Innovation Clusters in Technological Systems: A Network Analysis of 15 OECD Countries for the Mid‐1990s

Abstract

The paper aims to investigate how innovations cluster in different technological systems (TSs) when their “techno‐economic”, rather than “territorial” space, is considered. Innovation clusters of economic sectors are identified by referring to the innovation “potential” represented by their R&D expenditure and by applying social network analysis to the intersectoral R&D flows matrices of 15 OECD countries in the mid‐1990s. Different clusterization models are first tested in order to detect the way sectors group on the basis of the embodied R&D flows they exchange. Actual clusters are then mapped in the different TSs by looking for intersectoral relationships which can be qualified to constitute “reduced TSs” (ReTSs). In all the 15 TSs investigated the techno‐economic space appears organized in hierarchies, along which its constitutive sectors grouped into clusters with different density and composition. Once ReTSs are looked for, the 15 TSs display highly heterogeneous structures, but with some interesting similarity on the basis of which different clusters of TSs can be identified in turn.

Keywords:

• Innovation clusters
• technological systems
• R&D expenditure
• embodied innovation flows

Acknowledgements

Although the authors contributed equally in the development of this paper, and share the attribution of Sections 1 and 5, Sections 2, 4.2.1 and 4.2.3 could be attributed to Sandro Montresor, whereas Sections 3, 4.1 and 4.2.2 to Giuseppe Vittucci Marzetti. The authors would like to thank Bo Carlsson, Fulvio Castellacci and the other participants to the 2007 DRUID Summer Conference (Copenhagen, 18–20 June 2007), Maria Semitiel Garcia and the other participants to the 2007 Input–Output Conference (Istanbul, 2–6 July 2007), for their useful comments. The authors are particularly indebted to two anonymous referees for their suggestions on an earlier version of the paper. The authors acknowledge funding from the Italian Ministry for Education University and Research (MIUR)—PRIN 2005 on “Fragmentation and Local Development”. Giuseppe Vittucci Marzetti also gratefully acknowledges support for this research from the PAT (Provincia Autonoma di Trento, post‐doc scholarship 2006). The usual disclaimers apply.

Notes

1. This way of measuring embodied innovation flows suffers from a number of limitations, which range from the hypothesis that the whole R&D of a certain sector turns into actual innovations completely and instantaneously, to the hypothesis of a total embodiment of the same innovations in production flows, which excludes the presence of disembodied flows. Although problematic, these limitations can however be accommodated on the basis of both theoretical and practical arguments on which see, for example, Marengo and Sterlacchini (1990), Papaconstantinou et al. (1996) and Leoncini and Montresor (2003a).

2. It should be emphasized here that, although the Leontief inverse matrix can be seen as the sum of an infinite sequence of distinct production rounds (), the direct and indirect relationships that we capture through R should not be meant to represent a dynamic framework. In fact, as there is no “time” or propagation errors/lags within the input–output framework we rely on, the production rounds cannot be seen as “real”, but rather as purely logical. Although some attempts have been recently made to make the temporal/propagation dimension explicit (e.g. Dietzenbacher and Romero, 2007; Duchin and Levine, 2007), their practical implications are still diffuse and more relevant for other kinds of truly dynamic studies.

3. This technique tries to derive the properties of a network from the analysis of its local structures, most notably its sub‐networks of two (dyads) and three nodes (triads). In particular, the MAN analysis refers to the number of mutual (M), asymmetric (A) and null (N) dyads in each triad, along with that of the ties’ direction, whenever there is more than one triad with a given number of mutual, asymmetric and null dyads (Wasserman and Faust, 1994: Ch. 14). For an overview of the balance‐theoretic theory, in which these models were first proposed, see Wasserman and Faust (1994: Ch. 6).

4. In formal terms, in model (i) the candidate network structure is made up of no more than two “cliques”, where a clique is a maximal complete sub‐network containing three vertices or more. Model (ii) relaxes the assumption of two cliques of the first one and allows networks to consist of more cliques. A different version of model (iii), called transitivity model (Holland and Leinhardt, 1971), refers to transitivity for triples of distinct vertices and includes one type of triad in addition to those admitted in ranked clusters models. Although slightly more general than model (ii), it has been omitted from the conceptual description as it does not add substantial specifications. However, for the sake of completeness, it will be considered in the test of the present empirical application.

5. The SAD method is grounded on Davis and Leinhardt’s idea of ranked clusters models (1972). In formal terms, given a directed graph G = (V, E), that is, a set of vertices (or nodes or points) (V) and a set of ordered pairs of these vertices specifying their relations (E), a strong component of G can be formally defined as the maximal subset of vertices such that there is a directed path within P from x to y for each pair . A path from x ₁ to x ₂ is a graph M = (Z, F) of the form:

where x ₁,x ₂,…,x_k are all distinct (Diestel, 2005).

6. As usual, i′ denotes a (1×n) unit row vector. On the different methodologies for obtaining normalized R matrices, and on their different implications, see Montresor and Vittucci Marzetti (2007b).

7. In order to reduce aggregation biases as long as possible, matrix calculations have been carried out at the maximum attainable level of disaggregation and the results re‐aggregated as required only afterwards.

8. Defined as the ratio between the actual number of cells of C (s_C ) which are greater than t, and the maximum number of cells of the same matrix (n×(n−1)), D_t actually informs us about the role of t in shaping the structure of the correspondent TS‐network The TS’s arcs are in fact valued by the cells of C itself. In formal terms, D_t = s_C / (n(n−1)) with 0⩽D_t⩽1.

9. All the calculations were made using Pajek 1.09 (Batagelj et al., 2005).

10. By definition, the densities of the compared TSs are all exactly equal, to 1 and 0, respectively, only for the extreme cut‐offs (min and max) of the threshold distribution. In the neighbourhoods of these cut‐offs, however, the TSs degenerate in meaningless structures, either too dense to be investigated (around the minimum cut‐off), or too spare to have a techno‐economic meaning (around the maximum). Therefore, the best one can do is to look heuristically for density similarities, possibly pairwise or across groups of distributions, by referring to an appropriate intermediate cut‐off value. For t = 0.0111 we have been able to detect the maximum degree of similarity which is possible to have far from the extremes: see Montresor and Vittucci Marzetti (2007a) for detailed calculations.

11. In figure 3, SCs are identified by groups of shaded nodes while, whenever present, cliques have been “boxed” in a rectangle. Of course, the density of a TS and the number of its ReTSs are somehow related: a higher density implies a larger number of flows above the cut‐off, and thus a greater probability to have relationships which are able to fuel SCs in a network. However, this relationship is just potential and does not necessarily turn into actual. The illustrative comparison between the French and the Japanese graphs, with exactly the same density, confirms this point.

12. The three exceptions with more than one ReTS—that is, Australia (figure 3(a)), Czech Republic (figure 3(c)) and Italy (figure 3(g))—“confirm” this result, as they host just two ReTSs, still a low number.

13. In 2004, the Polish R&D intensity was 0.53 per cent, just above Mexico and Slovak Republic (OECD, 2006). However, since the mid‐1990s the local authorities have been trying to implement at least the downstream part of a still prospective system of innovation. The National SME Services Network (KSU) (http://ksu.parp.gov.pl/lista.html), a group of 180 cooperating business counselling centres all over the country for supporting development of the SME sector, is the most remarkable example of these efforts.Relatively smaller are the ReTSs which can be identified in the other countries, some of which—for example, Italy (figure 3(g)), the Netherlands (figure 3(l)), Norway (figure 3(m)) and the UK (figure 3(p))—have been recognized as having a low degree of connectivity in the correspondent systems of innovation (Ørstavik and Nås, 1998; Leoncini and Montresor, 2000b).