The “Political” Geography of Research Networks

Abstract

The production of scientific and technical knowledge is mostly concentrated in specific locations. Knowledge flows very easily within regions; however, scientific and technical knowledge does flow also between different regions. The aim of this article is to analyze knowledge flows between agglomerations of innovative inputs, and their effects on the innovative performance of regions. We estimate a regional knowledge production function and we test, through appropriate spatial econometric estimation techniques, the effect of both geographical and relational autocorrelation (as measured by participation to joint research networks in the European Union [EU] Sixth Framework Programme [FP6]). We adopt two selection criteria in order to define different relational “geographies” (hence spatial weights matrices), and we model the unobservable structure and link value of knowledge flows within these joint research networks. Our results confirm established evidence that knowledge flows within interregional research networks along a top-down nonsymmetrical and hierarchical structure. However, the EU enlargement - and a modified structure of incentive for collaboration activity of European institutions- changed the direction of knowledge flows toward a top-down dynamics of knowledge diffusion from coordinator to participants for “EASTWARD” research networks (whose coordinator is in the west and most participants in the east); while the contrary (a hierarchical bottom-up dynamic of knowledge transfer) is true for WESTWARD networks (whose coordinator is in the east and most participants in the west). FP6 is therefore a platform for knowledge barter exchange for EU15, while works as a mere one-way channel for knowledge diffusion from EU15 toward Central and Eastern European countries.

Keywords

innovation patents knowledge flows spatial econometrics spillovers

Introduction

Scientific and technical knowledge is generally concentrated in specific regions (Swann, Prevezer, and Stout 1998; Bresnahan, Gambardella, and Saxenian 2001; Maggioni 2002; Braunerhjelm and Feldman 2006). Knowledge flows readily within these geographical areas (and neighboring ones) via the mobility of inventors and highly qualified workers, the interactions among producers and subsuppliers of specialized inputs, and knowledge spillovers.

However, scientific and technical knowledge flows also across different areas and some breakthrough technologies have been developed through the explicit cooperative efforts of scientists and technicians working in different geographical locations.

This article draws on two literature streams. The first deals with the identification and study of innovation networks (Jaffe, Henderson, and Trajtenberg 1993; Paci and Usai 2000, 2009; Maurseth and Verspagen 2002; Cowan and Jonard 2003; Breschi and Lissoni 2004, 2009; Maggioni, Nosvelli, and Uberti 2007, 2014; Maggioni, Uberti, and Usai 2011; Maggioni, Breschi, and Panzarasa 2013; Maggioni and Uberti 2007, 2009, 2011; Hoekman, Frenken, and Oort 2009; Picci 2010; Broekel and Hartog 2013; Miguélez and Moreno 2013; Sebestyén and Varga 2013; Cassi and Plunket 2014); the second discusses the use of spatial econometric techniques to account for the existence of directly unmeasurable (or unmeasured) spillover effects (Audretsch and Feldman 1996; Acs, Anselin, and Varga 2002; Fischer and Varga 2003; Bottazzi and Peri 2003; Greunz 2003; Bode 2004; Moreno, Paci, and Usai 2005; LeSage and Pace 2009; Autant-Bernad and LeSage 2010; Mora and Moreno 2010; Usai 2011; Audretsch, Huelsbeck, and Lehmann 2012; Varga, Pontikakis, and Chorafakis 2014).

The article builds also on Maggioni et al. (2007, 2014) in assuming that knowledge can be diffused and exchanged either through unintentional patterns of diffusion based either on spatial contiguity (à la Acs et al. 2002) or on intentional relations based on aspatial networks (à la Cowan and Jonard 2004).

The first approach stresses the role played by “unintended” spatial spillovers; that is, that knowledge, generated by a given agent located in a specific locality, is diffused to other agents (i.e., firms, universities, and research centers) in neighboring areas (Acs et al. 2002). In this view, space is what matters and knowledge flows following a generally “purely’ geographical pattern.¹

The second approach emphasizes that the exchange of knowledge through aspatial networks occurs through voluntary “barter” exchanges, and interactions within specialized networks which have been intentionally established (Cowan and Jonard 2004). Technological and scientific knowledge, developed within a region, is exchanged within networks, which frequently arise from formal and contractual agreements between institutions. In this second case, it is relations that matter, and knowledge flows following intentional patterns, which may have little correlation with geographical proximity.

In previous papers (Maggioni et al. 2007, 2014; Maggioni and Uberti 2009), we used data on membership in joint research contracts (JRCs) financed by the EU’s Fifth Framework Programme (FP5) and applied different spatial econometric techniques in order to:

verify whether formal relationships based on aspatial networks between geographically distant regions prevail over unintended patterns of diffusion based on spatial contiguity,

measure the effects of different “spatial” weight matrices, which refer to both geographical and relational “proximity,”

devise a series of tests to identify the most likely organizational structure of knowledge flows, connecting European regions and financed by the EU through the FPs.

These analyses were conducted on samples composed of 109 (Maggioni et al. 2007) and 171 (Maggioni et al. 2014) EU15 regions, in which we used FP5 to proxy for the existence of knowledge barter exchanges. The present study includes an additional 54 regions from eleven Central and Eastern European countries (CEE11) which joined the EU in 2004 and 2007, and maps scientific collaborations using FP6 data.

The paper is organized as follows: in the second section we compare FP5 and FP6 and analyze the consequences of the CEE11 inclusion in the EU; in the third section, we describe the aggregation of FP6 contracts at the regional level in order to obtain relational weight matrices. In the fourth section, we formulate the main research questions (RQs), in the fifth section we describe the models to be estimated, and in the sixth section we present the results. The article concludes in the seventh section.

From FP5 to FP6: A Radical (Political) Change

The EU’s attempt to create a new integrated space for research and development (R&D) started in 1984 with the introduction of the FPs, which are a series of policy initiatives that have changed overtime in relation to the instruments adopted and their goals and budget (European Policy Evaluation Consortium [EPEC] 2011).

Over twenty-five years of FPs, the EU’s interest in what it calls “core themes” (energy and information and communication technologies) remained fairly stable, alongside the introduction of new “horizontal themes” to achieve better socioeconomic integration and cohesion across Europe. Different actors have been targeted in order to exploit scientific results on the markets, and explicit attempts have been made over time to involve private firms and institutions in the FPs.

In terms of funding, since the mid-1990s and the launch of FP4, EU has focused on four action categories: research and technological development (RTD), transfer of RTD knowledge, technology adoption and innovation, and support for policy making. An explicit intention in FP6 was the financing of research networks, for example: research infrastructures, networks of excellence, technology platforms or joint technology initiatives. The FPs have grown from an initial 2.75 billion ecu funding in FP1 (1984–88) to 51 billion euros in FP7 (2007–13; EPEC 2011) and are aimed at creating and strengthening the European R&D “fabric.”

Several contributions (Maggioni et al. 2007; Autant-Bernad et al. 2007; Maggioni and Uberti 2009; Paci and Usai 2009; Scherngell and Barber 2009, 2011; Balland 2012; Caloghirou, Ioannides, and Vonortas 2004; Caloghirou, Protogerou, and Vonortas 2011; Protogerou, Caloghirou, and Siokas 2010, 2013; Lata and Scherngell 2013; Marrocu, Paci, and Usai 2013b, 2014; Autant-Bernad, Fadairo, and Massard 2013; Maggioni et al. 2014) show that JRCs, financed under different FPs, are useful tools to trace knowledge flows across the EU.

In 2004 and 2007, there were several major structural political changes (i.e., accession of the new countries to the EU).² In this article, we compare data from FP5 referring to the EU15 countries with FP6 data which apply to a larger number of (very heterogeneous) regions (see Figure 1).

Figure 1.

Patents per million labor force (average value 2007–2008–2009).

FP6, which officially covered the period 2002–2006,³ was designed to “strengthen the scientific and technological bases of industry and encourage its international competitiveness while promoting research activities in support of other EU policies” (European Commission [EC] 2002, 1), and, following the principle of subsidiarity, projects have to be transnational. In other words, only consortia of partners from different members and associated countries can apply; for mobility and training actions, the fellows typically have to go to a country different from their country of origin or residence. Activities that can better be carried out at national or regional level, that is, without cooperation across borders will not be eligible under the FP (EC 2002).

Thus, for the regional analysis, we initially considered all contracts involving institutions located in the 225 Nomenclature of Territorial Units for Statistics, level 2 (NUTS2) regions belonging to the EU26 countries⁴ (see Online Appendix Table A1 for details of the distribution of JRCs in FP5 and FP6).

However, since it is reasonable to assume that the abovementioned political and policy changes induced a change in the networking behaviors of research institutions (in order to maximize the likelihood of receiving funding), and due to the vast differences existing in the innovative performance of “old” and “new” EU members (see Figure 1), in this article we run every step of the econometric analysis on different samples, which are based on Ex Ante and Ex Post sample selection criteria (see Online Appendix Figures A1 and A2 for a graphical depiction):

the Ex Post sample selection criterion identifies three different groups of regions based simply on the geographical location of networks members: the complete sample (225 regions belonging to the EU26), the EU15 sample (171 “western” regions), and the CEE11 sample (54 regions belonging to the eastern European countries which joined the EU in 2004 and 2007);

the Ex Ante sample selection criterion identifies four types of research contracts: “pure EU15” JRC which involves coordinators and participants exclusively located in the EU15; “pure CEE11” which include JRC whose coordinators and participants are exclusively located in the CEE11; EASTWARD research contracts which refer to JRC coordinated by institutions located in the EU15 that have at least one participant in CEE11; WESTWARD JRC which include networks coordinated by institutions located in the CEE11 with at least one participant from the EU15.

Although FP6 was aimed explicitly at fostering the creation of “research networks,” not all financed projects involved more than one participant; hence, in this analysis, from a total of 10,100 research projects, we selected 6,167 projects⁵ with an average 11.5 membership.

The spatial pattern of the contracts financed under the FP6 shows a rather uneven distribution of coordinating institutions among the EU15 and the CEE11 (see Table 1). From a total of 5,837 contracts related to EU26 countries, only 251 (4.2 percent) were managed by institutions located in CEE11 with at least one institution located in the EU15 (i.e., WESTWARD networks), and only 51 (0.8 percent) refer to contracts with exclusive CEE11 membership (i.e., pure CEE11). The spatial distribution of participants is also skewed, with larger absolute numbers in core EU15 regions. However, in relative terms, CEE11 regions host a larger ratio of participants/coordinators (i.e., 6,504/322 for CEE11 regions, against 47,823/5,515).

Table 1.

Geographical Distribution of FP6 Institutions (Ex Ante and Ex Post Criteria).

Ex Ante sample selection criterion
	JRC members	Number of institutions	Average membership	Regions
Pure EU15	d (Coord EU15)	2,913	8.33	144 (of 171)
	e (Partic EU15)	17,736		169 (of 171)
	x1 (Aliens)	3,628
EASTWARD	f (Coord EU15)	2,602	15.84	143 (of 171)
	c (Partic CEE11)	5,715		54 (of 54)
	g (Partic EU15)	29,036		170 (of 171)
	x2 (Aliens)	3,862
WESTWARD	a (Coord CEE11)	251	8.41	37 (of 54)
	m (Partic CEE11)	612		47 (of 54)
	h (Partic EU15)	1,051		135 (of 171)
	x3 (Aliens)	196
Pure CEE11	i (Coord CEE11)	51	4.71	16 (of 54)
	b (Partic CEE11)	177		43 (of 54)
	x4 (Aliens)	12
Ex Post sample selection criterion
	JRC members		Number of institutions
EU15	Coord		5,515
EU15	Partic		47,823
CEE11	Coord		322
CEE11	Partic		6,504
EU26	Coord		5,837
EU26	Partic		54,327
Aliens	Coord		330
Aliens	Partic		10,514
Total	Coord		6,167
Total	Partic		64,841

Note: JRC = joint research contract; CEE11 = eleven Central and Eastern European countries; EU = European Union; FP6 = Sixth Framework Programme; Aliens = institutions located outside EU26 countries.

Evidence of policy-induced effects on these spatial distributions is provided by the average membership of “pure,” or autarchic contracts (see Online Appendix Table 1): pure EU15 research projects show an average membership of 8.33, while for projects involving exclusively institutions located in CEE11 regions, the average membership is only 4.71. These proportions are replicated if we compare EASTWARD and WESTWARD research contracts: the average membership of EASTWARD networks is 15.84, while the average membership of WESTWARD networks is only 8.41. Western coordinators seem therefore able to manage a larger research network compared to eastern ones.

Table 2 presents the growth process of the number of scientific and technological institutions and organizations involved in JRC financed under FP5 and FP6 and the geographical distribution of the evolving European Research Area (ERA).

Table 2.

Coordinators and Participants in FP5 and FP6: Absolute Numbers, Shares, and Variations of Shares.

	FP5		FP6		FP5		FP6		ΔS (FP6-FP5)
Country	Coord	Partic	Coord	Partic	SCoord	SPartic	SCoord	SPartic	Coord	Partic	ΔSCoord—ΔSPartic
Austria	170	958	217	1,574	2.66%	2.68%	3.72%	2.90%	1.05%	0.21%	0.84%
Belgium	301	1,424	337	2,501	4.72%	3.99%	5.77%	4.60%	1.06%	0.61%	0.44%
Denmark	179	1,046	145	1,501	2.80%	2.93%	2.48%	2.76%	−0.32%	−0.17%	−0.15%
Finland	134	983	114	1,177	2.10%	2.75%	1.95%	2.17%	−0.15%	−0.59%	0.44%
France	834	4,658	802	6,332	13.07%	13.05%	13.74%	11.66%	0.67%	−1.40%	2.07%
Germany	1,068	5,509	1,042	8,912	16.73%	15.43%	17.85%	16.40%	1.12%	0.97%	0.15%
Greece	273	1,670	222	1,812	4.28%	4.68%	3.80%	3.34%	−0.47%	−1.34%	0.87%
Ireland	90	531	84	627	1.41%	1.49%	1.44%	1.15%	0.03%	−0.33%	0.36%
Italy	703	4,062	622	5,641	11.02%	11.38%	10.66%	10.38%	−0.36%	−1.00%	0.64%
Luxembourg	11	61	13	98	0.17%	0.17%	0.22%	0.18%	0.05%	0.01%	0.04%
Netherlands	517	2,149	434	3,364	8.10%	6.02%	7.44%	6.19%	−0.67%	0.17%	−0.84%
Portugal	87	912	69	956	1.36%	2.56%	1.18%	1.76%	−0.18%	−0.80%	0.61%
Spain	527	2,878	410	4,058	8.26%	8.06%	7.02%	7.47%	−1.23%	−0.59%	−0.64%
Sweden	183	1,435	220	2,219	2.87%	4.02%	3.77%	4.08%	0.90%	0.06%	0.84%
United Kingdom	1,184	4,857	784	7,051	18.55%	13.61%	13.43%	12.98%	−5.12%	−0.63%	−4.49%
Bulgaria	7	172	24	396	0.11%	0.48%	0.41%	0.73%	0.30%	0.25%	0.05%
Czech Republic	23	432	24	1,007	0.36%	1.21%	0.41%	1.85%	0.05%	0.64%	−0.59%
Estonia	10	127	17	337	0.16%	0.36%	0.29%	0.62%	0.13%	0.26%	−0.13%
Hungary	15	416	53	1,042	0.24%	1.17%	0.91%	1.92%	0.67%	0.75%	−0.08%
Latvia	13	93	5	177	0.20%	0.26%	0.09%	0.33%	−0.12%	0.07%	−0.18%
Lithuania	3	91	10	294	0.05%	0.25%	0.17%	0.54%	0.12%	0.29%	−0.16%
Malta	0	20	4	94	0.00%	0.06%	0.07%	0.17%	0.07%	0.12%	−0.05%
Poland	26	599	128	1,686	0.41%	1.68%	2.19%	3.10%	1.79%	1.43%	0.36%
Romania	6	199	22	534	0.09%	0.56%	0.38%	0.98%	0.28%	0.43%	−0.14%
Slovakia	7	161	20	381	0.11%	0.45%	0.34%	0.70%	0.23%	0.25%	−0.02%
Slovenia	11	249	15	556	0.17%	0.70%	0.26%	1.02%	0.08%	0.33%	−0.24%
Total EU26	6,382	35,692	5,837	54,327	100.00%	100.00%	100.00%	100.00%	—	—	—

Note: FP = Framework Programme.

Columns 1–8 record the absolute numbers and relative shares of each type in the whole FP; columns 9–11 show the variation in the shares of coordinators and participants from FP5 to FP6 and the differences in these variations. Columns 9–11 show that, in general, the participation of CEE11 increases more than EU15 participation and that this EASTWARD expansion of research networks is mainly due to an increase in the number of participants rather than coordinators.⁶

When we consider the CEE11 regions and the spatial distribution of coordinators and participants—depicted in Figure 2 and based on Ex Ante sample selection criterion (see Table 1 and Online Appendix Figures A1 and A2 for details)—we observe that most coordinators of WESTWARD networks (panel a in Figure 2) are localized in weakly innovative (measured by patent intensity) regions as compared to average CEE11. This suggests that they are seeking access to external competences through the establishment of networks involving western partners located in highly innovative regions.

Figure 2.

Geographic distribution of CEE11 coordinators and participants in the FP6. Note: Malta (part of CEE11) has not been represented in the maps. CEE11 = eleven Central and Eastern European countries; FP6 = Sixth Framework Programme.

The spatial distribution of participants suggests that, despite the similarity between the two typologies, participants in EASTWARD networks (panel c) tend to be located in regions showing higher levels of innovative activity compared to participants in WESTWARD networks (panel m). This may be due to western coordinators using stricter selection criteria when choosing eastern participants.

Figure 3 depicts the distribution of coordinators and participants according to the Ex Ante selection criterion for the EU15.

Figure 3.

Geographic distribution of EU15 coordinators and participants in FP6. Note: EU = European Union; FP6 = Sixth Framework Programme.

Panels d and f show an extreme similarity of the geographical distribution of coordinators of “pure EU” and EASTWARD networks and a striking difference with respect to geographical distribution of patenting intensity (Figure 1). In particular, German regions appear to be underrepresented in the coordination of JRC with respect to their innovative performance. The main difference concerns the geographic distribution of participants in WESTWARD networks. As shown in panel h, thirty-six regions are missing from these networks. This may suggest that eastern coordinators choose good performing western participants with the aim of drawing on their scientific and technological knowledge. Panels g and e show no significant differences and highlight a higher degree of spatial autocorrelation.

Some Empirical Issues Related to the Geography of Innovation Networks

Before performing the econometric analysis to investigate how scientific and technological networks are built across European regions, we need to peer inside the black box of networks financed by FP6 JRC. We believe that the regional level is the best to conduct an empirical analysis of innovative processes since it allows consideration of interagent spillovers which would be overlooked where the analysis performed at the individual agent or institution levels.

Regional innovation performance, proxied by patenting intensity, is determined by region-specific innovative inputs combined according to a knowledge production function (KPF) and influenced by the innovative performance of “neighboring” regions (neighborhood being defined, à la Maggioni, Nosvelli, and Uberti 2007, in both geographical and relational terms).

The innovation literature includes several spatial econometric analyses of regional innovative performance (as detailed in the introduction) and uses alternative measures of technological, institutional, social, and organizational neighborhoods (Torre and Gilly 2000; Boschma 2005; Cantner and Meder 2007; Boschma and Frenken 2009; Ponds, Van Oort, and Frenken 2007, 2010; Marrocu et al. 2013a), here we extend the line of research initiated by Maggioni et al. (2007) which considers JRC as network structures channeling interregional knowledge flows.

In order to detect how knowledge flows within these projects ideally requires a field experiment. If this is not feasible, then hypotheses can be formulated about the most probable network structure and tested using appropriate econometric tools. In what follows we provide a brief discussion of the hypotheses on different networks structures and layouts, which reflect how research contracts are organized internally; how interregional links are weighted, how knowledge flows between regions are measured, and how the directions of the links from sender to recipient of a specific piece of scientific and technological knowledge are identified.

Network Structures and Flows Directions

Each FP6 JRC includes a list of members and specifies different organizational roles, that is, coordinator (or principal investigator) and participants. No other information is provided on the internal structure of the network or the structure and direction of knowledge flowing within the research network. Therefore, in line with Maggioni and Uberti (2011), we conceived a number of logically consistent network structures compatible with knowledge flowing within these collaborative research networks by combining two orthogonal dimensions (i.e., the direction of links and the structure of the network) in order to consider their most meaningful combinations (Figure 4).

Figure 4.

Network structures of collaborative research contracts. Source: Maggioni et al. (2014).

Figure 4 depicts the case of a very small and simple research network composed of one coordinator and four participants. Thus, knowledge can flow in four different ways within the network, meaning four different relational structures can emerge.

Links (i.e., knowledge flows) can be reciprocal, and the underlying network structure may be hierarchical if there are mutual, egalitarian, but exclusive ties between the coordinator and each participant (Figure 4 panel A, Star). In this case the network displays a “star-like” structure, with a very high centralization value, but a symmetry of relations that guarantees mutual exchange of knowledge, filtered by the pivotal player.

However, knowledge can flow among the set of agents, irrespective of their structural position (Figure 4 panel B, Complete). This structure reflects the absence of hierarchy within this “flat” network (indeed all centralization indices have values equal to zero), and the potential for knowledge to flow among all the actors. There is no coordination and/or brokerage of knowledge and information, and all agents have an equal status of “member.”

The assumption of reciprocity of ties can be relaxed if we assume the existence of different knowledge stocks between coordinator and participants in terms of knowledge diffusion and absorptive capacity.

If knowledge flows involve an exclusive relation between the coordinator and each individual participant, as in a star-like structure, but different from star (Figure 4 panel A), there is no mutual or balanced exchange of knowledge between them. In this case, we can consider two alternative structures: an inward structure, that is, from participants to coordinator (Bottom-up, as in Figure 4 panel C); or an outward structure, that is, from coordinator to participants (Top-down, as in Figure 4 panel D).

Finally, there can be a network structure with no reciprocity of links and no hierarchy (Figure 4 panels E and F, Clockwise circle and Anticlockwise circle): in this case every member exchanges knowledge locally and exclusively with his/her nearest neighbor (in clockwise or counterclockwise direction), and a wheel-like knowledge flow structure emerges, in which all members are interchangeable and there is no central node. A wheel-like structure, by definition, achieves global transmission of knowledge only through multiple passages via local links. Wheel structures may provide microeconomic advantages, as shown by Jackson (2008), but since they are the most unlikely to achieve effective knowledge flows within research networks we exclude them from the econometric analysis.

How to Weight Knowledge Flows?

The second issue concerns the values of the links within a research network and the use of binary versus weighted networks to measure the existence and amount of knowledge exchanged (and/or transferred) within the network. This is related to a more general issue in social network analysis which has been addressed by the more recent literature (Fagiolo, Reyes, and Schiavo 2007; Fagiolo 2010; Opsahl and Panzarasa 2009; Opsahl, Agneessens, and Skvoretz 2010; Barigozzi, Fagiolo, and Garlaschelli 2010).

Based on the use of network membership data contained in the FP6 database, to represent actual knowledge flows, we formulated the following hypotheses:

we give the value 1 to each link described by the given network structure irrespective of the number of network nodes. Thus, we assume the amount of knowledge exchanged and/or transferred within a larger network to be higher than that exchanged in a smaller network and, indirectly, that there are no “budget constraints” on the relational capacity of a node. We indicate this modality with subscript 1;

we count 1/N (where N is the total number of nodes in a given network) for each link in the chosen network structure in order to take account of the limited relational capacity of the nodes within a network. We indicate this modality with subscript N;

we count 1/L (where L is the number of links of a given network) for each link described by the chosen network structure to take account of the limited relational capacity of a network which may depend nonlinearly on the number of nodes. We indicate this modality with subscript L;

we count F/N (where F is the amount of funds received from the EU and N is the total number of nodes of a given network) for each link described by the chosen network structure, to take account of both the different financial “sizes” of different networks and the limited relational capacity of a node within a network. We indicate this modality with subscript F.

On the bases of the above assumptions, we built sixteen different knowledge flow layouts (four structures × four link weights) for each JRC funded by FP6.⁷

We transformed each layout into a weight matrix to be used in the subsequent econometric analysis.

As discussed in the second section, we use two alternative selection criteria to identify two different geographical samples and, consequently, derive different weight matrices referring to EU15, CEE11, and EU26.

The first method, following an Ex Post sample selection criterion, consists of considering all JRCs and eliminating data (by deleting appropriate rows and columns of the weight matrix) referring to countries not included in each sample. This method produced three geographical samples: EU26, EU15, and CEE11 (see Table 1 and Figure A1, for details).

According to the second method, following an Ex Ante sample selection criterion, we obtain four alternative geographical samples: pure EU15, pure CEE11, EASTWARD, and WESTWARD (see Table 1 and Figure A2 for details). Pure JRCs are defined as those which have both the coordinators and participants in either CEE11 (i and b) or EU15 (d and e); EASTWARD refers to networks with a western coordinator (f) and at least one eastern participant (c), and WESTWARD referring to contracts with an eastern coordinator (a) and at least one western participant (h).⁸

The final results for each network layout (i.e., the combination of network structure and links weights), are squared matrices Z^m where a generic element Z^m _ij measures the extent of the knowledge flows between region i and region j within an m generic network layout.⁹

Since we are interested in both the relational and the geographic dimensions of knowledge flows, in the spatial econometric analysis of a regional KPF, we use the abovementioned sixteen Z^m matrices, as relational weight matrices, and three different geographical weight matrices (see the “Model and Estimation Strategy” section for details).

RQs and Hypotheses

The main hypothesis in this article is that region i’s innovative output, measured by patents, is explained by regional innovative inputs and structural characteristics, and by some spatial autocorrelation effects which may arise from geographic knowledge spillovers and/or relational knowledge barter exchanges mediated by a specific network layout, and weight links.

The economic literature on European research networks (Breschi and Cusmano 2004; Caloghirou et al. 2004; Caloghirou et al. 2011; Scherngell and Barber 2009, 2011; Protogerou et al. 2010, 2013; Lata and Scherngell 2013) financed under the FPs—despite the high variance in the data used, level of analysis, and estimation methods—shows the existence of an oligopolistic structure in which a restricted number of institutions localized in central and high-income regions play a major role along a core-periphery pattern.

In a previous paper (Maggioni et al. 2014), we showed that relational autocorrelation (proxied by FP5) influences the innovative performance of European NUTS2 regions. Although, theoretically, relational autocorrelation could apply to all hypothesized layouts, only one typology of contract structure (i.e., the Bottom-up layout C) appeared to be relevant for influencing the patenting activity of a relationally defined neighboring region. This result suggested that, on the one hand, intentional knowledge exchange mainly follows a hierarchical network structure, probably for efficiency reasons, and, on the other hand, that FPs may be used as policy instruments to sustain the innovative performance of certain regions, but do not foster regional cohesion, since most coordinators are located in core regions.

In this analysis, through a series of spatial econometric exercises presented in the succeeding sections, we test the existence and extent of relational autocorrelation, as mapped by FP6 JRCs, and distinguish different institutional dimensions of the EU26.

More generally, our empirical analysis tests the following RQs:

RQ1: Is the interregional network structure obtained by aggregating individual JRCs independent of the geographical sample coverage of the analysis?

While differences in innovative inputs and outputs between CEE11 and EU15 have been well documented (Azagra-Caro, Pontikakis, and Varga 2013; Hoekman et al. 2013; Sebestyén and Varga 2013), based on attributional data (patents see Figure 1), a relational perspective shows whether FP6 is an effective channel for the diffusion of scientific and technological knowledge across the entire EU26 or this function is limited to a subsample of regions. Different sample selection criteria are proposed in this article to highlight this issue.

RQ2: Is there a significant structural difference between JRCs coordinated by institutions located in CEE11 and EU15? Is there a “political geography” of research networks such that knowledge flows follow different patterns according to the geographic localization of coordinators and participants?

As discussed in previous Sections, we use Ex Ante sample selection criteria to identify four network typologies (pure EU15, pure CEE11, EASTWARD, and WESTWARD) which contribute to highlighting the effects of EASTWARD enlargement on the ERA and on innovation performance.

RQ3: Do actual knowledge flows in FP6 JRCs follow a hierarchical and asymmetric pattern as previously shown for the FP5?

The theoretical literature on network structure arising from the micro-based game theoretical approach (recently surveyed by Goyal 2007; Vega-Redondo 2007; Jackson 2008) or from the heterogeneous agents, simulation, and/or experimental approach (à la Cowan and Jonard 2003, 2004; Maggioni 2004; Morone and Taylor 2004; Callander and Plott 2005; Cassi and Zirulia 2008; Goeree, Riedl, and Ule 2009), and the large literature on the effects of network structure on the innovative performance of individual nodes (i.e., individual scientists, firms, regions) discusses the advantages arising from different network structures. In particular, while a small-world structure is shown to be the most efficient layout for maximizing the average content in a scientific network, it may be unfit for equity reasons and rejected by voluntary aggregations of research institutions. A star-like structure has been shown to be an efficient layout and is easily implemented if the balance of power between network members is very unequal.¹⁰ Due to the lack of consensus in the literature, we have no clear expectations about the prevailing structure. We discuss this further in the Results section.

RQ4: Is there a trade-off between the size of a region’s scientific and technological network and its effectiveness in influencing innovative performance in that region?

While it is clear that a large network provides an advantage with respect to the number of knowledge sources that can be accessed, there is theoretical evidence that network size may be constrained by time and the relations among network members and by the network coordinator who is at the hub of the network (Jackson and Wolinsky 1996; Goyal and Joshi 2003). Based on this intuition, a hierarchical network with decreasing returns to the number of nodes (or links) should perform better in depicting the effects of knowledge flows within regional JRC on innovation performance. We address this further in the Results section.

Model and Estimation Strategy

The empirical analysis consists of testing a KPF (Grilliches 1979) where the innovative output of a region (measured by patenting intensity) is modeled as a function of three different sources of R&D (to proxy for a number of other unobservable innovative inputs) and other variables characterizing the innovative and productive structure of each region.

Formally:

P A T_{i}^{t} = f (B i z R D_{i}^{τ}, G o v R D_{i}^{τ}, U n i R D_{i}^{τ}, P R O D_{i}^{τ}, I N N_{i}^{τ}, A C C E S S_{i}^{τ}, B E T W_{i}^{τ}),

where the dependent variable (PAT_i ^t ) is the number of patent applications to the European Patent Office per million labor force registered at time t in region i. This variable is the average value for period t, that is, 2007, 2008, and 2009 (source: Organization for Economic Cooperation and Development [OECD] 2013).

The first three regressors are different sources of knowledge, that is, business R&D (BizRD_i ^τ), government R&D (GovRD_i ^τ), and university R&D (UniRD_i ^τ) expenditure (source: Eurostat 2013a; OECD 2013). Following Glaeser et al. (1992), we also consider the positive influence of the specialization or differentiation of the region’s productive and innovative structures on its innovative output. To do this, we include PROD_i ^τ and INN_i ^τ, which are location quotients calculated, respectively, for local units in high-tech sectors and for high-tech patents¹¹ (source: Eurostat 2013a, 2013b).

ACCESS_i ^τ is the multimodal accessibility index, a measure of the region’s combined (air, road, and rail) accessibility measured as deviation from the European average which is set equal to 100 (source: Espon 2013).

In order to control for the different average innovative performance of CEE11 regions, in EU26, EASTWARD, and WESTWARD samples, we introduced a dummy variable.

All regressors are computed for each region i at time τ, that is, the average for the period 1999–2005. We expect this time lag to: take into account the delay between R&D expenditure and patent application, to reduce endogeneity, and to cope with the problem of missing values at the regional level.

To estimate the relevance of the structural position of regions within FP6 research networks, we included another regressor, BETW_i ^τ, for the betweenness centrality¹² of each region i. This variable proxies for the ability of the region’s bridging position to control the diffusion of scientific and technical knowledge across research networks stretching across Europe.¹³

The theoretical and empirical economics of innovation literature shows that innovative activity, similar to several other economic phenomena, is characterized by agglomeration and spillovers, with the result that ordinary least square (OLS) estimations could be biased, which requires the use of spatial econometric techniques. Hence, it is crucial to define the dimension of the “space” captured by the spatial weight matrices adopted to identify the presence of spatial dependence (Anselin 1988 and 2010; LeSage and Pace 2009; Elhorst 2010).

Traditionally, spatial econometric techniques have drawn on the basic definition of weight matrices as contiguity matrices, that is, binary matrices indicating whether two regions share a border or not. In this analysis, we apply first-order geographical or local contiguity, LCont. In this binary matrix, a value equal to 1 identifies adjacent neighbors.

In order to define different effects of geographical proximity not captured by simple binary matrix, we include other geographic dimensions such as a global contiguity matrix (GCont) and geographical proximity (Prox). GCont is based on the shortest contiguity path between region i and region j and is a discrete matrix computed as the geodesic proximity.¹⁴ Finally, we include a measure of geography that has been widely used in spatial econometric exercises, that is, the geographical distance between centroids of regions. Since all other weight matrices record higher values for more relevant neighbors, we transform distance into proximity (Prox) by computing the inverse of the geographical distance.

To summarize, our econometric analysis involves nineteen weight matrices: three dimensions of geography, that is, LCont, GCont, and Prox, and sixteen relational layouts computed according to both Ex Ante and Ex Post selection criteria, as discussed in the third section. This means that spatial weight matrices (usually denoted by W) are, alternatively, the geographical matrices, labeled GEO, and the relational matrices labeled REL.

Before discussing the estimation results, recall that the econometric analyses were carried out on three samples, the full sample of 225 regions (i.e., EU26, EASTWARD, and WESTWARD networks) and two subsamples of 171 regions (i.e., EU15 and pure EU15) and 54 regions (i.e., CEE11 and pure CEE11) in order to identify the most relevant determinants of knowledge spillovers in EU. We proceed by estimating a double-log specification of the explicit versions of model 1 and detecting the presence of autocorrelation in the residuals with respect to each of the weight matrices and to quantify it.

Starting with the seminal contribution from Anselin (1988), the spatial econometric literature has proposed several models to deal with the problem of spatial spillovers in the context of cross sectional data.¹⁵ A very general spatial autoregressive model including both a spatially autoregressive error term (indicated by the coefficient λ) and the spatial lag on the dependent variable (indicated by the coefficient ρ) can be defined as follows:

y = ρ W_{1} y + X β + u and u = λ W_{2} u + ∊,

where W₁ and W₂ are squared spatial weight matrices which might be the same, and ∊ is the error term. Imposing some restrictions on the weights in the previous model, that is, W₁ = 0 or W₂ = 0, two different spatial autoregressive models can be tested, a spatial error model (SEM), if W₁ = 0:

y = X β + u and u = λ W_{2} u + ∊

or a spatial autoregressive model¹⁶ (SAR), if W₂ =0:

y = ρ W_{1} y + X β + u

Since the objective is to consider the joint presence of geographical and relational effects we estimate a two weights SAR model, as described in Lacombe (2004) and LeSage and Pace (2009), defined as follows:

y = ρ_{G E O} W_{G E O} y + ρ_{R E L} W_{R E L} y + X β + ∊,

where W_GEO and W_REL are the geographical and the relational weights and we can estimate both geographic lag (ρ _GEO ) and relational lag (ρ _REL ) jointly.

The model strategy is depicted in Figure 5. First, we test an OLS estimation of model (1); second, we identify the presence of spatial autocorrelation in the residuals and accordingly the model to be estimated (i.e., a SEM or a SAR). Significance of at least the ρ _GEO and ρ _REL coefficients suggests the estimation of a Lacombe model, including both REL and GEO spatial weights matrices.

Figure 5.

Estimation stages.

In order to compare the two different nonnested models (SAR and Lacombe), following Burnham and Anderson (2002) we compute Akaike weights, prob_j , as follows:

p r o b_{j} = \frac{\exp [- \frac{1}{2} (A I C_{C j} - A I C_{C M I N})]}{\sum_{r = 1}^{R} \exp [- \frac{1}{2} (A I C_{C r} - A I C_{C M I N})]},

where j is the model, R is the number of models, and Akaike Information Criterion (AIC)_C is the bias-adjusted AIC value.

Corrado and Fingleton (2012) stress that the selection of spatial weight matrices is crucial for an accurate detection of the presence of spatial spillovers. In this article, we draw on the innovation networks literatures in order to select the appropriate weight matrices.

Results

Note first that the independent variables, which usually appear in a KPF, act in this estimations as controls; thus we do not focus on the value of their parameters. In relation to the spatial autocorrelation analysis, we are interested mainly in the significance of ρs since this should distinguish which, among the different network layouts, is conducive to flows of relevant knowledge. More specifically, in order to identify and deal with the issue of spatial autocorrelation, we follow a straightforward, four-stage tree-trimming procedure (Figure 5), which we apply to each data sample, geographical specification, and layout and subsequently restrict its application to the case selected based on the results.

Stage 1: Table 3, which is based on the explicit versions of model (1) for all samples, geographies, networks, and layouts, presents the diagnostic tools, Lagrange multiplier (LM) and robust LM, used to detect the presence of a spatially autoregressive error term (and its coefficient λ) or the spatial lag on the dependent variable (and its coefficient ρ), according to Hendry’s procedure. Following Florax, Folmer, and Rey (2003) and adopting a “hybrid specification strategy” based on the robust LM values, the tests indicate the presence of spatial autocorrelation for all geographic weight matrices (LCont, GCont, and Prox) when the analysis is performed on all samples¹⁷ according to an Ex Post sample selection criterion (see Table 3 for EU26, EU15, and CEE11), and according to an Ex Ante sample selection criterion (see Table 3, for EASTWARD and WESTWARD).

Table 3.

LM test on SAR and SEM.

EU26, n = 225 (continued)
	LM lag value	Pr	LM robust lag	Pr	LM error value	Pr	LM robust error value	Pr
LCont	9.249	(0.002)	4.323	(0.038)	8.015	(0.005)	3.089	(0.079)
GCont	9.830	(0.002)	9.753	(0.002)	0.088	(0.766)	0.011	(0.917)
Prox	10.381	(0.001)	8.472	(0.004)	4.637	(0.031)	2.727	(0.099)
A₁	0.002	(0.962)	0.300	(0.584)	0.446	(0.504)	0.744	(0.388)
A_F	0.001	(0.974)	0.345	(0.557)	0.277	(0.599)	0.621	(0.431)
A_L	0.026	(0.873)	0.467	(0.494)	0.247	(0.619)	0.689	(0.407)
A_N	0.023	(0.879)	0.412	(0.521)	0.225	(0.635)	0.613	(0.434)
B₁	0.012	(0.912)	0.007	(0.932)	0.096	(0.756)	0.092	(0.762)
B_F	0.042	(0.837)	0.255	(0.614)	0.913	(0.339)	1.126	(0.289)
B_L	0.001	(0.978)	0.240	(0.624)	0.273	(0.601)	0.512	(0.474)
B_N	3.669	(0.055)	2.053	(0.152)	1.696	(0.193)	0.080	(0.778)
C₁	0.002	(0.966)	0.517	(0.472)	0.465	(0.495)	0.980	(0.322)
C_F	0.000	(0.983)	0.282	(0.595)	0.170	(0.680)	0.452	(0.502)
C_L	0.051	(0.821)	0.516	(0.473)	0.177	(0.674)	0.642	(0.423)
C_N	0.045	(0.832)	0.500	(0.480)	0.188	(0.665)	0.643	(0.423)
D₁	2.365	(0.124)	1.990	(0.158)	0.468	(0.494)	0.092	(0.762)
D_F	1.493	(0.222)	1.032	(0.310)	0.476	(0.490)	0.015	(0.903)
D_L	2.351	(0.125)	1.739	(0.187)	0.651	(0.420)	0.039	(0.844)
D_N	2.333	(0.127)	1.829	(0.176)	0.563	(0.453)	0.060	(0.807)
EU15, n = 171 (continued)
LCont	10.180	(0.001)	6.606	(0.010)	4.434	(0.035)	0.860	(0.354)
GCont	1.012	(0.314)	1.210	(0.271)	0.911	(0.340)	1.109	(0.292)
Prox	11.102	(0.001)	9.430	(0.002)	3.555	(0.059)	1.883	(0.170)
A₁	0.582	(0.445)	1.317	(0.251)	0.008	(0.930)	0.743	(0.389)
A_F	0.207	(0.649)	0.028	(0.867)	0.534	(0.465)	0.355	(0.551)
A_L	1.240	(0.265)	1.505	(0.220)	0.123	(0.726)	0.387	(0.534)
A_N	1.214	(0.271)	1.509	(0.219)	0.105	(0.745)	0.401	(0.527)
B₁	0.000	(0.991)	0.007	(0.934)	0.019	(0.891)	0.025	(0.873)
B_F	0.421	(0.517)	0.003	(0.957)	0.926	(0.336)	0.508	(0.476)
B_L	0.221	(0.638)	0.174	(0.677)	0.061	(0.805)	0.014	(0.907)
B_N	2.883	(0.090)	2.208	(0.137)	0.960	(0.327)	0.284	(0.594)
C₁	0.789	(0.374)	2.144	(0.143)	0.045	(0.832)	1.401	(0.237)
C_F	0.329	(0.566)	0.075	(0.784)	0.270	(0.603)	0.017	(0.898)
C_L	1.593	(0.207)	2.471	(0.116)	0.055	(0.815)	0.933	(0.334)
C_N	1.539	(0.215)	2.473	(0.116)	0.038	(0.846)	0.971	(0.324)
D₁	3.348	(0.067)	3.956	(0.047)	0.132	(0.717)	0.740	(0.390)
D_F	2.706	(0.100)	3.643	(0.056)	0.064	(0.800)	1.001	(0.317)
D_L	3.396	(0.065)	4.063	(0.044)	0.169	(0.681)	0.835	(0.361)
D_N	3.383	(0.066)	4.094	(0.043)	0.138	(0.710)	0.850	(0.357)
CEE11, n = 54 (continued)
	LM lag value	Pr	LM robust lag	Pr	LM error value	Pr	LM robust error value	Pr
LCont	5.470	(0.019)	9.630	(0.002)	1.080	(0.299)	5.240	(0.022)
GCont	6.063	(0.014)	7.686	(0.006)	1.228	(0.268)	2.852	(0.091)
Prox	5.600	(0.018)	6.208	(0.013)	0.033	(0.855)	0.642	(0.423)
A₁	1.019	(0.313)	0.053	(0.818)	2.492	(0.114)	1.526	(0.217)
A_F	0.509	(0.475)	0.007	(0.936)	1.376	(0.241)	0.873	(0.350)
A_L	1.479	(0.224)	0.032	(0.858)	3.834	(0.050)	2.386	(0.122)
A_N	1.433	(0.231)	0.040	(0.842)	3.667	(0.055)	2.274	(0.132)
B₁	0.329	(0.567)	0.263	(0.608)	0.190	(0.663)	0.124	(0.724)
B_F	0.352	(0.553)	0.277	(0.599)	0.681	(0.409)	0.605	(0.437)
B_L	2.729	(0.099)	0.577	(0.447)	2.160	(0.142)	0.008	(0.930)
B_N	0.000	(0.983)	0.007	(0.932)	0.000	(0.986)	0.007	(0.933)
C₁	0.757	(0.384)	0.035	(0.852)	1.853	(0.173)	1.131	(0.288)
C_F	0.451	(0.502)	0.017	(0.897)	1.037	(0.309)	0.603	(0.437)
C_L	1.433	(0.231)	0.087	(0.768)	3.089	(0.079)	1.743	(0.187)
C_N	1.331	(0.249)	0.079	(0.779)	2.937	(0.087)	1.685	(0.194)
D₁	2.445	(0.118)	2.918	(0.088)	0.010	(0.919)	0.483	(0.487)
D_F	2.426	(0.119)	2.757	(0.097)	0.034	(0.853)	0.365	(0.546)
D_L	2.141	(0.143)	3.094	(0.079)	0.034	(0.854)	0.986	(0.321)
D_N	2.212	(0.137)	3.075	(0.079)	0.017	(0.895)	0.881	(0.348)
EASTWARD, n = 225 (continued)
LCont	9.249	(0.002)	4.323	(0.038)	8.015	(0.005)	3.089	(0.079)
GCont	9.830	(0.002)	9.753	(0.002)	0.088	(0.766)	0.011	(0.917)
Prox	10.381	(0.001)	8.472	(0.004)	4.637	(0.031)	2.727	(0.099)
A₁	0.169	(0.681)	1.014	(0.314)	0.165	(0.685)	1.010	(0.315)
A_F	0.081	(0.777)	0.159	(0.690)	0.463	(0.496)	0.542	(0.462)
A_L	0.366	(0.545)	1.045	(0.307)	0.015	(0.902)	0.694	(0.405)
A_N	0.351	(0.554)	0.976	(0.323)	0.015	(0.904)	0.640	(0.424)
B₁	0.002	(0.966)	0.080	(0.778)	0.105	(0.746)	0.183	(0.669)
B_F	0.001	(0.970)	0.528	(0.468)	0.949	(0.330)	1.475	(0.225)
B_L	0.560	(0.454)	1.529	(0.216)	0.120	(0.729)	1.089	(0.297)
B_N	0.363	(0.547)	0.783	(0.376)	0.057	(0.812)	0.477	(0.490)
C₁	0.166	(0.683)	1.200	(0.273)	0.222	(0.637)	1.256	(0.263)
C_F	0.045	(0.832)	0.073	(0.787)	0.239	(0.625)	0.267	(0.606)
C_L	0.407	(0.524)	1.273	(0.259)	0.025	(0.874)	0.891	(0.345)
C_N	0.378	(0.539)	1.204	(0.272)	0.030	(0.862)	0.857	(0.355)
D₁	4.503	(0.034)	4.001	(0.046)	0.514	(0.474)	0.012	(0.912)
D_F	3.902	(0.048)	3.674	(0.055)	0.247	(0.620)	0.019	(0.890)
D_L	4.545	(0.033)	4.066	(0.044)	0.482	(0.488)	0.002	(0.965)
D_N	4.542	(0.033)	4.093	(0.043)	0.450	(0.502)	0.001	(0.971)
WESTWARD, n = 225
	LM lag value	Pr	LM robust lag	Pr	LM error value	Pr	LM robust error value	Pr
LCont	9.249	(0.002)	4.323	(0.038)	8.015	(0.005)	3.089	(0.079)
GCont	9.830	(0.002)	9.753	(0.002)	0.088	(0.766)	0.011	(0.917)
Prox	10.381	(0.001)	8.472	(0.004)	4.637	(0.031)	2.727	(0.099)
A₁	4.417	(0.036)	7.192	(0.007)	0.103	(0.748)	2.877	(0.090)
A_F	3.232	(0.072)	6.876	(0.009)	0.490	(0.484)	4.134	(0.042)
A_L	4.495	(0.034)	8.318	(0.004)	0.236	(0.627)	4.059	(0.044)
A_N	4.520	(0.034)	8.147	(0.004)	0.212	(0.645)	3.839	(0.050)
B₁	3.410	(0.065)	5.889	(0.015)	0.170	(0.680)	2.648	(0.104)
B_F	2.503	(0.114)	5.387	(0.020)	0.451	(0.502)	3.335	(0.068)
B_L	3.337	(0.068)	6.417	(0.011)	0.284	(0.594)	3.365	(0.067)
B_N	3.369	(0.066)	6.345	(0.012)	0.264	(0.607)	3.240	(0.072)
C₁	5.944	(0.015)	9.854	(0.002)	0.210	(0.647)	4.121	(0.042)
C_F	4.651	(0.031)	9.045	(0.003)	0.486	(0.486)	4.880	(0.027)
C_L	5.663	(0.017)	10.250	(0.001)	0.325	(0.568)	4.913	(0.027)
C_N	5.746	(0.017)	10.210	(0.001)	0.303	(0.582)	4.767	(0.029)
D₁	2.182	(0.140)	2.743	(0.098)	0.184	(0.668)	0.746	(0.388)
D_F	2.230	(0.135)	3.064	(0.080)	0.450	(0.502)	1.285	(0.257)
D_L	2.110	(0.146)	2.959	(0.085)	0.494	(0.482)	1.343	(0.246)
D_N	2.136	(0.144)	2.904	(0.088)	0.405	(0.524)	1.173	(0.279)

Note: Probabilities reported in parentheses. LM = Lagrange multiplier; SEM = spatial error model; SAR = spatial autoregressive model; EU = European Union; CEE11 = eleven Central and Eastern European countries.

In terms of the relational weight matrices, the presence of autocorrelation depends on the different network layouts and on the geographical samples used in the analysis.

When adopting the Ex Post sample selection criterion (for EU15 and CEE11), the classical approach for network structure D and all geographic specifications suggests that the model to be estimated should include a spatial lag term, while for the remaining relational weights matrices (A, B, and C), the robust LM tests do not detect any spatial dependence to be corrected by a SAR model (or SEM). In other words, if knowledge flows are described in terms of A (i.e., hierarchical Star structure with mutual exchanges of knowledge), B (i.e., totally a flat non hierarchical structure with no core region), or C (i.e., a Bottom-up hierarchic structure with flows of knowledge stemming from the participants and flowing toward the coordinator of the joint research network contract), we do not observe any influence of the relational knowledge barter exchange phenomenon on the level of regional innovative activity. However, if the analysis is performed on the complete sample (225 regions from the EU26), no relational autocorrelation is present and the model can be estimated using OLS (see Table 3).

The results are similar for the Ex Ante sample selection criterion which identifies pure EU15 and pure CEE11 research contracts (see Online Appendix Table A2 and A3) and shows that the ERA under FP6 has radically changed the geography of European innovation, and significant pure EU15 and pure CEE11 knowledge flows are no longer evident.¹⁸ This suggests that autarchic scientific networks (despite their numerical relevance within FP6: 2,964 out of 5,837 contracts) do not significantly convey knowledge.

For EASTWARD and WESTWARD research networks, we find evidence of spatial autocorrelation and the LM test suggests that it should be corrected for using a SAR model for all three geographies. For EASTWARD networks, the D network structure is the only valid relational weight matrix, while for WESTWARD networks all network structures (apart from D) perform equally well.

Therefore, the selected hierarchical structure would seem to suggest a clear organizational principle for EASTWARD networks, in which knowledge flows from the core to the periphery. So, why would an established scientific institution located in a core western region look for participants located in the CEE11 countries in order to diffuse its scientific knowledge? If this is a network for knowledge exchange, what is the incentive for such a behavior? The most feasible answer is that the institutions located in core EU15 regions look for potential participants in CEE11 countries, in order to increase the probability of being funded.

Stage 2: Following Marrocu et al. (2013b), for substantive reasons, we do not consider a spatial Durbin model specification.¹⁹ This would require neighbors’ R&D investments to be productive across regions. Since such an assumption would be unrealistic in the European context, as shown by Deltas and Karkalokos (2013), we consider it more reasonable to assume that innovation spillovers work through the effective level of knowledge achieved by neighboring regions, proxied by the level of patenting intensity. Therefore, in the succeeding empirical analysis we employ a SAR specification, described in general terms in equation 4, and applied to the KPF specification:

\begin{array}{l} P A T_{i}^{t} = β_{0} + β_{1} B i z R D_{i}^{τ} + β_{2} G o v R D_{i}^{τ} + β_{3} U n i R D_{i}^{τ} + β_{4} D U M M Y C E E 11 \\ + β_{5} P R O D_{i}^{τ} + β_{6} I N N_{i}^{τ} + β_{7} A C C E S S_{i}^{τ} + β_{8} B E T W_{i}^{τ} + ρ_{Z} W P A T_{i}^{t} + ν_{i}^{τ} \end{array}

1bis

where ρ _Z is the coefficient of the spatially lagged dependent variable Patents which can be alternatively computed for three geographic (LCont, GCont, and Prox) and four “surviving” relational weight matrices (D₁ , D_N , D_L _, and D_F ), and where $ν_{i}^{τ}$ is the disturbance.

Based on the independent variables (see Table A4), both the direct and indirect effects of BizRD are generally positive and significant; while the coefficient of GovRD is always not significantly different from zero, probably because government R&D mostly focuses on precompetitive basic research which is not directly patentable. The significance and sign of the direct and indirect effects of UniRD are heavily dependent on the model specification and sample selection. In relation to the other coefficients, the direct and indirect effects of ACCESS are generally positive.

The coefficient of innovative specialization (INN), but not PROD, is positive and significant for EU15. For the CEE11 subsample, the direct and total effects of PROD are significant and negative. We can interpret these results as follows: innovative specialization matters only after a certain threshold of innovative activity, as in the EU15 subsample. Below this threshold, as in the CEE11 subsample, productive specialization is more relevant for determining the region’s innovative performance. These results suggest that differentiation rather than specialization is a source of innovation advantage for emerging regions in line with Jacobs’ (1964, 1969, 1984) argument.

The BETW of a region, computed for D layouts of the knowledge flows networks, displays significant positive effects only for the CEE11 sample. This result contrasts with previous results based on FP5 data (Maggioni et al. 2007, 2014) and hints at a reduced role of FP6 JRC as effective channels for knowledge flows aimed at the production of commercially exploitable innovation results (i.e., patents). It points also to the important role of FP6 research networks in the innovative activity of CEE11 regions and the advantage of being a local information broker bridging between western coordinators and other eastern partners (Wasserman and Faust 1994).

When the analysis is performed on the complete EU26 sample, the direct and indirect effects of the CEE11 dummy are negative and significant, confirming that the innovative performance of Central and Eastern European regions is still lagging behind.

The analysis performed on the ρ _Z coefficients shows the most important results. For the whole sample EU26, geography plays the usual role of unintentional barter exchange; for EU15 subsample almost all D weight matrices confirm the existence of relational autocorrelation and similar results apply to LCont and Prox. When the analysis is conducted on CEE11 subsample—and spatial autocorrelation is computed on geographical weight matrices for LCont and Prox—ρ _Z are negative, while for the GCont matrix it is positive, suggesting that the role of geographical proximity in innovation activities is different in eastern and western regions. While in the EU15 subsample, innovation shows a significant level of spillovers, in the CEE11 subsample, innovation hot spots are located far apart and their relations are mediated by innovation hubs (i.e., coordinators) located in the EU15 (as shown by the ρ _Z computed on the relational weight matrix).

These results are in line with the descriptive statistics already presented and hints at the existence of a differential role played by FP6 JRCs in the context of EU15 where they are channels of intentional knowledge barter exchange and in EU26 sample where knowledge flows mostly from western coordinators toward eastern participants.

The above results were obtained mainly on the basis of Ex Post sample selection criteria; further insights can be obtained by considering an Ex Ante selection criterion. In particular, for EASTWARD networks both relational and geographical ρ _Z are positive and significant. Also, almost all independent variables display similar behaviors to the EU15 subsample; however, the coefficient of the dummy CEE11 variable loses its significance when relational weights matrices are used.

WESTWARD networks display completely opposite behaviors when relational proximities are considered. ρ _Z are negative and significant for almost all relational matrices, signaling that JRCs coordinated by CEE11 institutions and financed by FP6 tend to connect eastern coordinators located in not very innovative regions, to western participants located in highly innovative regions. Thus, WESTWARD networks act more as a knowledge diffusion tool between old and new EU members than as a collaborative environment involving similarly performing regions.²⁰

Stage 3: We have shown that, at least in some cases, regional innovative performance is influenced by geographical and relational neighboring regions. We now move a step forward by testing the joint effect of these two weight matrices and limiting the analysis to those matrices whose ρ show the existence of significant effects of (both geographical and relational) autocorrelation. We estimate a Lacombe (two weights SAR, i.e., equation 6) model including both weight matrices (a geographical matrix and a relational matrix described by layouts D₁, D_L, and D_N ) as follows:

\begin{array}{l} P A T_{i}^{t} = β_{0} + β_{1} B i z R D_{i}^{τ} + β_{2} G o v R D_{i}^{τ} + β_{3} U n i R D_{i}^{τ} + β_{4} D U M M Y C E E 11 \\ + β_{5} P R O D_{i}^{τ} + β_{6} I N N_{i}^{τ} + β_{7} A C C E S S_{i}^{τ} + β_{8} B E T W_{i}^{τ} + ρ_{D} W_{R E L} P A T_{i}^{t} \\ + ρ_{G} W_{G E O} P A T_{i}^{t} + ∊_{i}^{τ}, \end{array}

1ter

where ρ _DW_RELPAT^t _i and ρ _GW_GEOPAT^t _i represent respectively the spatial lags of the dependent variable for both the relational and geographical weight matrices, and $∊_{i}^{τ}$ is the disturbance.

Although the main objective of the paper is to investigate the internal structure of scientific and technological knowledge flows within the regional networks activated in Europe by the FP6 and to test the relevance of knowledge spillovers via geographical and relational channels for enhancing regional innovative performance, the sign and significance of coefficients of the covariates is also interesting (see Table A5).²¹ The values and significance of the coefficients are quite similar to those in Table A4: BizRD is almost always the only research input displaying positive effects and, therefore, positively influencing regional innovative activity; the effects of PROD and INN are in line with the results from stage 2 and suggest that scientific and technological specialization is positively related to innovative performance in the advanced EU15 regions.

An exception is the BETW variable estimated for EU15, which shows negative and significant effects. This result²² can be explained if we consider this variable as signaling (as suggested within a bibliometric context by Leydesdorff 2007) the “degree of interdisciplinarity” of the regional scientific and technological institutions (e.g. universities, research institutions, firms, etc.)²³ in western regions.

Finally, the values of the spatially lagged dependent variables, ρ _G and ρ _D are the most interesting for this article. The analysis performed on EU15 shows that while ρ _D are positive and significant only if geography is taken into account through the local contiguity matrix, the ρ _G values are much more stable.

If the analysis is performed on EASTWARD networks, we find that when geography and relations are considered together, ρ _D is positive and significant if geography is measured as local and global contiguity matrices. This does not apply if geography is measured as proximity. In WESTWARD networks, the result of negative and significant coefficients on the different relational ρ, obtained in stage 2, is confirmed.

Table 4 summarizes the ρs coefficients obtained from the empirical analysis, based on Ex Ante and Ex Post sample selection criteria, and the sample and subsamples analyzed.

Table 4.

Synoptic Table (SAR and Lacombe Models) ρ Coefficients.

			Ex Post				Ex Ante
			EU26	EU15		CEE11	Pure EU15	Pure CEE11	EASTWARD		WESTWARD
SAR	ρ geo	LCont	0.157***	0.188***		−0.325***	0.185***	−0.330***	0.157***		0.157***
		GCont	0.272***			0.362*			0.272***		0.272***
		Prox	0.259***	0.278***		−0.440*	0.264***	−0.469***	0.259***		0.249***
	ρ rel	A₁						SEM			−0.139**
		A_F						SEM			−0.111*
		A_L						SEM			−0.137**
		A_N						SEM			−0.135**
		B₁									−0.112*
		B_F
		B_L						SEM			−0.106*
		B_N									−0.109*
		C₁						SEM			−0.148**
		C_F						SEM			−0.124**
		C_L						SEM			−0.141**
		C_N						SEM			−0.143**
		D₁		0.315**				SEM	0.213**
		D_F						SEM	0.188**
		D_L		0.314*				SEM	0.209**
		D_N		0.313*				SEM	0.212**
				ρ rel	ρ geo				ρ rel	ρ geo	ρ rel	ρ geo
Lacombe	A₁	LCont									−0.162**	0.146***
	A_F	LCont									−0.134**	0.149***
	A_L	LCont									−0.158**	0.145***
	A_N	LCont									−0.160**	0.145***
	C₁	LCont									−0.188***	0.145***
				ρ rel	ρ geo				ρ rel	ρ geo	ρ rel	ρ geo
	C_F	LCont									−0.163**	0.149***
	C_L	LCont									−0.178***	0.144***
	C_N	LCont									−0.180***	0.144***
	D₁	LCont		0.325*	0.171***				0.177*	0.144***
	D_L	LCont		0.336*	0.169***				0.180*	0.144***
	D_N	LCont		0.334*	0.170***				0.179*	0.144***
	A₁	GCont									−0.143**	0.238***
	A_F	GCont									−0.115*	0.246***
	A_L	GCont									−0.142**	0.239***
	A_N	GCont									−0.143**	0.238***
	C₁	GCont									−0.176**	0.231***
	C_F	GCont									−0.149**	0.239***
	C_L	GCont									−0.168**	0.233***
	C_N	GCont									−0.170**	0.233***
	D₁	GCont							0.178*	0.244***
	D_L	GCont							0.181*	0.244***
	D_N	GCont							0.181*	0.244***
	A₁	Prox									−0.165**	0.227***
	A_F	Prox									−0.134**	0.230***
	A_L	Prox									−0.162**	0.226***
	A_N	Prox									−0.163**	0.226***
	C₁	Prox									−0.188***	0.238***
	C_F	Prox									−0.159**	0.242***
	C_L	Prox									−0.178***	0.238***
	C_N	Prox									−0.180***	0.238***

Note: Probabilities reported in parenthesis. Significance level: ***1 percent, **5 percent, and *10 percent. SEM = spatial error model; SAR = spatial autoregressive model; EU = European Union; CEE = Central and Eastern European countries.

Stage 4: In order to compare the different ρ values (recorded in Tables A4 and A5) computed in different nonnested models, following Burnham and Anderson (2002), we compute Akaike weights, prob_j , as defined in equation (6).

In Tables 5 –7, the main diagonal displays the estimated spatial lag coefficients for SAR specifications (1bis) with one weight matrix, while the off-diagonal values are the estimated spatial lag coefficients for the Lacombe model (1ter) where two geographic and relational matrices are considered jointly. In Table 5, weighted average column allows comparison of ρs across different matrices.

Table 5.

Akaike Weighted Average for Comparing ρ in Nonnested Models.

EU15, n = 171
	LCont	D₁	D_L	D_N	Weighted average
LCont	0.188***	0.171***	0.169***	0.170***	0.188
D₁	0.325*	0.315**			0.315
D_L	0.336*		0.314*		0.314
D_N	0.334*			0.313*	0.314

Note: Main diagonal values are ρ coefficients displayed in Table A4 (model 1bis), while off-diagonal values are ρ coefficients displayed in Tables A5 (model 1ter). Significance level: ***1 percent, **5 percent, and *10 percent. EU = European Union.

Table 6.

Akaike Weighted Average for Comparing ρ in Nonnested Models.

EASTWARD, n = 225
	Lcont	Gcont	D₁	D_L	D_N	Weighted average
LCont	0.157***		0.144***	0.144***	0.144***	0.158
GCont		0.272***	0.244***	0.244***	0.244***	0.272
D₁	0.177*	0.178*	0.213**			0.213
D_L	0.180*	0.181*		0.209**		0.209
D_N	0.179*	0.181*			0.212**	0.212

Table 7.

Akaike Weighted Average for Comparing ρ in Nonnested Models.

WESTWARD, n = 225
	Lcont	Gcont	Prox	A₁	A_F	A_L	A_N	C₁	C_F	C_L	C_N	Weighted average
Lcont	0.157***			0.146***	0.149***	0.145***	0.145***	0.145***	0.149***	0.144***	0.144***	0.158
Gcont		0.272***		0.238***	0.246***	0.239***	0.238***	0.231***	0.239***	0.233***	0.233***	0.272
Prox			0.259***	0.227***	0.230***	0.226***	0.226***	0.238***	0.242***	0.238***	0.238***	0.241
A₁	−0.162**	−0.143**	−0.165**	−0.139**								−0.139
A_F	−0.134**	−0.115*	−0.134**		−0.111*							−0.111
A_L	−0.158**	−0.142**	−0.162**			−0.137**						−0.137
A_N	−0.160**	−0.143**	−0.163**				−0.135**					−0.135
C₁	−0.188***	−0.176**	−0.188***					−0.148**				−0.148
C_F	−0.163**	−0.149**	−0.159**						−0.124**			−0.124
C_L	−0.178***	−0.168**	−0.178***							−0.141**		−0.141
C_N	−0.180***	−0.170**	−0.180***								−0.143**	−0.143

It should be noted that, in accordance with the log-likelihood values, the Akaike weights suggest that the contemporaneous inclusion of two weights matrices (one relational, the other geographical) does not improve the goodness of fit, thus, the weighted average presented in Table 5 is almost the same as the coefficient of a simple SAR model in which only one weight matrix (either relational or geographical) is computed.

These results show that, when the EU15 subsample based on an Ex Post criterion is considered, the effect of relational autocorrelation is stronger than the only surviving geographical effect (measured by local contiguity), thus, confirming that the effectiveness of FP6 as a knowledge exchange infrastructure is limited to western and more longstanding EU members. However, this result should be complemented by the evidence obtained from analyzing the two complementary networks based on application of an Ex Ante sample selection criterion. For EASTWARD networks, while local contiguity is less relevant, the coefficients of global contiguity are higher than those of relational proximity (irrespective of any weight link). For WESTWARD networks, the effect of geography is positive and significant while the coefficients of relational autocorrelation are negative and significant. This apparently puzzling result can be explained by considering the incentive structure of a coordinator located in a CEE11 region which is looking for potential participant in the EU15 regions in order to exchange knowledge (as in a star configuration, i.e., A in Figure 4) or “import” knowledge (as in a bottom-up configuration, i.e., C in Figure 4). Again, the formal asymmetry of the knowledge flows needs to be compensated by a higher than average probability of being financed,²⁴ suggesting the specificity of this policy induced network (low performers in the East connected to high performers in the West), which explains why the ρs coefficients of the relational components are negative.

Conclusion

Regional innovation activity is a complex phenomenon in which several forces are at play. A KPF, which relates regional innovative inputs to regional innovative output, should take into account the effects of both geographical and relational proximities. In this article, following Maggioni et al. (2007, 2014) and Maggioni and Uberti (2009), we modeled relational proximity in terms of FP6 research contracts, as a measure of interregional intentional knowledge exchange among research institutions and geographic proximity in terms of contiguity, geodesic, and linear proximity, as a measure of unintended knowledge spillovers.

As in Maggioni et al. (2014), we went further than identifying the presence of autocorrelation. The methodology used allowed us to identify which knowledge flow structure is more effective for the relational autocorrelation of innovative performance at regional level and employed a spatial econometric specification to consider the joint effects of geographical and relational autocorrelation (Lacombe 2004), and a statistical procedure to allow us to compare ρ across nonnested models (Burnham and Anderson 2002).

Note, however, that our research methodology may suffer from the following limitations. First, regional participation in FP6 JRCs does not account for every possible channel of scientific and technological knowledge transfer; second, the available data (on funded JRCs) did not allow us to disentangle “spontaneous” changes occurring in the networking strategies of individual research institutions from “policy-induced” changes in networking behaviors resulting from the EU selection process. We used data on FP6 and extended the analysis to 225 regions in the EU26, including the recent CEE11 members. The results show that both geographical and relational autocorrelations influence the innovative performance of European regions, which confirms previous results.

However, the contemporaneous inclusion of two weight matrices did not improve the goodness of fit of the estimation which may be better represented by a SAR model in which geography and relations are considered separately. However, the value of the Akaike weighted average of ρ for different weight links, for the EU15 sample, shows that the N and L weight links outperform the 1 weight links, thus confirming the role played by opportunity costs and budget constraints, when considering the optimal size of a research network. This result does not hold if we consider only EASTWARD networks: global continuity is very relevant followed by relational proximity computed according to either 1 and N weight links. WESTWARD networks are characterized by an even more peculiar pattern of a positive value of geographic ρs and negative values of relational ρs.

Second, in contrast to the results based on FP5, the only network layout which significantly relates the innovative performance of different regions is the hierarchical nonsymmetrical layout, a Top-down D structure, in which knowledge flows, through a hub and spoke structure, from coordinators to participants. This may be explained by the inclusion in the full sample of fifty-four regions belonging to the CEE countries whose level of technological and scientific knowledge (as measured both by innovative inputs and output levels) is lower compared to the EU15, and by the implicit rules of FP6, which indirectly encourage the inclusion in JRC of newly accessed countries.

These results suggest, on the one hand, that intentional knowledge exchanges mainly follow hierarchical network structures, most probably for efficiency reasons and, on the other hand, that the Eastward enlargement of the EU has generated a significant shift in innovative policies from FP5 to FP6. The change can be summarized as a shift from sustaining the excellence of already developed regions, to the fostering of regional cohesion, thereby encouraging greater inclusion of less scientifically and technologically advanced regions compared to previous FPs.

Our results show that geography matters in establishing scientific networks. Innovative activity, proxied by patent intensity, is influenced more by the private sector (BizRD) than by government and university R&D, and by the innovative regional specialisation (INN) in high-tech sectors.

CEE11 subsample is different. Their innovative activity is still limited and geographical spillovers are virtually nonexistent. FP6 JRCs act as long-distance channels for information diffusion moving EASTWARD, rather than virtual infrastructures for effective knowledge exchanges.

The use of an Ex Ante sample selection criterion in order to select different geographical samples further highlights the specific behavior of WESTWARD and EASTWARD research networks. In the first case, western coordinators transfer scientific and technological knowledge to eastern participants (and are probably compensated by a higher probability of being selected), which confirms structure D; in the second case, the eastern coordinator imports knowledge from western participants (which are compensated by a similar mechanism to the one described above), which results in a specific network pattern marked by a high degree of heterogeneity in which low innovating regions are tied to highly innovative regions.

If regional and national cohesion between old and new EU members is an explicit target of FP6, then the evidence shows it has been achieved, and a political geography effect is at work. However, it is yet difficult to test whether this result has been achieved at the expense of stronger innovative performance from core regions.

More data (on nonfinanced projects) and field experiments/research are needed to confirm the “compensation mechanism” at work and, possibly, to measure not only the effectiveness but also the efficiency of FPs as innovation policy instruments.

Footnotes

Authors’ Note

Previous versions of this paper have been presented in workshops and conferences in Barcelona, Utrecht, Milan and Pécs. The authors wish to thank the participants for their comments and observations. S. Beretta, E. Marrocu, S. Usai, and two anonymous referees, helped in making this paper better. Usual caveats apply.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: The authors acknowledge the support from UCSC D3.1 Research Funds.

Notes

Supplemental Material

The online appendices are available at .

References

Acs

Z. J.

Anselin

Varga

. 2002. “Patents and Innovation Counts as Measures of Regional Production of New Knowledge.” Research Policy 31:1069–85.

Anselin

1988. Spatial Econometrics: Methods and Models. Norwell, MA: Kluwer Academic Publisher.

Anselin

2010. “Thirty Years of Spatial Econometrics.” Papers in Regional Science 89:3–25.

Arenas

Cabrales

Danon

Díaz-Guilera

Guimerà

Vega-Redondo

. 2010. “Optimal Information Transmission in Organizations: Search and Congestion.” Review of Economic Design 14:75–93.

Audretsch

D. B.

Feldman

. 1996. “R&D Spillovers and the Geography of Innovation and Production.” American Economic Review 86:641–52.

Audretsch

D. B.

Huelsbeck

Lehmann

E. E.

. 2012. “Regional Competitiveness, University Spillovers, and Entrepreneurial Activity.” Small Business Economics 39:587–601.

Autant-Bernad

LeSage

. 2010. “Quantifying Knowledge Spillovers Using Spatial Econometric Tools.” Journal of Regional Science 51:427–652.

Autant-Bernad

Billand

Frachisse

Massard

. 2007. “Social Distance versus Spatial Distance in R&D Cooperation: Empirical Evidence from European Collaboration Choices in Micro and Nanotechnologies.” Papers in Regional Science 6:495–519.

Autant-Bernad

Fadairo

Massard

. 2013. “Knowledge Diffusion and Innovation Policies Within the European Regions: Challenges Based on Recent Empirical Evidence.” Research Policy 42:196–210.

10.

Azagra-Caro

J. M.

Pontikakis

Varga

. 2013. “Delocalization Patterns in University–Industry Interaction: Evidence from the Sixth R&D Framework Programme.” European Planning Studies 21:1676–701.

11.

Balland

P. A.

2012. “Proximity and the Evolution of Collaboration Networks: Evidences from R&D Projects Within the GNSS Industry.” Regional Studies 46:741–56.

12.

Barigozzi

Fagiolo

Garlaschelli

. 2010. “The Multi-network of International Trade: A Commodity-specific Analysis.” Physics Review E 81:046104.

13.

Bode

2004. “The Spatial Pattern of Localised R&D Spillovers: An Empirical Investigation for Germany.” Journal of Economic Geography 1:43–64.

14.

Borgatti

S. P.

2005. “Centrality and Network Flow.” Social Networks 27:55–71.

15.

Boschma

R. A.

2005. “Proximity and Innovation. A Critical Assessment.” Regional Studies 39:61–74.

16.

Boschma

R. A.

Frenken

. 2009. “The Spatial Evolution of Innovation Networks: A Proximity Perspective.” In The Handbook of Evolutionary Economic Geography, edited by Boschma

Martin

, 120–35. Cheltenham, UK: Edward Elgar.

17.

Bottazzi

Peri

. 2003. “Innovation and Spillovers in Regions: Evidence from European Patent Data.” European Economic Review 47:687–710.

18.

Braunerhjelm

Feldman

. 2006. Cluster Genesis. The Origins and Emergence of Technology-based Economic Development. Oxford, UK: Oxford University Press.

19.

Breschi

Cusmano

. 2004. “Unveiling the Texture of a European Research Area: Emergence of Oligarchic Networks under EU Framework Programmes.” International Journal of Technology Management 27:747–72.

20.

Breschi

Lissoni

. 2004. “Knowledge Networks from Patent Data: Methodological Issues and Research Targets.” In Handbook of Quantitative Science and Technology Research: The Use of Publication and Patent Statistics in Studies of S&T Systems, edited by Moed

Glänzel

Schmoch

, 613–43. Berlin, Germany: Springer.

21.

Breschi

Lissoni

. 2009. “Mobility of Inventors and Networks of Collaboration: An Anatomy of Localised Knowledge Flows.” Journal of Economic Geography 9:439–68.

22.

Bresnahan

Gambardella

Saxenian

. 2001. “‘Old Economy’ Inputs for ‘New Economy’ Outcomes: Cluster Formation in the New Silicon Valleys.” Industrial and Corporate Change 4:835–60.

23.

Broekel

Hartog

. 2013. “Explaining the Structure of Inter-organizational Networks Using Exponential Random Graph Models.” Industry and Innovation 20:277–95.

24.

Burnham

Anderson

D. R.

. 2002. Model Selection and Inference: A Practical Information-theoretic Approach. New York: Springer-Verlag.

25.

Callander

Plott

C. R.

. 2005. “Principles of Network Development and Evolution: An Experimental Study.” Journal of Public Economics 89:1469–95.

26.

Caloghirou

Ioannides

Vonortas

N. S.

. 2004. “Research Joint Ventures: A Survey in Theoretical Literature.” In European Collaboration in Research and Development: Business Strategies and Public Policies, edited by Caloghirou

Vonortas

N. S.

Ioannides

, 20–35. Cheltenham, UK: Edward Elgar.

27.

Caloghirou

Protogerou

Vonortas

. 2011. “Does Exposure to University Research Make a Difference? Evidence from Europe.” Workshop on Academic Entrepreneurship and Economic Development, San Sebastian, Spain, September 8–9.

28.

Cantner

Meder

. 2007. “Technological Proximity and the Choice of Cooperation Partner.” Journal of Economic Interaction and Coordination 2:45–65.

29.

Cassi

Plunket

. 2014. “Proximity, Network Formation and Inventive Performance: In Search of the Proximity Paradox.” Annals of Regional Science 53:395–422.

30.

Cassi

Zirulia

. 2008. “The Opportunity Cost of Social Relations: On the Effectiveness of Small Worlds.” Journal of Evolutionary Economics 18:77–101.

31.

Corrado

Fingleton

. 2012. “Where Is the Economics in Spatial Econometrics?” Journal of Regional Science 52:210–39.

32.

Cowan

Jonard

. 2003. “The Dynamics of Collective Invention.” Journal of Economic Behaviour and Organizations 52:513–32.

33.

Cowan

Jonard

. 2004. “Network Structure and the Diffusion of Knowledge.” Journal of Economic Dynamics and Control 28:1557–75.

34.

Deltas

Karkalakos

. 2013. “Similarity of R&D Activities, Physical Proximity, and R&D Spillovers.” Regional Science and Urban Economics 43:124–31.

35.

Elhorst

2010. “Applied Spatial Econometrics: Raising the Bar.” Spatial Economic Analysis 5:9–28..

36.

Elhorst

2014. Spatial Econometrics. From Cross-sectional Data to Spatial Panels. Heidelberg, Germany: Springer.

37.

EPEC (European Policy Evaluation Consortium). 2011. Understanding the Long Term Impact of Framework Programme, Final Report to the European Commission DG Research. Accessed September 28, 2012. https://ec.europa.eu/research/evaluations/pdf/archive/other_reports_studies_and_documents/long_term_impact_of_the_fp.pdf

38.

Espon. 2013. Espon Database. Accessed September 27, 2012. http://www.espon.eu/main/Menu_ScientificTools/

39.

EC (European Commission). 2002. The Sixth Framework Programme in Brief. Accessed September 28, 2012. https://ec.europa.eu/research/fp6/pdf/fp6-in-brief_en.pdf

40.

European Commission—CORDIS. 2012. Sixth Framework Programme Database. Brussels, Belgium: European Commission.

41.

Eurostat. 2013a. Series: Regions-science and Technology. Accessed September 25, 2012. http://epp.eurostat.cec.eu.int

42.

Eurostat. 2013b. Structural Business Survey. Accessed September 27, 2012. http://epp.eurostat.cec.eu.int

43.

Fagiolo

2010. “The International-trade Network: Gravity Equations and Topological Properties.” Journal of Economic Interaction and Coordination 5:1–25.

44.

Fagiolo

Reyes

Schiavo

. 2007. “International Trade and Financial Integration: A Weighted Network Analysis.” Quantitative Finance 10:389–99.

45.

Fischer

M. M.

Varga

. 2003. “Spatial Knowledge Spillovers and University Research: Evidence from Austria.” Annals of Regional Science 37:303–22.

46.

Florax

R. J. G. M.

Folmer

Rey

S. J.

. 2003. “Specification Searchers in Spatial Econometrics: The Relevance of Hendry’s Methodology.” Regional Science and Urban Economics 33:557–79.

47.

Glaeser

E. L.

Kallal

Scheinkman

Shleifer

. 1992. “Growth in Cities.” Journal of Political Economy 100:1126–52.

48.

Goeree

J. K.

Riedl

Ule

. 2009. “In Search of Stars: Network Formation among Heterogeneous Agents.” Games and Economic Behaviour 67:445–66.

49.

Goyal

2007. Connections: An Introduction to the Economics of Networks. Princeton, NJ: Princeton University Press.

50.

Goyal

Joshi

. 2003. “Networks of Collaboration in Oligopoly.” Games and Economic Behaviour 43:57–85.

51.

Greunz

2003. “Geographically and Technologically Mediated Knowledge Spillovers between European Regions.” Annals of Regional Science 37:657–80.

52.

Grilliches

1979. “Issues in Assessing the Contribution of Research and Development to Productivity Growth.” Bell Journal of Economics 10:92–116.

53.

Hoekman

Frenken

Oort

. 2009. “The Geography of Collaborative Knowledge Production in Europe.” Annals of Regional Science 43:721–38.

54.

Hoekman

Scherngell

Frenken

Tijssen

. 2013. “Acquisition of European Research Funds and Its Effect on International Scientific Collaboration.” Journal of Economic Geography 13:23–52.

55.

Jackson

M. O.

2008. Social and Economic Networks. Princeton, NJ: Princeton University Press.

56.

Jackson

M. O.

Wolinski

. 1996. “A Strategic Model of Social and Economic Networks.” Journal of Economic Theory 71:44–74.

57.

Jacobs

1964. The Death and Life of Great American Cities. Harmondsworth, UK: Penguin and Cape.

58.

Jacobs

1969. The Economy of Cities. New York: Random House.

59.

Jacobs

1984. Cities and the Wealth of Nations: Principles of Economic Life. New York: Random House.

60.

Jaffe

A. B.

Henderson

Trajtenberg

. 1993. “Geographic Localization of Knowledge Spillovers as Evidenced by Patent Citations.” Quarterly Journal of Economics 108:557–98.

61.

Lacombe

D. J.

2004. “Does Econometric Methodology Matter? An Analysis of Public Policy Using Spatial Econometric Techniques.” Geographical Analysis 36:105–18.

62.

Lata

Scherngell

. 2013. “Towards an Integrated European Research Area? Findings from Eigenvector Filtered Spatial Interaction Models Using European Framework Programme Data.” Papers in Regional Science 92:555–77.

63.

LeSage

Pace

R. K.

. 2009. Introduction to Spatial Econometrics. Boca Raton, FL: CRC Press/Taylor and Francis Group.

64.

Leydesdorff

2007. “Betweenness Centrality as an Indicator of the Interdisciplinarity of Scientific Journals.” Journal of the American Society for Information Science and Technology 58:1303–19.

65.

Maggioni

M. A.

2002. Clustering Dynamics and the Location of High-tech Firms. Heidelberg, Germany: Springer Verlag.

66.

Maggioni

M. A.

2004. “The Dynamics of Open Source Software Communities and Industrial Districts: The Role of Market and Non-market Interactions.” Revue d’Économie Industrielle 107:127–50.

67.

Maggioni

M. A.

Breschi

Panzarasa

. 2013. “Multiplexity, Growth Mechanisms and Structural Variety in Scientific Collaboration Networks.” Industry and Innovation 20:185–94.

68.

Maggioni

M. A.

Nosvelli

Uberti

T. E.

. 2007. “Space vs. Networks in the Geography of Innovation: A European Analysis.” Papers in Regional Science 86:471–93.

69.

Maggioni

M. A.

Nosvelli

Uberti

T. E.

. 2014. “Does Intentional Mean Hierarchical? Knowledge Flows and Innovative Performance of European Regions.” Annals of Regional Science 53:453–85.

70.

Maggioni

M. A.

Uberti

T. E.

. 2007. “Inter-regional Knowledge Flows in Europe: An Econometric Analysis.” In Applied Evolutionary Economics and Economic Geography, edited by Frenken

, 230–55. Cheltenham, UK: Edward Elgar Publishing Company.

71.

Maggioni

M. A.

Uberti

T. E.

. 2009. “Knowledge Networks across Europe: Which Distance Matters?” Annals of Regional Science 43:691–720.

72.

Maggioni

M. A.

Uberti

T. E.

. 2011. “Networks and Geography in the Economics of Knowledge Flows.” Quality and Quantity 45:1031–51.

73.

Maggioni

M. A.

Uberti

T. E.

Usai

. 2011. “Treating Patents as Relational Data: Knowledge Transfers and Spillovers across Italian Provinces.” Industry and Innovation 18:39–67.

74.

Marrocu

Paci

Usai

. 2013a. “Productivity Growth in the Old and New Europe: The Role of Agglomeration Externalities.” Journal of Regional Science 53:418–42.

75.

Marrocu

Paci

Usai

. 2013b. “Proximity, Networks and Knowledge Production in Europe.” Technological Forecasting and Social Change 80:1484–98.

76.

Marrocu

Paci

Usai

. 2014. “The Complementary Effects of Proximity Dimensions on Knowledge Spillovers.” Spatial Economic Analysis 9:9–30.

77.

Maurseth

Verspagen

. 2002. “Knowledge Spillovers in Europe: A Patent Citations Analysis.” Scandinavian Journal of Economics 104:531–45.

78.

Miguélez

Moreno

. 2013. “Research Networks and Inventors’ Mobility as Drivers of Innovation: Evidence from Europe.” Regional Studies 47:1668–85.

79.

Mora

Moreno

. 2010. “Specialisation Changes in European Regions: The Role Played by Externalities across Regions.” Journal of Geographical Systems 12:311–34.

80.

Moreno

Paci

Usai

. 2005. “Spatial Spillovers and Innovation Activity in European Regions.” Environment and Planning A 37:1793–12.

81.

Morone

Taylor

. 2004. “Knowledge Diffusion Dynamics and Network Properties of Face to Face Interactions.” Journal of Evolutionary Economics 14:327–51.

82.

OECD (Organization for Economic Cooperation and Development). 2013. Science, Technology and Patents. Accessed September 26, 2012. http://stats.oecd.org/Index.aspx?QueryId=33210

83.

Opsahl

Agneessens

Skvoretz

. 2010. “Node Centrality in Weighted Networks: Generalizing Degree and Shortest Paths.” Social Networks 32:245–51.

84.

Opsahl

Panzarasa

. 2009. “Clustering in Weighted Networks.” Social Networks 31:155–63.

85.

Paci

Usai

. 2000. “Technological Enclaves and Industrial Districts. An Analysis of the Regional Distribution of Innovative Activity in Europe.” Regional Studies 34:97–104.

86.

Paci

Usai

. 2009. “Knowledge Flows Across the European Regions.” Annals of Regional Science 43:669–90.

87.

Picci

2010. “The Internationalization of Inventive Activity: A Gravity Model Using Patent Data.” Research Policy 39:1070–81.

88.

Ponds

Van Oort

Frenken

. 2007. “The Geographical and Institutional Proximity of Research Collaboration.” Papers in Regional Science 86:423–44.

89.

Ponds

Van Oort

Frenken

. 2010. “Innovation, Spillovers and University-industry Collaboration: An Extended Knowledge Production Function Approach.” Journal of Economic Geography 10:231–55.

90.

Protogerou

Caloghirou

Siokas

. 2010. “Policy-driven Collaborative Research Networks in Europe.” Economics of Innovation and New Technology 19:349–72.

91.

Protogerou

Caloghirou

Siokas

. 2013. “Research Networking and Technology Fusion through EU-funded Collaborative Projects.” Science and Public Policy 40:576–90.

92.

Scherngell

Barber

M. J.

. 2009. “Spatial Interaction Modelling of Cross-region R&D Collaborations: Empirical Evidence from the 5th EU Framework Programme.” Papers in Regional Science 88:531–46.

93.

Scherngell

Barber

M. J.

. 2011. “Distinct Spatial Characteristics of Industrial and Public Research Collaborations: Evidence from the Fifth EU Framework Programme.” Annals of Regional Science 46:247–66.

94.

Sebestyén

Varga

. 2013. “Research Productivity and the Quality of Interregional Knowledge Networks.” Annals of Regional Science 51:155–89.

95.

Swann

G. M. P.

Prevezer

Stout

. 1998. The Dynamics of Industrial Clustering. Oxford, UK: Oxford University Press.

96.

Torre

Gilly

J. P.

. 2000. “On the Analytical Dimension of Proximity Dynamics.” Regional Studies 34:169–80.

97.

Usai

2011. “The Geography of Inventive Activities in OECD Regions.” Regional Studies 45:711–31.

98.

Varga

Pontikakis

Chorafakis

. 2014. “Metropolitan Edison and Cosmopolitan Pasteur? Agglomeration and Interregional Research Network Effects on European R&D Productivity.” Journal of Economic Geography 14:229–63.

99.

Vega-Redondo

2007. Complex Social Networks. Cambridge, UK: Cambridge University Press.

100.

Wasserman

Faust

. 1994. Social Network Analysis: Methods and Applications. Cambridge, UK: Cambridge University Press.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

1.01 MB