Geography Reference
In-Depth Information
imity, and showed that a mixture of low and high proximity was best for i rms. This has
also been done in the case of geographical proximity: some have suggested a mixture of
local and non-local linkages to be best for i rms, and a combination of local buzz and
global pipelines to be best for the long-term evolution of clusters (Bathelt et al., 2004).
The second strategy is to classify all relations along a continuum and to assess the success
of each particular relation separately. Then, by testing its ef ect and its quadratic ef ect,
one can assess whether an optimal level of proximity exists (the linear ef ect should then
be positive and the quadratic ef ect should be negative). For example, making use of
patent data, Gilsing et al. (2007) assessed the ef ect of technological distance between
i rms in alliance networks in high-tech industries on the exploration innovative perform-
ance of i rms. As expected, they found an inverse U-shaped function between technologi-
cal distance and exploration.
4. Networkdynamics
A key empirical insight from studies on networks, be it in the context of innovation and
knowledge production or in other contexts, holds that networks have very pronounced
structures (Newman, 2003). What we mean by structure is that the set of links between
nodes in a network are very dif erent from the properties of a random network, that is,
the properties one obtains by randomly connecting nodes to create a network structure.
Structured (or 'organized') networks require a true explanation, while random networks
can simply be 'explained' stochastically.
Random networks are characterized by two important features. First, the degree
distribution follows a normal distribution, where the degree of a node stands for the
number of links of a node. Since a random network is constructed by assigning links
between two randomly selected nodes, the degree of nodes will follow a normal distribu-
tion. Second, in a random network, there is no clustering: the probability of two nodes
being linked is totally independent of whether these two nodes are linked indirectly via
a third node. These two properties of random networks - normal degree distribution
and absence of clustering - are never observed in social networks or inter-i rm networks.
Empirically, one typically observes that the degree distribution is skewed, with few nodes
having a high degree and many nodes having a low degree. Apparently, some nodes are
more 'popular' to link with than other nodes. And, one observes that clustering is a very
signii cant phenomenon ('friends of friends are often friends with one another'). That is,
many nodes participate in triangle relationships. Yet, some nodes do so much more than
other nodes. The extent to which a node is clustered can be indicated by the number of
triangles divided by the number of possible triangles.
At the level of single nodes, these observations lead to two questions: (1) How can one
explain dif erences in the degree of nodes? and (2) How can one explain the clustering
of nodes? Below, we discuss these features using the concepts of preferential attachment
and closure, respectively. Then we propose an industry lifecycle perspective on network
evolution and regional lock-in.
Preferential attachment
One key conceptual breakthrough in the study of dynamic networks has been the paper
by Barabasi and Albert (1999). In this paper, the authors start from the observation that
many networks are characterized by scale-free degree distributions where degree stands
