their name or whether their name is ambiguous, c.f. [98]. Third, the primary independent variable (liberal democracy) is measured at the country level and as applied to individuals would be invariant within each country-year unless individuals change affiliations. While such an analysis would be an important contribution, it is not within the scope of the present article.

Since there are already a large proportion of binary research collaboration ties between many countries in the international system, much of the variability in the data resides in the edge weights. The data were used to produce symmetrical, undirected, valued matrices for each year between 2008 and 2017. The exact number and order of nodes across years were maintained for comparability between years using the final year (2017) as the node-set, resulting in a sample of 200 countries. However, after eliminating country columns/rows for those with missing data in the primary independent variable, the final node sample included 170 countries.

The primary independent variable of interest is the Varieties of Democracy Institute’s (V-Dem) liberal democracy index [1,2]. The V-Dem data represents the most comprehensive and current characterization of democracy across nations in the international system and is more granular than other extant measures such as the Economist Democracy Index or the Polity Project. The liberal democracy index combines an index of liberalism with an electoral democracy index. The liberal component is composed of three sub-components: 1) equality before the law and individual liberties, 2) judicial constraints on the executive, and 3) legislative constraints on the executive. The electoral democracy index is based on Robert Dahl’s [58,59] characterization of polyarchy and is composed of five democracy components measuring 1) freedom of association, 2) free and fair elections, 3) freedom of expression and alternative sources of information, 4) elected official, 5) and suffrage [2]. The index is standardized on a 0–1 scale, where higher on the scale represents a greater realization of the “ideal” of liberal democracy. Fig 1 shows two plots: Fig 1A on the left shows the time frame average liberal democracy score of each country listed in order from highest to lowest score, Fig 1B shows the difference in liberal democracy score between 2017 and 2008 listed in order of largest increase in score to largest decrease in score. In Fig 1B countries are colored by their average liberal democracy score, while in Fig 1B countries are colored by their 2017 score. Notably no country scored higher than 0.9 on the index. Fig 1B shows many less developed countries such as TUN, GMB, MMR, LKA, GEO, BFA, FJI, KGZ, VUT, NPL, and LBY had the greatest increases in democracy, while more developed countries like TUR, HUN, POL, UKR had the greatest declines. Meanwhile, developed countries like FRA, GBR, and CHN have virtually no change, while USA declined slightly. However, end point differences of course mask non-linear changes that occurred between 2008 and 2017.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Liberal Democracy among 170 Countries 2008–2017.

A. The time frame average liberal democracy score of each country listed in order from highest to lowest score. B. The difference in liberal democracy score between 2017 and 2008 listed in order of largest increase in score to largest decrease in score.

https://doi.org/10.1371/journal.pone.0287058.g001

This research employed a select number of control variables to account for systematic differences between countries that might also drive international research collaboration. These include three measures from World Bank Indicators, including the natural logarithm of gross domestic product per capita, the natural logarithm of population size, and the level of urbanization. The inclusion of these variables as controls is relatively straightforward. The wealth and population size of the nation likely enhances its ability to participate in a variety of collaborative research activities, while urbanization may reduce the need to collaborate externally due to a concentration of domestic collaboration partners [99]. The natural logarithm of the number of authors aggregated at the country level, was used as a measure of national research capacity to control for higher tie probability due to larger numbers of scholars available for collaboration. This variable may provide a more granular control than other commonly used measures, such as the number of Ph.D. granting institutions or secondary school enrollment. These data were captured separately from Elsevier’s Scopus data. Other control variables were considered, such as R&D/GDP and gross domestic expenditure on research and development (GERD). However, the data coverage for these variables is far smaller and therefore were not included due to the broad nature of the hypotheses. Finally, it is important to control for geography since policy and proximity effects are known to affect collaboration [80,100,101]. Geography was controlled for in two ways: 1) region was included as a fixed factor to account for proximity and political effects of geography by using the V-Dem data’s geopolitical region variable, which includes ten regions; and 2) a country-country distance matrix was used as an edge covariate. The distance matrix was generated from the R ‘cshapes’ package. Distances were calculated by closest border-to-border distance using the Correlates of War (COW) coding of countries. Lastly, homophily/heterophily terms were also included for all node attribute controls.

Network models in the exponential family offer a wide variety of endogenous terms which model the various sources of dependency in relational data. Scholarship using ERGMs has produced a relatively standard set of model terms that tend to be used to account for social processes, such as density (edges), preferential attachment (gwdegree–geometrically weighted degree distribution), and transitivity (gwesp–geometrically weighted edgewise shared partner distribution). For VERGMs, a different logic is applied. VERGM uses the term “sum”, which is similar to the term “edges”. The sum term represents the intercept and is used to account for the endogenous effect of tie strength. The term “nodesqrtcovar” is used to account for actor heterogeneity in tie strength (a proxy for preferential attachment). Following, Pilny and Atouba [102], nodesqrtcovar is interpreted similarly to gwdegree. Lastly, the model term “nonzero”, that estimates the amount of zero inflation, is included in the VERGMs.

Some variables had missing data. ERG models require no missing data on node covariates. Thus, the node file and the network matrixes must contain complete data on the same set of nodes. First, the V-Dem data was used as the basis for the data merge within the node file, which contained data on liberal democracy for 179 countries. Second, after merging all other data with the V-Dem data, the countries SSD, SML, XKX, and ZZB were removed for completely missing data on the controls. Third, the V-dem data also contained two country codes for Palestine: PSE appears to be the most used, so PSG was dropped. This left six countries with missing data: ERI, PRK, SOM, SYR, TWN, and VEN. Therefore, the following imputation strategy was taken. Fourth, TWN, VEN, and PRK did not contain World Bank GDP data, which were available in the V-Dem data and were imputed manually. TWN did not contain data on population size or urbanization from the World Bank which were instead gathered from https://www.statista.com/ and imputed manually; urbanization data was partially missing across 5-year intervals and was carried forward across missing years. Fifth, ERI, SOM, and VEN had partially missing data on GDP, population size, and urbanization, so their values were imputed from the average of the existing country data. Sixth, at this point, the node file was matched to the matrixes, which dropped COG, GNB, GNQ, and HKG from the V-dem Data, and 30 countries were dropped from the matrixes (AND, ANT, ATG, BHS, BLZ, BMU, BRN, FSM, GIB, GLP, GRD, GRL, GUF, KNA, KOS, LCA, LIE, MCO, NCL, PLW, PYF, REU, SMR, SSD, TOA, VAT, VCT, WAS, WSM, ZRN). This left 170 countries matching in both the node file and the matrixes. Finally, the country-country distances for Israel were inputted in the distance matrix for Palestine because the geographical distance data did not include Palestine as a country.

Limitations

There are a few limitations to the current study. While the V-Dem data does have a very broad coverage of countries in the international system, there were at least thirty countries contained in the bibliometric data which were not covered. In this case, having very broad data coverage is important because much of the variability over time is between less developed countries. Another limitation is that there currently are no valued network options for temporal inferential network models. Another limitation was the failure to find a model to converge on the valued network with raw edge weights using the Poisson distribution. This may be due to the extreme density of the networks and the wide dispersion in the edge weight values [90]. There are also well-known limitations of the Web of Science database used for the networks. Finally, as with all observational studies, in the absence of a control group and randomization, causal inference cannot be conclusively established regarding hypothesized effects.