Specialization, Diversity And Geographical Diffusion Of Knowledge

1. Specialization, Diversity And Geographical Diffusion OfKnowledge Corinne AUTANT-BERNARD, University of Saint-Etienne James P. LESAGE, University of Texas XREAP Workshop Barcelona - July 1st, 2008

2. OUTLINE • Motivations • Model • Empirical implementation • Conclusion

3. Regions where firms are specialized in a particular industry should benefit from increasing returns and produce more innovation output Knowledge externalities are enhanced by cross-fertilization and complementarities between firms MOTIVATIONS (1) Industrial structure and knowledge diffusion: Theoretical background • The question of the intra- or inter-sectoral origin of knowledge externalities is crucial for understanding innovation mechanisms and the dynamic of growth and localization that derive from them (Duranton and Puga, 2001, Duranton 2006) • Widespread debate, in growth analysis: Marshall-Arrow-Romer (MAR) vs Jacobs externalities

4. MOTIVATIONS (2) Industrial structure and knowledge diffusion: Empirical background • A number of empirical studies distinguish between the sectoral origin of technological externalities (Jaffe 1986, Glaeser et al. 1992, Audretsch and Feldman 1999, Rosenthal and Strange 2001, Fung and Chow 2002, Henderson 2003, Greunz 2004…) • Mostly conclude that both kinds of externality exert a significant impact on productivity growth or innovation • But these studies do not account for the spatial dependence due to knowledge spillovers and potential omitted variables

5. The aim of the study is to provide a model likely to test the spatial as well as the technological dimension of knowledge spillovers. MOTIVATIONS (2) Industrial structure and knowledge diffusion: Empirical background • A few recent studies incorporate a measure of technological similarity between regions in conjunction with spatial proximity measures (LeSage, Fischer and Scherngell 2007, Parent and LeSage 2008) • Conclude that both kinds of proximity matter • But, because these studies did not use sample data that include industry-specific information, they cannot provide a great deal of insight into the nature of interindustry versus interregional influences of innovation.

6. MODEL (1) A spatial Durbin model: We start from a traditional knowledge production function which takes the following form: (1) where the vector Y represent a (logged) vector of N regions innovation ouput, across T time periods and M industries. X represents the (logged) vector of observable/measurable private and public R&D inputs. X* denotes unmeasured inputs that are excluded from the explanatory variable set in X. If both the measurable variable X and the unmeasurable excluded variable X* exhibit spatial dependence, then a spatial regression relationship will result.

7. MODEL (1) A spatial Durbin model: Let the spatial autoregressive processes in (2) and (3) govern spatial formation of observable and unobservable inputs to the knowledge production process. (2) (3) (4) u,  and  are zero mean, constant variance disturbance terms. W is an N by N spatial weight matrix reflecting the connectivity structure of the regions. The scalar parameters  and  reflect the strength of spatial dependence in X and X*.

8. MODEL (1) A spatial Durbin model: Replacing with (2) and (3) in equation (1), we find: (5) That can be rewritten: (6)

9. MODEL (1) A spatial Durbin model: In this spatial Durbin model, calculation of the response of innovation measured by industry u patents in region i, Yiu to the vth knowledge production input in region j from the matrix of inputs X, e.g., Yiu=xjv will differ from conventional non-spatial regression models. Changes in R&D inputs in regions i impact directly on Yi but also indirectly, through its effect on Yj. Therefore, the coefficients estimated from eq (1) cannot be used directly to assess the impact of R&D inputs and WY on innovation. Pace and LeSage (2007a) have proposed scalar summary measures for the n by n matrix of impacts arising from changes in the x¡th explanatory variable on the dependent variable vector representing regional innovation Y. We rely on this approach to assess the direct and indirect (spillovers effects) of changes in R&D inputs.

10. MODEL (2) Spatial Tobit model: In our exploration we use industry-specific patents granted as the dependent variable, we observe zero counts of industry-specific patents granted in regions during some time periods. Specifically, in our sample of size TMN = 9, 306, we have 1,243 cases of zero patents, or 13.35% of the sample. For this reason, we rely on a Bayesian spatial Tobit regression model (LeSage, 2000) that accommodates zero observations of the dependent variable, treating them as latent observations reflecting negative utility associated with the patenting decision. The zero-valued observations are sampled conditional on the non-zero values in y, and the other parameters of the model.

11. Empirical implementation (1) Data: • 11 industries over 9 years and 94 French regions • Patents granted as dependent variable • Private R&D expenditure • Public R&D proxied by scientific publications • Turnover to control for economic size of an industry in each region • Time- and industry-fixed effects

12. Empirical implementation (2) Direct, indirect and total effect estimates

13. Empirical implementation (2) Direct, indirect and total effect estimates The direct effects measure how industry specific patenting activity at the regional level responds to intra-sectoral R&D activity versus inter-sectoral R&D activity. This should be the focus of regional policy makers.

14. Empirical implementation (2) Direct, indirect and total effect estimates The indirect effects reflect spatial spillovers effects arising from intra- and inter-industry private and public research. This should be the focus of national policy makers.

15. Empirical implementation (2) Direct, indirect and total effect estimates The total effets estimates give the overall impact of private and public R&D activity, from both a regional and a national perspective. Addind the total effects estimates from private and public R&D actvity suggests that Jacobs externalities are twice MAR when taking both own and other effects into account.

16. Empirical implementation (3) The spatial profile of indirect impacts The impacts decay to zero as we move farther in space. The profile of decay is very rapid, with the impacts on second order neighbors falling to less than 1/6 that found for the first-order neighbours.

17. CONCLUSION • Methodological contribution: • Bayesian spatial Tobit regression model to deal with the fact that our space-time panel regional industry-level patenting activity contains zero observations (LeSage 2000) • Correct interpretation of the direct and indirect effects that arise from changes in own- and other-industry R&D activity on regional patenting activity (LeSage and Pace 2007).

18. CONCLUSION • New insights on MAR vs Jacobs externalities • All types of research activities, private, public, and own- as well as other-industry activity have a positive effect on industry-level patenting activity at the regional level • The largest direct and indirect effects are associated with private R&D activity that spills across industry boundaries, supporting Jacobs externalities • However, Jacobs externalities decrease more drastically with distance than MAR externalities, pointing to different optimal strategies for regional versus national officials

19. Specialization, Diversity And Geographical Diffusion OfKnowledge Corinne AUTANT-BERNARD, University of Saint-Etienne James P. LESAGE, University of Texas XREAP Workshop Barcelona - July 1st, 2008