You are on page 1of 14

Research Policy 51 (2022) 104081

Contents lists available at ScienceDirect

Research Policy
journal homepage: www.elsevier.com/locate/respol

Knowledge complexity and the mechanisms of knowledge generation and


exploitation: The European evidence
Cristiano Antonelli a, *, Francesco Crespi b, Francesco Quatraro a
a
Dipartimento di Economia e Statistica Cognetti de Martiis, Università di Torino, BRICK Collegio Carlo Alberto, Italy
b
Dipartimento di Economia, Università Roma Tre, BRICK Collegio Carlo Alberto, Italy

A R T I C L E I N F O A B S T R A C T

JEL classification: A knowledge complexity trade-off can be identified if and when the complexity of the stock of knowledge en­
O31 genders positive effects in the recombinant generation of new technological knowledge but negative ones in its
O32 exploitation in terms of productivity gains. On the one hand, the complexity of the stock of knowledge increases
R11
the scope for recombination and hence the amount of knowledge that each firm is able to generate with a given
Keywords: budget. On the other, the complexity of the stock of knowledge has controversial effects on productivity: the
Knowledge composition
indirect effects -via the larger amount of knowledge generated upstream- are positive but the direct ones can be
Knowledge complexity
Knowledge generation function
negative if they generate difficulties in the exploitation of a highly heterogeneous stock of knowledge. The
Knowledge exploitation econometric test on the European regions in the years 1997–2009, provides strong empirical evidence about the
Technology production function relevance of the composition of the stock of quasi-public knowledge and its twin positive effects in terms of the
Knowledge complexity trade-off direct support of the knowledge generation function and of the indirect one on productivity growth. Moreover,
the analysis highlights the negative direct effects on productivity dynamics, though with substantial heteroge­
neity across different groups of regions.

1. Introduction technological knowledge actually shapes the working of economic


systems.2
Technological knowledge plays a central role in shaping the working This paper implements an empirical study to investigate the new
of an economic system and its growth. The stock of technological knowledge complexity trade-off that takes place if and when the effects
knowledge available in the system is one of its crucial endowments. Its of knowledge composition at the regional level are positive on the
effects, however, are far from being automatic. Because of the special generation but negative in the exploitation of new knowledge. The
features of knowledge as a good and an economic activity, the under­ proposed analysis builds upon recent advances of the economics of
standing of the mechanisms by means of which it affects the working of knowledge that have identified the intrinsic heterogeneity of knowl­
an economic system requires an articulated framework able to appre­ edge. Knowledge should be regarded as a highly differentiated bundle of
ciate and integrate the joint effects of its limited appropriability and knowledge items, so that the composition of the stocks of knowledge
exhaustibility, and hence extendibility and cumulability, not only with within economic systems plays an important role that adds to its sheer
respect to the size but also to its composition, both in its generation and size. The analysis on the composition of the stocks of knowledge have
in its exploitation.1 This undertaking builds upon the path-breaking highlighted knowledge complexity as a relevant dimension to explain
contributions of Zvi Griliches (Griliches, 1979, 1986, 1992, 1995) and the dynamics of knowledge generation and the evolution of production
helps overcoming the limits of the new growth theory (Romer, 1990; patterns for all the other goods (Hidalgo and Hausman, 2009; Petralia
Grossman and Helpman 1991; Aghion and Howitt, 1992) highlighting et al., 2017; Antonelli et al., 2017; Balland et al., 2019; Rigby et al.,
the variety of interdependent mechanisms by means of which 2019; Broekel, 2019). In parallel, in the context of the ‘new growth

* Corresponding author.
E-mail address: cristiano.antonelli@unito.it (C. Antonelli).
1
The concept of knowledge extendibility dates back to Popper (1969), and it refers to the feature of knowledge constructs of being fruitfully “generalizable to other
settings or amenable to new theoretical developments” (Wunsch, 1984: p.8). See also Rochowiak (1988).
2
We thank for the useful comments on the previous versions of the paper by three anonymous referees and the guest editors of the special issue on Economic
Complexity.

https://doi.org/10.1016/j.respol.2020.104081
Received 1 March 2019; Received in revised form 13 May 2020; Accepted 24 July 2020
Available online 6 August 2020
0048-7333/© 2020 Elsevier B.V. All rights reserved.
C. Antonelli et al. Research Policy 51 (2022) 104081

theory’, the complexity of knowledge and its impact on the dynamics of downstream production of goods. In so doing, we aim to assess how the
the aggregate economy has been recently analysed by a few authors (see composition – next to the size – of the stock of quasi-public knowledge
for instance Magalhães and Afonso, 2017; Sequeira et al., 2018; Bucci affects the working of the system and the mechanisms by means of
et al., 2019). which it affects the productivity growth of economic systems. Our main
Such a concept has been mainly applied in the analysis of the hypothesis is that knowledge complexity yields different effects on
knowledge generation function, by recognizing that next to the effects of knowledge generation and exploitation and eventually on productivity,
the size of the stock of quasi-public knowledge, its composition has because of the differential channels and mechanisms at work.
distinctive consequences that affect the production of new knowledge The empirical analysis focuses on 189 European regions, over the
(Antonelli, 2011). Because of the dual role of knowledge as both an time span 1997–2009 for which we calculate a regional Knowledge
input of the upstream recombinant knowledge generation activity and Complexity Index (KCI) based on the information contained in patent
its output and an input in the downstream productivity of all the other applications. The results confirm our hypotheses on the existence of
goods, knowledge complexity appears to have strong potential in the different channels through which the size and the composition of the
context of productivity dynamics explanations, i.e. in the analysis of the stock of available knowledge shape the generation and exploitation of
productivity equation.3 new technologies and affect the economic productivity performances of
In this context of analysis, an array of empirical investigations at the European regions.
firm, industrial and regional levels have tested separately the role of the The rest of the paper is organized as it follows. Section 2 provides a
knowledge composition respectively in the knowledge generation synthetic account of the different approaches to the role of the compo­
function and in explaining productivity growth, confirming that the sition of the knowledge stock in the generation and exploitation of
characteristics of the stock of knowledge is a relevant factor in both knowledge elaborated in the literature and presents the hypotheses.
cases. In particular, previous evidence suggests that the related variety Section 3 introduces the empirical analysis and elaborates a consistent
of the stock of knowledge helps increasing, with a given budget, the econometric approach. Section 4 presents the results of the empirical
levels of knowledge output. On the contrary, the impact of variety on tests. The conclusions summarize the main results of the analysis.
productivity dynamics seems to be more controversial as indicated in
the systematic review of the literature provided by Content and Frenken 2. Literature background and research hypotheses
(2016). In particular, the introduction of new varieties of knowledge
inputs can generate a tension between productivity-gains due to more Previous literature investigated on the one hand the role of knowl­
specialization and productivity-losses associated with the presence of edge as an input of the knowledge generation function and, on the other
more complex knowledge bases to be applied to production processes hand, knowledge as a driver of productivity growth. In particular, the
(Jones, 1995; Bucci, 2013; Ferrarini and Scaramozzino, 2016; Bucci literature on the knowledge generation function analyzed the effects of
et al., 2019). the stock of existing knowledge on the generation of new knowledge,
Moreover, previous empirical literature, did not specifically which is now viewed as the outcome of an intentional and dedicated
addressed the simultaneous effects of the composition of the knowledge activity characterized by the recombination of existing knowledge items
stock on the knowledge generation function and on the productivity. (Weitzman, 1996, 1998; Fleming and Sorenson, 2001). Because of
Hence, the present contribution aims at reconcile these different strands knowledge indivisibility the whole body of existing knowledge is a
of investigation in a single analytical framework able to assess jointly the necessary and indispensable complementary input next to research and
effects of the composition of the stock of knowledge both in the development activities to generate new knowledge. Hence both the
knowledge generation function and in the technology production stock of knowledge internal to each agent and external to it are expected
function. to play a role (Jaffe, 1986; Coe and Helpman, 1995; Bottazzi and Peri,
In particular, building on the work by Hidalgo and Hausman (2009), 2007; Ang and Madsen, 2015).
and following the methodology proposed by Balland and Rigby (2017) In particular, a large body of empirical literature at the regional and
we focus on the notion of knowledge complexity, which reflects both the local levels has showed the relevance of knowledge externalities in
variety and rarity of the knowledge stock available in a given economic generating innovation, but also that both geographical and technolog­
area. On the one hand, we can interpret this measure as an indicator of ical proximity matter (Antonelli and Colombelli, 2015a,b; Boschma,
qualified variety, so that increasing complexity is expected to have 2005; Drivas et al., 2016). Among others, Bottazzi and Peri (2003) es­
positive impact on recombinant knowledge dynamics as it supports the timate the effect of knowledge externalities in generating innovation, by
scope for the recombinant generation of new knowledge and eventually looking at R&D and patent data for European Regions in the 1977–1995
leads to augmented rates of introduction of innovations. On the other period. They find that spillovers are very localized and rapidly decay
hand, as suggested by Balland and Rigby (2017) the proposed measure with geographical distance. On the same line, Greunz (2005) estimate a
of knowledge complexity also reflects the difficulty with which different regional knowledge production function, finding that knowledge spill­
kinds of knowledge can be exploited directly into the technology pro­ overs exist especially between geographically and technologically close
duction function (TPF). For this reason, high levels of complexity can regions. Similarly, Moreno et al. (2005) highlight the relevance of in­
hamper productivity dynamics. ternal regional factors (R&D expenditure and agglomeration economies)
More specifically, the proposed analysis investigates the relation­ beside the role of spatial spillovers due to innovation activities per­
ships between knowledge complexity, innovation and productivity in formed by other regions. Moreover, the interaction between internal and
order to unveil whether the positive effects of knowledge complexity external technological knowledge with (both internal and external) so­
take place only in the upstream generation of knowledge and/or in the cioeconomic and institutional conditions has been found to play a role
for the regional innovation performance (Rodríguez-Pose and Cre­
scenzi, 2008). Interestingly, such a literature also shows that threshold
3
effects and complex interactions effects might be relevant (Charlot et al.,
Following the CDM literature (See Crépon et al., 1998) we use the term 2015), and that the magnitude of different sources of knowledge spill­
“productivity equation” to identify the “technology production function” used
overs might be different. In particular, in the case of French regions, it
by Griliches (1979) to indicate the extended version of the traditional pro­
duction function, in which knowledge adds to the usual inputs to production, i.
e. capital and labour. Following Griliches we stick to the use of “knowledge
generation function” instead of the CDM innovation equation to denote the
activities that concern the generation of new knowledge (See Antonelli and
Colombelli, 2105a).

2
C. Antonelli et al. Research Policy 51 (2022) 104081

has been found that spillover effects are associated with private R&D Beside the analyses on the knowledge generation function, the role of
activity that spills across industry boundaries, suggesting that Jacobs size and composition of knowledge stocks has been investigated in
externalities decrease more drastically with distance than Marshallian relation to the analysis of productivity drivers. An abundant body of
externalities (Autant-Bernard and LeSage, 2011). However, even Mar­ literature has shown that the levels of (total factor) productivity are
shallian externalities have been found to differ by industry and to have positively and significantly associated with the size of the stock of
evolved over time (Figueiredo et al., 2015; Diodato et al., 2018). knowledge (see for instance the seminal contributions by Griliches,
The understanding of the heterogeneity of knowledge enables 1979, 1984, 1992) .4 Moreover, recent contributions have specifically
important progress with the exploration of a new dimension: the role of analyzed the interactions between complexity and the quantity of
its composition (Quatraro, 2010; Rigby, 2015). The composition of the knowledge or technological varieties in the system with respect to TFP
regional stock of knowledge is considered as the source of augmented dynamics, suggesting that possible trade-offs between productivity gains
opportunities for the recombinant generation of new knowledge both (due to more specialization) and productivity losses (due to more
within and between firms: i) the larger the variety of the knowledge complexity) may arise (Sequeira et al., 2018; Bucci et al., 2019).
stock internal to each firm and the larger its knowledge output; ii) the In parallel, the analyses conducted at local and regional levels,
larger the variety of the knowledge stock at the system level and the highlight the influence of agglomeration effects and local externalities
larger the knowledge output of each firm. Much attention in this context on regional economic performance, within the urban economics
has been paid -once more- to the effects of spillovers and related ex­ approach (see De Groot et al., 2009 and Puga, 2010 for extensive re­
ternalities. So far, the analysis had privileged the role of Marshallian views). In particular, the New Economic Geography suggests that the
knowledge externalities: the new approach explores the role of Jacobs presence of open economies and localized spillovers is able to spur the
(1969) knowledge externalities (Glaeser et al., 1992; Feldman and growth of output and productivity (Martin and Ottaviano, 1999; Bald­
Audretsch, 1999; Frenken et al., 2007; Galliano et al., 2015). In partic­ win and Martin, 2004; Ottaviano and Thisse, 2004 among others).
ular, Rigby (2015) and Boschma et al. (2014) show that the probability Accordingly, Dettori et al. (2012) find a positive effect of human capital
of filing a patent in a given technological class in a US city is strongly and patent stock on total factor productivity, associated with significant
influenced by its relatedness with the local technological base. The evidence about dynamic increasing returns triggered by diachronic
analysis of the co-occurences of patents in the different technological pecuniary knowledge externalities (Antonelli et al., 2011). Such results
classes marks an important progress that enables to appreciate the levels have been further qualified, on the one hand, by research efforts
of diversity articulated in coherence and complementarity of the stock of enquiring the role of both size and composition of knowledge on pro­
knowledge available in a given local context (Antonelli et al., 2010; ductivity dynamics, and, on the other hand by empirical analyses
Quatraro, 2010 and 2016; Colombelli and Quatraro, 2018). looking at the differentiated effects that knowledge and innovation
A step further in this research avenue has been made possible by might have across different regions.
recent contributions that applied complexity theory to the analysis of With respect to the first issue, working on OECD data Saviotti and
economic development. In particular, Hidalgo and Hausman (2009) Frenken (2008) found that related export variety tend to stimulate GDP
developed a measure of the complexity of products and of countries, and growth per capita and labor productivity, while the effects of unrelated
showed how this measure is able to capture the complexity of the set of export variety can be seen only in the medium-long run. This result
capabilities available in different countries. In their perspective, not suggests that while related variety eases knowledge recombination into
only the number of the activities and the related knowledge bases new products, with direct growth effects, knowledge unrelated variety is
matter, but also, their relative scarcity. Building on this analysis, Balland harder to recombine, but if successful, can sustain long-term growth.
and Rigby (2017) used patent records from the US to identify the Similarly, Falcioglu (2011) looked at productivity growth in Turkish
technological profile of US metropolitan areas between 1975 and 2010. regions, and found that only related variety positively affects regional
Their findings suggest that knowledge complexity is unevenly distrib­ productivity. Consistent evidence has been found for Dutch regions by
uted across US cities and that more complex technological structures do Boschma and colleagues (2011), by distinguishing between
not necessary lead to higher rates of patenting. Interestingly, more manufacturing and service sectors. However, the empirical evidence on
complex patents are found to be less likely to be cited than less complex the role of knowledge composition does not appear to be fully consoli­
patents when citing and cited patents are located in different metro­ dated as highlighted by the evidence provided in Quatraro (2010) and in
politan areas, suggesting that the highest levels of the Knowledge Cortinovis and Van Oort (2015).
Complexity Index reveal that the knowledge developed in a given ter­ On the other hand, while the bulk of previous studies have generally
ritory is so sophisticated that it can only be (re)produced in a few areas. focused on the effect of knowledge and innovation for the average re­
Following the same approach, Balland et al. (2019), show that EU re­ gion, more recent attempts have tried to identify eventual differentiated
gions tend to technologically grow more if they specialize in complex impact across regions (Marrocu et al., 2013; Capello and Lenzi, 2015 and
technologies related to existing technologies in the region. Similarly, 2014). Interestingly, the key drivers of productivity gains have been
Antonelli et al. (2017), find evidence that knowledge complexity posi­ found to differ across different groups of regions, with a closer rela­
tively contributes to the knowledge generation performance of Euro­ tionship between scientific knowledge and productivity dynamics in
pean regions, suggesting that the composition of the knowledge base regions with highest knowledge capabilities, but also with a strong link
represents a complementary input -next to its size- into knowledge between knowledge embedded in capabilities and higher productivity
generation. Finally, Petralia et al. (2017) show that technological performance in less knowledge intensive regions. This result suggests
diversification is higher for technologies related to countries’ compe­ that accounting for regional heterogeneity is important to properly ac­
tences, but also that countries, along their development paths, tend to count for differentiated pathways from knowledge and innovation
move towards more complex and valuable technologies, (Fig. 1). leading to higher productivity growth, and also claim for a more refined

4
The economic literature largely discussed both the theory behind the notion
of total factor productivity (TFP) and to what extent it finds empirical support.
Such debate showed that, among other factors, TFP may result from the existing
stock of knowledge, endogenously accumulated or adopted by each region or
country, but may also be linked to the externalities arising from the existing
stock of human capital -its quantity and quality-, and to an array of institutional
factors (Jones, 2009).

3
C. Antonelli et al. Research Policy 51 (2022) 104081

Fig. 1. Regional distribution of the focal variables (average values 2001–2006).

4
C. Antonelli et al. Research Policy 51 (2022) 104081

analysis on the differentiated effects of different types of dynamic dynamics can be negative when it limits the scope of knowledge appli­
increasing returns triggered by the special features of knowledge. cation and exploitation in the productivity equation (the downstream
Building on the literature discussed so far, we propose that the technology production function of all the other goods that includes the
relevance of both the size and the composition of regional knowledge stock of knowledge as a key input). The intuition behind this research
bases in explaining regional performances needs to be investigated both hypothesis is that the complexity of the stock of knowledge tends to
with respect to the generation of technological knowledge and to the reduce its extensibility, i.e. the opportunity for the use of
production of all the other goods. While the empirical literature at the knowledge-as-it-stands. Hence, firms based in regions with high levels of
firm level building on the CDM model (Raymond et al., 2015) exten­ complexity of the stock of knowledge have lesser opportunities to
sively investigated the twin role of knowledge as an input in the gen­ benefit from the limited appropriability of knowledge and to imitate it
eration of new technological knowledge and as a driver of productivity freely.
dynamics, the relationships previously identified at the regional level In conclusion, the contribution to productivity growth of the size of
have been mainly investigated as separated issues, and their simulta­ the knowledge stock is expected to be positive both in the knowledge and
neous appreciation is still largely unexplored. in the productivity equations via: i) the pecuniary knowledge external­
As previously discussed, the extant literature showed that both the ities generated upstream in the knowledge generation activities that
size and the composition of the available stock of knowledge are rele­ provide knowledge, as an input of the downstream technology produc­
vant in shaping innovation and economic performances as they exert tion function, at costs that are below equilibrium; and ii) the pecuniary
significant effects both in the knowledge production function (Jaffe, knowledge externalities in the knowledge generation activities where
1986) and in explaining productivity dynamics (see for instance Gri­ the size of the knowledge base helps supporting the generation of new
liches, 1979; Zachariadis, 2003; Ha and Howitt, 2007). However, the knowledge by means of the reduction of the access and use cost of the
introduction of new varieties of knowledge inputs can generate a tension stock of knowledge as an input. On the contrary, knowledge complexity is
between productivity-gains due to more specialization and supposed to enhance productivity dynamics via the pecuniary knowl­
productivity-losses associated with the presence of more complex edge externalities provided to downstream firms that use knowledge as
knowledge bases to be applied to production processes (Jones, 1995; an input in the production of all other goods -i.e.- the innovation
Bucci, 2013; Ferrarini and Scaramozzino, 2016; Bucci et al., 2019). channel- but is expected to yield a negative contribution to the pro­
In view of the discussion articulated so far, we are able to spell out or ductivity equation in its exploitation.
main working hypothesis. We hypothesize that a direct and an indirect The implications of the analysis are important. All the elements of a
effect of complexity on productivity can be identified. The direct effect of new possible trade-off emerge as the ultimate outcome of our hypoth­
complexity pertains to the technology production function, while the eses. If knowledge complexity exerts positive effects on the generation of
indirect effect is due to the influence that knowledge complexity exerts in new technological knowledge, but negative ones on its exploitation, it is
the upstream knowledge generation equation. This latter is based on the clear that a knowledge complexity trade-off takes place. Too much
role exerted by knowledge complexity in the knowledge generation complexity hampers the productivity with respect to all the other goods
function, whereby the stock of knowledge available in the system is used but supports the output levels in terms of knowledge. An “equilibrium”
as an input in the production of new knowledge and in the eventual level of knowledge complexity can be identified according to the bal­
introduction of innovations. Here, though knowledge complexity might ance between negative and positive effects. In this respect, the specific
increase the technological distance and entropy in the system (Peretto national, regional, and industrial (market and institutional) conditions
and Smulders, 2002), knowledge complexity is expected to display a into which generation and exploitation, respectively, take place are
(net) positive effect triggered by the limited exhaustibility and appro­ likely to have relevant effects on the actual levels of knowledge
priability of knowledge that enable perspective users to access and use complexity that firms are able to manage effectively.
-again- existing knowledge at costs that are below equilibrium. Hence,
the provision of cheaper knowledge generated upstream with positive 3. The empirical analysis
effects in terms of pecuniary knowledge externalities leads to increasing
of total factor productivity. 3.1. Dataset
On the contrary, regarding the direct effect, firms can be better able
to exploit a specialized knowledge base with low levels of complexity, as The dataset used for the present analysis consists of a large collection
the burden of knowledge increases with its heterogeneity and applies of patent applications at the European Patent Office.5 Patent applica­
more to its exploitation than to its generation. The complexity of the tions have been collected and elaborated through the last release of the
stock of knowledge limits the advantages triggered by the mechanisms REGPAT database (February 2018). REGPAT is built upon the World­
of extensibility: the larger the complexity and the lower the output upon wide Statistical Patent Database (PATSTAT) published by the European
which each knowledge item can be applied reducing the relevant Patent Office (EPO) twice a year and the OECD Patent Database, which
portion of the downward slope of average knowledge costs (Bucci et al., relies on the EPO Bibliographic Database and Abstracts (EBD) which
2019). also integrates pre-regionalized data by the KU Leuven. Each record
This argument has two main implications for the analysis of the ef­ provides detailed information concerning the application and priority
fects of the size and the composition of the stock of knowledge. On the dates of the corresponding patent, the assignee and inventor’s names,
one hand, both the size and the complexity of knowledge bases posi­ their addresses, the country identifier, the region codes, the application
tively influence innovation performances, via knowledge exchanges and
recombinant innovation dynamics (Antonelli et al., 2017). On the other
hand, while the size of the stock of knowledge increases exploitation
5
opportunities, its composition has controversial effects due to burden of The use of patent data in innovation studies is affected by some well-known
heterogeneous knowledge. Hence, our hypothesis is that the effects of limitations. In particular, we are aware that the use of patent data might lead to
knowledge complexity are strong and positive in knowledge generation, an underestimation of the contribution to the overall innovation output in a
specific region from smaller companies relying on informal incremental inno­
while the contribution of knowledge complexity to productivity
vation models, firms operating in the service sectors as well as firms relying on
alternative approaches to protect the value of innovation.

5
C. Antonelli et al. Research Policy 51 (2022) 104081

and publication numbers, along with additional input data of the from the Cambridge Econometrics’ European Regional Database (ERD),
regionalization algorithms.6 Patents have been assigned to regions on which is a highly disaggregated dataset across both sectoral and sub-
the basis of the addresses of the inventors.7 regional dimensions. Eurostat’s REGIO database is the primary source
We dropped regions belonging to the first quintile, in terms of patent of data for the ERD but is supplemented with data obtained from
applications, during the observed period.8 Moreover, we have retained AMECO, a dataset provided by the European Commission’s Directorate
regions that allowed us to have a balanced panel. The final database General Economic and Financial Affairs (DG EcFin).
covers the inventive activity of 189 geographical regions, located in 18
members states of the European Union (EU) over the 1997–2009 time
3.2. Variables and summary statistics
interval (see Table 1 for details) .9
Each patent application reports one or more International Patent
3.2.1. The dependent variables
Classification (IPC) codes. Such codes are the basis for our analysis of the
This paper investigates the impact of the complexity of the knowl­
technological composition of the regional portfolios. We used 4-digit IPC
edge base on productivity growth at the regional level. To do so, we
codes obtaining a mapping of all patents into 623 technology areas.
implement a two-step approach in which we estimate two related
Hence, for each region we have computed the number of patent appli­
equations, the ‘knowledge generation equation’, and the ‘productivity
cation in each of the 623 technology areas by year. Patents with more
equation’, so as to single out the direct effect of the complexity of the
than one IPC codes have been double counted.
knowledge base on regional productivity growth, from the indirect ef­
Data about the economic performances of regions have been drawn
fect channeled by the regional knowledge stock.
The dependent variable of the knowledge equation is the standard
Table 1 output measure of the knowledge production function, i.e. the regional
List of sampled countries. knowledge stock per employee (KSTOCK). The knowledge stock is
Country Freq. Percent Cum. calculated by applying the permanent inventory method to regional
AT 117 4.76 4.76 patent applications, implementing a 15% depreciation rate.10 The
BE 130 5.29 10.05 dependent variable of the productivity equation is the regional labor
BG 65 2.65 12.7 productivity (LP), calculated as the ratio between regional value added
CZ 104 4.23 16.93
and total hours worked by local employees.
DE 351 14.29 31.22
DK 65 2.65 33.86
ES 195 7.94 41.8 3.2.2. The complexity of the regional knowledge base
FI 13 0.53 42.33 Following an established tenet of literature, we adopt an indicator of
FR 260 10.58 52.91 the degree of complexity of the regional knowledge based on the in­
HU 78 3.17 56.08
formation contained in patent documents i.e. IPC (International Patent
IE 26 1.06 57.14
IT 156 6.35 63.49 Classification) codes (Antonelli et al., 2017; Balland and Rigby, 2017).
NL 156 6.35 69.84 In particular, the methodology is based on the adaptation of a method
PL 130 5.29 75.13 originally developed by Hidalgo and Hausmann (2009). The steps for
PT 65 2.65 77.78
the computation for the region-level indicators has been presented in
RO 78 3.17 80.95
SK 52 2.12 83.07 details in Balland and Rigby (2017) .11 Below we provide a synthetic
UK 416 16.93 100 illustration of the properties of such indicators.12
The Hidalgo and Hausmann (2009) method has been applied to
Total 2457 100 regional patent portfolios data based on the location of the inventors,
with the aim of obtaining a fine-grained analysis of the technological
composition of innovation activities carried out in a specific area. In
particular, the first step consists of identifying. whether region r = 1, …,
R is specialized in technology t = 1, …, T or not. This is achieved by
6
The address is linked to one or multiple regions on the basis of the postal implementing the revealed technology advantage indicator (RTA). The
code, the town name or using a mixed approach and manual correction. A
RTA is based on the patent applications that are observed in each region
specific field value designates all the records for which the regionalization
every year. It compares the relative frequency of patenting in a given
procedure has been unsuccessful.
7
Patents can be allocated to a single or multiple geographical zone according
technology t in region r, with the relative frequency of patenting in the
to the corresponding addresses of either their applicants or inventors. The two same technology t at the European level. If this index is larger than 1,
competing methodologies produce entirely different results and have distinct
economic interpretations: the former often indicates where the owner of the
10
intellectual property right has its headquarters or legal residence, whereas the We follow the established body of literature based on Hall et al. (2005) that
latter generally refers to the place in which the inventive process occurred. The applies to patent applications the same depreciation rate as the one applied to
second approach seems to provide a better approximation of the human capital R&D expenditures, i.e. 15% (see for example McGahan and Silverman 2006;
embedded in the local workforce, as it allows both the geographical and the Nesta 2008; Laitner and Stolyarov 2013; Rahko 2014).
11
temporal distribution of the competences and technological specializations to The index used in our analyses has been computed exploiting the R package
be measured consistently. ‘EconGeo’, described in Balland and Rigby (2017).
8 12
We chose to exclude the regions in the first quintile given that the number Broekel (2019) and Rigby et al. (2019) maintain that the application of the
of patents in these regions is extremely limited and this could bias the analysis. Hidalgo and Hausman methodology to European patent data does not give
For instance, the region with the highest number of patents among the excluded useful results, as the resulting complexity values of technologies differ sub­
regions is ES53 (Balearic Islands) which accounts for 0.01% of the total patent stantially from the US evidence. This is due to the fact the methodology rests on
count (188 patents out of almost 1.8 million in the 1994-2008 period). Table A2 the spatial diffusion of technologies, which tend to be more dispersed in the EU,
in the appendix provides information on the number of regions per country hence showing a higher ubiquity. Another issue could concern the underesti­
included in the analysis. It must be noted that this does not represent a selection mation of complexity. The basic question remains as to what extent the
bias, as it would do instead in microeconometric settings based on firm-level complexity of a technology has to be place-independent. There are alternative
data (Hewitt Dundas et al., 2019). approaches, which however are not free of limitations. While a comparative
9
Although the Cambridge Econometrics dataset as well as the availability of study of the performance of alternative indicators in different contexts is
patent data allow for a much wider coverage of countries and years, our sample beyond the scope of the present paper, this is certainly an issue worth of further
is constrained by the availability of data on regional R&D expenditure. investigation.

6
C. Antonelli et al. Research Policy 51 (2022) 104081

then the region is said to be specialized in the generation of technology t. strand, the knowledge generation equation can be specified as it
Once RTA is calculated for any technology in all regions, it is possible follows14:
to calculate: Kr, 0(i), which indicates variety, i.e., the number of tech­ ∑
→ KSTOCKi,t = α +β1 R&DSTOCKi,t− 1 +β2 KCIi,t− 1 +β3 MANi,t− 1 + φt + εi,t
nologies in which region i is specialized; K t,0 , which represents the
ubiquity of a specific technology, i.e., the number of regions specialized (1)
→ →
in technology j. Given the starting values, K r,0 and K t,0 , an iterative The technology production function (TPF) is as it follows:
formulae is applied to derive the Knowledge Complexity Index (KCI) that ̂ i,t−
LPi,t = a + b1 LPi,t− 1 + b2 Li,t− 1 + b3 KCIi,t− 1 + b4 KLi,t− 1 + b5 KSTOCK
synthetically measures the complexity of regional knowledge base.13 ∑
1

We report in the Appendix the ranking of technologies according to + β6 R&DSTOCKi,t− 1 + μt + ei,t


the values taken by the KCI variable. It can be seen that most complex (2)
technologies are those that may be understood as related to the phar­
The subscript i refers to region, while t refers to the year. By including
maceutical sector, while less complex technologies are related to the
in the productivity equation the predicted values of KSTOCK from the
manufacturing of machineries and parts thereof.
knowledge equation (Eq. (1)), we are able capture the differential effect
The following econometric analysis will use of the KCI indicator
of KCI on knowledge output and productivity, so as to ascertain to what
computed for each European region along the observed years. Such in­
extent the effect on productivity is indirect and driven by the effect on
dicator provides a consistent measure of the evolution of the degree of
knowledge, rather than direct.
complexity of the innovation activities carried out in a certain region, as
All variables in the Eqs. (1) and (2) have been transformed by
reflected by the related patent portfolio.
applying the inverse hyperbolic sine transformation, which can be
Note that the method derives implicitly complexity patterns from the
interpreted as a logarithmic transformation, but it is defined also for
empirical observation of the distribution of patenting activities across
values less than, or equal to, zero.15
regions. Hence, it is the degree of geographical clustering of different
Both the knowledge generation equation (KGE) and the technology
activities that reveals the properties of knowledge.
production function have been estimated by implementing the GMM
system estimator (Arellano and Bond, 1998; Blundell and Bond, 1998).
3.2.3. Other variables
Following the extant literature within the knowledge production
4. Results
function (KPF) framework, we include the following explanatory vari­
ables in the knowledge generation equation (KGE) (Moreno et al., 2005).
The results of the econometric estimations are reported in Tables 4, 5
First, we introduce an input measure of innovation, i.e. the stock of R&D
and 6. Table 4 concerns the innovation (knowledge generation) equa­
per employee (R&DSTOCK). This is obtained by applying the perma­
tion. The first column presents the baseline model. As expected the co­
nent method to the regional gross domestic expenditure on research and
efficient of R&DSTOCK is positive and significant, in line with the
experimental development (GERD), and using a 15% depreciation rate.
established literature within the knowledge generation function
Second, we include the share of manufacturing employees on the total
framework. Actually, R&D expenditures are considered as an input
level of regional employment (MAN) to control for the weight of sectors
measure proxying the quantity of knowledge available as a production
with higher propensity to patenting, provided that the literature has
factor in the equation in which patents proxy the output. The larger are
stressed that patenting is very rare in service sectors.
the investments in R&D activities, the higher is the output in terms of
The TPF also includes the ratio between the stock of fixed capital
patents. As for the control variables the share of manufacturing
formation and employment (KL), KSTOCK and the regional level of
employment (MAN) show the expected sign, as it is featured by a posi­
employment (L). All regressions also include time and region fixed
tive and statistically significant coefficient.
effects.
In the second column we introduce the proxy for the complexity of
Table 2 reports the descriptive statistics of the variables included in
the regional knowledge base, KCI. In line with the previous evidence
the empirical analysis, while Table 3 shows the pairwise correlation
reported in Antonelli et al. (2017), the coefficient of this variable is
coefficients amongst the variables. In some cases, these coefficients are
positive and significant, providing support to our hypothesis according
significant at the 5% level. However, it must be noted that, as far as the
to which not only the size but also the composition of the knowledge
variables that are used as regressors are concerned, the values of the
stock does contribute to the local generation of new technologies. From
coefficients do not raise particular concerns about multicollinearity.
a theoretical viewpoint this result suggests that knowledge complexity
enhances the effectiveness of innovation activities by providing local
3.2.4. The econometric model
economic agents with a larger set of potentially combinable knowledge.
The econometric model follows the logic behind the seminal model
Finally, the lagged values of MAN and GDP are characterized by positive
proposed by Crepon, Duguet and Mairesse (1997), by adapting it to the
and significant coefficients.
regional level. This level of aggregation does not present some of the
Finally, column (3) includes KCI and R&DSTOCK together in the
concerns arising at the micro-level, especially those related to selection
right-hand-side of the estimated equation. The results are consistent
bias and missing values of R&D expenditure. We are therefore left with
with those discussed so far. Actually, R&DSTOCK shows a positive and
two key equations, i.e. the knowledge (innovation) and the productivity
equation.
In line with the extant literature in the geography of innovation
14
It must be noted that existing studies implement the innovation equation by
using patent flows per capita as a dependent variable, and R&D flows per capita
13
The KCI has been calculated by using the eigenvector method. The values as an input measure. In this paper we opt for the original formulation stemming
obtained are the eigenvector associated with the second largest eigenvalue of from the seminal edited book by Griliches (1984), and use patent and R&D
the projected region-technology matrix. stocks. Nonetheless, in Table 5 we show the results obtained in regressions
implemented using patent and R&D flows, for the sake of comparison.
15
This is an alternative to the Box-Cox transformations, defined by the
1/2
following formula: inverse y = log[yi + (yi2 + 1) ]. Except for very small
values of y, the inverse sine can be interpreted as a standard logarithmic var­
iable. However, unlike a logarithmic variable, the inverse hyperbolic sine is
defined at zero and at negative values.

7
C. Antonelli et al. Research Policy 51 (2022) 104081

Table 2
Descriptive statistics.
variable N min max mean St.dev. skewness kurtosis

lnPAT 2457 0.001 3.526 1.044 0.863 0.528 2.425


LP 2457 0.002 0.076 0.029 0.013 − 0.559 2.503
lnKCI 2457 − 0.047 0.009 − 0.002 0.006 − 1.774 9.071
GDP 2457 0.003 0.098 0.049 0.020 − 0.778 2.703
lnGERD 2268 0.005 3.791 1.396 0.912 0.255 2.119
man 2457 0.055 0.400 0.189 0.070 0.533 2.802
K_L 2457 0.054 3.749 2.296 0.651 − 1.253 4.127
L 2457 5.295 9.744 7.700 0.759 − 0.405 3.389
lnGERD 2268 0.005 3.791 1.396 0.912 0.255 2.119

Table 3
Pairwise correlation matrix.
KSTOCK LP KCI GDP R&DSTOCK MAN KL L

KSTOCK 1
LP 0.6315 1
KCI 0.4800 0.3744 1
GDP 0.6676 0.9582 0.4153 1
R&DSTOCK 0.8026 0.7459 0.3734 0.7624 1
MAN − 0.1497 − 0.4899 0.0625 − 0.4345 − 0.3489 1
K/L 0.4364 0.7749 0.2988 0.7711 0.6014 − 0.3661 1
L 0.1484 − 0.1473 0.1933 − 0.0877 0.1058 0.2114 − 0.2056 1

Table 4 Table 6
Knowledge Generation Function (KGE). TPF.
(1) (2) (3) (1) (3) (5)

KSTOCK t-1 0.7766*** 0.8580*** 0.7724*** LP t-1 0.6796*** 0.6794*** 0.6767***


(0.0051) (0.0026) (0.0053) (0.0100) (0.0102) (0.0076)
KCIt-1 0.0097** 0.0286*** KCI t-1 − 0.0015*** − 0.0013*** − 0.0016***
(0.0040) (0.0056) (0.0002) (0.0002) (0.0002)
R&D_FLOW t-1 0.0439*** 0.0442*** K/L t-1 0.0001 0.0001 0.0001
(0.0023) (0.0022) (0.0001) (0.0001) (0.0001)
MAN t-1 0.1302*** 0.0781*** 0.1233*** L t-1 0.0029*** 0.0025*** 0.0028***
(0.0042) (0.0044) (0.0039) (0.0003) (0.0003) (0.0003)
N 1675 1893 1667 R&DSTOCK t-1 − 0.0000 0.0002*** − 0.0000
AR(1) − 4.2963 − 4.5792 − 4.2439 (0.0001) (0.0001) (0.0001)
AR(2) − 0.9818 − 0.6423 − 0.9510 KSTOCK
̂ 1 t-1
0.0019***
Sargan statistic p<0.01 p<0.01 p<0.01
(0.0001)
GMM-system estimations KSTOCK
̂ 2 t-1
0.0011***
..Standard errors in parentheses (0.0001)
* p < 0.10. KSTOCK
̂ 3 t-1
0.0019***
** p < 0.05. (0.0001)
*** p < 0.01. N 1540 1675 1540
AR(1) − 5.5678 − 5.7115 − 5.5754
AR(2) 0.1313 0.5860 0.1373
Table 5 Sargan statistic p<0.01 p<0.01 p<0.01
Knowledge Generation Function (KGE, patent flows). GMM-system estimations.
(1) (2) (3) Standard errors in parentheses.
* p < 0.10.
PAT_flow t-1 0.8421*** 0.8693*** 0.8405***
(0.0053) (0.0032) (0.0055) ** p < 0.05.
KCIt-1 0.0127*** 0.0042 *** p < 0.01.
(0.0043) (0.0067)
R&D_FLOW t-1 0.0097** 0.0124***
significant coefficient, and the same applies to MAN. Also, our hypoth­
(0.0041) (0.0038)
MAN t-1 0.1178*** 0.0943*** 0.1101***
esis about the role of Jacobs’ externalities in local innovation dynamics
(0.0053) (0.0024) (0.0053) finds support, as the coefficient of KCI is positive and significant.
N 1675 1893 1667 Table 5 provides an additional set of estimations including the flows
AR(1) − 4.3576 − 4.6169 − 4.3197 instead of the stocks. As anticipated in Section 3, although a relevant
AR(2) − 1.0180 − 0.6988 − 1.0067
number of empirical studies in the geography of innovation literature
Sargan statistic p<0.01 p<0.01 p<0.01
implement the knowledge production function by including patent and
GMM-system estimation. R&D flows as dependent and independent variables respectively, we do
Standard errors in parentheses. prefer to stick to the original formulation stemming from Griliches
* p < 0.10.
(1984) that relies on the stocks of both patents and R&D expenditures.
** p < 0.05.
The results seem to be completely in line with those obtained in
*** p < 0.01.
Table 4. Indeed, R&DSTOCK shows a positive and significant coefficient
both in columns (1) and (3). The positive and significant coefficient of

8
C. Antonelli et al. Research Policy 51 (2022) 104081

KCI is also confirmed in columns (2) and (3). For what concerns the Lenzi (2013) .18 The authors propose a taxonomy of European regions
control variables, MAN is featured by a positive and significant reflecting the differential territorial patterns of innovation. Instead of
coefficient. relying on the simple notion of knowledge intensity, five clusters are
The estimation of the knowledge generation equation (KGE) allows identified, which reflect five different modes of approaching the inno­
us to validate our hypothesis about the positive role of knowledge vation process. The clusters are labelled as it follows: i) Imitative
complexity in the innovation dynamics grasped by the knowledge gen­ innovation areas; ii) Smart and creative diversification areas; iii) Smart
eration function. Moreover, this represents a necessary step to imple­ technological application areas; iv) Applied science areas; v) European
ment a test of our hypothesis about the effects of knowledge complexity science-based areas. Details about the features of the clusters can be
in the downstream technology production function, which requires to found in the original paper. For our purposes, it is worth stressing that
estimating the impact of KCI on local productivity dynamics. the last cluster is featured by highly formalized innovative efforts. These
The results of the GMM estimation of the technology production become less and less formalized as one proceeds towards the first cluster,
function are reported in Table 6. Each of the three columns includes the while the weight of tacit knowledge increases. Also, the intensity of
predicted value of KSTOCK based on the estimation reported in the innovation efforts decreases from the last to the first cluster. Table 7
correspondent column in Table 5. The lagged level of the dependent reports the distribution of the observations in our dataset across the 5
variable (LP), is characterized by the expected positive and significant clusters.
coefficient. Moreover, it must be noted that the coefficient is signifi­ In order to ascertain the differential effect of knowledge complexity
cantly lower than one, suggesting that in our sample regions we observe across the five clusters we decided to proceed in the following way. We
convergence of labor productivity. The coefficient of R&DSTOCK is have first generated five dummy variables, taking value 1 if the region
positive and significant in two out of three estimations. The estimated belongs to the relevant cluster, 0 otherwise. Then we have interacted
values of KSTOCK are also featured positive and significant coefficients each dummy with the KCI variable. We finally obtained five interaction
in all of the three models. This is in line with the established literature on variables, each taking the value of KCI when the dummy is equal to one,
knowledge complexity and productivity and confirms the effectiveness zero otherwise. In this way we cover the entire distribution of the KCI
of the indirect channel through which knowledge complexity affects variable and we can directly appreciate the differential effect of KCI
labor productivity. across the five clusters of regions. As a consequence, the KCI variable per
For what concerns KCI, one can observe, instead, that the coefficient se is not included in the regression (and in any case it would not be
of this variable is negative and significant in all of the three. Such a identified). In so doing the provided econometric estimates are able to
result suggests, that when the positive indirect effect of knowledge capture the effects of knowledge complexity on productivity dynamics
complexity on productivity via the generation function is accounted for within each cluster, thus taking into account structural differences in
(evidenced by the positive and significant coefficient of the KSTOCK ̂ terms of absolute levels of knowledge complexity.19
variables), the residual impact of knowledge complexity becomes The results of these additional estimations are reported in Table 8.
negative, due to reduced knowledge exploitation potential. This is We only show the coefficients of the interaction variables, the other ones
consistent with our hypothesis about the relationship between techno­ being consistent with the previous estimations. Column (1) shows the
logical complexity and productivity spelt out in Section 3. In other differential effect of KCI on the innovation dynamics of the regions in the
words, technological complexity is an opportunity as far as innovation five clusters. Let us recall that all the variables have undergone a
dynamics are concerned, but it can likely be an obstacle when economic transformation that approximates a logarithmic transformation. Though
agents choose to use complex knowledge in their production processes with some caution, the coefficients can be read as elasticities. KCI seems
as an input-ready-to-be used, rather than in input in the recombinant to yield negative and significant effects on the innovative dynamics of
knowledge generation activities.16 regions in ‘Imitative’ areas, while no significant effects emerge for what
concerns ‘Smart and creative diversification’ areas. In these areas
innovative dynamics are very weak, mostly based on imitation or the
4.1. Sensitivity leveraging of tacit knowledge and learning-by-doing. The coefficient of
KCI is positive and significant for regions in “Smart technological
4.1.1. Estimations by macro-areas application” and “Applied science” qreasa, which are strongly involved
The results of the econometric estimations reported in the previous
section provide support to the idea that the composition of the knowl­
Table 7
edge stock affects in different ways the generation of innovation, and its
Distribution of sampled observations across the 5 clusters as in Capello and
actual exploitation in the production process. Lenzi (2013).
The overall results may however hide some important differences
Cluster Freq. Percent Cum.
amongst European regions, due to the heterogeneity of both innovation
and economic performances. For this reason, in this section we carry out Imitative innovation areas 325 13.23 13.23
two distinct and yet complementary sensitivity analyses to ascertain the Smart and creative diversification areas 793 32.28 45.50
Smart technological application areas 689 28.04 73.54
differential impact of knowledge complexity across different types of Applied science areas 442 17.99 91.53
regions.17 European science-based area 208 8.47 100.00
The first test is based on the split of our sample of European regions Total 2457 100.00
in five clusters, according to the taxonomy proposed by Capello and

16
It is worth stressing that this effect might also be due to other relevant costs
18
faced by firms in presence of excess variety, like coordination, search and The authors wish to thank Roberta Capello and Camilla Lenzi for sharing
transaction costs (see Grimpe and Kaiser, 2010; Gkypali et al., 2017). the details of the taxonomy.
17 19
The heterogeneity of European regions in terms of economic and techno­ Note that we show interaction effects in the form of direct effects (i.e., we
logical conditions raises an important issue related to the link between complex show the effects for KCI interacted with each of the n relevant interaction
dynamics and evolutionary patterns in terms of initial conditions and path- categories). This is a strategy algebraically equivalent to showing the effects for
dependence (see Gkypali et al., 2019). However, this goes beyond the scope the non-interacted KCI and its interactions with only n − 1 interaction cate­
of the present paper. gories (i.e., in the form of differential impacts). We prefer displaying results in
this way as they directly show the impacts for each typology of region without
needing additional algebraic computations to get direct effects.

9
C. Antonelli et al. Research Policy 51 (2022) 104081

Table 8 coefficient along all of the observed quantiles. Also, it must be noted that
Differential effect of KCI across regional clusters. the magnitude of the coefficient is pretty stable up to the 50th quantile,
(1) (2) while it is slightly higher at the 75th and the 90th quantile. For what
KGE TPF concerns our main regressor of interest, KCI, the estimated coefficient
Imitative innovation areas − 0.5301*** 0.0258*** appears to be positive and significant across all of the reported quantiles.
(0.0574) (0.0038) Also in this case, the coefficients seem to be higher for regions at the
Smart and creative diversification areas − 0.1199 0.0053** 75th and the 90th quantile.
(0.0750) (0.0025) Table 10 instead concerns the technology production function. Also
Smart technological application areas 2.1832*** − 0.0918***
(0.1877) (0.0065)
in this case, each column reports the coefficients of the variables esti­
Applied science areas 0.3843** − 0.0557*** mated on the quantile of the distribution of the dependent variable
(0.1526) (0.0061) indicated in the header. The results show that R&DSTOCK is positive
European science-based area − 0.0992 − 0.0321** and significant along all the distribution, and the magnitude of the co­
(0.3846) (0.0130)
efficient increases as one moves upwards in the distribution. KCI shows
GMM-system estimations. interesting dynamics, as its coefficient turns out to be negative and
Standard errors in parentheses. significant across all of the quantiles. However, it can be noted that the
* p < 0.10. magnitude of the coefficients increases up to the median region, and
** p < 0.05. then it declines.
*** p < 0.01. Overall, our results provide further support to our working hypoth­
esis, and suggest that our estimates are quite reliable. When compared
in the application and generation of knowledge. with the results we have discussed in Section 4.1.2, these estimate
In column (2) we report instead the effect of KCI on the productivity suggest that at the aggregate level the positive effects of KCI is relevant
dynamics of the regions in the five clusters. The coefficient is positive for the generation of new technologies. The negative effect instead
and significant in the case of the ‘Imitative areas’ as well as of ‘Smart and dominates in the technology production function, due to the limits that
creative diversification areas’ (though weakly significant in this latter). excess knowledge heterogeneity raises to the extendibility and appli­
The effect turns out to be negative and significant for the Smart tech­ cability of existing knowledge to local economic activities. On the con­
nological application’ and ‘Applied science-based’ areas, while it seems trary, differential patterns have been identified, according to different
not yield significant effects on the productivity of regions in the ‘Euro­ modes of innovation that feature the sample region. This suggests that
pean science-based’ area. This evidence would suggest that the hy­ our hypothesis concerning the negative effect of KCI on productivity
pothesis on the negative effect of knowledge complexity on productivity dynamics needs to be qualified not according to the performance of
needs to be qualified. In regions strongly relying on imitation dynamics regions in quantitative terms, but in terms of dominant patterns of
for the preservation of competitiveness, excess specialization can hinder innovation dynamics. In this sense, our hypothesis that the negative
diversification and hence competition, while technological diversifica­ effect stems from constraints to the extensibility and applicability of too
tion increases the chances to develop distinctive comparative advan­ much heterogeneous knowledge receives further support by the com­
tages. On the contrary, the more regional economic agents rely on bination of these sensitivity tests.
genuine innovation dynamics eventually characterized by high absolute
levels of knowledge complexity, the more difficult is for local imitating 4.1.3. Spatial econometrics estimations
agents to leverage highly diversified local competences. As it has been shown by previous studies, the empirical analysis of
regional-level dynamics of innovation and productivity can be inter­
4.1.2. Fixed effects quantile estimation ested by spatial dependence amongst the units of observation. In these
To further investigate possible heterogeneities in how regions’ pro­ situations, linear regressions can yield biased results. Spatial econo­
ductivity dynamics are affected by KCI, we implement a fixed effects metrics models provide an empirical framework able to account for both
quantile regression for panel data (Canay, 2011). The results are re­ spatial dependence and heterogeneity (Anselin, 1988; LeSage, 1999).
ported in Tables 9 and 10. Many different empirical models could be implemented. However,
Table 9 reports the results concerning the KGE. Each column reports following Elhorst (2014), we decided to run a Spatial Durbin Model
the coefficients of the variables estimated on the quantile of the distri­ (SDM), i.e. an estimator including the spatial lag of both the dependent
bution of the dependent variable indicated in the header. Our results are and the exogenous variables, as it is the most comprehensive empirical
in line with previous estimations, and they are consistent with our hy­ framework (Varga, 1998; Elhorst, 2003, 2010).
potheses. In particular, R&DSTOCK shows a positive and significant Table 11 reports the results of SDM estimations. Column (1) refer to
the KGE. The results show that the spatial autoregressive component (i.
Table 9
e. neighbor regions’ knowledge production) is always positive and sig­
KGE (fixed-effects quantile estimation). nificant. The spatial lag of the focal regressor, KCI, is also featured by a
positive and significant coefficient.
(2) (3) (4) (5) (6)
In order to ascertain the impact of spatial dependence on the rela­
R&DSTOCKt-1 0.305*** 0.306*** 0.309*** 0.313*** 0.316*** tionship between the innovation outcomes and KCI, we report the
(0.00648) (0.00610) (0.00349) (0.00707) (0.00600)
marginal effects of the variables. As is common in the implementation of
KCI t-1 1.320** 0.853*** 1.563*** 2.594*** 2.014***
(0.561) (0.286) (0.348) (0.435) (0.399) spatial econometrics estimator, we show total, direct and indirect ef­
GDP t-1 2.318*** 2.641*** 2.853*** 2.769*** 3.044*** fects. The marginal effects indicate that KCI persistently feature a pos­
(0.272) (0.222) (0.167) (0.279) (0.233) itive and significant effect on regional innovation dynamics. Overall,
MAN t-1 1.332*** 1.310*** 1.275*** 1.199*** 1.230***
even after controlling for spatial dependence, our key hypothesis about
(0.0369) (0.0298) (0.0210) (0.0382) (0.0497)
Constant 0.150*** 0.225*** 0.302*** 0.397*** 0.437***
the positive impact of KCI on regional innovation cannot be rejected.
(0.0292) (0.0124) (0.0137) (0.0229) (0.0178) Column (2) of Table 11 reports the result of the SDM estimation of
N 2457 2457 2457 2457 2457 the technology production function. We only show the coefficients of the
Standard errors in parentheses. most relevant variables. The estimated coefficient of KCI is negative and
* p < 0.1. significant. The same applies to its spatial lag. Excess heterogeneity
** p < 0.05. seems to limit the extensibility and the applicability not only of
*** p < 0.01. knowledge generated within the region, but also of that generated in

10
C. Antonelli et al. Research Policy 51 (2022) 104081

Table 10
TPF (fixed-effects quantile estimation).
(2) (3) (4) (5) (6)

K/L t-1 0.000126* 0.000168*** 0.000312*** 0.000472*** 0.000610***


(0.0000663) (0.0000473) (0.0000459) (0.0000581) (0.0000767)
KSTOCK
̂ t-1
0.00583*** 0.00540*** 0.00490*** 0.00463*** 0.00426***
(0.000419) (0.000299) (0.000221) (0.000352) (0.000551)
KCI t-1 − 0.0128*** − 0.00930** − 0.00979** − 0.0125** − 0.0143*
(0.00413) (0.00446) (0.00404) (0.00507) (0.00816)
R&DSTOCK t-1 0.000721*** 0.000948*** 0.00114*** 0.00123*** 0.00137***
(0.000159) (0.0000997) (0.0000825) (0.000118) (0.000155)
L t-1 − 0.00323*** − 0.00330*** − 0.00330*** − 0.00331*** − 0.00338***
(0.0000531) (0.0000374) (0.0000237) (0.0000377) (0.0000774)
Constant 0.0447*** 0.0461*** 0.0468*** 0.0475*** 0.0485***
(0.000601) (0.000453) (0.000307) (0.000454) (0.000760)
N 2268 2268 2268 2268 2268

Standard errors in parentheses.


*
p < 0.1.
**
p < 0.05.
***
p < 0.01.

neighbor regions. To be sure, marginal effects indicate that such nega­


Table 11
tive effect holds with respect to both direct, and indirect and total ef­
Spatial econometrics estimation (SDM).
fects. Overall, also our hypothesis concerning the negative effect of KCI
(1) (2) on regional productivity dynamics cannot be rejected even after
KGE TPF
explicitly accounting for the spatial dependence of regional productivity
Main data.
KCI t-1 0.6333* − 0.0325**
The implications of the results are important. The empirical evidence
(0.3955) (0.0165)
R&DSTOCK 0.2752*** − 0.0004
confirms that knowledge complexity exerts positive effects on the gen­
t-1
(0.0103) (0.0006) eration of new technological knowledge, but negative ones on its
KSTOCK
̂ t-1 0.0064*** exploitation. The case of knowledge complexity trade-off is confirmed:
(0.0015) too much complexity hampers the productivity with respect to all the
Spatial lag regressor other goods but bolster up the output levels in terms of knowledge. The
KCI t-1 14.6594*** − 0.2879*** specific national, regional, and industrial market and institutional con­
(2.7702) (0.1114)
ditions into which knowledge generation and knowledge exploitation
R&DSTOCK t-1 − 0.1197*** 0.0066**
(0.0372) (0.0026) respectively take place have relevant effects on the levels of knowledge
Spatial autoregressive component complexity that firms are able to manage. From this viewpoint the evi­
rho 0.9265*** 0.8733*** dence confirms that the actual levels of knowledge complexity, practiced
(0.0199) (0.0381) by firms at the regional level, are -partly- endogenous.
Variance
sigma2 0.0040*** 0.0000***
(0.0001) (0.0000) 4.1.4. Other robustness checks
Direct effects A further set of robustness checks concerns the sensitivity of the
KCI t-1 1.9022*** − 0.0509*** results to the way one measures the composition of the knowledge stock.
(0.7152) (0.0172)
For this reason we report in Table 12 the results of the estimation of the
R&DSTOCK t-1 0.2875*** − 0.0001
(0.0108) (0.0006) knowledge generation equation and of the technology production
KSTOCK
̂ t-1 0.0068*** function by substituting a measure of knowledge variety for KCI. The
(0.0015) knowledge variety measure is obtained by calculating the information
Indirect effects entropy index for each region, constrained to those the technologies in
KCI t-1 229.1572** − 2.6334** which the regions shows a revealed technology advantage. This indi­
(110.1488) (1.2951) cator therefore captures the extent to which patents in the regional
R&DSTOCK t-1 2.0474* 0.0527*
(1.2234) (0.0301)
KSTOCK
̂ t-1 0.0492* Table 12
(0.0288) Innovation and TPF equations – Variety of RTAs.
Total Effects (1) (2)
KCI t-1 231.0595** − 2.6843** KGE TPF
(110.7390) (1.3003)
R&DSTOCK t-1 2.3349* 0.0526* Variety t-1 0.0023** − 0.0001***
(1.2281) (0.0303) (0.0009) (0.0000)
0.0560* R&DSTOCK t-1 0.0313*** 0.0001**
KSTOCK
̂ t-1
(0.0056) (0.0001)
(0.0297) N 2141 1533
N 2457 1430 AR(1) − 4.4765 − 5.6529
AIC − 6508.0789 − 15,600.7450 AR(2) − 0.5613 0.1004
BIC − 6415.1718 − 15,500.7018 Sargan statistic P<0.01 P<0.01
Standard errors in parentheses. GMM-system estimations.
*
p < 0.1. Standard errors in parentheses.
**
p < 0.05. * p < 0.10.
***
p < 0.01. ** p < 0.05.
*** p < 0.01.

11
C. Antonelli et al. Research Policy 51 (2022) 104081

patent portfolio are clustered or dispersed across the technologies for Table 13
which the region shows an RTA. Innovation and TPF equation (top 20% innovating regions).
The result indicates that even with a different measure of variety, the (1) (2)
evidence of a positive effect on innovation and of a negative one on KGE TPF
productivity seems to be robust. This provides further support to our KCI t-1 0.0930* − 0.0096***
working hypothesis about the negative impact of the hetherogeneity of (0.0512) (0.0017)
the regional knowledge base on the application of local technologies for R&DSTOCK t-1 0.0332 0.0045***
production purposes. (0.0451) (0.0011)
N 389 389
Another source of bias in our results could be due to the intensity of AR(1) − 4.1119 − 3.7331
patenting in the sampled regions. Actually, for regions with a small AR(2) − 0.6179 0.8831
number of patents, the calculation of RTA, which is at the basis of the Sargan statistic P<0.01 P<0.01
KCI index, might turn out to be misleading. For this reason, in Table 13 GMM-system estimations.
we provide the results of the estimation of the knowledge generation Standard errors in parentheses.
equation and of the technology production function carried out on the * p < 0.10.
regions in the top 20th centile of the distribution of patent applications. ** p < 0.05.
The results are consistent with those obtained in the baseline esti­ *** p < 0.01.
mations. In particular, KCI show a positive and (weakly) significant
coefficient in the knowledge generation equation (column (1)), while it a method originally developed by Hidalgo and Hausmann (2009). With
is featured by a negative and significant coefficient in the technology respect to this choice, our findings suggest that, despite the limitations of
production function. Overall, the sensitivity analyses carried out in this this measure of complexity when applied to European patent data
Section suggest that our results are robust to different specification of (Broekel, 2019; Rigby et al., 2019), we were able to capture substantial
the models and to the check for the most relevant sources of biases. relative heterogeneity in terms of complexity across European regions,
which in turns affects their innovative and productivity performances.
5. Conclusions Moreover, we showed that the evidence of a positive effect on innova­
tion and of a negative one on productivity is confirmed when a direct
The stock of knowledge available in a system is a key endowment. measure of knowledge variety is considered. Nevertheless, further
The mechanisms by means of which the size and composition of this research is needed to test the validity of our results when different
endowment contribute the working of economic systems, however, are complexity indicators and different geographical contexts are
far more complex than assumed by standard new growth theory. considered.
Technological knowledge is quite a special good, actually an activity, The other sensitivity analyses carried out on econometric results
characterized by the limited appropriability and exhaustibility that lead showed that one can find significant differences in the effect of KCI
to the accumulation of a stock of quasi-public knowledge. This stock can across European regions, based on their differential degree of speciali­
be regarded as an endogenous endowment that deploy differentiated zation in knowledge-based activities and on different levels of produc­
effects in its generation and exploitation. tion efficiency. These results yield important implications for regional
The mechanisms shaping knowledge generation differ from those innovation policies. First of all, consistently with an increasing body of
related to its exploitation. The size and composition of the stock of quasi- literature, one-size-fits-all innovation policies can hardly work for the
public knowledge available in a system enable to access (due to its improvement of innovation and productivity performances at the
limited appropriability), and use again (due to its limited exhaust­ regional level (Capello and Lenzi, 2014). Regional innovation policies
ibility), existing knowledge inputs at costs that are lower than their should be instead fine-tuned according to the structural features of
reproduction. The limited appropriability and exhaustibility of knowl­ regional innovation and economic activities, by explicitly accounting for
edge trigger pecuniary knowledge externalities that affect both the the distinction between knowledge generation and application. Second,
knowledge generation function and the technology production function. and related to the previous point, our results are connected to the
In this respect, the paper makes three contributions to the literature: ongoing policy debate on smart specialization strategies (Rigby et al.,
i) it highlights the pervasive role of the composition of the stock of 2019). In this framework, it is advocated that regional policies should
knowledge together with its size, ii) it widens the scope of the analysis consider the composition of the existing bundle of economic and tech­
on the role of knowledge complexity both in the upstream generation of nological activities to successfully orient and renew the regional pat­
knowledge and in the downstream production of all the other goods; iii) terns of industrial and technological specialization. Our results suggest
it identifies the working of three distinct mechanisms, by means of that this process should take into account the intrinsic difficulties that
which the complexity of the stock of quasi-public knowledge affects the agents in specific places can have in the command of too much variety of
working of the economic system. The evidence provided in this paper knowledge when the application to production processes is at stake.
suggests that complexity of the stock of quasi-public knowledge: i) Constraints to extensibility, application and imitation risk to yield the
qualifies the amount of knowledge that can be used as an input in the unintended effect of depressing overall local productivity dynamics.
recombinant knowledge generation function, ii) increases the amount of Overall, our results provide further support to the quest for place-based
knowledge that enters as an input in the downstream technology pro­ regional development policies based on coordinated efforts aiming at
duction function; iii) hamper the ready-to-be-used exploitation of the valorising untapped potential of regions by leveraging their structural
knowledge stock. This latter result indicates that the use of a knowledge features (Barca et al., 2012).
stock with high levels of complexity as an input ready to be used in the In conclusion, the analysis of the divergent effects of knowledge
production process of all the other goods can be made more difficult by complexity on, respectively, productivity and knowledge output and the
the excess array of technological domains that need to be mastered results of the empirical investigations confirm that, after the knowledge
which limit the benefits of knowledge extensibility. appropriability trade-off, a knowledge complexity trade-off takes place.
The analysis of these three effects has been carried out in two steps, The knowledge complexity trade-off has important policy implications.
by estimating the knowledge generation and the technology production The social and private returns of knowledge complexity may differ.
functions separately. The econometric model tests the equations on 189 Economic systems might benefit from higher levels of knowledge
European regions in the years 1997–2009 and adopts an indicator of the complexity that firms are not able to manage because of its negative
degree of complexity of the regional knowledge base that use the in­ effects on productivity. Market systems may engender, not only the -well
formation contained in patent documents building on the adaptation of

12
C. Antonelli et al. Research Policy 51 (2022) 104081

known Arrovian- undersupply of the quantity of knowledge, but also its Coe, D., Helpman, E., 1995. International R&D spillovers. Eur. Econ. Rev. 39, 859–887.
Colombelli, A., Quatraro, F., 2018. New firm formation and regional knowledge
wrong composition. An important target for economic policy is then
production modes: italian evidence. Res. Policy 47 (1), 139–157.
identified: to help the system to achieve levels of knowledge complexity Content, J., Frenken, K., 2016. Related variety and economic development: a literature
that go beyond the limits of private returns and are closer to its social review. Eur. Plann. Stud. 24 (12), 2097–2112.
returns. Cortinovis, N., Van Oort, F., 2015. Variety, economic growth and knowledge intensity of
European regions: a spatial panel analysis. Ann. Reg. Sci. 55 (1), 7–32.
Crépon, B., Duguet, E., Mairesse, J., 1998. Research and development, innovation and
Authors’ contribution productivity: an econometric analysis at the firm level. Econ. Innov. New Technol. 7
(2), 115–158.
De Groot, H.L.F., Poot, J., Smit, M.J., 2009. Agglomeration, innovation and regional
The authors equally contributed to all the steps of the research development: theoretical perspectives and meta-analysis. In: Capello, R., Nijkamp, P.
process. (Eds.), Handbook of Regional Growth and Development Theories. Edward Elgar,
Cheltenham, pp. 256–281.
Dettori, B., Marrocu, E., Paci, R., 2012. Total factor productivity, intangible assets and
Declaration of Competing Interest spatial dependence in the European regions. Reg. Stud. 46 (10), 1401–1416.
Diodato, D., Neffke, F., O’Clery, N., 2018. Why do industries coagglomerate? How
Marshallian externalities differ by industry and have evolved over time. J. Urban
The authors of this paper, Cristiano Antonelli, Francesco Crespi and Econ. 106, 1–26.
Francesco Quatraro, declare that they do not have any conflict of Drivas, K., Economidou, C., Karkalacos, S., Tsionas, E.G., 2016. Mobility of knowledge
and local innovation activity. Eur. Econ. Rev. 85, 39–61.
interest.
Elhorst, P., 2003. Specification and estimation of spatial panel data models. Int. Reg. Sci.
Rev. 26, 244–268.
References Elhorst, P., 2010. In: Fischer, M.M., Getis, A. (Eds.). Springer, New York, pp. 377–408
eds.
Elhorst, J.P., 2014. Spatial Econometrics. From Cross-Sectional Data to Spatial Panels.
Aghion, P., Howitt, P., 1992. A model of growth through creative destruction.
Springer, Berlin Heidelberg.
Econometrica 60 (2), 323–352.
Falcioglu, P., 2011. Location and determinants of productivity: the case of the
Ang, J.B., Madsen, J.B., 2015. What drives ideas production across the world.
manufacturing industry in Turkey. Emerg. Mark. Finance Trade 47 (Suppl. 5),
Macroecon. Dyn. 19, 79–115.
86–96.
Anselin, L., 1988. Spatial Econometrics: Methods and Models. Kluwer, Dordrecht.
Feldman, M.P., Audretsch, D.B., 1999. Innovation in cities: science-based diversity,
Antonelli, C. (Ed.), 2011. Handbook on the Economic Complexity of Technological Change.
specialization and localized competition. Eur. Econ. Rev. 43, 409–429.
Edward Elgar Publishing, Cheltenham.
Ferrarini, B., Scaramozzino, P., 2016. Complexity specialization and growth. Struct.
Antonelli, C., Krafft, J., Quatraro, F., 2010. Recombinant knowledge and growth: the
Change Econ. Dyn. 7, 52–61.
case of ICTs. Struct. Change Econ. Dyn. 21, 50–69.
Figueiredo, O., Guimarães, P., Woodward, D.P., 2015. Industry localization, distance
Antonelli, C., Colombelli, A., 2015a. External and internal knowledge in the knowledge
decay, and knowledge spillovers: following the patent paper trail. J. Urban Econ. 89,
generation function. Ind. Innov. 22, 273–298.
21–31.
Antonelli, C., Colombelli, A., 2015b. The cost of knowledge. Int. J. Prod. Econ. 168 (C),
Fleming, L., Sorenson, O., 2001. Technology as a complex adaptive system: evidence
290–302.
from patent data. Res. Policy 30, 1019–1039.
Antonelli, C., Crespi, F., Mongeau Ospina, C., Scellato, G., 2017. Knowledge complexity,
Frenken, K., Van Oort, F., Verburg, T., 2007. Related variety, unrelated variety and
Jacobs externalities and innovation performance in European regions. Reg. Stud. 51
regional economic growth. Reg. Stud. 41 (5), 685–697.
(11), 1708–1720.
Galliano, D., Magrini, M.B., Triboulet, P., 2015. Marshall’s versus Jacobs’ externalities in
Antonelli, C., Patrucco, P.P., Quatraro, F., 2011. Productivity growth and pecuniary
firm innovation performance: the case of French industry. Reg. Stud. 49 (11),
knowledge externalities: an empirical analysis of agglomeration economies in
1840–1858.
European regions. Econ. Geogr. 87 (1), 23–50.
Gkypali, A., Filiou, D., Tsekouras, K., 2017. R&D collaborations: is diversity enhancing
Autant-Bernard, C., LeSage, J.P., 2011. Quantifying knowledge spillovers using spatial
innovation performance. Technol. Forecast. Soc. Change 118, 143–152.
econometric models. J. Reg. Sci. 51, 471–496.
Gkypali, A., Kounetas, K., Tsekouras, K., 2019. European countries’ competitiveness and
Baldwin, R.E., Martin, Philippe, 2004. Agglomeration and Regional Growth. In:
productive performance evolution: unraveling the complexity in a heterogeneity
Henderson, J.V., Thisse, J.F. (Eds.), Handbook of Regional and Urban Economics:
context. J. Evol. Econ. 29 (2), 665–685.
Cities and Geography. Elsevier, Amsterdam, pp. 2671–2712.
Glaeser, E., Kallal, H., Scheinkman, J., Shleifer, A., 1992. Growth of cities. J. Polit. Econ.
Balland, P.A., Rigby, D.L., 2017. The geography of complex knowledge. Econ. Geogr. 93,
100, 1126–1152.
1–23.
Greunz, L., 2005. Intra- and inter-regional knowledge spillovers: evidence from European
Balland, P.A., Boschma, R., Crespo, J., Rigby, D.L., 2019. Smart specialization policy in
regions. Eur. Plann. Stud. 13 (3), 449–473.
the European Union: relatedness, knowledge complexity and regional
Griliches, Z., 1979. Issues in assessing the contribution of research and development to
diversification. Reg. Stud. 53, 1252–1268.
productivity growth. Bell J. Econ. 10 (1), 92–116.
Barca, F., McCann, P., Rodríguez-Pose, A., 2012. The case for regional development
Griliches, Z., 1986. Productivity, R&D, and basic research at the firm level in the 1970s.
intervention: place-based versus place-neutral approaches. J. Reg. Sci. 52 (1),
Am. Econ. Rev. 77 (1), 141–154.
134–152.
Griliches, Z., 1992. The search for R&D spillovers. Scand. J. Econ. 94, 29–47
Boschma, R., 2005. Proximity and innovation: a critical assessment. Reg. Stud. 39,
(Supplement).
61–74.
Griliches, Z., 1995. R&D and productivity: econometric results and measurement issues.
Boschma, N., Stam, E., Schutjens, V., 2011. Creative destruction and regional
In: Stoneman, P. (Ed.), Handbook of the Economics of Innovation and New
productivity growth: evidence from the Dutch manufacturing and services
Technology. Basil Balckwell, Oxford, pp. 52–89.
industries. Small Bus. Econ. 36 (4), 401–418.
Grimpe, C., Kaiser, U., 2010. Balancing internal and external knowledge acquisition: the
Boschma, R., Balland, P.A., Kogle, D.F., 2014. Relatedness and technological change in
gains and pains from R&D outsourcing. J. Manage. Stud. 47, 1483–1509.
cities: the rise and fall of technological knowledge in US metropolitan areas from
Grossman, G.M., Helpman, E., 1991. Quality ladders in the theory of growth. Rev. Econ.
1981 to 2010. Ind. Corpor. Change 24, 223–250.
Stud. 58, 43–61.
Bottazzi, L., Peri, G., 2003. Innovation and spillovers in regions: evidence from European
Ha, J., Howitt, P., 2007. Accounting for trends in productivity and R&D: a
patent data. Eur. Econ. Rev. 47, 687–710.
Schumpeterian critique of semi-endogenous growth theory. J. Money Credit Bank.
Bottazzi, L., Peri, G., 2007. The international dynamics of R&D and innovation in the
39, 733–774.
long-run and in the short-run. Econ. J. 117, 586, 511.
Hall, B., Jaffe, A., Trajtenberg, M., 2005. Market value and patent citations. Rand J.
Broekel, T., 2019. Using structural diversity to measure the complexity of technologies.
Econ. 36 (1), 16–38.
PLoS ONE 14 (5), e0216856.
Hidalgo, C.A., Hausmann, R., 2009. The building block of economic complexity. Proc.
Bucci, A., 2013. Returns to specialization competition population and growth. J. Econ.
Natl. Acad. Sci. U.S.A. 106, 10570–10575.
Dyn. Control 37 (10), 2023–2040.
Hewitt-Dundas, N., Gkypali, A., Roper, S., 2019. Does learning from prior collaboration
Bucci, A., Carbonari, L., Trovato, G., 2019. Variety, competition, and population in
help firms to overcome the ‘two-worlds’ paradox in university-business
economic growth: theory and empirics. CEIS Res. Pap. 16 (2), 2. No. 428.
collaboration. Res. Policy 48 (5), 1310–1322.
Canay, I.A., 2011. EL inference for partially identified models: large deviations
Jacobs, J., 1969. The Economy of Cities. Jonathan Cape, London.
optimality and bootstrap validity. J. Econ. 156 (2), 408–425.
Jaffe, A., 1986. Technological opportunity and spillovers of R&D: evidence from firms’
Capello, R., Lenzi, C., 2013. Territorial patterns of innovation: a taxonomy of innovative
patents, profits, and market value. Am. Econ. Rev. 76 (5), 984–1001.
regions in Europe. Ann. Reg. Sci. 51 (1), 119–154.
Jones, C.I., 1995. R&D-based models of growth. J. Polit. Econ. 103 (4), 749–784.
Capello, R., Lenzi, C., 2014. Spatial heterogeneity in knowledge, innovation, and
Jones, B.F., 2009. The burden of knowledge and the “death of the renaissance man”: is
economic growth nexus: conceptual reflections and empirical evidence. J. Reg. Sci.
innovation getting harder. Rev. Econ. Stud. 76 (1), 283–317.
54 (2), 186–214.
Laitner, J., Stolyarov, D., 2013. Derivative ideas and the value of intangible assets. Int.
Capello, R., Lenzi, C., 2015. Knowledge innovation and productivity gains across
Econ. Rev. (Philadelphia) 54 (1), 59–95.
European regions. Reg. Stud. 49 (11), 1788–1804.
LeSage, J.P., 1999. Spatial Econometrics. The Web Book of Regional Science. Regional
Charlot, S., Crescenzi, R., Musolesi, A., 2015. Econometric modelling of the regional
Research Institute. West Virginia University, Morgantown, WV.
knowledge production function in Europe. J. Econ. Geogr. 15 (6), 1227–1259.

13
C. Antonelli et al. Research Policy 51 (2022) 104081

Magalhães, M., Afonso, O., 2017. A multi-sector growth model with technology diffusion Raymond, V., Mairesse, J., Mohnen, P., Pal, F., 2015. Dynamic models of R&D
and networks. Res. Policy 46, 1340–1359. innovation and productivity: panel data evidence for Dutch and French
Marrocu, E., Paci, R., Usai, S., 2013. Productivity growth in the old and new Europe: the manufacturing firms. Eur. Econ. Rev. 78, 285–306.
role of agglomeration externalities. J. Reg. Sci. 53 (3), 418–442. Rigby, D., 2015. Technological relatedness and knowledge space: entry and exit of US
Martin, P., Ottaviano, G.I.P., 1999. Growing locations: industry location in a model of cities from patent classes. Reg. Stud. 49, 1922–1937.
endogenous growth. Eur. Econ. Rev. 43, 281–302. Rigby, D., Roesler, C., Kogler, D., Boschma, R., Balland, P.A., 2019. Do EU regions benefit
McGahan, A.M., Silverman, B.S., 2006. Profiting from technological innovation by from smart specialization? Pap. Evol. Econ. Geogr. 19.31.
others: the effect of competitor patenting on firm value. Res. Policy 35 (8), Rochowiak, D., 1988. Extensibility and completeness: an essay on scientific reasoning.
1222–1242. J. Speculative Philos. 2 (4), 241–266.
Moreno, R., Paci, R., Usai, S., 2005. Spatial spillovers and innovation activity in Rodríguez-Pose, A., Crescenzi, R., 2008. Research and development spillovers innovation
European regions. Environ. Plann. A: Econ. Space 37 (10), 1793–1812. systems and the genesis of regional growth in Europe. Reg. Stud. 42 (1), 51–67.
Nesta, L., 2008. Knowledge and productivity in the world’s largest manufacturing Romer, P., 1990. Endogenous technological change. J. Polit. Econ. 98 (5), S71–102 part
corporations. J. Econ. Behav. Organ. 67 (3–4), 886–902. 2.
Ottaviano, G.I.P., Thisse, J.F., 2004. Agglomeration and Economic Geography. In: Saviotti, P.P., Frenken, K., 2008. Trade variety and economic development of countries.
Henderson, J.V., Thisse, J.F. (Eds.), Handbook of Regional and Urban Economics: J. Evol. Econ. 18 (2), 201–218.
Cities and Geography. Elsevier, Amsterdam, pp. 2563–2608. Sequeira, T., Gil, P., Afonso, O., 2018. Endogenous growth and entropy. J. Econ. Behav.
Peretto, P., Smulders, J.A., 2002. Technological distance, growth and scale effects. Econ. Organ. 154, 100–120.
J. 112 (481), 603–624. Varga, A., 1998. University Research and Regional Innovation: a Spatial Econometric
Petralia, S., Balland, P.A., Morrison, A., 2017. Climbing the ladder of technological Analysis of Academic Technology Transfers. Kluwer Academic Publishers, Boston.
development. Res. Policy 46 (5), 956–969. Zachariadis, M., 2003. R&D innovation and technological progress: a test of the
Puga, D., 2010. The magnitude and causes of agglomeration economies. J. Reg. Sci. 50, Schumpeterian framework without scale effects. Can. J. Econ. 36, 566–586.
203–219. Weitzman, M.L., 1996. Hybridizing growth theory. Am. Econ. Rev. 86, 207–212.
Quatraro, F., 2010. Knowledge coherence, variety and economic growth: manufacturing Weitzman, M.L., 1998. Recombinant growth. Q. J. Econ. 113 (2), 331–360.
evidence from Italian regions. Res. Policy 39, 1289–1302. Wunsch, G.J., 1984. Theories, models, and knowledge the logic of demographic
Rahko, J., 2014. Market value of R&D, patents, and organizational capital: finnish discovery. Genus 1–17.
evidence. Econ. Innov. New Technol. 23 (4), 353–377.

14

You might also like