http://www.jrc.ec.europa.eu/ Knowledge for Growth – Industrial Research & Innovation (IRI)
REGIONAL CLUSTERS OF
ECONOMIC ACTIVITY IN
EUROPE: ARE SOCIAL AND
GEOGRAPHICAL PROXIMITY
THE KEY DETERMINANTS?
CONTRIBUTED PAPER FOR THE 2007 CONFERENCE ON
CORPORATE R&D (CONCORD)
R&D IN THE ECONOMY
File name: Author: Status: Last updated: Organisation: <PAPER_DE DOMINICIS ET AL.> <DE DOMINICIS L., FLORAX, R.J.G.M., DE GROOT, H.L.F.> <Draft> <28 August 2007> <VRIJE UNIVERSITEIT AMSTERDAM> Page 1 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? TABLE OF CONTENTS
1 ­ Introduction ...............................................................................................................3 2 ­ Data and exploratory analysis...................................................................................4 2.1 Measuring relational proximity ...............................................................................7 2.2 Global and local spatial autocorrelation.................................................................9 3 ­ Determinants of innovation activity in European regions: a knowledge production function approach ..............................................................................................................13 3.1 Knowledge spillovers: the role of geographical and social proximity ...................14 4 ­ Empirical results......................................................................................................15 5 ­ Conclusions.............................................................................................................19 References ........................................................................................................................20
Page 2 of 21
REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? 1 ­ Introduction Converting technological knowledge into economic growth and welfare is one of the keys for boosting the competitiveness of any country or region in the modern economy. Technological innovation is universally considered as an important driver for long run production and a necessary condition for sustainable economic growth. In knowledge based economies the competitive advantages of countries, regions and firms derive – among other factors – from their success in innovating. In 2005, the European Council has decided to re­launch the Lisbon Strategy. Knowledge and innovation have been declared one of the main areas for action in the new Lisbon partnership for growth and jobs. To achieve the goals in the agenda, research and innovation should be put at the heart of EU policies, EU funding and business activity. The production of innovation is closely related to factors both internal and external to the firms or institution in which they generate. The type of industry, its size, location, ownership type, they are all factors that affect the rate of innovation of a firm from the inside. Next to internal factors, there are external factors that are related to the environment in which the firms operate. These variables are responsible for creating the fertile soil ­ the so­called innovative milieu ­ which makes easier the road to innovation (Camagni, 1995). One of the paradigms of the knowledge based economy is the recognition that the diffusion of knowledge is just as significant as its creation, leading to increased attention to the concepts of knowledge spillovers, knowledge distribution networks, and national and regional systems of innovation. In a knowledge based economy, firms search for linkages to promote inter­firm interactive learning and for outside partners and networks to provide complementary assets. These relationships help firms to spread the costs and risk associated with innovation among a greater number of agents, to have access to new research results, and to acquire key technological components of a new product or process. Scholars stressed the importance that geography has in promoting the process of knowledge diffusion. Spatial proximity may facilitate learning processes through mechanisms of knowledge spillovers and especially sticky knowledge. However, it is not only geographical distance that matter. Boschma (2005) indicates at least four other forms of proximities that coexist with geographical proximity; namely, cognitive, organizational, social and institutional proximity. Recent theories of innovation and regional economic development recognize the importance of intangible factors in explaining the success of a firm, a region or a country. The socio­cultural context in which firms operate influence the propensity to turn ideas and inventions into new products and processes. Capello (2002) suggests that this socio­cultural context may be determined by what the author defines “relational proximity” or “relational capital”, a concept very close to the notion of social capital (Putnam, 1993) and which describes the pattern and intensity of networks among people and the shared values which arise from those networks. Social capital has been found to be an important determinant in explaining differences in European regional economic growth (Beugelsdijk and van Schaik, 2005), and innovation is an important channel through which social capital improves economic growth (Akçomak and ter Weer, 2006). Social capital is a community characteristic that facilitates the type of innovative,
Page 3 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? risk­taking behaviour that is essential to entrepreneurs to be innovative (Westlund and Bolton, 2003). Therefore, innovativeness and entrepreneurship can be seen as products of regions with social capital. Building on the literature on the knowledge production function (Griliches, 1979 and Jaffe, 1986) the aim of the paper is to explain the observed differences in the production of innovative output across European regions. Our main research question is whether geographical proximity and social (or relational) proximity are important veichles of tranmission driving the production of innovative output in Europe. The paper is organized as follows. Section 2 presents the data used for the empirical analyses, visualize their spatial distributions, and discover the underneath patterns of spatial association. The knowledge production function approach is introduced in section 3. We extend the basic specification of the knowledge production function to investigate the role that geographical and social proximity have on the creation of new ideas. Section 4 presents the empirical results. Section 5 concludes. 2 ­ Data and exploratory analysis The empirical analyses in this paper are based on a sample of 147 NUTS2 European regions, covering a total of 10 countries. 1 Where not differently stated, data are from the Eurostat REGIO database. Innovation activity is measured using patent applications to the European Patent Office. To allow for the size of an economy, we use the number of patents for 100.000. Averages over a three­year period (2000 to 2002) is used to smooth out transient effects and approximate long­run values (Griliches, 1979). The pitfalls associated with the use of equating patent applications to innovation activity are widely recognized. Researchers enumerate a number of faults with using patent data as a proxy for knowledge creation (Jaffe, 1989, Varga, 1997). One major drawback is that not all inventions are patented and not all patents have the same value. Another drawback is that only some of the patents granted are applied commercially and/or lead to major technological improvements. Patent data have the main advantage in that most countries have national patent systems organised on centralised databases, the data cover almost all technological fields, and patent documents contain a large amount of information concerning the invention, technology, inventor, etc. To conclude, they still provide a good proxy for the rate of innovation in that: “[…] In this desert of data, patent statistics loom up as a mirage of wonderful plentitude and objectivity. They are available; they are by definition related to inventiveness, and they are based on what appears to be an objective and only slowly changing standard” (Griliches 1990, p. 1661). Figure (1) shows the spatial distribution of patent applications across European regions. The figure gives a clear picture of the spatial concentration of patenting in a small number of regions, with a clear tendency towards a core­periphery structure. In general, Germany, Denmark, and the South­Eastern part of UK show the highest scores. These countries are leading in innovative capacity. On the other side, Mediterranean countries and the 1 Austria, Denmark, France, Germany, Ireland, Italy, Spain, Portugal, the Netherlands, and United Kingdom. Due to the lack of available data, analyses for the UK are at the NUTS1 level.
Page 4 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? Northern regions of the UK exhibit the smallest values. Only two regions in Italy (Emilia Romagna and Lombardia) are above the average. Europe_nuts2_03.shp ­1 ­ 0 Std. Dev. Mean 0 ­ 1 Std. Dev. 1 ­ 2 Std. Dev. 2 ­ 3 Std. Dev. > 3 Std. Dev. Figure 1: spatial distribution of patent applications (2000­2002) Data on Research and Development are used to establish the amount of research in each region. Expenditure in R&D represents one of the major drivers of economic growth in a knowledge based economy. Research and development spending is essential for making the transition to a knowledge based economy as well as for improving production technologies and stimulating growth. The variable we use is the total intramural R&D expenditure (GERD) as a percentage of GDP (so called research intensity). We consider sectoral aggregated expenditure as well we distinguish between expenditure in R&D performed by the private sector, the public sector and the higher education sector. There are significant differences among countries. Figure 2 shows that R&D intensity varies considerably between regions within individual countries and is often concentrated at a national level in a small number of regions. Especially in the private and the governmental sector, there is clear evidence of a spatial pattern in the distribution of the variable.
Page 5 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? Europe_nuts2_03.shp ­2 ­ ­1 Std. Dev. ­1 ­ 0 Std. Dev. Mean 0 ­ 1 Std. Dev. 1 ­ 2 Std. Dev. 2 ­ 3 Std. Dev. > 3 Std. Dev. Aggregated sectors Europe_nuts2_03.shp ­2 ­ ­1 Std. Dev. ­1 ­ 0 Std. Dev. Mean 0 ­ 1 Std. Dev. 1 ­ 2 Std. Dev. 2 ­ 3 Std. Dev. > 3 Std. Dev. Private sector Europe_nuts2_03.shp ­2 ­ ­1 Std. Dev. ­1 ­ 0 Std. Dev. Mean 0 ­ 1 Std. Dev. 1 ­ 2 Std. Dev. 2 ­ 3 Std. Dev. > 3 Std. Dev. Europe_nuts2_03.shp ­1 ­ 0 Std. Dev. Mean 0 ­ 1 Std. Dev. 1 ­ 2 Std. Dev. 2 ­ 3 Std. Dev. > 3 Std. Dev. Public sector Higher education sector Figure 2: spatial distribution of investments in R&D (1999­2001) The structural characteristics of the regional economy are a very important factor in explaining regional innovative capacity. Innovative capacity is higher in areas with a strong presence of high­technology industries (Audretsch, 1998, and Acs, 2002). Employment in medium­high and high­technology manufacturing sectors is an indicator of the manufacturing economy that is based on continual innovation through creative, inventive activity. The share of employment in medium­high and high­technology sectors is here used to capture potential localization economies. We also include in our analyses data on employment in the high­technology service sector. The high­technology service sector provides services directly to consumers, such as telecommunications, and provide inputs to the innovative activities of other firms in all sectors of the economy. The latter can increase productivity throughout the economy and support the diffusion of a range of innovations, in particular those based on ICT. Later, when presenting the model specification, we will describe a number of other variables that in our opinion are able to explain the observed variation in the distribution of innovative activity across European regions.
Page 6 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? 2.1 Measuring social proximity Measures of social capital are not without controversy. 2 There is no widely held consensus on how to measure social capital, which is one of its weaknesses. The most common source of information for measuring social capital across European regions is the European Values Survey, which contains a number of questions that can be used to assess social capital. However, this survey is not conducted every years, and geographical aggregation are possible only at the NUTS1 level. Differently, we introduce a measure of social capital based on information contained in the Standard Eurobarometer Survey. The Standard Eurobarometer Survey is a cross­national longitudinal study designed to compare and gauge trends within member states of the European Union. This database offers several advantages: it covers the whole of the European Union, it is conducted twice per year and it is the only survey at the European level where individual respondents are coded at NUTS2 geographical level of aggregation. Although the range of questions has been expanded over the years, the programme aims to keep most of the survey constant, so that data is comparable over time. Our indicators are derived using information contained in the Standard Eurobarometer Survey. The Standard Eurobarometer is a series of surveys regularly performed to monitor the public opinion in the European Union. The surveys are conducted on behalf of the European Commission at least two times a year in all member states of the European Union. Since the early 1970s they are providing regular information on social and political attitudes in the European publics. We created regional average values for five indicators gathered from two editions of the survey. The first three indicators come from the Eurobarometer Survey 55.1 (EB_55.01) carried out between April and May 2001. The first indicator, OPINION LEADERSHIP, is based on the answers to the following two questions: “When you, yourself hold a strong opinion, do you ever find yourself persuading your friends, relatives or fellow workers to share your views? If so, does this happen often, from time to time or rarely?” and “When you get together with your friends, would you say you discuss political matters frequently, occasionally or never?”. The variable forms an indicator of the individual’s potentials to take an active and leading role in the political science. A good leadership is required to achieve the coordination required to benefit from social capital (Durlauf and Fafchamp, 2005). This indicator takes values from 1 to 4 with an increasing intensity of the leadership. The second indicator, DAILY NEWSPAPER USE, measures the level of newspaper readership and is a mark of interest of the individuals in community life. Researchers report a positive relationship between commercial newspaper readership and social capital (Putnam, 1993). We build this variable using the question “About how often do you read the news in daily papers?” The variable takes values from 1 to 5 with a decreasing intensity in readership. To allow comparability with the other indicators of social proximity, we recoded this variable from 1 to 5 with an increasing intensity of readership. The third variable, LIFE SATISFACTION, is based on the question “On the whole, are you very satisfied, fairly satisfied, not very satisfied or not at all satisfied with the life you lead? Would you say you are...?” giving the choice to the respondent to answer on a scale from 1 to 4 whether they are very satisfied, fairly satisfied, not very satisfied, and not at all satisfied. As before, to allow for comparability we recoded the variable from 1 to 4 with an increasing level of satisfaction. 2 In this paper, we will use interchangeably the terms social capital and social proximity.
Page 7 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? A second group of variables is considered to measure the attachment of individuals to their own town or region. The sense of community is a psychological construct and a correlate of social capital (Pooley et al. 2005). This variables reflects the feelings of attachment and belonging that an individual has towards a community. To gather data, we relied on the Eurobarometer survey 54.01 (EB_54.1) carried out between November and December 2000. In two separate queries the respondents are asked to answer to the question “People may feel different degrees of attachment to their town or village, to their region, to their country or to Europe. Please tell me how attached you feel to your region (country)?”. Answers are coded from 1 to 4 with a decreasing intensity of attachment to the territory. We recoded the answer from 1 to 4 with an increasing level of attachment. The five variables were processed with principal component analysis suggesting the use of two factors. The first factor, SOC1, includes the three variables indicating the level of participation and satisfaction of the individuals to social and civic life. The second factor, SOC2, contains the remaining two variables reflecting the attachment of the individuals to the territory. The highest values of the variable SOC1 are found in the regions Noord­ Holland, Groningen, Noord­Brabant, and Flevoland in the Netherlands, Trentino Alto Adige in Italy, and Münster and Weser­Ems in Germany. The lowest values are in the regions Alentejo and Algarve in Portugal, Haute­Normandie and Picardie in France, and Murcia and Castilla­La Mancha in Spain. For the indicator associated to the level of attachment to the local territory, SOC2, the highest values are found in the regions Friesland, Gelderland, and Flevoland in the Netherlands, Luxembourg in Belgium and in the Ile de France. We observe the lowest levels of attachment to the local territory in the regions Oberpfalz and Trier in Germany, Burgerland in Austria and Galicia and Baleares in Spain. Figure 3 below shows the spatial distribution of the two variables across European regions. We observe a rather scattered picture for the variable associated to the attachment to the local territory, with a tendency to cluster across regions in France, the Netherlands, and Northern Italy. The first indicator of social proximity shows a stronger tendency to concentrate in space, in particular across regions in Germany, Denmark, the Netherlands and Northern Italy. Europe_nuts2_03.shp ­3 ­ ­2 Std. Dev. ­2 ­ ­1 Std. Dev. ­1 ­ 0 Std. Dev. Mean 0 ­ 1 Std. Dev. 1 ­ 2 Std. Dev. 2 ­ 3 Std. Dev. Europe_n uts2_03.shp ­3 ­ ­2 S td. Dev. ­2 ­ ­1 S td. Dev. ­1 ­ 0 Std. Dev. Mean 0 ­ 1 Std. Dev. 1 ­ 2 Std. Dev. 2 ­ 3 Std. Dev. > 3 Std. Dev. Figure 3: spatial distribution of social capital: participation to and satisfaction in social life SOC1 (left) and sense of attachment to the territory SOC2 (right)
Page 8 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? 2.2 Global and local spatial autocorrelation This section analyzes in more detail the spatial distribution of patents and investments in R&D using Exploratory Spatial Data Analysis (ESDA), which has been defined as a set of “techniques to describe and visualize spatial distributions, discover patterns of spatial association, suggest different spatial regimes or other forms of spatial instability and identify atypical observations or outliers” (Anselin, 1995). Central to ESDA is the analysis of spatial association or spatial autocorrelation between observations. Positive spatial autocorrelation occurs when high or low values of a variable tend to cluster together in space and negative spatial autocorrelation when high values are surrounded by low values and vice­versa. A crucial issue in the definition of spatial autocorrelation is the notion of “location similarity”, or the determination of those locations for which the values of the variable are correlated. This is formally expressed in a spatial weight matrix. The nature of the spatial interaction may be defined in several ways, such as simple contiguity (i.e. common border), distance contiguity, inverse distance (to account for distance­decay effects). Both these weights are closely linked to the physical feature of the spatial units on a map. When the spatial interaction is determined by factors linked to economical variables, authors have proposed the use of weights with a more direct relation to the particular phenomenon under study (i.e. travel time, social or economical distances). 3 We use a spatial weight matrix based on the inverse of the squared distance between pair of locations, with critical cut­off points at the first quartile of the arc­distance distribution, which reads as: 4
ìïw =
W = í i , j
ïîw i , j =
2 -1 i,j
(d )
0
if d i , j < Q(1) (1) otherwise where di,j is the distance between centroids of region i and region j, and Q(1) is the cut­off point at the first quartile of the arc­distance distribution. The critical cut­off distance implies that we expect spatial interaction above this distance to be negligible. A standard measure to check for the presence of spatial autocorrelation is the Moran’s I statistics (Moran, 1950). Formally, for each variable of interest, the Moran’s I is:
n
I =
åå w ( x
ij
N n
i =1 j =1
i
- x ) ( x j - x )
i =1 j =1 n
åå w
n (2) n ij
å(x
i - x )
2 i =1 where N is the sum of observations, wij is the element in the spatial weight matrix corresponding to the observation pair i, j (with i ¹ j ), xi and xj are observations for the 3 It is important not to forget that the standard estimation and testing approaches assume the weight matrix to be exogenous. Therefore, indicators for the socioeconomic weights should be chosen with great care to ensure the exogeneity, unless their endogeneity is considered explicitly in the model specification (Anselin and Bera, 1998). 4 Other types of spatial weight matrix – contiguity matrix and inverse distance matrix – have been used and they produced similar results.
Page 9 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? locations i and j (with mean x), and the first term is a scaling constant. The former is the traditional approach to spatial autocorrelation, in which the overall pattern of dependence is summarized into a single indicator. Values of I larger than the expected value
E ( I ) = -1 ( n - 1 ) indicate positive spatial autocorrelation and vice­versa. 5 Table 1 lists the Moran’s I statistic and associated z­ and p­values for five variables: Patent applications (2000­2002), and research and development intensity in the period 1999 to 2001 as aggregate and also distinguishing between the private, public and higher education sectors. In three out of five cases the z ­values for Moran’s I are positive and significant, indicating the presence of positive spatial autocorrelation. The highest level of spatial autocorrelation is found in the variable that measure patent applications. An interesting result is in the spatial autocorrelation in the R&D intensity for the different sectors. As for research intensity, only research efforts made in the private and the public sectors appears to be spatially correlated. In particular, the Moran’s I coefficient relative to the private sector is rather high if compared to the one for the public sector, corroborating the hypothesis the firms tend to cluster in space, taking advantage of the presence of localization economies (Marshall, 1890). VARIABLE Moran’s I ST.DEV. Z­VALUE PROB Patent (ln) 0.721 0.039 18.679 0.000 R&D (ln) 0.208 0.039 5.498 0.000 R&D private (ln) 0.384 0.039 10.073 0.000 R&D higher education (ln) 0.030 0.038 0.969 0.332 R&D government (ln) 0.083 0.039 2.286 0.022 Table 1: Moran’s I measure of spatial autocorrelation Figure 4 shows the Moran’s I scatterplot map for the three variables for which we found a significant value of positive spatial autocorrelation. The Moran’s I scatterplot map provides a visual exploration of global spatial autocorrelation. in which the global Moran’s I is decomposed into four categories. These four categories identify four types of spatial association between a location and its neighbours. Two of these categories imply positive spatial association: the first one where a location with an above­average value is surrounded by neighbours whose values are also above average (high­high, in red), or where a location with a below­average value is surrounded by neighbours whose values are also below average (low­low, in pink). The other two categories imply negative spatial association: the first category where a location with an above­average value is surrounded by neighbours with below average values (high­low, in dark blue), or where a location with a below­average value is surrounded by neighbours with above average values (low­high, in light blue). This map is the visual counterpart of the Moran’s I scatterplot graph frequently used in papers performing exploratory spatial data analysis. We first comment the map for patent applications. The Moran’s I statistic in table 1 already indicated a high degree of spatial autocorrelation. The inspection of the map reveals one interesting result: a significant part of Mediterranean countries belong to the low­low category, while the high­high category is monopolized by regions in central and northern Europe. A similar pattern is found also in the maps relative to expenditure in R&D, although with a more 5 Under the null hypothesis of absence of spatial autocorrelation.
Page 10 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? scattered pattern (also confirmed by the low level of global spatial autocorrelation in table 1). Q_PATAVG High­High Low­Low High­Low Low­High Patent applications 2000­2002 Q_R_D High­High Low­Low High­Low Low­High R&D expenditure (aggregated sectors) Q_RDPR High­High Low­Low High­Low Low­High R&D expenditure (private sector) Figure 4: Moran scatterplot maps The traditional Moran’s I measure of spatial autocorrelation is global, in a sense that it captures the overall spatial pattern in the data and summarizes it in a single statistic. While global measures allow us to test for spatial patterns over the whole study area, it may be the case that there is significant autocorrelation in only a smaller section. A further problem with global measures of spatial autocorrelation is that – when positive and significant – they are not able to distinguish between situations where the index is determined by closeby positive values or by closeby negative values. On the contrary, the local indicators of spatial association (or LISA; Anselin 1995) are designed specifically to find evidence of local spatial patterns in the empirical data. In what follows, we measure local spatial dependence using the local version of the Moran’s I statistic described before. The local Moran’s I produces a measure of spatial autocorrelation for each individual location and is designed to test whether the distribution of values around that specific location deviates from spatial randomness. Local indicators of spatial association can be used for the detection of significant local spatial clusters (also called ‘‘hot spots’’) as well as for diagnostics of local instability, significant outliers and spatial regimes. The use of local indicators of spatial association offers two main advantages in the analysis of the spatial distribution of economic activities, namely [i] they provide precise information on the exact
Page 11 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? location of the identified innovation clusters and [ii] they allow to assess the statistical significance of the local patterns identified. The local Moran statistics for an observation i is defined as (Anselin 1995):
Ii = zi å j ¹i w ij z j (3) For ease of interpretation, the weights wij are row standardized and by convention the elements on the main diagonal are set equal to zero. As before, the spatial ordering is defined using the squared inverse distance with cut­off points at the first quartile of the arc­ distance distribution. Figure 5 is map showing the regions where the local Moran’s I is significant. M_PATAVG not significant High­High Low­Low Figure 5: Local Moran’s I scatterplot map of patent applications (2000­2002) We observe a significant hot­spot with a strong territorial component in Central Europe. Almost all regions in the cluster belong to Germany, with the exception of 2 regions respectively from the Netherlands (region Noord­Brabant) and France (region Alsace). To summarize, our exploratory analysis revealed the presence of a strong spatial pattern in the production of innovative output – as well in investments in research and development – across European regions. In the next section, we further explore this spatial component. We introduce the standard model of knowledge production and we extend the basic formulation in a way that incorporates proximity, as geographical proximity as well social proximity.
Page 12 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? 3 ­ Determinants of innovation activity in European regions: a knowledge production function approach The concept of the knowledge production function (KPF) has been introduced by Griliches (1979) for measuring the impact of R&D and knowledge spillovers to productivity growth. Taking the logarithmic transformation of a Cobb­Douglas production function as a framework, the basic formulation of the KPF links knowledge inputs to innovative outputs according to:
n ( RD output ) = ln a + ln ( RD input )
(4) Jaffe (1989) modifies the Griliches’ knowledge production function by using a measure of innovation – such as patents or new product introductions – as the dependent variable, and industry and university research and development expenditures as two independent variables:
n ( Pi ) = b 0 + b1 ln(RDI ,i ) + b 2 ln(RDU ,i ) + b 3 éëln(RDU , i )ln(C r ) ùû + e i (5) Where P is a measure of innovation output in region i, RDI is research and development performed by the industry in regions i, RDU is university research in the region and C is a measure of geographic coincidence of university and industrial research. Variants of the model in equation (5) have been proposed in the literature, considering factors other than research and development in industries and universities as determinants of the creation of new knowledge. 6 In this paper we extend the formulation in equation (5) to include social and geographical proximity as important factors in explaining the observed differences in innovation output across European regions. The basic specification the model we want to estimate is:
n ( Pi ) = b 0 + b1 ln(RDi ) + b 2 ln(SPi ) + b ln( Z i ) + e i (6) Where the dependent variable is the natural logarithm of patent applications to the European Patent Office observed in the NUTS2 region i, β0 is a constant, RD is expenditure in Research and development in the region, SP is a set of variables measuring social capital in region i , and Z is a vector of regional economic characteristics. Expenditure in research and development represents one of the major drivers of economic growth in a knowledge based economy. The variable is expressed as total intramural expenditure in R&D as percentage of regional GDP (R&D intensity). We consider expenditure in R&D at the aggregate level, as well we distinguish between investment in R&D performed by the private sector, the higher education sector and the governmental sector. As before, we assume a time lag between investments in research and the production of new ideas. 7 6 See Moreno et al. (2005) for a recent analysis of the spatial distribution of innovation activities and its determinants in Europe. 7 Time series of regional data on expenditure in R&D in Europe are very sparse. Therefore, we use averages calculated over the period 1999 to 2001. For Belgium we use data at the NUTS1 level.
Page 13 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? We already mentioned that the propensity to patent differs between types of industry. Structural regional characteristics are included in the model using employment in the high­ tech manufacturing sector. These sectors are viewed as the most innovative within the manufacturing economy (Audretsch, 1998). The proportion of the workforce employed in these fields is an indicator of the capacity of the economy as a whole to exploit the results of R&D and innovation. We also include a variable to control for the percentage of employment in the high­tech service sector. The latter can increase productivity throughout the economy and support the diffusion of a range of innovations, in particular those based on ICT. As we assume that innovation needs time to be produced, we use data for 2000, assuming a time lag of two years between innovation input and innovation output. A central characteristic of a knowledge economy is continual technical development and innovation. Under these conditions, individuals need to continually learn new ideas and skills ­ or to participate in life­long learning. The current interest and support for life long learning is based on the belief that continuous learning is required to address several economic, technological and demographic changes in modern economies. These include a shift in the modes and distribution of knowledge due to information and communication technologies (ICT), a shift in employment from manufacturing to services, and a speeding up of the rate of technical change. For this reason, the relationship between life long learning and innovation should be of great interest to innovation policy, although to date this relationship has attracted considerably less attention than the benefits of life long learning to social goals such as reducing social exclusion and income inequality, or counteracting demographic ageing. We include in our estimation a variable that measures the participation in long­life learning as a percentage of 25­64 years age class. There is a long standing debate about the effectiveness of EU funding programmes in promoting the economic development of European lagging regions. The main objectives of the Structural Fund programmes in the 1994­99 period under Objective 1 were those of reducing the disparities in GDP and unemployment between the regions of Europe. Incentives in investments in R&D have been emphasized in countries like Austria and the Netherlands. Among the strategic aims of the 1994­99 Objective 2 programmes, job creation was the most common overall objective. Strategies have mainly been focused on the types of intervention used by regions tackling industrial decline and re­conversion. This has included support for the business environment (mainly aid to business for industrial investments and business infrastructure), investment in infrastructure, land recovery, environmental protection, and human resources development. Many programmes have also included interventions for R&D and technology transfer. Previous results indicate that EU funding has no direct effect on economic outcomes (Akçomak and ter Weer, 2007). The variable STRUCTURAL FUND is the total Structural Fund expenditure related to Regional Development and Productive Infrastructure (Obj. 1, 2 and 6 ERDF) in the period 1994­1999. Data are from the ESPON Database (2006). Finally, we include per capita regional GDP to assess the possible presence of some sort of local market effect. 3.1 Knowledge spillovers: the role of geographical and social proximity Information related to innovative output flows more easily when agents are located at a close distance, thanks to frequent face­to­face interaction and to social bonds that foster reciprocal trust (Breschi and Lissoni, 2001). Potential knowledge spillovers are included in
Page 14 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? the knowledge production function in equation (4) adding an additional regressor which measure the level of innovativeness in neighbouring regions. This approach assumes that the production of new ideas in one region does not only depend on the values of the explanatory variables in that region, but is influenced as well by the level of innovativeness in closeby regions, subject to decay distance.
ln ( Pi ) = b 0 + rW ln(Pj ) + b1 ln(RDi ) + b 2 ln(SPi ) + b ln(Z i ) + e i (7) Where W is a pre­defined spatial weight matrix which provides the structure of the assumed spatial relationship between regions i and j (with i≠j). Estimation of the model in equation (7) cannot be performed using Ordinary Least Squares, due to the presence of the spatially­lagged dependent variable on the right hand side, which is endogenous and therefore correlated with the error term εi. Instead, the model can be estimated using maximum likelihood or instrumental variables estimation (Anselin, 1988). The inclusion of the two measures of social proximity introduced above allows controlling for the impact of social proximity on innovation. 4 ­ Empirical results The results of the estimation of the knowledge production function equation are shown in Table 2. The dependent variable is the natural logarithm of number of patent applications to the European Patent Office in each NUTS2 region. The number of patent applications is an average of three years’ data (2000 to 2002), to smooth out possible transient effects. We assume that innovative output needs some time to be produced. In order to appropriately model the relationship between innovation input and output, the following input variables enter into the model with a time lag of around 2 years: employment in high­ tech manufacturing and service sector, expenditure in research and development, percentage of workers participating in long­life learning. The values for these variables are observed in the year 2000. 8 Information on structural funds are extracted from the ESPON Database (2006). Data refer to Structural Fund expenditure related to Regional Development and Productive Infrastructure (Objective 1, 2 and 6 ERDF) over the period 1994­1999. Gross domestic product is measured in 2002 and gives a measure of the size of the economy. Finally, the two indicators of social capital are observed in the period 2001 (SOC2) and 2000 (SOC1). All explanatory variables are expressed in natural logarithms. We start estimating a traditional KPF equation where no spatial effects are included. Columns (1) and (2) report the OLS results. In column (1) we control for the impact of expenditure in R&D on innovation without distinguishing among private, public and higher education sectors. In column (2), we verify the impact on innovation of R&D investments in distinguishing between private, public, and higher education sectors. Commenting the results, we find that participation in long­life training does not seem to be a net contributor in the innovation process. 8 Due to problems of data availability, data for expenditure in R&D are averages over the period 1999­2001
Page 15 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? Table 2: Regression results Constant (1) (2) (3) (4) (5) –17.565 *** (2.72) –14.975 *** (2.586) –0.015 (0.051) 0.631 *** (0.102) 0.089 (0.165) 1.588 *** (0.293) 0.458 *** (0.100) –0.024 (0.049) 0.399 *** (0.107) 0.102 (0.158) 1.369 *** (0.278) –10.234 *** (2.159) 0.520 *** (0.056) 0.068 * (0.039) 0.423 *** (0.081) –0.013 (0.124) 0.906 *** (0.227) 0.382 *** (0.075) –9.014 *** (2.087) 0.473 *** (0.054) 0.052 (0.037) 0.283 *** (0.083) 0.004 (0.121) 0.817 *** (0.221) –8.661 ** (4.039) 0.309 * (0.167) 0.064 (0.072) 0.520 *** (0.157) –0.361 (0.278) 0.355 (0.475) 0.333 ** (0.139) Spatial lag patent Long life learning) High­tech manufacturing High­tech service GDP R&D R&D private –0.038 *** (0.012) 2.436 *** (0.571) 0.235 (0.349) 147 0.80 28.116 *** 9.208 64.450 ** 0.477 *** (0.068) –0.027 (0.047) –0.043 (0.076) –0.032 *** (0.012) 2.654 *** (0.533) 0.191 (0.327) 147 0.82 9.285 *** 13.811 71.206 6.907 *** 32.824 *** 2.055 81.580 *** 50.809 *** 7.398 *** 35.874 *** 5.714 ** 70.816 *** 40.656 *** R&D public R&D higher education Structural funds SOC1 SOC2 Obs Adj. R­squared Jarque­Bera Koenker Bassett White Moran’s I (error) LM (error) Robust LM (error) LM (lag) Robust LM (lag) Likelihood ratio (lag) –0.022 ** (0.010) 1.207 *** (0.450) –0.206 (0.263) 147 0.361 *** (0.054) 0.001 (0.036) –0.027 (0.058) –0.020 ** (0.009) 1.457 *** (0.432) –0.193 (0.252) 147 71.225 *** 63.544 *** –0.052 ** (0.024) 3.649 * (1.898) 3.962 * (2.070) 147 (6) –8.563 *** (2.384) 0.424 *** (0.079) 0.052 (0.044) 0.277 *** (0.096) –0.074 (0.148) 0.626 ** (0.262) 0.372 *** (0.063) –0.007 (0.041) –0.055 (0.068) –0.025 ** (0.012) 2.503 *** (0.888) 0.806 (0.732) 147 The dependent variable is the natural logarithm of patent application as defined in the text. The dependent variables are expressed in natural logarithmic form. Standard errors in parenthesis. The statistic significance of the parameters is indicated by ***, **, *, referring respectively to the 1%, 5% and 10% level. Instruments in (5) and (6): percentage of population aged over 65 and human resources in science and technology.
Page 16 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? Regression results show that the creation of new ideas is more likely to occur in regions where the percentage of employment in medium­high and high­tech manufacturing industries is high. The estimated coefficient is positive and highly significant. We start estimating a traditional KPF equation where no spatial effects are included. Columns (1) and (2) report the OLS results. In column (1) we control for the impact of expenditure in R&D on innovation without distinguishing among private, public and higher education sectors. In column (2), we verify the impact on innovation of R&D investments in distinguishing between private, public, and higher education sectors. Commenting the results, we find that participation in long­life training does not seem to be a net contributor in the innovation process. Regression results show that the creation of new ideas is more likely to occur in regions where the percentage of employment in medium­high and high­ tech manufacturing industries is high. The estimated coefficient is positive and highly significant. The results for the impact of the high­tech service sector on innovation does not support our hypothesis that service sector may help to create a fertile ground for innovation. Investment in research and development is one of the most decisive inputs in the knowledge production function. We find that not all research sectors are equally productive in terms of innovation production. Investments in research performed by the private sector seem to be the only ones able to move into new patented products. A plausible explanation can be found in the fact that research performed in the private sector is in general commercial­oriented and companies often try to protect the results through patenting. On the opposite, research in public and higher education sectors is usually less applied, resulting in a weaker (and in our case insignificant) impact on the number of new patents (Bilbao­Osorio and Rodriguez­Pose, 2004). As for the structural funds, our results confirm previous findings that subsidies from the EU Commission does not always translate into regional capacity to innovate. However, we should take this results with caution, before affirming that EU funds are not effective. If we take into consideration the fact that structural funds are mostly devoted to lagging regions, it is natural to think that these regions may need some time to fill the gap with the technological leaders and start being competitive. We find a significant value of social capital on innovation only for the variable measuring the participation in social life and the level of satisfaction. This confirms the hypothesis that innovation is a product of regions with high levels of social capital. Spatial diagnostics are presented in columns (1) and (2). The tests are computed with a distance based spatial weight matrix as defined before. The Moran’s I on the residuals is positive and significant, indicating that there is some form of spatial dependence in the data that deserve to be taken into account in the estimations. The results of the Lagrange Multiplier tests are used to decide whether a spatial lag or a spatial error model of spatial dependence is the most appropriate to control for the presence of spatial dependence in the OLS residuals. Following the decision rule suggested in Anselin and Florax (1995), if LM­Lag is more significant than LM­Error and robust LM­Lag is significant but robust LM­ Error is not (or is less significant), then the appropriate model is the spatial lag model. Conversely, if LM­Error is more significant than LM­Lag and robust LM­Error is significant but robust LM­Lag is not (or is less significant), then the appropriate specification is the
Page 17 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? spatial error model. After applying the rule we conclude that in both models in columns (1) and (2) a spatial lag­based specification should be preferred. The spatial lag model extends the basic model in columns (1) and (2) to add the spatially lagged dependent variable. Furthermore, this variable can be interpreted as a measure of knowledge spillovers coming from neighbouring regions. Spatial­lag models can be estimated using maximum likelihood or instrumental variables estimation methods. Estimation via maximum likelihood requires normality of the error term. Looking at the diagnostic tests, the Jarque­Bera test rejects the assumption of normality of the residuals (p–value<0.001). Therefore, columns (3) and (4) report the results for the spatial lag model estimated via maximum likelihood only for a sake of comparison. We will focus and comment in detail the results obtained using 2SLS instrumental variables estimation in columns (5) and (6). Two sets of instruments will be considered. The choice of optimal instruments is an important consideration. The first set of instruments is used to deal with the endogeneity of the spatially lagged dependent variable. The application of instrumental variables to the spatial lag model was initially outlined in Anselin (1988) where the use of the spatially lagged exogenous variables is suggested as instrument for the spatially lagged dependent variable. A second set of instruments is used to deal with the measurement error in the variables that measure the regional level of social capital. Social capital is a broad term encompassing the social norms and networks facilitating collective action for mutual benefit. .Empirical measures of social capital are not without problems. In column (5) and (6) we instrument the social capital variables – SOC1 and SOC2 – to correct for possible measurement error. Specifically, we instrument for social capital through some of the variables that have been found to be related to social capital in the work of Glaeser et. al (2000), viz. the educational level and the percentage of the percentage of population older than 65. education levels are not available at the regional level. We may have used the number of student enrolled in the tertiary education, but this is a measure of the potential education level of region. Therefore, we use data on human resources in science and technology as a proxy for the regional level of education. They are defined as individuals who fulfil at least one of the following conditions: [1] have successfully completed tertiary­ level education in a Science and Technology field of study and/or [2] work in a Science and Technology occupation as professionals or technicians. Our assumption is that the older is the regional population, the lower is the incentive in investing in social capital. As for the education level, in most countries high social capital is often associated with years of formal education. Even holding constant other factors, including race, income, gender, ethnicity, occupation, and many others, more educated people have wider, deeper, stronger social networks and participate more in social, community, and political life. Instrumenting for SOCI and SOC2 with these two variables and all of the other right­hand­ side variables, SOC1 remain a significant predictor for innovation, and in column (5), also the coefficient associated to SOC2 is significant. Also the signs are as expected, corroborating our initial hypothesis that social proximity is a major force in the process of creation of a knowledge­based economy, and that innovation is indeed a product of regions with social capital.
Page 18 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? 5 ­ Conclusions The paper has focused on geographical proximity and social proximity as key factors in explaining differences in the production of innovation across European regions. Our exploratory analysis showed evidence of the existence of drastic differences in the production of innovation across European regions. Building on the knowledge production function literature, we estimated a model in which, among other factors, geographical proximity and social proximity are controlled for in the estimations. As in previous studies, employment in high­tech industries, and investments in research and development in the private sector are important factors in explaining why some regions innovate more than others. We found that geography matters fro innovation. Regions surrounded by other regions that outperform in terms of innovation are likely to show a high capacity to introduce new products or processes. We found that regions with high levels of social capital are also characterized by a high number of patent applications. This last observation has important policy implications. If the amount of social capital available in one regions is able to promote innovation, there is need of policy decisions addressed to encouraging associative activities among the business community, fostering links between institutions (i.e. university/ research institutions and the private sector), and encouraging linkages among companies, between industries and between firms and supporting institutions.
Page 19 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? References Acs (2002) Innovation and the Growth of Cities. Northampton, MA: Edward Elger. Acs, Z.J., Anselin, L., and Varga, A. (2002) Patents and Innovation Counts as Measures of Regional Production of New Knowledge, Research Policy, 31: 1069–85. Akçomak, S. and ter Weer, B. (2006) Social Capital, Innovation and Growth: Evidence from Europe, UNU­MERIT Working Paper Series no. 2006­040. Akçomak, S. and ter Weer, B. (2007) How Do Social Capital and Government Support Affect Innovation and Growth? Evidence from the EU regional Support Programmes, UNU­MERIT Working Paper Series no. 2007­009. Anselin, L. (1988) Spatial Econometrics: Methods and Models: Dordrecht: Kluver: Anselin L (1995) Local Indicators of Spatial Association – LISA, Geographical Analysis 27: 93–115. Anselin, L. and 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, pp. 237­289. New York: Marcel Dekker. Anselin L, and Florax RJGM (1995) Small Sample Properties of Tests for Spatial Dependence in Regression Models. In Anselin L, Florax R (eds.) New Directions in Spatial Econometrics. Springer­Verlag: Berlin. Audretsch, D.B. (1998) Agglomeration and the Location of Innovative Activity, Oxford Review of Economic Policy, 14(2): 18–29. Bilbao­Osorio, B. and Rodríguez­Pose, A. (2004) From R&D to Innovation and Economic Growth in the EU, Growth and Change, 35(4): 434–55. Boschma, R.A. (2005), Proximity and Innovation. A Critical Assessment, Regional Studies, 39(1): 61–74. Breschi, S. and Lissoni, F. (2001) Knowledge Spillovers and Local Innovation Systems: A Critical Survey, Industrial and Corporate Change, 10(4): 975–1005 Beugelsdijk, S., and van Schaik, T. (2005) Social Capital and Growth in European Regions: An Empirical Test, European Journal of Political Economy, 21: 301–24 Camagni R. (1995) The concept of innovative milieu and its relevance for public policies in European lagging regions, Papers in Regional Science 74(4): 317–40. Capello, R. (2002) Spatial and Sectoral Characteristics of Relational Capital in Innovation Activity, European Planning Studies, 10(2): 177–200. Durlauf, S.N., and Fafchamps, M.(2005) Social Capital. In Philippe Aghion and Steven Durlauf (Eds.), Handbook of Economic Growth, volume 1: Elsevier. Griliches, Z. (1979) Issues in Assessing the Contribution of Research and Development to Productivity Growth, Bell Journal of Economics, 10: 92–116. Griliches, Z. (1990) Patent Statistics as Economic Indicators: a Survey, Journal of Economic Literature, 28: 1661–707. Jaffe, A.B., (1986) Technological Opportunity and Spillovers of R&D: Evidence from Firms' Patents, Profits, and Market Value, American Economic Review, 76(5): 984–1001. Jaffe A.B., (1989) Real Effects of Academic Research, American Economic Review, 79(5): 957­70. Marshall A (1890) Principles of economics. MacMillan: London. Moran P.A.P. (1950) Notes on continuous stochastic phenomena, Biometrika 37: 17–23. Moreno, R., R. Paci, et al. (2005) Innovation Clusters in European Regions, CRENOS Working Papers, no. 2005/12. Pooley, J.A., Cohen L., and Pike, L.T. (2005) Can sense of community inform social capital?, The Social Science Journal, 42 (1).
Page 20 of 21 REGIONAL CLUSTERS OF ECONOMIC ACTIVITY IN EUROPE: ARE GEOGRAPHICAL AND SOCIAL PROXIMITY THE KEY DETERMINANTS? Putnam (1993) Making Democracy Work: Civic Traditions in Modern Italy. Princeton University Press: Princeton. Varga, A. (1997) Regional Economic Effects of University Research: A survey, Unpublished manuscript, West Virginia University, Regional Research Institute, Morgantown, WV. Westlund, H. and Bolton, R. (2003) Local Social Capital and Entrepreneurship, Small Business Economics 21(2): 77–113.
Page 21 of 21