Treating Patents as Relational Data: Knowledge Transfers and Spillovers across Italian Provinces

Abstract

The paper applies a relational perspective to patent data in order to investigate the characteristics of innovation flows within and across 103 Italian NUTS3 regions (province).

In this way it is possible to use the CRENoS database on regional patenting—built on EPO data spanning from 1978 to 2003—to investigate the scientific and technological “relations” among “invention-creating ” and “invention-adopting ” territories. In particular, patents are used as relational data connecting inventors and applicants along a dual interpretation of a “knowledge production” and a “knowledge utilization” function. In addition a gravity model is used to identify frictions and attractions of the Italian innovation system.

Analytical tools, such as social network analysis, spatial econometrics and negative binomial estimation procedures, are used to map and measure the structure and the evolution of a series of innovation sub-systems, both at territorial level (i.e. province) and at the industry level (i.e. five specific industries, chosen according to the Pavitt's taxonomy, Footwear, Textiles, Machinery, Personal Computers and Chemicals).

Keywords:

• Patents
• network analysis
• spatial econometrics
• relational data
• regional innovation system
• Italy

Acknowledgements

The authors would like to thank for useful comments Mario Nosvelli, Barbara Dettori and participants of workshops and conferences (Cagliari, Jena, Karlsruhe Sophia, Utrecht) where previous versions of this paper were presented, and two anonymous referees for useful suggestions and remarks. In this version, the paper (and in particular Section 3) has also greatly benefited from the comments of Uwe Cantner. The usual caveats apply.

Notes

¹ For complementary approaches to the study of innovation networks, see also Doloreux and Mattson (2008), Staber (2008), Love and Roper (2009) and Visser (2009).

² See also Gallie (2009) and Massard and Mehier (2009).

³ One may also note that, along with the contribution of Cantner and Guerzoni (2009), a relational approach to patents considers the demand-pull and the technology-push perspectives as well as the interaction of market-related incentive mechanisms and knowledge inner dynamics.

⁴ This subsample refers almost exclusively to freelance agents and accounts for less than 14 per cent of all links and refers obviously to intra-provincial flows.

⁵ It is worth noting that the great majority of our patents are applied either by firms or individuals and only a small fraction by research institutes, either public or private.

⁶ Information can also be provided by professional representatives, a compulsory element when applicants do not have either their residence or their principal place of business in a contracting state.

⁷ One should remember that multi-location manufacturing firms with an interprovincial diffusion in Italy (those which are under our scrutiny) are just a small quota, that is 2 per cent. Such a quota is however variable across sectors and ranges from a maximum of around 15 per cent in the petrochemical sector and a minimum of 0.5 per cent in the wood and furniture sector.

⁸ It is difficult to find an adequate measure of the commuting phenomenon across provinces for work-related reasons (see Deyle and Grupp, 2005 for an application to the German case). One way is to compare employed people who are resident in a province but not necessarily work within its borders and those who actually work in a plant within the province. The difference between these two values indicates if a province is a “net exporter” or a “net importer” of workforce. However, at the national level this difference is zero. One way to assess the phenomenon is, therefore, to take the absolute value of all such differences and to compute the average. We find that the average difference, at the national level, is just 8 per cent of the total number of employees.

⁹ See footnote 3.

¹⁰ The use of functional units, such as local labour systems, would be more adequate, but unfortunately most economic and social data are gathered exclusively at the administrative level (NUTS3).

¹¹ CRENoS (Centre for North–South Economic Research) has a long tradition in the construction of databases on innovation in Italy and Europe (see Paci and Usai, 2000b).

¹² Usually for each patent there is one applicant, but there could be some exceptions with more than an applicant.

¹³ To identify these manufacturing sectors we used conversion tables from IPC classification to NACE. The procedure is based on Evenson et al. (1991), who developed a concordance table between patent and industrial classifications, commonly refereed to as Yale Technology Concordance (YTC). This concordance was later refined by Johnson (2002).

¹⁴ These totals are not the same because we are not limiting our analysis to those patents which have either an applicant resident in Italy (when inventors are computed) or an inventor resident in Italy (when applicants are counted). The fact that the equivalent number of patents in Italy is higher when counting inventors rather than applicants suggests that an important quota of Italian inventions are transformed into patents by foreign agents and firms.

¹⁵ This total is lower than the two counts used for the KUF and the KPF because the flow of the matrix is defined considering only patents which are invented and applied by Italian residents.

¹⁶ In particular, we selected two categories, Prof. 2 and Prof. 3.1 according to ISCO 1988.

¹⁷ Betweenness centrality is normally calculated as the fraction of shortest paths between node pairs that pass through the node of interest. For a detailed discussion of different measures of network centrality (see Freeman, 1979).

¹⁸ We selected Ateco sectors DL30, DL32, DL33.

¹⁹ See, for a recent application to European regions, Moreno et al. (2005).

²⁰ For a definition of the QAP procedure, which compares two matrices and computes the significance of their similarity through a large number of random permutations of rows and columns, see Krackhardt (1988); for its application to the network of innovators, see Cantner and Graf (2006).

²¹ We compute these correlations for original networks and for binary ones, and results are confirmed.

²² Self-loops quota is 81 per cent in Footwear, 72 per cent in Textiles and 75 per cent in Machinery.

²³ In fact the PC self-loops quota is equal to 56 per cent, while the Chemicals one is even smaller, at 50 per cent.

²⁴ Density, d, in a digraph is calculated as , where L indicates the total number of links present in the network and n the number of nodes. This value ranges between 0 (empty network) and 1 (full network).

²⁵ More formally, centralization value is calculated as where is the centrality value of the most central region in the system, C ⁱ _g is the degree of centrality of generic node i indicating the number of direct links, and the denominator reflects the maximum level of centrality obtainable in a system of n regions (see Freeman, 1979).