When Do Global Pipelines Enhance the Diffusion of Knowledge in Clusters?

abstract

Recent studies have stressed the role played by global pipelines in fostering the growth of clusters and innovativeness. In this article, we develop a formal model to investigate when global pipelines contribute to an increase in local knowledge, depending on various characteristics of clusters such as size, knowledge endowment, and the ease of transmission of internal knowledge. This model is an extension of C owan and J onard’s () model in which we introduce the concept of cluster and a role for spatial proximity in the diffusion of knowledge. Our results reveal that there is a natural tendency of actors within global pipelines to act as external stars, rather than gatekeepers of knowledge. Global pipelines are beneficial for the accumulation of knowledge only if the cluster is either characterized by a high-quality local buzz or is small and weakly endowed in terms of knowledge.

• knowledge gatekeepers
• clusters
• knowledge diffusion

Acknowledgments

The authors thank the editor and the referees for their comments. The paper has benefitted from the remarks received at the DRUID Summer Conference 2008 and in seminars at KITeS, Bocconi University and URU, Utrecht University. Financial support from Progetto Alfieri–Fondazione CRT is gratefully acknowledged.

Notes

Notes

However, it should be acknowledged that multinationals could locate in clusters to acquire local knowledge instead of sharing their own knowledge with local firms (Tan and Meyer ).

Coherently with our definition of a cluster as composed of specialized firms, the categories of knowledge are those relevant to the specific industry present in the cluster.

Representing knowledge as a vector of real numbers is clearly a simplification with respect to the appreciative theorizing on knowledge specificities and properties (Foray ). However, this is common in other models, being functional to consider the average level of knowledge in the cluster as a simple measure of a cluster’s performance.

Spatial proximity enters the model in two different ways: the probability of forming a link within a cluster is greater than the probability of forming a link between clusters and the effect of local buzz captured by α _W ≥ α _B. It is worth stressing that these two effects of spatial proximity do not simply reinforce each other; they have different consequences on knowledge exchanges, as shown in our results.

If an agent can learn in several categories, learning takes place in randomly chosen (with uniform probability) categories.

Implicitly, it is assumed that the access to knowledge has a value in itself and therefore that knowledge is not considered an input of an innovation process through recombination.

For a further discussion of knowledge specialization, see footnote 14. We thank an anonymous referee for having raised this point.

These are the same values as in Cowan and Jonard ().

For a discussion of the use of regression analysis in simulation experiments, see Kleijnen ().

The assumption that parameters are independent is introduced to explore more efficiently the whole parameter space in the econometric exercise. In the real world, we could expect that at least some of them are indeed correlated. For instance, the level of “social capital” in a cluster is likely to be positively correlated with p _EEW, p_ENEW, and p _NENEW. If that is the case, it would imply that some vectors p are more likely to be observed than others, which does not affect our results, however.

By construction, p _EEB * α _W and p _EEB * α _B are highly correlated, which creates a problem for both variables being in the same regression. For this reason, we consider three variants for each model, including each interaction in turn, as well as the two interactions simultaneously (all the regressions are available on request). In each model, the formulation chosen is the one for which (1) the variable included is statistically significant when included in isolation, (2) the magnitude of the coefficient for the included variable varies less than the coefficient of the excluded variable when the other variable is added, and (3) the adjusted R ² is the largest.

A particular situation, not considered in our model, is when experts are subsidiaries of multinational companies. In this case, a negative impact of global pipelines can be explained referring to a “branch plant effect” (Phelps ). Multinational branches extract local knowledge, which is conveyed to their headquarters located outside the cluster. This is an increasingly common phenomenon among multinationals from emerging countries that invest in developed countries with the primary aim of accessing strategic resources. For a recent study on Chinese companies investing in Italian clusters see, Pietrobelli, Rabellotti, and Sanfilippo ().

Since the direct and the mediated effects go in opposite directions, there are values of α _W for which the two are offset so that the marginal effect of external connections on the performance of a cluster is negligible. Solving -0.84 + 1.5α_W= 0, we find that in a neighborhood of α_W = 0.56 the marginal effect of p_EEB is statistically insignificant, with the size of the neighborhood depending on the chosen level of significance. We thank an anonymous reviewer for having raised this point.

As we discussed earlier, we can expect external connections to be more beneficial (or even unambiguously beneficial) if clusters are significantly specialized in a set of knowledge fields that are relevant for a specific sector. To estimate the effect of knowledge specialization, we ran a numerical experiment assuming that Cluster 1 is endowed with experts in knowledge categories 1, 2, and 3, and Cluster 2 is endowed with experts in categories 3, 4, and 5. In line with expectations, we found that the coefficient for p _EEB is positive (although not statistically significant), while the coefficient of the interaction term p _EEB *α _W is positive and statistically significant. Moreover, we found that when clusters are specialized, the importance of within-cluster interactions is reduced (p _ENEW and p _NENEW are no longer significant); in fact, the diffusion of local knowledge is not sufficient when the cluster lacks knowledge in some relevant categories. The estimation results are available on request.

The coefficient of p _NENEB is larger in both Models 2 and 3 than in Model 1 because in the former two models, the dependent variable is the average knowledge of a single cluster (Cluster 2). Therefore, when knowledge is exchanged between agents belonging to different clusters, the increase in the knowledge of the agent outside the cluster does not have an impact on the dependent variable. Clearly, such an effect is not present in Model 1, in which the dependent variable is the average knowledge of the whole economy.

The direct effect of the size of a cluster on the diffusion of knowledge, mentioned earlier, does not lie at the core of our analysis. However, our model reproduces the positive impact the size of a cluster on the diffusion of knowledge shown in the empirical literature. By comparing the average knowledge of cluster 1 (the large “cluster”) and cluster 2 (the small “cluster”) in Scenarios B and D, we found that in Scenario B, the cluster 1/cluster 2 ratio of average knowledge is 0.727 at t = 0, while it increases up to 0.980 at t = 100000; for Scenario D, the values are, respectively, 0.980 and 1.389.

Competition can also involve local firms in the cluster and foreign global corporations. While competition may have direct negative consequences, we observed that these firms are not likely to be involved in the barter of knowledge even without competition, as exemplified by the coefficient of p _ENEB .