Micro-behaviors and structural properties of knowledge networks: toward a ‘one size fits one’ cluster policy

ABSTRACT

The economic returns of cluster policies have been recently called into question. Based on a “one size fits all” approach consisting in boosting R&D collaborations and reinforcing network density, cluster policies are suspected to have failed in reaching their objectives. The paper proposes to go back to the micro foundations of clusters in order to disentangle the links between the long run performance of clusters and their structural properties. We use a simple agent-based model to shed light on how individual motives to build knowledge relationships can give rise to emerging structures with different properties, which imply different innovation and renewal capacities. The simulation results are discussed in a micro-macro perspective, and motivate suggestions to reorient cluster policy guidelines towards more targeted public-funded incentives for R&D collaboration.

KEYWORDS:

• cluster policy
• networks
• micro-behaviours
• structural properties
• simulation

JEL CLASSIFICATION:

• B52
• O32
• R12
• Z13

View correction statement:

Corrigendum

Notes

1. Other papers have also found positive effects of small worlds on individual performance (Verspagen and Duysters 2004; Uzzi and Spiro 2005; Schilling and Phelps 2007).

2. However, the work of Fleming, King, and Juda (2007) fails in finding such evidence.

3. Scale-free networks echo one of the forgotten results of Milgram experiments in small-world analysis: the role of super-connectors.

4. Some authors have shown that thresholds in the small worldness may exist (Uzzi and Spiro 2005).

5. If this condition cannot be matched, the disrupted relationship is not recreated and the tie definitively dies.

6. Density of a network refers to the ratio between the existing number of ties and the number of potential ties in a network.

7. Along each simulation step, entry and rewire mechanisms are simultaneously activated, but in a cumulatively unbalanced proportion. On the one side, the number of entries decreases as the population comes close to its maximum threshold (M). Therefore, preferential or random attachment mechanisms play only few times in each step. On the other side, the number of rewired links, defined as proportion λ of the existing population, increases when the population comes close to M. Consequently, the number of times bridging and closure are activated becomes higher. Since this unbalance is reproduced at each step, the effect of entry mechanisms tends to smooth, while the effect of rewire tends to be enforced.

8. The results presented are the average values of these 20 runs for each parameter setting.

9. When β goes from 0.9 to 1, the increase in clustering coefficient accelerates. Once again, the splitting of the network of several components explains this pattern.

10. When rewiring strategies are not considered (λ = 0, not displayed here), no triangles can appear and so clustering is 0.

11. Degree distribution is computed in absolute terms, so higher (lower) values mean more (less) hierarchical networks.

12. As for density and reachability, the effect of β on degree distribution also exhibits a trend reversal when above a very high level of closure, the network splits into several components.

13. Recall that assortative networks are characterized by positive degree correlation. Disassortative networks are characterized by negative degree correlation.

14. The same limitation concerns the locational attributes. The model could be extended introducing a distinction between local and non-local nodes in networks in order to study how and for what purpose organizations shape knowledge relationships with local and non-local organizations. See Fitjar and Rodríguez-Pose (2011) and Balland, Suire, and Vicente (2013).