Papers in Evolutionary Economic Geography
Knowledge networks in the Dutch aviation industry:
the proximity paradox
Tom Broekel and Ron Boschma
Knowledge networks in the Dutch aviation industry:
the proximity paradox
Tom Broekel and Ron Boschma1
Department of Economic Geography, Faculty of Geosciences, Utrecht
University, The Netherlands
The importance of geographical proximity for interaction and knowledge sharing has been
discussed extensively in economic geography. There is increasing consensus that it is one out of
many types of proximities that might be relevant. We argue that proximity may be a crucial
driver for agents to connect and exchange knowledge, but too much proximity between these
agents on any of the dimensions might harm their innovative performance at the same time. In a
study on knowledge networks in the Dutch aviation industry, we test this so-called proximity
paradox empirically. We find evidence that the proximity paradox holds to some degree. Our
study clearly shows that cognitive, social and geographical proximity are crucial for explaining
the knowledge network of the Dutch aviation industry. But while it takes cognitive, social and
geographical proximity to exchange knowledge, we found evidence that proximity lowers firms’
innovative performance, but only in the cognitive dimension.
Keywords: geographical proximity, knowledge networks, proximity paradox, social network
analysis, aviation industry
JEL Codes: R11, R12, O18, O33
1 The authors would like to thank Matté Hartog for his help.
In Economic Geography, few issues have been studied more frequently as the question of what
role geographical proximity plays for knowledge sharing and innovation. Backed by the
argument that the exchange of tacit knowledge requires face-to-face contacts, it has long been
emphasized that knowledge sharing is highly sensitive to geographical distance (Audretsch and
Feldman, 1996; Gertler, 2003). This view on the role of geographical proximity has recently
been challenged theoretically (see e.g., Boschma, 2005; Lagendijk and Oinas, 2005; Broekel and
Binder, 2007). This critical view has been initiated by the French school of proximity dynamics
(Rallet, 1993; Kirat and Lung, 1999; Rallet and Torre, 2005). Their critical voices particularly
emphasize that geographical proximity is just one dimension among a number of other proximity
dimensions that can explain interaction between geographically proximate actors.
Boschma (2005) proposed five dimensions of proximity that impact on the likelihood of
knowledge exchange between actors and their innovative performance. His claim is that
geographical proximity is neither a necessary nor a sufficient condition for inter-organizational
learning and innovation. Boschma (2005) also argued that geographical proximity is more likely
to become effective rather indirectly through the other types of proximity. Breschi and Lissoni
(2003) and Ponds et al. (2007), among others, have confirmed this empirically for social,
institutional and cognitive proximity.
Extending these ideas, in a recent paper, Boschma and Frenken (2009) introduced what
they describe as the so-called proximity paradox. While proximity may be a crucial driver for
agents to connect and exchange knowledge, too much proximity between these agents on any of
the dimensions might harm their innovative performance. So, while a high degree of proximity
may be considered a prerequisite to make agents connected, proximity between agents does not
necessarily increase their innovative performance, and may possibly even harm it. Following
Nooteboom’s work on optimal cognitive distance (Nooteboom, 2000), Boschma and Frenken
(2009) claim it depends on the (optimal) level of proximity whether a connection between agents
will lead to higher innovative performance or not.
This issue of the proximity paradox is put central in an empirical study on the knowledge
network of the Dutch aviation industry. The Dutch aviation industry is an interesting case,
because it lost its flagship, the Fokker Company, in 1996, after which it went through a major
restructuring and reorientation process. The question then is how the knowledge network looks
like in the post-Fokker period, and what are its main drivers. Among other things, we will test
whether a shared past in the Fokker company (as a proxy for social proximity) increased the
probability of two aviation firms to connect. This study draws on own data that were collected
through semi-structured interviews of 59 profit and non-profit organizations that are active in
manufacturing activities and engineering services in the Dutch aviation sector.
Our paper has two objectives. The first objective is to assess empirically the extent to
which the different forms of proximity affect the technical knowledge network in the Dutch
aviation industry. Employing social network analysis, our study confirms the importance of
cognitive, organizational, and social proximity for the structure of the technical knowledge
network. We also found geographical proximity to be a driver of network formation, even when
controlling for the other proximities. The second objective is to determine which proximities
determine the innovative performance of aviation firms, while controlling for the usual suspects.
Our study provides empirical evidence for the proximity paradox with respect to the cognitive
dimension, not the geographical and social dimension. That is, proximity is required to connect
firms, but it does not necessarily yield above average innovative performance of these firms.
The paper is structured as follows. In Section 1, the different dimensions of proximity are
discussed. We specify how they influence the likelihood that actors are linked and what that
means for their innovative performance. Section 2 provides a short description of the Dutch
aviation industry, the data and the variables we constructed. Section 3 will briefly present the
methodological tools (QAP and network autocorrelation regression) we employed. Results of the
analyses are presented and discussed in Section 5. Section 6 concludes.
Firms’ embeddedness in knowledge networks has increasingly been recognized as an important
determinant of their economic and innovative performance (see, e.g., Powell et al., 1996). Given
limited resources firms can invest into research and development, their ability to collaborate and
make use of external knowledge becomes crucial for their success. It is well known that firms
have different absorptive capacities, which matter for their usage of external knowledge (Cohen
and Levinthal, 1990). This determines not only their likelihood to engage in knowledge sharing
but also the likelihood that obtained knowledge can be successfully used and implemented.
It is widely accepted that, in addition to their absorptive capacity, other factors influence
economic actors’ decisions to become engaged into knowledge sharing activities. An argument
frequently put forward in the literature is that geographical proximity facilitates knowledge
transfer (Feldman and Florida, 1994). This view on the role of geography for knowledge
exchanges has recently been challenged with the argument that it is not mere co-location that
matters for knowledge exchanges, but membership in knowledge networks (Castells, 1996). In
this sense “geographical proximity only creates a potential for interaction, without necessarily
leading to dense local relations” (Isaksen, 2001, p. 110).
The French school of proximity dynamics (see e.g. Rallet and Torre, 1999; Rallet and
Torre, 2005) has played a prominent role in this debate. They claim that geographical proximity
is just one among a number of proximity dimensions. In this context, Boschma (2005) claimed
that geographical proximity is neither a necessary nor a sufficient condition for knowledge
sharing and innovation. He proposed five dimensions of proximity (cognitive, social,
organizational, institutional, and geographical) that may impact on the likelihood of knowledge
exchange between actors and their innovative performance. We will briefly discuss each of these
below (for an extensive treatment, see Boschma, 2005), with the exception of institutional
proximity, for which we have no variance in our study.
Cognitive proximity refers to the degree of overlap between two actors concerning their
knowledge bases. Actors need to have a sufficient absorptive capacity to identify, interpret and
exploit knowledge of other actors (Cohen and Levinthal, 1990). However, if two actors’
knowledge bases are too similar, the likelihood of an innovative recombination is lower than
when dissimilar knowledge bases are merged. According to Nooteboom (2000), there exists a
tradeoff “…. between cognitive distance, for the sake of novelty, and cognitive proximity, for the
sake of efficient absorption” (p. 152). The relationship between the cognitive distance between
two actors and their innovation performance is therefore to be expected to take an inverted u-
shape (Cohendet and Llerena, 1997). In other words, both very proximate and very distant actors
are likely to gain little from cooperating in innovation activities. The optimal level of cognitive
proximity follows from the need to keep some cognitive distance (to stimulate new ideas through
recombination) and to secure some cognitive proximity (to enable effective communication and
knowledge transfer). Moreover, high cognitive proximity generally implies that two firms have
very similar competences, which means that when they engage in knowledge exchange, they run
a serious risk of weakening their competitive advantage with respect to the network partner. It is
also for this reason that one expects that excessive cognitive proximity may be harmful to
performance (see Nooteboom et al., 2007; Boschma et al., 2009; Boschma and Frenken 2009).
Consequently, it is not so much the quantity of contacts and intensity of knowledge exchanges
that matters for firms’ success, but rather the type of knowledge exchanged, and how that
matches the existing knowledge base of the firms. In this respect, cooperation is most fruitful
when network partners have technologically related, not similar knowledge bases.
Actors may also be close or not in organizational terms. Boschma (2005) defined organizational
proximity “… as the extent to which relations are shared in an organizational arrangement, either
within or between organizations” (p. 65). It can be seen as a continuous scale going from
autonomy to control. It is very low for totally independent actors and very high for actors that are
part of the same hierarchical system. According to Boschma (2005), organizational proximity
helps to manage knowledge exchange and reduce transactions costs. However, excessive
organizational proximity may also hamper interactive learning, as it constrains flexibility.
An alternative way to define organizational proximity is the degree to which organizations
have similar routines and incentive mechanisms (Metcalfe, 1994). In innovation studies,
researchers tend to make a distinction between profit and non-profit organizations. To put it
simply, profit organizations have an interest to keep their knowledge away from competitors,
while non-profit organizations like universities have a public mission and, therefore, are more
open to exchange knowledge with others. Because of these different routines, a profit and a non-
profit organization have a low degree of organizational proximity, which lowers their probability
to connect and collaborate. This is in line with the problematic relationship between universities
and private firms, which has been documented extensively. According to Broekel and Binder
(2007), actors’ search biases make it more likely that non-profit organizations will interact with
other non-profit organizations. This might also be true for profit-oriented organizations,
especially when the firms are not direct competitors.
Social proximity refers to the social embeddedness of actors in terms of friendship, kinship, and
experience at the micro-level (Boschma, 2005). This has to be seen as being distinct from
institutional proximity, which refers to institutions (like ethnic and religious values) at the
macro-level. Of particular interest is the role of trust, which is likely to be positively influenced
by social proximity. Trust has frequently been argued to foster knowledge exchange (Maskell
and Malmberg, 1999). In particular with respect to secrecy and the dangers of free riding, trust-
based relations are often depicted as superior to anonymous or newly established relations.
Hence, social proximity should be a strong predictor of the existence of a link between two
actors. The existence of “old boys networks”, for instance, is likely to influence knowledge
sharing activities. The same can be argued for actors with a shared history, like being at the same
school, university or company, which generates a sense of belonging to the same community.
However, too much social proximity may also be harmful for innovative performance, because
of an overload of loyalty and commitment in social relationships (Boschma, 2005).
None of the previously introduced proximities require co-location. It can be argued, though, that
geography plays a role by facilitating the other types of proximities. Broekel and Binder (2007)
have argued that geographical proximity may also directly impact on the likelihood that actors
will exchange knowledge. Through various mechanisms, geography influences individuals’
motivation and search heuristics and can bias them towards spatially close knowledge sources.
Hence, geographical proximity influences the other types of proximity but also increases the
likelihood of two actors to engage in knowledge exchange more directly. However, while
geographical proximity may offer certain advantages to knowledge sharing activities, there is
increasing evidence that a dominance of local linkages may also reduce the innovative
performance of a firm (Boschma, 2005; Broekel and Meder, 2008).
In sum, there are good reasons for why these different types of proximities should impact on
actors’ knowledge networks. Although empirical research is still primarily focused on the role of
geographical proximity (see e.g. Jaffe, 1989; Audretsch and Feldman, 1996), there is a growing
number of empirical approaches that aim to disentangle the different types of proximity and their
effects on firms’ behavior (Fleming et al., 2007; Ter Wal, 2009). Breschi and Lissoni (2003) for
example, using patent citations show that geographical proximity looses its predictive power for
two actors being linked when controlling for social proximity. Ponds et al (2007) find that
geographical proximity is of smaller relevance for research collaborations between academic
organizations, as opposed to collaborations between academic and non-academic organizations.
However, they suggest that geographical proximity is still helpful to overcome institutional
barriers between different types of organizations. Cantner and Meder (2007) find that cognitive
and social proximity are relevant for cooperation activities. They demonstrate that the
technological overlap between two actors (cognitive proximity) positively influences the
likelihood of these actors to engage in cooperation. This is also true for having cooperation
experience with the same cooperation partner. In this latter respect, past cooperation experiences
may have led to a decrease of the social distance between the two partners, which positively
impacts on future cooperation. Mowery et al. (1998) found similar results and, moreover,
suggested an inverted U-shape relationship between the probability to cooperate and the
technological (cognitive) similarity of two actors. However, while most of these studies take into
account at least two types of proximity, they hitherto have ignored or overlooked the relevance
of the other types.
The proximity paradox
Another critique on existing studies is that in such a proximity framework, it remains unclear
whether in the end, the choice of network partner matters for firms’ economic performance. In a
recent paper, Boschma and Frenken (2009) introduced what they describe as the so-called
proximity paradox. While proximity may be a driver for agents to connect and exchange
knowledge, too much proximity between these agents on any of the dimensions might not
necessarily increase their innovative performance, and may possibly even harm it.
Following Nooteboom’s work on optimal cognitive distance (Nooteboom, 2000), Boschma
and Frenken (2009) suggest it might depend on the (optimal) level of proximity whether a
connection between agents will lead to a higher innovative performance or not. While the
cognitive dimension has drawn most attention in the literature, optimal levels of proximity may
exist for the other forms of proximity as well (Boschma, 2005). With respect to geographical
proximity, it has been suggested that a mixture of local and non-local linkages to be best for
firms, and a combination of local buzz and global pipelines to be best for the long-term evolution
of clusters. With respect to social proximity, the optimal social distance might consist of a
balance between embedded relationships within cliques and strategic ‘structural hole’
relationships among cliques (Fleming et al. 2007). Uzzi (1996) found evidence of a mixture of
low proximity (defined as arm’s length ties) and high proximity (depicted as embedded ties) to
be best for firms. An optimal level of organizational proximity may be accomplished by loosely
coupled networks with weak ties between autonomous agents, which combine advantages of
organizational flexibility and coordination (Grabher and Stark 1997). However, the different
forms of proximity may also interact in this respect. Excessive proximity in one dimension that is
compensated by some degree of distance on another dimension may still enhance the innovative
performance of a firm. For instance, a firm may have primarily relationships with other local
firms (meaning too much geographical proximity), but when these provide access to a range of
different but related knowledge bases, this might still positively affect firm performance.
Our paper aims to add to this literature in two ways. First, we test the influences of four
types of proximity (i.e. cognitive, social, organizational, and geographical) on the likelihood of
two firms to exchange knowledge in the Dutch aviation industry. Following the literature on
knowledge networks (Giuliani and Bell, 2005; Boschma and Ter Wal, 2007; Sammarra and
Biggiero, 2008), we focus on the exchange of technological knowledge, which is regarded as
most relevant for firms’ innovation activities in this sector. Second, we examine whether
cognitive, social, organizational and geographical proximity matter for firms’ innovation
performance, controlling for factors like absorptive capacity. We also test whether there is a
curvilinear relationship between the proximity dimensions and innovative performance. Doing
so, we determine whether the proximity paradox holds in the Dutch aviation industry.
The aviation industry is known for being highly agglomerated and clustered in a few locations
worldwide, like Seattle and Toulouse (Hickie, 2006). Large multinational firms like Airbus and
Boeing dominate this industry. Their headquarters and main facilities function as attractors for
other businesses, e.g., specialized suppliers, sub-contractors, and service companies. This caused
the emergence of the typical hub-and-spoke type structure in this industry (Gray et al. 1996).
There is no large company in the aviation sector that has its headquarters or a major
production facility in the Netherlands. In the past, this was different. Since its establishment in
1919, Fokker B.V. dominated the industry in the Netherlands for almost eighty years. Fokker
used to be a crucial player in the aviation related knowledge networks in the Netherlands (van
Burg et al., 2008). Its core business was the production of 50-100 seater planes, most famously
the F 27, as well as the more recent Fo 50 and Fo 100. At its peak in mid 1991, about 13,000
people worked for Fokker, before employment declined steadily to 7,141 in early 1996
(Ligterink, 2001). In 1996, Fokker had to declare bankruptcy for three core units. This meant the
biggest single job cut in the modern history of the Netherlands by putting 5,600 people out of
work (Reuters, 1996). While about 950 jobs have been saved through the founding of a new
firm, “Fokker Aviation”, which was eventually taken over by Stork B.V., the Dutch aviation
industry lost its sole aircraft producer and one of its technological flagships.
More than a decade later, the Dutch aerospace industry consists of about 80 firms, most of
which are SMEs (NAG, 2008). These firms contribute to 0.9 percent (275 million Euros) of the
EU-25 value added in manufacturing of air- and spacecrafts in 2003 (EUROSTAT, 2002). Total
employment is about 5,000 employees. These numbers exclude maintenance and overhaul of
aircrafts. Including these, the number of employees increases to 15,000, which generate a
turnover of 2.2 billion Euro (NAG, 2008). The industry surely has regained strength by
successfully filling niches in the aviation market (Heerkens, 1999). Nevertheless, the Dutch
aviation sector can be regarded as marginal in comparison to countries like Germany and France.
A study of the knowledge network in the Dutch aviation sector is interesting for many
reasons. First of all, aviation is known to be a highly knowledge intensive industry, in which
access to external knowledge might be crucial (Niosi and Zhegu, 2005). Secondly, it is
interesting to investigate how this knowledge network is shaped, in the absence of a dominant
player. Thirdly, it is intriguing to find out whether the Fokker Company, despite its bankruptcy a
long time ago, still affects the nature of this knowledge network. Many of the current
entrepreneurs and top-managers in the Dutch aviation industry may have been former employees
of Fokker. Given the strong exposure to the very distinct Fokker identity (Kriechel, 2003), their
knowledge searching and sharing activities are still likely to be shaped by these experiences and
biased towards their former co-workers (see, e.g., Broekel and Binder, 2007). This enables us to
construct a social proximity indicator, which might affect the knowledge network in the industry.
Our empirical study is based on own data collection. In late 2008-early 2009, we interviewed 59
organizations that belong to the aviation industry in the Netherlands. Most of these organizations
are members of the Netherlands Aerospace Group (NAG), which is the most important trade
organization. Their members account for about 95 percent of the total turnover generated by
Dutch firms in the aviation industry (NAG, 2008). In 2008, the organization had 83 members.
We interviewed only those members that were active in manufacturing and/or engineering, since
for these activities, innovation and the exchange of technological knowledge is likely to be of
utmost importance.2 This applies to 40 firms, of which we interviewed 37. The three firms that
were not interviewed do not show any eye-catching features. In the course of the interviews, five
additional firms were named as being relevant, which were not member of the NAG, but clearly
active in the aviation industry. All of these have been interviewed as well. This increased our
total population to 45, of which 42 have been interviewed by extensive semi-structured
interviews on the spot. Accordingly, our response rate is 93 percent.
The list of the NAG also includes non-profit organizations, which we interviewed in a
different way. These organizations, as well as additional non-profit organizations named during
the firm interviews as relevant knowledge sources, were asked to indicate the intensity of
interaction with the other relevant non-profit organizations. This applied to 17 organizations
2 We interviewed also 2 firms that were not active in manufacturing or engineering. These firms confirmed the low
importance of technological exchange for their firms’ competiveness. Moreover, none of the interviewed firms
mentioned any maintenance-oriented firm as a relevant knowledge source.
which increased our sample to 59 organizations.3 The intensity level ranged from 1 to 3, with 1
indicating no interaction, and 2 and 3 medium and very high intensity.
Figure 1 shows the technological knowledge network of the Dutch aviation industry, based
on the data of these 59 profit and non-profit organizations. The 59 vertices account for 146 edges
and the network has a density of 0.085. This density is half as big as the density observed by
Sammarra and Biggiero (2009) for the technological knowledge network in the Italian aerospace
cluster of Rome. This cluster consists, however, of only 33 firms. Nevertheless, knowledge
network relations are rather sparse in the Netherlands, which is also indicated by the large
number of isolates. Non-profit organizations, most noticeably the Technical University of Delft
and the successor of the Fokker Company (Stork Aerospace B.V.), represent important contacts
for firms and thus, take up prominent positions in the network.
- Figure 1 here -
As explained in Section 2, the first part of our analysis aims to assess the impact of the various
forms of proximity on the structure of the technological knowledge network of the Dutch
aviation industry. More in particular, we estimate the importance of the different types of
proximities on the likelihood that two actors are linked. Our dependent variable LINKTEC is
dichotomous and indicates whether actor i or j mentions the other as a relevant source of
technological knowledge. We assume that knowledge exchanges are always reciprocal in nature.
This implies that if i is naming j as relevant contact, we also assume that i is a relevant contact of
j. In other words, we assume an undirected network. As independent variables, we have included
the 4 dimensions of proximity, as discussed in Section 2, and we have measured these as follows.
3 While most of the interviewed firms do not show strong connections to the space industry, this is not true for the
non-profit organizations. However, our main focus is on firms.
To assess the effect of geographical proximity, we calculate the logarithm of the geographical
distance in kilometers between two actors, which results in a continuous positive variable.4
While other studies use travel time (Ejermo and Karlsson, 2006), the spatial scale of the network
of Dutch aviation industry is rather small. Few contacts are not located in the direct neighboring
countries of the Netherlands (including Great Britain), for why the use of travel distances is
unlikely to change the results. The logarithm ensures that outliers in form of trans-continental
relations do not disturb the estimations. We also estimate a squared term to control for non-linear
effects (GEO2). In order to reduce multicollinearity problems, we subtract the mean in advance.
We refer to cognitive proximity as the technological similarity of two actors’ knowledge bases.
We have constructed two measures. For the first measure, we rely on the technology classes that
are assigned by the Netherlands Aerospace Agency (NAG). The NAG defines 15 technologies of
which 13 are relevant for the firms considered in this study. The technological fields and the
according number of organizations are listed in Table 1. In case the interviewed organization is
not a member of the NAG, the profile was created on the basis of the organization’s webpage.
The variable TECNAG is defined dichotomously, with a value of one if both organizations are
active in the same technology, and zero otherwise.
- Table 1 here -
The second variable is based on 3-digit NACE codes assigned to each organization. The
assignment of the NACE codes has been done as follows. First, if an organization named another
organization being a relevant knowledge source, it was asked to provide information about the
content of the knowledge exchange, i.e. which technologies this exchange concerned, and to
what 3-digit NACE codes this corresponded. Second, we asked each organization to mention the
three most important sectors (3-digit NACE codes) from which they recruit their key personnel.
Lastly, we searched for information about the organizations on the Internet. This included the
4 We add 0.001 km to all distances to ensure a positive logarithm.
organizations’ own websites as well as the company information webpage:
www.mintportal.bvdep.com, which also classifies firms according to the NACE scheme. This
last option was particularly relevant for organizations that have not been interviewed, but which
were named as relevant knowledge sources. Based on these sources of information, each
organization has been characterized by a number of 3-digit NACE codes.
In order to define a similarity indicator for two organizations, we first had to define the
similarity between two technologies (NACE codes). In a first step, following Breschi and Lissoni
(2003), the similarity between technologies is estimated on the basis of their co-occurrence at the
level of the organization. If technology A has frequently been assigned to organizations that are
also characterized by technology B, the technologies are perceived being related. In addition to
this direct relation, we also consider the indirect relation between two technologies. Accordingly,
if technology A is frequently assigned to the same organizations as technology C, and the same
is true for technologies B and C, A and B are also comparatively similar. In practice, we estimate
the Cosine index, as given in Ejermo (2003):
with t as the number of technologies and g,k,z as indices of technologies under consideration. In
this equation, wzk is the number with which technologies z and k coincide at the organization
level. In total, 72 different technologies appear in our data set.
When organizations have multiple technologies, we do not know the technologies’ relative
importance, i.e. no information is available on the share of turnover or employees attributed to
each technology. We therefore estimate similarity in two ways. In the first, we search for the
most similar pair of technologies in the firms’ technology vectors. More precisely, we compare
two organizations’ (i,j) vectors of technologies (Ti and Tj). Next, we identify for each technology
z (z ! ti) of organization i, the maximal rizg within organization j’s technologies. The same is done
for the technologies of organization j. The rizg are added up and divided by the sum of the
number of technologies assigned to both organization i and j. The latter ensures that the resulting
similarity index sij is not biased positively towards diversified organizations (those with many
technologies assigned to). The estimation shows the following.
Because the values of the Cosine index rzg are between 0 and 1, the similarity index ranges
from 0 and 1 as well, with 1 indicating perfect technological similarity. In extreme cases, all of
organization i’s technologies are compared to one technology of organization j. The rationale
behind this is that lacking information on the relative importance of a technology for an
organization we assume that access to just one particular technology makes two organizations
similar. This is because their knowledge bases overlap, which enables efficient communication.
In the estimations we also consider a quadratic term of the similarity indicators to check for
non-linear effects. Because these variables are likely to cause multicollinearity with the
similarity indicators, we subtract the mean of the variable before the squaring.
Hence SIM2 will be large for small and large values of similarity.
We have constructed a dichotomous variable to account for the (likely) existence of social
relations between organizations. The variable (FOK) amounts to one if former employees of
Fokker B.V. are members of the top management of both firms and zero otherwise. As pointed
out in Section 2.1, having a shared past in Fokker may reflect a community feeling that might
still affect the structure of the network, even after its collapse. The importance of Fokker for the
Dutch aviation industry has been immense in the past (see van Burg et al., 2008). Hence, “old-
boys” networks may still be in place and give exclusive knowledge sharing opportunities. The
data for this variable were collected in the interviews by means of the following question: “Do
you or anyone else of the firm's top-management have a personal relation to the former Fokker
N.V., i.e. has been a former employee of Fokker N.V.?”
We approximate organizational proximity by differentiating between profit and non-profit
organizations (universities, research institutes, associations, and trade organizations). With few
exceptions, these non-profit organizations turn out to be highly connected, and are also
frequently named by firms as important technological knowledge sources. A dichotomous
variable (PUB) is constructed being one when both organizations are universities, research
institutes, trade organizations, or associations, and zero when otherwise. In order to keep the
number of variables small, we decided to treat all these actors equally although we are aware that
they might provide very different functions. A similar variable has been constructed for
interactions between profit-oriented organizations, i.e. firms, (PRIVATE).
We have included some control variables that might also affect the likelihood of being linked in
the technological knowledge network. First, we have taken the logarithm of the absolute size
difference of organizations (SIZE), which controls for variations in the cooperation behavior
related to the size of organizations (see, e.g. Beise and Stahl, 1999; Graf, 2007). Because the
organizations in our sample are heterogeneous, we also control for that fact that some
organizations are more focused on the aviation industry than others. More precisely, the variable
(AERO) indicates if two organizations are mainly active in the aviation industry. For firms, this
implies that the share of their turnover attributed to aviation is above average. In case of other
organizations, we define them to be “dedicated” to aviation if their focus is mainly on this sector.
To create this dichotomous variable, we primarily rely on information derived from the
organizations’ websites. Lastly, organizations may also differ with respect to their openness
towards external knowledge. Two organizations that perceive external knowledge as being
highly relevant, can be expected to have a higher likelihood to be linked than two organizations
that rely more on internal knowledge. The variable OPEN therefore is defined to be one if the
relative importance of organization i and j attribute to external knowledge is above average. This
information is collected by the following question we posted during the interviews: “Please
indicate in terms of percentage the relative importance of: a) knowledge acquired inside the
company; b) knowledge acquired outside the company (adding up to 100%)”.
All variables are summarized in Table 2, which also includes some descriptives. Table 3
shows the QAP-correlations among these variables. Most variables turn out to be weakly
correlated. Hence, we can include all of them into the network regression. Only the quadratic
terms as well as the different versions of the similarity indicators show considerable correlation
with their non-squared counterparts. For this reason, these are tested separately.
- Table 2 here –
- Table 3 here –
In order to test the importance of the types of proximities on the likelihood that two actors are
linked, we employ network regression techniques. These techniques allow the use of relational
variables. Relational variables describe the “relationship” between two actors, i.e. the extent to
which they are distinct, similar, or share certain characteristics. A particular value xij (i=1…n and
j=1…n) indicates the relation between firm i and j with n as the number of observations.
With this type of data at hand, social network analysis employs linear or logit regressions.
The difference to standard OLS and logit techniques is that the dependent and independent
variables are not vectors but n*n (adjacency) matrices. For the application of the standard
regression tools, the matrices are vectorized in the sense that the columns are stringed together to
form one vector with n2 elements, i.e. the first elements are the relations of actor 1 to all others,
next are those of actor 2, and so on. In networks without loops, the diagonal of the adjacency
matrices is meaningless and is eliminated. It reduces the vector length to n*(n-1). In this paper
we treat all relations as undirected, for which the upper and lower half of the adjacency matrices
are identical. These redundant elements are cut and the number of elements is n*(n-1)/2.
The dependent variable is then regressed with a standard logit model on the independent
variables. The logit model is chosen because the dependent variable is 0/1 variable with 1
indicating the existence of a link between two actors and 0 for the absence of a link. Such
network data are, however, characterized by frequent row/column/block autocorrelation and
therefore standard tools of inference are problematic (Krackhardt 1987). A solution is provided
by the Quadratic Assignment Procedure (Hubert 1987, Krackhardt 1987, Krackhardt 1988). The
idea is to compare the estimated model statistics to the distribution of such statistics resulting
from large numbers of simultaneous row/column permutation of the considered variables (before
the vectorization). If, for example, the originally estimated positive coefficient is larger than 95
percent of the coefficients estimated from the permutated samples, it represents a significance
level of 0.05. More precise for the QAP, Dekker’s “semi-partialling plus” procedure (Dekker et.
al. 2003) is used that is implemented in the SNA package of the statistical software R. This
method is known to be more robust with respect to multicollinearity (Dekker et al. 2003).
The regression results for the technological knowledge network of the Dutch aviation industry
are summarized in Table 4. The pseudo R2 reveals that our model performs well in explaining
the network structure. More interestingly are the fraction of correctly predicted “1s” and “0s”,
which amount to 0.578 and 0.936 respectively. Hence, the model seems to be better in explaining
the absence of a link than the presence. Nevertheless, both values are considerably larger than
zero, indicating a satisfactory model fit. Overall, the findings confirm our theoretical predictions:
all types of proximity (organizational, cognitive, social and geographical) are found to impact on
the likelihood of being linked to another organization as far as technological knowledge
exchange is concerned.
- Table 4 here -
Organizational proximity is highly relevant for the existence of links whereby, in particular,
public organizations are highly connected among each other. In fact, the variable PUB is positive
and highly significant, indicating the high degree of connectedness of non-profit organizations in
the knowledge network. This also explains its very high odd ratio.5 The negative significant
coefficient of PRIVAT is not surprising, as we observe only few links between firms.
With respect to cognitive distance, we find a positive and significant impact of
technological similarity expressed by SIMNACE. While the significance level of SIMNACE is just
0.1, the odd ratio is very high. This indicates that organizations tend to link more with
technologically similar organizations, a result which confirms findings in other studies (e.g.
5 We checked for multicollinearity but the odd ratio kept its size when removing correlated variables.
Mowery et al. 1998; Sorenson and Singh, 2007; Canter and Meder, 2007). The squared term
(SIMNACE2) is insignificant, implying that the likelihood of being connected increases
monotonically with increasing levels of technological similarity. As Canter and Meder (2007),
we therefore cannot confirm an inverted-U relationship between the likelihood to cooperate and
cognitive proximity. This is also true for the SIMNAG variable.
Our study shows that social proximity (FOKKER) also influences positively the likelihood
of two organizations to exchange knowledge. So, as expected, having a shared past in the former
Fokker Company matters. Apparently, long after its bankruptcy in 1996, the Fokker Company
still affects the Dutch aviation industry: a shared past in this company seems to help persons,
now working for different organizations, to establish and maintain knowledge sharing links.
Geographical distance (DIST) has a strong and negative effect on the likelihood of having
a knowledge link between two organizations. Its squared version is insignificant, suggesting a
rather linear effect. So, even for controlling for the other forms of proximity, we found a positive
effect of geographical proximity. While this finding is as expected, it has frequently been shown
that the effect of geographic proximity is often caused by (unobserved) social relationships, i.e.
social proximity (see, e.g., Breschi and Lissoni, 2003; Ponds et al., 2007). Since our FOKKER
variable may not fully cover the social proximity dimension, it cannot be excluded that our
geographical proximity variable may still account for some effect of social proximity.
In sum, we found considerable evidence for the first part of the proximity paradox, namely
that proximity (in whatever form) is needed (or helpful) for the establishment of links in the
technological knowledge network of the Dutch aviation industry.
The second objective of the paper is to determine which proximity dimensions (while controlling
for usual suspects) determine the innovative performance of firms in the Dutch aviation industry.
The focus is now on firms’ attributes, not their relations to other firms. For this analysis, only the
42 firms are considered. As common in innovation studies (see, e.g., Sternberg and Arndt, 2001),
we approximate the innovative performance of firms by means of the share of significantly
improved products / processes on a firm’s turnover (INN).
We have constructed attribute variables approximating firms’ network positions, based on
all links reported related to technological knowledge, i.e. including knowledge sources outside
the aviation industry and outside the Netherlands. The 42 firms reported 158 links to 100
organizations inside and outside the Netherlands. As our first network variable, we described the
firm’s position in a network by the number of links it has (LINKS). This corresponds to the well-
known network measure of degree centrality (Freeman, 1979). We also accounted for the
internationalization of a firm’s ego network (FOREIGN), measured by the share of links with
organizations outside the Netherlands to the total number of links a firm has.
With respect to geographical proximity, we measured the variable DIST as the average
distance between a firm and the organizations it is exchanging knowledge with. It will also be
tested in its squared and mean subtracted form DIST2, to account for potential non-linear effects.
In order to assess the effect of cognitive proximity, we constructed a variable that
determined the similarity of a firm’s technological profile to the combined technological profiles
of the organizations it is linked to, based on the similarity measure (SIMNACE) described in
Section 3.1,. A technological similarity variable (SIM) can be defined as the sum of similarity
measures (SIMNACE) of those organizations a firm is directly connected to (geodesic distance
equal to one). This sum was divided by the firm’s total number of links, in order to reduce the
correlation with the latter. SIM thus represents the average technological similarity of a firm to
the technological knowledge base of its direct contacts. Again, subtracting the mean and
squaring this variable accounts for a possible non-linear effect (SIM2).
With respect to organizational proximity, we measured the share of contacts with non-
profit organizations (ORG) for each firm. Social proximity is accounted for by a dummy variable
(FOK), which amounts to 1 if the top management of the firm are former Fokker employees, and
zero otherwise. It has to be pointed out that this variable may also capture effects related to the
experience of the top management in this particular industry. As mentioned before, Fokker went
bankrupt in 1996, and its employees had more than a decade to collect experiences in the
We also constructed a number of control variables. First, we accounted for the absorptive
capacity of firms, which might positively impact on their innovative performance (see e.g.
Boschma and Ter Wal, 2007). This has been approximated by two variables: the share of R&D
employees in total employment (R&D), and the share of employees with at least an university
bachelor degree (SKILL). Since our dependent variable is measured as a share of turnover, we
have ensured that these independent variables are also independent of firm size. Second, we have
included the variable AERO that measures the dedication of a firm to the aviation industry (see
Section 3.1). Third, we have included firm’s age (AGE) and the number of employees to account
for any firm size effects (EMPL).
The descriptives of the variables are shown in Table 5. Since we have 42 observations in
our sample, not all explanatory variables can be considered at the same time. We check their
correlation structure first and exclude redundant variables. Table 6 shows the correlation among
the independent and dependent variables. The correlation shows that R&D is highly correlated
with firms’ innovativeness (r=0.58***). Not surprisingly is also the high correlation between
R&D and the share of employees with at least a university bachelor degree (r=0.61***). Old
firms also have higher levels of employment (r=0.51**). Even when excluding some redundant
variables, the number of variables is still too high to be included into one model. For this reason,
we define a base-line model, with which we test the influence of the other variables.
- Table 5 here –
- Table 6 here –
In order to determine which proximity dimensions (while controlling for usual suspects)
determine the innovative performance of firms in the Dutch aviation industry, we focus on firms’
attributes. For the analysis of attribute data, standard regression tools can be used. A central
assumption is however that firms exchange knowledge through networks, and for their
innovative performance it matters with whom they are connected. In other words, firms’
attributes are not independent observations, as they depend on the characteristics (attributes) of
the linked actors. This violates the assumption of independence in OLS regressions. This is even
more problematic when network measures are used as independent variables. In such settings,
standard tools of inference are not valid.
The use of so-called network autocorrelation models allows circumventing this problem
(see, e.g., Anselin, 1988; Leenders, 2002). Here the regression model is specified by
with y as the response, and X the covariance matrix. The error term nu has the usual
characteristics. W1 and W2 are defined as
and Eq. 5
with W1i and W2i as the elements of one or two network adjacency matrices. In this sense W1
and W2 describe the relationships between the actors, i.e. the technological knowledge network.
Rho1 can be regarded as an autoregression parameter (AR) that parameterizes the autoregression
of each y value on its neighbors in the network W1. In the context of this paper, this accounts for
the potential of knowledge spillovers between the actors and W1. Rho2 captures the moving
average (MA) and parameterizes the autocorrelation of each disturbance in y on its neighbors in
network W2. It accounts for an incorrect or mis-specified unit of analysis (Anselin and Bera,
1998). In addition, Rho2 may also take into account effects when certain events or shocks diffuse
through he entire network. In our setting Rho2 is of minor importance as the unit of analysis are
firms (which are well defined) and relevant knowledge is very unlikely to diffuse beyond
geodesic distances of 1. Because of the limited number of observations and the fact that Rho1
and Rho2 represent additional parameters that reduce the available degrees of freedom, we
decided to exclude Rho2. The models are therefore estimated considering only the autoregession
parameter Rho1 (W1), which corresponds to the technological knowledge network. We
nevertheless checked the autocorrelation of the residuals. It turned out that the observed network
autocorrelation in the residuals is generally low (Geary’s C 0.7 =[0.7,1.35] and Moran’s I < 0.22
at low geodesic distances).
The model is estimated by a maximum likelihood procedure using the lnam function
implemented in the SNA package of the statistical software R. Our dependent variable is
however a proportion (share of turnover) for which a standard linear regression approach is not
valid. We therefore use a logit transformation on the dependent variable:
For zero values, we follow Petrie and Sabin (2000) and set these to 1/(2*n), with n being the
number of observations. The coefficients have to be interpreted as odd ratios.
In order to test the impact of the different types of proximity on firms’ innovation performance,
we use a linear knowledge production function model. First, we define a base-line model to
explain the innovativeness of the 42 firms on the basis of the control variables R&D, EMPL,
AERO, AGE and SKILL. The results are shown in the first column of Table 7.6
- Table 7 here -
Only the share of turnover attributed to aviation (AERO) turns out to be negative and
significant: aviation oriented firms seem to be less successful in innovation. Another finding is
that R&D is insignificant. This changes, however, when we exclude the insignificant control
variables and include measures that take into account a firm’s knowledge network (Model 1).7
As expected, R&D becomes positively significant and remains so in all subsequent models,
while AERO loses its significance level. We find that the number of links a firm has (LINKS) to
be negatively related to innovative performance. We have to point out, though, that its
significance level is low, and LINKS loses its significance in subsequent model specifications.
An explanation might be that it is not sufficient to be connected to many persons, but that it
matters rather to whom one is connected. The share of international linkages does not affect the
innovative performance of Dutch aviation firms. Relations with international partners are not
more or less important than relations with national partners.
Model 2 shows however that the average geographical distance to knowledge network
partners (DIST) has a negative and significant impact on innovative performance. Including this
variable also eliminates the significance of the number of linkages. Consequently, local linkages
seem to be more beneficial to Dutch aviation firms. DIST2 is not significant, suggesting a
6 The results are based on the estimations when controlling for autocorrelation with respect to the technological
7 Note that the following models are estimated on 34 cases, which have a positive number of knowledge links, so
excluding the firms that did not indicate any knowledge links.
monotone relationship. Combined with the result in Section 3.3, it seems that firms not only
prefer local partners, but these also bring more economic benefits to the firms. In other words,
the proximity paradox does not seem to hold in purely geographical terms: geographical
proximity is both a driver of knowledge network formation and a stimulus for innovative
performance of firms through their geographically close partners. However, as Boschma (2005)
mentioned, the effect of geographical proximity may just grasp the relevance of the other types
of proximity. This is explored in the subsequent models.
In Models 4 and 5, we included the organizational proximity variable (ORG). Our analysis
clearly shows that the share of knowledge linkages a firm has with non-profit organizations does
not matter for innovative performance. In other words, while organizational proximity was a
driver behind the formation of knowledge network relationships (see Section 3.3), it does not
positively affect the innovative performance of firms. The same holds for its squared version
(ORG2). This result tends to provide some evidence for the proximity paradox.
In Model 6, the geographical distance measure is confronted with the social proximity
variable FOKKER. Both variables are significant. Having a positive effect, a firm’s
embeddedness in the “old-boys” network related to the former Fokker Company tends to
increase its innovative performance. This result was also confirmed in some of the interviews.
Consequently, the proximity paradox does not seem to hold in the case of social proximity: a
shared past in the Fokker company did not only increase the likelihood to establish a knowledge
link, but it also enhanced the effectiveness of knowledge exchange.
In Models 7 and 8, we assess the economic effect of cognitive proximity through the ego
networks of firms. In addition to the negative effect of geographical distance and the positive
influence of social proximity, the technological similarity indicator shows a negative and
significant effect. In other words, as expected, the more the technological profile of the partners
in a firm’s ego network overlap with the technological profile of the firm, the more cognitive
proximity between the firm and its partners, and the lower the innovative performance of the
firm. In Model 8, the squared term turns insignificant indicating a rather linear effect. This does
not meet our expectations of an inverted u-shape relationship. A possible explanation for this
might be found in the construction of the cognitive proximity indicator. We consider only a
subsample of all existing industries’ NACE codes (and technologies), namely those that are
linked to the aviation industry. The subsample is therefore likely to be biased towards aviation
related industries. In other words, the subsample is more related to aviation than the average of
the entire spectrum of industries and technologies. Our analysis’ negative coefficient may hence
reflect the decreasing slope of the inverted u-shape while the increasing part is not captured by
our measure. This deserves more attention in future research.
In sum, our analysis demonstrates that geographical, social and cognitive proximity matter
for the innovative performance of firms in the Dutch aviation industry. The first two impact on
innovative performance positively, while the latter has a negative effect. All three have also been
shown to positively influence network formation. In other words, the proximity paradox seems to
hold for the cognitive dimension, and to a lesser extent for organizational proximity.
In this paper, we tested the so-called proximity paradox, as proposed by Boschma and Frenken
(2009). This is about the fact that proximity is required to connect to knowledge networks, but
proximity does not necessarily yield superior innovative performance. We have collected data on
59 profit and non-profit organizations that are active in the Dutch aviation industry, and
employed QAP and network autocorrelation regression models to test this paradox. In our
analyses, we distinguished between four forms or proximity (cognitive, geographical, social and
organizational) that might impact on network formation and the innovative performance of firms.
Our analyses provided strong evidence for the first part of the proximity paradox, which
concerns the forms of proximity as drivers of knowledge network formation. We found indeed
that cognitive, organizational, geographical and social proximity between organizations
increased their likelihood to connect and exchange knowledge. Most interestingly, we also found
geographical proximity to be a driver of network formation, even when controlling for the other
proximities. The study provided, however, mixed evidence for the second part of the proximity
paradox, which concerns the effects of proximity on the innovative performance of firms while
controlling for the usual suspects. It did seem to hold for the cognitive dimension and to some
extent to the organizational dimension of proximity. However, the proximity paradox did not
hold for the geographical dimension: geographical proximity is both a driver of knowledge
network formation and a catalyst for innovative performance through local network partners. The
same is true for social proximity: it helps to establish and maintain knowledge network linkages,
but it also enables a more effective utilization of this network fostering innovative performance.
Our analyses also showed that there is an interesting relationship between cognitive and
geographical proximity. In the light of the first part, the results show that firms chose their
knowledge partner because of geographical closeness as well as technological similarity. The
second analysis rather suggests that networking with local knowledge partners as well as with
technologically distant actors is beneficial for firms’ innovation performance. Put differently,
being linked to geographically close partners that have divergent knowledge bases is likely to
increase firms’ innovation performance. This result calls for further research.
Further research is also needed concerning the effect of geographical proximity. We cannot
exclude in our analysis that geographical proximity might still capture some other effects, like
social proximity (see, Breschi and Lissoni, 2003), as the latter was quite broadly defined in our
study With the data at hand, we cannot disentangle any further these two effects. This certainly
needs to be explored in future work. Finally, we are in need of more dynamic analyses,
accounting for knowledge network formation over time, and examine how that is not only
affected by the various proximities, but also how proximities change over time due to the
evolution of networks (Balland, 2009; Boschma and Frenken, 2009; Ter Wal, 2009).
Anselin, L. (1988), Spatial Econometrics: Methods and Models, Norwell, MA: Kluwer.
Anselin, L., Bera A. (1998), Spatial Dependence in Linear Regression Models with an
Introduction to Spatial Econometrics, In: A. Ullah and D. Giles (Eds.), Handbook of Applied
Economic Statistics, 237-289, New York: Marcel Dekker.
Audretsch, D. B., Feldman, M. (1996), Spillovers and the Geography of Innovation and
Production, American Economic Review, 86:630–640.
Balland, P.A. (2009), Proximity and the Evolution of Collaboration Networks - Evidences from
R&D Projects within the GNSS Industry, Papers in Evolutionary Economic Geography 0914.
Beise, M., Stahl, H. (1999), Public Research and Industrial Innovations in Germany, Research
Boschma, R. A. (2005), Proximity and Innovation: A Critical Assessment, Regional Studies,
Boschma, R.A., Frenken, K. (2009), The Spatial Evolution of Innovation Networks. A Proximity
Perspective, In: Boschma, R.A., Martin, R. (Eds.), Handbook of Evolutionary Economic
Geography, Cheltenham: Edward Elgar, chapter 5, forthcoming.
Boschma, R. A., ter Wal, A. L. J. (2007), Knowledge Networks and Innovative Performance in
an Industrial District: The Case of a Footwear District in the South of Italy, Industry and
Boschma, R., Eriksson, R., Lindgren, U. (2009), How does Labour Mobility Affect the
Performance of Plants? The Importance of Relatedness and Geographical Proximity, Journal of
Economic Geography, 9(2):169-190.
Breschi, S., Lissoni, F. (2003), Mobility and Social Networks: Localised Knowledge Spillovers
Revisited, CESPRI Working Paper 142.
Broekel, T., Binder, M. (2007), The Regional Dimension of Knowledge Transfers - A
Behavioral Approach, Industry and Innovation, 14(2):151–175.
Broekel, T., Meder, A. (2008), The Bright and Dark Side of Cooperation for Regional Innovation
Performance, Papers in Evolutionary Economic Geography 08.11.
Cantner, U., Meder, A. (2007), Technological Proximity and the Choice of Cooperation Partner.
Journal of Economic Interaction and Coordination, 2(1):45-65.
Castells, M. (1996), The Information Age: Economy, Society and Culture Volume 1, The Rise of
the Network Society, Oxford UK: Blackwell Publishers.
Cohen, W., Levinthal, D. (1990), Absorptive Capacity: A New Perspective on Learning and
Innovation, Administrative Science Quarterly, 35(1):128–152.
Cohendet, P., Llerena, P. (1997), Learning, Technical Change and Public Policy: How to Create
and Exploit Diversity, In: C. Edquist (Eds.) Systems of Innovation, London and Washington:
Dekker, D.; Krackhardt, D.; Snijders, T.A.B. (2003), Mulicollinearity Robust QAP for Multiple
Regression, CASOS Working Paper, Carnegie Mellon University.
Ejermo, O. (2003), Patent Diversity as a Predictor of Regional Innovativeness in Sweden,
CESPRI Working Paper 140.
Ejermo, O., Karlsson, C. (2006), Interregional Inventor Networks as Studied by Patent
Coinventorships, Research Policy, 35:412–430.
Feldman, M. P., Florida, R. (1994), The Geographic Sources of Innovation: Technological
Infrastructure and Product Innovation in the United States, Annals of the Association of
American Geographers, 84(2):210–229.
Fleming, L., C. King III, A.I. Juda (2007), Small Worlds and Regional Innovation, Organization
Science, 18(6): 938-954.
Freeman, L. C. (1979), Centrality in Social Networks: Conceptual Clarification, Social
Gertler, M.S. (2003), Tacit Knowledge and the Economic Geography of Context or the
Undefinable Tacitness of Being (there), Journal of Economic Geography, 3:75-99.
Giuliani, E., Bell, M. (2005), The Micro-Determinants of Meso-level Learning and Innovation:
Evidence from a Chilean Wine Cluster, Research Policy, 34(1):47-68.
Grabher, G., Stark, D. (1997), Organizing Diversity: Evolutionary Theory, Network Analysis
and Postsocialism, Regional Studies, 31:533-544.
Graf, H. (2007), Gatekeepers in Regional Networks of Innovation, Jenaer Economic Research
Gray, M., Golob, E., Markusen, A. (1996), Big Firms, Long Arms, Wide Shoulders: The ’Hub-
and-Spoke’ Industrial District in the Seattle Region, Regional Studies, 30(7):651-666.
Heerkens, H. (1999), Dutch Industry in Holding Pattern, Interavia Business & Technology.
Hickie, D. (2006), Knowledge and Competitiveness in the Aerospace Industry: The Cases of
Toulouse, Seattle and North-west England, European Planning Studies, 14(5):687–716.
Isaksen, A. (2001), Building Regional Innovation Systems: Is Endogenous Industrial
Development Possible in the Global Economy?, Canadian Journal of Regional Science,
Jaffe, A. (1989), Real Effects of Academic Research, American Economic Review, 79(5):957-
Kirat T., Lung Y. (1999), Innovation and Proximity: Territories as Loci of Collective Learning
Processes, European Urban & Regional Studies, 6(1), 27-39.
Krackhardt, D. (1987), QAP Partialling as a Test of Spuriousness, Social Networks, 9 171-186.
Krackhardt, D. (1988), Predicting With Networks: Nonparametric Multiple Regression Analyses
of Dyadic Data, Social Networks, 10:359-382.
Kriechel, B. (2003), Heterogeneity Among Displaced Workers, Maastricht: ROA.
Lagendijk, A., Oinas, P (2005), Proximity, External Relations, and Local Economic
Development, In: Arnoud Lagendijk and Päivi Oinas (Eds.), Proximity, Distance and Diversity,
Issues on Economic Interaction and Local Development, Ashgate, 2005, 3-22
Leenders, T.Th.A.J. (2002), Modeling Social Influence Through Network Autocorrelation:
Constructing the Weight Matrix, Social Networks, 24(1):21-47.
Ligterink, J. E. (2001), Corporate Financial Risk Management, Amsterdam: Thesis publishers.
Maskell, P., Malmberg, A. (1999), Localized Learning and Industrial Competitiveness,
Cambridge Journal of Economics, 23:167–186.
Metcalfe, S. (1994), The Economic Foundations of Technology Policy: Equilibrium and
Evolutionary Perspectives, In: Dodgson M., Rothwell R (Eds.), The Handbook of Industrial
Innovation, Cheltenham: Edward Elgar, 409-512.
Mowery, D. C., Oxley, J. E., Silverman, B. S. (1998), Technological Overlap and Interfirm
Cooperation: Implications for the Resource-based View of the Firm, Research Policy,
NAG (2008), Your Guide to the Netherlands Aerospace Industry, Netherlands Aerospace Group.
Niosi, J., Zhegu, M. (2005), Aerospace Clusters: Local or Global Knowledge Spillovers?,
Industry and Innovation, 12(1):1–25.
Nooteboom, B. (2000), Learning and Innovation in Organizations and Economies. Oxford:
Oxford University Press.
Nooteboom, B., Van Haverbeke, W., Duysters, G., Gilsing, V., A. van den Oord (2007), Optimal
Cognitive Distance and Absorptive Capacity, Research Policy, 36:1016- 1034.
Ponds, R., F.G. van Oort, K. Frenken (2007), The Geographical and Institutional Proximity of
Research Collaboration, Papers in Regional Science, 86:423-444.
Powell, W. W., Walter, W., Koput, K. W., Smith-Doerr, L. (1996), Interorganizational
Collaboration and the Locus of Innovation: Networks of Learning in Biotechnology,
Administrative Science Quarterly, 41(1).
Rallet A. (2003), L’économie de Proximité. Propos D’étapes. Etudes et Recherche sur les
Systèmes Agraires et le Développement, 33:11-23.
Rallet, A., Torre, A. (1999), Which Need for Geographical Proximity in Innovation Networks at
the Era of global Economy?, ERSA Conference Papers ersa99pa128, European Regional Science
Torre A., Rallet, A. (2005), Proximity and Localization, Regional Studies, 39:47-59.
Reuters (1996), International Business; Core Units Are Closed by Fokker, The New York Times.
Sammarra, A., Biggiero, L. (2008), Heterogeneity and Specificity of Inter-Firm Knowledge
Flows in Innovation Networks, Journal of Management Studies, 45(4):800-829.
Sternberg, R., Arndt, O. (2001), The Firm or the Region: What Determines the Innovation
Behavior of European Firms?, Economic Geography, 77(4):364-382.
Sorenson, O., Singh, J. (2007), Science, Social Networks and Spillovers, Industry & Innovation,
Ter Wal, A.L.J. (2009), The Spatial Dynamics of the Inventor Network in German
Biotechnology, Geographical Proximity versus Triadic Closure, Utrecht: Department of
Torre A., Gilly J-P. Gilly, 2000, On the Analytical Dimension of Proximity Dynamics, Regional
Uzzi, B. (1996), The Sources and Consequences of Embeddedness for the Economic
Performance of Organizations: The Network Effect, American Sociological Review, 61:674-698.
van Burg, E., van Raaij, E., Berends, H. (2008), The Fibres that Hold an Innovation Network: An
Analysis of Knowledge-Sharing in the Glare Network, In: de Man, A.-P. (Eds.), Knowledge
Management and Innovation in Networks, Cheltenham: Edward Elgar.
Figure 1: Dutch aviation knowledge network
Technological field according to NAG
Airframe subsystems & components
Propulsion & engine components
Avionics, simulation & control
Education & training
Engineering & R&D
Space subsystems & components
Maintenance & overhaul
Table 1: NAG technological fields
Number of firms
Relational variables Type Share
Size difference (SIZE)
Non-firm links (PUB)
Firm links (PRIVATE)
Geographic distance (DIST)
Technological similarity NAG (TECNAG)
Technological similarity NACE code (SIMNACE)
Related variety effect (SIMNACE)2
Shared Fokker history (FOK)
Dedicated towards aerospace (AERO)
Importance of external knowledge (OPEN)
Table 2: Relational variables
2 FOK. AERO
Table 3: QAP-correlation relational variables
Chi-Squared test of fit
Total Fraction Correct:
Fraction Predicted 1s Correct:
Fraction Predicted 0s Correct:
Numbers in parentheses are based on models not reported estimations. Because the other variables’ coefficients did
not change significantly they are not listed.
Table 4: QAP-logit network regression technological knowledge
INN EMPL R&D
nbr.val 42 42 42
nbr.null 14 0 7
Min 0 5 0
Max 90 1500 4
median 9 51.5 0.09
Mean 21.96 121.67 0.33
std.dev 27.54 241.82 0.66
LINKS SIM SIM2
nbr.val 42 42 42
nbr.null 8 0 0
Min 0 0.51 0.00
Max 15 1 1
median 2 0.82 0.01
Mean 3.60 0.80 0.20
std.dev 3.89 0.13 0.39
SKILL CENT DIST
nbr.val 42 42 42
nbr.null 8 12 0
Min 0 0 0.01
Max 100 22 6.94
median 9.5 2 2.69
Mean 21.02 4.52 2.751
std.dev 28.61 5.45 1.79
Table 5: Descriptives
1657.960 on 10 degrees of freedom,
EMPL R&D AERO AGE FOREIGN LINKS SIM
Table 6: Correlation matrix
34 Download full-text
Base-line (1) (2) (3) (4) (5) (6) (7) (8)
Moran I res.
Geary C res.
Table 7: Network autocorrelation regression: determinants of innovation performance
1.23 1.12 1.19 1.22 1.21 1.29 1.16 1.12 1.14