Connected from the outside: The role of U.S. regions in promoting the integration of the European research system

Achieving strong integration between member countries is a primary goal for the European Union (EU). In research & development (R&D), specific policies have been implemented (Nedeva & Stampfer, 2012; Scherngell & Barber, 2011). The Framework Programs for Research and Technological Development are an example of such policies.

The EU R&D system has been analyzed in depth in the literature, with contrasting results. Hoekman, Frenken, and Tijssen (2010) and Miguelez and Moreno (2013) have found that the bias to collaborate within the same EU country has diminished over time. Morescalchi, Pammolli, et al. (2015) have underlined that this decrease has stopped since the mid-1990s. Chessa, Morescalchi, et al. (2013), moreover, have highlighted that the EU integration growth might have been driven by trends toward globalization of research more than by the aforementioned EU-specific efforts.

In this paper we study a related, though different, problem. In fact, we aim to understand which countries and regions contribute most to the integration of the European R&D system. Our study includes both the EU and the U.S., to shed light on the role that relevant external agents play in European integration. To the best of our knowledge, ours is the first attempt to tackle this issue.

Networks of innovators (Orsenigo, Pammolli, & Riccaboni, 2001; Owen-Smith, Riccaboni, et al., 2002; Powell & Grodal, 2005) can be analyzed to assess interregional connections. We focus on the patent coinventorship network (Chessa et al., 2013; Morescalchi et al., 2015), where nodes are regions, and edges are weighted by the number of coinventions occurring between regions.

We employ the resistance distance (Klein & Randić, 1993) to measure distances within the network. Resistance distance takes into account the path(s) that must be covered on the network to join two nodes. Also, the resistance distance between two nodes of a network represents the expected time that a random walk needs to move from the first node to the second one (von Luxburg, Radl, & Hein, 2010). In our case, this measure can be considered as a proxy of the velocity of the information flow (Stephenson & Zelen, 1989) along the network, which takes into account not only the shortest paths between the nodes (Goddard & Oellermann, 2011) but also longer ones, because information may flow indirectly on the network also on these paths (Bozzo & Franceschet, 2013).

To evaluate the contribution of individual countries and regions to EU integration (i.e., their integration capability), we first define an indicator of EU integration on the basis of the closeness centrality between EU regions belonging to different countries in the technological collaboration network. Then, the integration capability of a country or region is quantified by measuring the difference in the indicator value when that same country or region is removed from the network. Our analyses are focused on patent data and therefore, as discussed in Arora, Belenzon, and Patacconi (2018), Arora, Belenzon, et al. (2019), and Arora, Fosfuri, and Gambardella (2004), are biased toward development activities rather than toward research activities. As a consequence, the knowledge flows that we are investigating are more related to technological knowledge than to scientific knowledge.

Our main findings are the following:

• • 

  The countries exhibiting the largest contribution to EU R&D integration are Germany and the United States, with the latter being more relevant than most EU countries.

• • 

  In this context, we find that a considerable fraction of the regions that are most relevant for EU R&D integration are located in the United States, rather than within EU borders.

• • 

  The smallest EU countries turn out to be those benefiting most from the U.S. contribution to establish an indirect connection to other EU countries.

The paper is organized as follows. Section 2 summarizes the previous studies on the border and distance effects on the intensity of collaborations. Section 3 describes the Regpat data set that has been employed in this work, and introduces the indicator we use to measure the integration capability. Section 4 shows a set of analyses on the coinventor network. First, we propose some descriptive statistics and pictures, to provide an initial understanding of the structure of the network (section 4.1). Second, the integration capability of countries (section 4.2) and individual regions (section 4.3) is analyzed. Third, the previous results are deepened to understand which EU countries rely most on the United States to connect to other EU countries (section 4.4). Finally, section 5 concludes the paper.

In recent years, several studies have analyzed the effects of geography on R&D collaborations. In particular, the intensity of R&D collaborations between regions (i.e., their “R&D closeness”) has been studied based on geographical distance and on belonging to the same country. The intuition suggests that in a globalized world, where low transport costs, ICT facilities, and widespread knowledge of the English language are making communication between widely separated people easier, geographical factors should play a marginal role in determining the collaboration intensity between two regions (Frenken, Hoekman, et al., 2009; Singh & Marx, 2013). However, the analyses proposed so far in the literature, relying on different data and tools, have produced conflicting conclusions.

Among the papers supporting a decrease in the importance of geographical factors over time, Brun, Carrère, Guillaumont, et al. (2005) consider the trade scenario; the authors propose a gravity model generated from data of the United Nations Commodity Trade Statistics, where the effect of physical distance on the trade volume between countries is shown to diminish over time. Waltman, Tijssen, and van Eck (2011), in contrast, consider Web of Science (WoS) data on scientific publications, and compute for each paper the greatest distance between the addresses of the authors; they observe that, in spite of differences between scientific sectors, there is a clear trend of increasing distance over time.

Other studies, however, claim alternative evidence. Ponds (2009) studies international collaborations employing a probit regression on copublication data involving Dutch institutions; he finds that these collaborations grow, but at the same pace as the national ones. Maisonobe, Eckert, et al. (2016) build a copublication network between cities using data from the Science Citation Index Expanded, and find that in most countries domestic collaborations grow faster than international ones.

In the EU, an increase in collaborations between countries might be favored not only by the trend toward globalization of research, but also by the specific policies undertaken. Hoekman et al. (2010) apply a gravity model to copublication data from WoS, finding that the bias toward collaborating with partners from the same EU country decreases over time, while the bias toward cooperators that are geographically close does not. Miguelez and Moreno (2013) employ a gravity model to study the patent regional coinventor network; similar to Hoekman et al. (2010), they find that the importance of belonging to the same country diminishes over time, while the distance effect actually grows. Chessa et al. (2013) propose difference-in-differences estimates on four regional networks, concluding that integration between EU countries is growing, but no more than one would expect due to research globalization trends. Morescalchi et al. (2015) claim, through a gravity model on patent regional networks, that distance and country effects within the EU decreased only until the mid-1990s. Another gravity model for patent data is introduced by Cappelli and Montobbio (2016), who share the view that the effects of distance and national borders within the EU are decreasing over time. Finally, Doria Arrieta, Pammolli, and Petersen (2017), using publications data, show that the 2004/2007 EU enlargement has had a negative impact on cross-border collaborations.

The above results, though conflicting in some respects, show signs of a pattern toward high R&D integration within the EU, while the effectiveness of European policies has not been fully demonstrated. In this work we study EU integration from a different point of view. We introduce a different way to measure the R&D closeness between regions, not relying only on the intensity of direct collaborations (e.g., number of coinventorships or coauthorships) but considering also indirect connections. The introduction of an indirect measure allows us to understand which countries and regions provide the greatest contribution to EU integration, fostering the connection of EU countries and regions.

The data employed in this study are drawn from the OECD Regpat database (March 2018 version), containing all patent applications filed with the European Patent Office (EPO). Each patent is associated with its inventors, whose geographic location, in terms of NUTS3 regions, is also known. Table 1 reports some basic statistics related to the Regpat data set. Note that only 9.6% of the patents that are coinvented by inventors coming from different regions are also coassigned to multiple institutions. This happens because many coinventor relations are referred to inventors working in different subsidiaries of multinational firms. Therefore, the coinventorship network reflects for a relevant part the organization of work between firms and their subsidiaries. Still, we are interested in patterns of knowledge flows between regions, so we maintain that the embedding of new knowledge in the collaborating regions is relevant irrespectively of institutional boundaries.

The data set globally contains 5,520 regions, but in our analyses we consider just those belonging to the EU-15 countries¹ (1,067 regions) and the United States (3,144 regions). We restrict the analysis to the EU-15 countries because they have been part of the European Union for the longest time, and thus have been more significantly involved in its policies. The remaining EU countries have been in the Union only since 2004 or later (i.e., no more than 17.1% of the total time span considered in this work), which to us appears to be too little to include them in a study on EU R&D integration.² In Figure 1 we plot the number of patents by year from 1980 to 2014, considering patents including at least one U.S. inventor and patents including at least one EU-15 inventor; in both cases the number is steadily increasing.

These data are used to build the coinventor geographic network, referred to specific time periods. In these networks nodes are constituted by NUTS3 regions, while the weight w_ij(t) of the edge joining the nodes i and j in the network at time t is given by the number of coinventions happened between the i and j regions in the time period t. In our work the time period t will be represented by one of the intervals 1980–1989, 1990–1999, 2000–2009, and 2010–2014, or by individual years. Table 2 shows the number of edges and the sum of the weights in the networks related to 1980–1989, 1990–1999, and 2000–2009; we omit the latest period (2010–2014) because it is shorter. Like the number of patents, the values of these indicators are remarkably also growing with time.

In this section we illustrate the main methods and techniques used to carry out the analyses. First, the resistance distance is introduced (section 3.2.1). Then, we describe our measures of integration capability (section 3.2.2) and how we use null models to support our claims (section 3.2.3). Finally, changepoint detection is explained (section 3.2.4).

We remark again the importance of evaluating the closeness between pairs of nodes with a measure, like the inverse of the resistance distance, which takes into account multiple paths joining the nodes, and not just first-order interactions. First, using a measure considering multiple paths joining the nodes recognizes that knowledge may also flow on the network in an indirect, mediated way. Second, as we will explain in section 3.2.2, this allows us to measure the contribution that nodes (i.e., regions) and sets of nodes (i.e., countries) provide to the closeness of other nodes.

Note that a possible alternative closeness measure taking paths into account, as mentioned in the introduction, is the inverse of the shortest-path distance, where the shortest-path distance between two nodes is the sum of the inverses of the weights of the edges lying on the shortest path joining the two nodes. However, this measure is less suitable than the inverse of the resistance distance in our scenario, because it considers only the shortest paths, neglecting the fact that information may also flow on the network on other, longer paths.

To assess the contribution of a subset of the nodes of the network to the average cross-border closeness (i.e., to the integration), we compute the percentage closeness loss that happens when this subset is excluded from the network. We will exclude sets of nodes to evaluate the contribution of the countries, and single nodes to evaluate the contribution of specific hubs (intended as very relevant regions, characterized by many connections). The percentage closeness loss associated with a region/country represents the integration capability of that region/country.

The quantity $c¯$^Sj, in contrast, denotes the value obtained measuring $c¯$ once the subset S has been removed from the network, therefore excluding S from both the average and the paths used in the computation of the c_ijs. $c¯$^Sj can be computed through Eq. (3), but determining the c_ijs exploiting only the paths not transiting from the nodes in S. For instance, if the subset S again contains the German nodes, the German nodes are not considered in the average and are not allowed to appear in the paths used in the computation of the closeness between the pairs of nodes.

Finally, to analyze in greater depth the contribution of the United States to European integration also shortest paths are computed. The shortest path between two nodes s and t of a network is the path between s and t such that the sum of the weights of the constituent edges is minimized (Goddard & Oellermann, 2011). In more detail, we will consider the shortest paths between EU nodes belonging to different countries, counting how many American nodes are contained in these paths.

Notice that, in summary, the percentage closeness loss represents the ability of a set of nodes to make other nodes of the network closer. Therefore, it is a measure for sets of nodes that is related to two other traditional centrality measures defined instead for individual nodes: betweenness and current-flow betweenness. The betweenness of a node is the number of shortest paths crossing that node, while current-flow betweenness measures the extent to which a node lies on paths between other nodes. Betweenness considers only the shortest paths while current-flow betweenness takes into account all the paths, although longer paths give a lesser contribution. Percentage closeness loss considers all the paths, not only the shortest ones; therefore, it is a measure referred to sets of nodes that is more similar to current-flow betweenness.

To better appreciate the percentage closeness losses obtained on our networks we sometimes compare them with those measured on null models, where the null model of a network is another network obtained by keeping some elements constant and randomizing other ones. In more detail, we will use three classes of null models:

• • 

  Gravity-based null model (Expert, Evans, et al., 2011). This model is used to check whether the detected patterns are simple effects of gravitylike forces, depending on spatial distance and on a concept of mass. Let g_ij be the geographical distance between the regions associated with nodes i and j, and M_i and M_j the masses of the nodes. The weight $wijNM$ of the edge (i, j) in the null model is defined as $wijNM$ = M_iM_jf(g_ij), with f(g) = (⁠$∑i,j∣gij=g$w_ij)/(⁠$∑i,j∣gij=g$M_iM_j). The weight of the edge (i, j) grows with the mass of i and j, and with the weights that in the real network are associated with the nodes geographically at the same distance of i and j. This null model preserves neither the weights of the edges nor the strength of the nodes, but maintains the total weight of the network. In our framework the spatial distances are continuous, so it is necessary to divide them into bins. We consider the mass of a node as the total number of patents produced in the corresponding region in the time frame of interest.

• • 

  Null model where the edges within the United States are randomly reshuffled. We consider two variants of this null model: one that does not preserve the strength of the nodes, and one that approximately preserves it (Rubinov & Sporns, 2011). For instance, if in the real network the edge between Boston and Los Angeles is 100, in the null model this weight can be assigned to the edge between Portland and Memphis. In the first variant all the U.S. nodes obtain approximately the same strength. Both the variants preserve the weights of the edges, which are reshuffled. The two variants of this null model are dubbed US-INT and US-INT-STR, respectively.

• • 

  Null model where the EU-US connections are randomly reshuffled. We consider two variants of this null model: The first preserves the strength (of the EU-US connections) just for the EU nodes, while the second (approximately) preserves it for both EU and U.S. nodes (Rubinov & Sporns, 2011). For instance, if in the real network the edge between Paris and Santa Clara has weight 200, in the null model this weight can be assigned to the edge between Paris and Anchorage. Both variants preserve the weights of the EU-US edges, which are reshuffled. The two variants of this null model are dubbed EU-US and EU-US-STR, respectively.

In order to better appreciate the yearly variations of the percentage closeness loss we use a technique named changepoint detection (Killick, Fearnhead, & Eckley, 2012), which identifies the time instants (changepoints) corresponding to abrupt changes in a function. Identifying the changepoints splits the function in sections, and in particular we split the yearly percentage closeness loss function where the regression line changes the most. This is achieved by finding the sections of the function such that the sum of the residual errors of the regressions in each section is minimized.

In this section we show the results of our analyses on the coinventorship network. We begin by providing some preliminary statistics and pictures (section 4.1), and then we analyze the integration capability of countries (section 4.2) and of individual regions (section 4.3). Finally, we study in more detail the impact of the United States on the closeness of the individual EU countries to the other EU countries (section 4.4).

In this section we show some preliminary statistics and illustrations to provide an initial understanding of the structure of the coinventor network. In particular, we want to highlight the features of the national subnetworks and of the hubs.

Table 3 contains some statistics related to the national subnetworks in the 2000–2009 period. For each country we report the number of nodes, the number of nodes with 250 or more patents, and the average closeness (computed as the inverse of the resistance distance) between the pairs of nodes associated with 250 or more patents. We have computed this average considering only the nodes associated with a certain number of patents because the other ones are not likely to appear in effective paths joining EU regions belonging to different countries. The average closeness between U.S. nodes is 78% greater than that of the EU (see the last two lines of the table).

We can also see the greater number of connections existing in the U.S. subnetwork with respect to the EU ones pictorially in Figure 2, where the network contains a node for each country, but we have decided to decompose the United States into the component states; only the edges representing at least 500 coinventions are included. The edge thickness is proportional to the number of coinventions and the size of the nodes is proportional to the sum of the weights in the national subnetwork. The figure allows us to appreciate the links between the U.S. states, whereas the EU countries are less connected.

Finally, Figure 3 represents the shortest paths connecting EU nodes belonging to different countries in the 2000–2009 period. The figure shows the edges that are part of at least one shortest path joining two regions belonging to different EU countries. The edge thickness grows with the edge weight, while the node size grows with the betweenness, computed considering only the shortest paths joining EU nodes belonging to different countries. The figure is quite difficult to read due to the size of the network, but in any case allows us to understand that, even when considering only the paths joining different EU regions, there emerge a significant share of relevant hubs that are within the United States and not in the EU itself.

Summarizing, this preliminary analysis suggests that the U.S. subnetwork contains many nodes associated with a relevant number of patents, and that these nodes are more connected between each other than happens in the European national subnetworks. So it seems plausible that the U.S. subnetwork as a whole might provide a faster, though indirect, connection between EU regions. Moreover, we can note that there are several nodes from both the EU and the United States that are crossed by many shortest paths joining EU nodes belonging to different countries. Again, it seems worthwhile to investigate the relative importance of U.S. and EU hubs in making the EU R&D system more integrated.

In this section we will use the procedure described in section 3.2 to show which countries (among the EU countries and the United States) have the greatest integration capability; that is, contribute the most to increasing the closeness between the EU regions belonging to different countries.

To measure the integration capability of a country, the formulas of section 3.2 are applied considering the subset S to be excluded as the set of the nodes belonging to that country. In this way the value resulting from Eq. (4) gives the percentage closeness loss due to the removal of the country; the greater this percentage, the greater the integration capability of the country.

Table 4 shows the integration capability for each EU country and for the United States. The analyzed years are divided into four periods: 1980–1989, 1990–1999, 2000–2009, and 2010–2014. The main feature that emerges from the table is that Germany and the United States are by far the countries with the greatest capability to connect the EU countries. Notice that the United States exhibits a large contribution to European R&D integration, greater than that shown by the European countries themselves. On the one hand this is due to the larger population—the population of the United States is, for instance, almost five times that of France—providing more possibilities to establish collaborations, but on the other hand it indicates that the United States plays a fundamental role in the European R&D system. Indeed, a large population alone is not enough to develop joint R&D projects.

We want now to understand whether the measured integration capabilities are just the result of simple gravitylike forces (i.e., mass and distance effects), or are due to more complex dynamics (e.g., different propensity to long-distance collaborations). To this aim, the gravity null model, introduced in section 3.2, is employed. The integration capabilities observed using the gravity model are in Table 5, reported as differences with respect to the values observed in the real network, while Figure 4 provides a pictorial representation of the comparison between the real network and the gravity null model referred to the 2000–2009 period. We note that when the gravity model is used, almost all the European countries show an integration capability that is greater than that observed in the real network (i.e., they contribute to European integration less than is expected due to simple mass-distance effects); for instance, Germany would be 2.5 times more important if R&D collaborations only depended on mass and distance. In contrast, the United States exhibits an integration capability that is greater than the amount due to gravity.

In order to further investigate the role of the United States in the EU R&D system and try to understand the nature of the connections linking Europe and the United States, we make use of a more straightforward measure of distance along the network (i.e., shortest paths) and we employ null models disrupting some portions of the network (internal U.S. network and EU-US connections) to assess their relative importance.

In the first place, we have determined all the shortest paths joining EU regions belonging to different countries, and classified them on the basis of the number of U.S. regions that are included. The second column of Table 6 reports the shortest paths statistics related to the real network in the 2000–2009 period (for brevity, the other periods follow a similar pattern). U.S. regions participate in more than half of the intra-EU cross-border shortest paths, thus confirming the importance of the United States in the EU R&D system. Interestingly, as we sensed from Figure 2, many shortest paths include several U.S. nodes (24% of the shortest paths include more than three U.S. nodes), and this confirms that since the U.S. subnetwork contains many internal connections, the most convenient way to link two EU regions is often to move to the U.S. subnetwork, cover “cheap” paths inside this subnetwork, and find the most appropriate node to exit.

Let us consider the two classes of nongravity null models introduced in section 3.2: The first class randomizes the connections inside the United States, while the second one randomizes the connections between EU and U.S. nodes. Thus, the first class allows us to evaluate the relevance to the EU integration of the connections internal to the United States, while the latter permits us to assess the importance of the EU-U.S. connections. Columns 3 to 6 of Table 6 contain the results of the shortest path analysis for the null models (2000–2009 period), while Table 7 shows the percentage closeness loss for the null models. All the values reported for the null models are obtained by repeating the random generation of the models 100 times and then averaging the measurements. The tables indicate the differences with respect to the results obtained with the real network. We have performed a t-test to evaluate the statistical significance of the differences; all the values in Tables 6 and 7 are statistically significant with p-values ≪ 0.01, except for one value in Table 6.

Regarding the class of null models reshuffling the U.S. internal connections, the first variant preserves just the weights of the network, while the second one also maintains the strengths of the nodes (i.e., it preserves the hubs within the U.S. subnetwork). First, we note that the U.S. integration capability in terms of resistance distance remains almost constant in both variants. This happens because the U.S. subnetwork has many nodes and edges, and therefore reshuffling the connections leaves, in both cases, good paths between the pairs of U.S. nodes; this confirms that the great number of connections in the U.S. subnetwork helps join the EU regions. However, the second null model exhibits a performance that is more similar to that of the real case, thus suggesting that the presence of strong U.S. hubs facilitating the links is also important. When we analyze the effect of null models on shortest paths, which are more sensitive than the resistance distance to changes in the network, these considerations are reinforced: On the first null model, the number of shortest paths transiting from the United States falls with respect to the real situation, while on the second null model it even grows. This behavior seems to indicate again the importance of the hubs, confirming the intuitive evidence of Figure 3. The growth of the number of shortest paths with U.S. nodes in the second null model is probably due to the fact that once the hubs are preserved, a more balanced distribution of the weights to the edges helps find better paths.

The relevance of the U.S. hubs is confirmed also by the last class of null models: those reshuffling the EU-US connections. The first null model of this class does the reshuffling by preserving the strength of the transatlantic connections just for the EU nodes, while the second one preserves this strength also for the U.S. nodes. The second null model behaves similarly to the real situation, while in the first one the U.S. contribution to EU integration decreases; this suggests that it is not enough to connect to the U.S. network, it must be done through the right access points.

Finally, we conduct a finer-grained temporal analysis to shed further light on the variations of the U.S. contribution that emerged in the four decades by using yearly networks. We want to assess whether the U.S. integration capability has evolved over time following a steady trend, or the tendency has changed through time. To do this, we employ changepoint detection analysis. In brief, this method allows us to retrieve the optimal set of linear slope changepoints to model the observed data (see section 3.2.4 for details), thus discovering the possible changes in the trend of the magnitude of the U.S. integration capability. We apply the method to the pattern of yearly percentage closeness loss in the EU network due to collaborations with the United States, with the aim of identifying the years in which the trend of growth or decrease of the U.S. contribution to the EU integration has changed significantly. We have considered the years from 1981 to 2014, omitting 1980, which is associated with few data. The resulting plot is in Figure 5(a). The changepoint associated with the greatest reduction of the residual error of the regression is detected in 1997 (highlighted in red in Figure 5); then, two more changepoints are identified in 1983 and 1987. Interestingly, before 1997 the U.S. contribution shows, globally, a positive trend, while after 1997 there is a long period with a clearly negative trend. Figure 5(b) shows the R² and p-values of the discovered regressions; note that the last two regressions, which are those of greatest interest, are significant at p < 0.05.

It can be seen that the latter result is consistent with the evidence shown by Chessa et al. (2013), who have highlighted that EU integration, in the same case of coinventorship, has experienced growth starting in the years before 2000. Also, they find that the integration level has subsequently stabilized. Our changepoint detection analysis identifies a clear inversion of the tendency in the U.S. contribution to EU integration in the same period, which then started to decrease. Therefore, the growth of the EU integration level found by Chessa et al. (2013) seems to be reflected in a progressive emancipation of the EU from the U.S. R&D system. In this respect, we point out the possible role of EU policies, characterized by increasing financing of R&D programs, fostering intra-EU collaborations.

Summarizing this section, we find that Germany and the United States provide the highest contribution to connect the EU countries. In particular, the United States has a more significant impact than most European countries. Our analyses indicate that two important factors that make the United States able to help connect the EU regions are represented by the high number of links in the internal U.S. network and the presence of U.S. hubs: To connect two EU nodes it is enough that these two nodes are close to two distinct American nodes, which may then usually be easily linked through a path within the U.S. subnetwork especially, due to the help of effective internal hubs. Finally, we observe that the U.S. contribution to EU integration seems to have been decreasing since 1997.

In this section we appraise the integration capability of individual hubs. Studying individual nodes is interesting, because they are much more similar in terms of population than the countries, thus leading to less biased analysis results.

It must be noticed that the integration capability of a node may derive from two different factors: the ability to connect foreign regions, and the ability to connect regions of the same country with the outside. The American hubs can benefit only from the first factor, since we are considering just cross-border EU links.

To evaluate the integration capability of an individual region, the procedures of section 3.2 are applied considering this region as the subset S to be excluded in Eqs. (3)–(4). Eqs. (3)–(4) actually result in evaluating both the ability to connect foreign regions and the ability to connect regions of the same country to the outside. We are also interested in evaluating the first factor alone, and to this end we consider, in the numerator and denominator of Eq. (3), only the pairs of regions not belonging to the same country of the hub. We begin by conducting a comparison between EU and U.S. hubs in terms of the ability to connect foreign regions, and then we analyze the European hubs considering also their ability to connect regions of the same country to the outside. We have considered for each analyzed time period an initial set of nodes with 30 EU regions and 30 U.S. regions chosen as those with the greatest current-flow betweenness, where the current-flow betweenness has been computed considering only the paths joining regions belonging to different EU countries.

Table 8 shows for the four time periods mentioned above the percentage closeness loss for the EU and U.S. main hubs considering only the ability to connect foreign regions, while Table 9 repeats the evaluation only for the EU hubs appraising also the ability to connect nodes of the same country of the hub to the outside.

The main evidence that arises from Table 8 is that the effect of the U.S. hubs is comparable to that of EU hubs, and even stronger in the period 1990–1999. This further supports our previous considerations regarding the importance of the United States in the EU R&D system: The strong American hubs may act as entry and exit points in the U.S. subnetwork, and then also facilitate the connections inside the subnetwork.

The most recurrent European hub is Munich, while other important regions are Berlin and Aachen. Regarding the United States, the main hub seems to be Cambridge, MA, with an important role played by San Jose, CA; a very relevant integration capability is shown also by Houston, TX in 1980–1989 and Rockville, MD in 1990–1999. As a further insight about the U.S. hubs, we can analyze their main IPC patent classes. The most frequent class is Medical/Veterinary for Cambridge and Rockville, Computing for San Jose and Drilling/Mining for Houston. Therefore, with the exception of Houston in 1980–1989, it appears that the integration capability of the U.S. has been driven by regions focused on ICT and life science fields.

Table 9, instead, takes into account also the ability to connect regions of the same country with the outside. In this table new regions emerge, for instance Milan and Vienna. These regions play an important role in their national subnetworks. It is interesting to note that the contribution of Milan has diminished over time, suggesting that the other Italian regions may have become more capable to connect autonomously with foreign EU countries.

The previous section has highlighted the very relevant role of the United States in strengthening the connections between European countries. In this section we look further into this issue by trying to understand which European countries most need the United States to become close to the other ones.

In order to evaluate the impact of the United States on the closeness of a specific EU country to the other EU countries we evaluate again the percentage closeness loss. In this case Eq. (3) must consider only the pairs of nodes involving a region of the country of interest, and the subset S of the network to be excluded is represented by the American nodes. In this way Eq. (4) results in the contribution of the United States to the closeness of the country under analysis to the other EU countries.

Table 10 shows the numerical results in the usual four time periods, while Figure 6 gives a graphical intuition of the proportion of the U.S. contribution to the different countries. It is possible to notice that the countries benefiting the most from U.S. collaborations are the smallest ones: Actually, these countries, due to their size, need more external collaborations to carry out R&D projects. This confirms the findings of Waltman et al. (2011), according to which peripheral countries are more prone to start long-distance collaborations. An exception is represented by the UK, but such an exception was expected given the well-known strong relationship of this country with the United States.

Figure 7 deepens the results shown in Table 10 and Figure 6, by proposing a heatmap of the percentage closeness loss due to the United States between pairs of EU regions. In more detail, the heatmap showing the percentage closeness loss is in Figure 7(b), while for the sake of completeness Figure 7(a) reports the starting situation, with the absolute closenesses computed on the network including only the EU. As in Table 10, it can be noticed that the most peripheral countries benefit most from the U.S. contribution; in addition, Figure 7(b) underlines that the connections between pairs of peripheral countries are the ones helped most. Another peculiarity highlighted by Figure 7(b) is that the diagonal of the heatmap is very light, meaning that the connections between pairs of regions within the same EU country do not need the mediation of the U.S. to be established. Finally, note that Germany, which provides a very relevant contribution to EU integration (see the previous sections), does not seem to benefit that much from the help of the United States to become connected to the other EU countries. This may be due to the fact that Germany, as is clear from Figure 7(a), has a high closeness to the other countries also on the network containing only the EU countries, and so does not need the U.S. to create connections.

In this work we have studied the patent coinventor network and used indirect distance measures to investigate the contribution of individual countries and regions to European R&D integration, that is, their integration capability. The analysis has been carried out on a network encompassing both the EU and United States, in order to ascertain also possible contributions from the United States to European integration.

After having proposed some descriptive statistics and pictures for the coinventor network, we have analyzed the contribution to EU integration provided by countries and regions by computing the amount of network-based closeness between European regions belonging to different countries that is lost when specific subsets of the nodes of the network are removed. We can summarize the main conclusions of this work as follows:

• • 

  The countries that contribute most to connecting regions across EU countries are Germany and the United States. In particular, the United States proves to be more relevant in joining EU countries than most of the EU countries themselves. The integration capability of the United States is more than the country would have had if the collaborations were driven exclusively by gravitylike effects, while the integration capability of almost all the EU countries is lower than that due to gravity. Moreover, our analyses indicate that an important factor that makes the United States able to foster the connection of EU regions is represented by the high number of links in the U.S. subnetwork: To connect two EU nodes on the coinventor network it is enough that these two nodes are close to two distinct American hubs, which may then usually be easily linked through a “fast” path within the United States. Also, the connections within the U.S. subnetwork are facilitated by the presence of strong internal hubs. In addition, the U.S. contribution to EU integration seems to have been decreasing since 1997, in conjunction with renewed efforts by the EU in support of European technological collaborations.

• • 

  There are strong regional hubs in terms of integration capability in both Europe and the United States. Some European hubs, especially German ones, are able to connect regions of foreign countries. Other European hubs, the most notable example being Milan, have a remarkable effect on the average cross-border closeness, but their contribution is especially in the connection of nodes of their same country with the outside.

• • 

  The role of the United States in promoting integration with other EU countries is stronger for the smallest EU countries, probably because these countries, due to their size, need more external collaborations to carry out R&D projects.

A first natural development of our work consists in using the techniques introduced in this paper to analyze the role that other external countries besides the United States, such as Japan, play in EU integration.

Moreover, in the introduction we claimed that patents are more related to development activities than to research activities, and thus our coinventorship network is biased toward the flow of technological knowledge. It would be interesting to delve into the more scientific part of knowledge flow, building a collaboration network using scientific publication data.

It would be interesting also to repeat the study using the regional coassignment network, which shows the connections between the regions where the headquarters of companies and institutions are located. This alternative analysis might highlight different trends, and comparing these results with those obtained on the coinventorship network—which more naturally describes the relationships of knowledge exchange between regions—might lead to further intriguing insight. In addition, the coassignment network might also be used to investigate the contribution of different institutional types (e.g., companies vs. public research organizations) to EU R&D integration.

Finally, another possible extension to our study regards the evaluation of the integration capability of a region through a mathematical model with weights to be learned from data; such weights would allow us to understand the relevance of various effects (e.g., the size of the nodes), to the integration capability.

Emanuele Rabosio: Data curation, Formal analysis, Investigation, Methodology, Software, Visualization, Writing—original draft, Writing—review & editing. Lorenzo Righetto: Formal analysis, Methodology, Visualization, Writing—review & Editing. Alessandro Spelta: Methodology, Writing—Review & editing. Fabio Pammolli: Conceptualization, Methodology, Supervision, Writing—review & editing.

The authors have no competing interests.

No funding has been received for this research.

This research work was carried out using the Regpat data set (March 2018 version), publicly available from OECD.