Geographic proximity and firm–university innovation linkages: evidence from Great Britain Purchased

1. Introduction

This article provides new evidence on the role of geographic proximity in firm–university innovation linkages for Great Britain and investigates the existence of spatially mediated knowledge transfer from university research. There has been wide policy and academic interest in the impact that higher education institutions have on innovative activity in their host regions. Differences in innovation rates are often considered an important factor underlying regional disparities in economic performance and constitute a policy concern in the UK and elsewhere (HM Treasury, 2001). Indeed, some countries have implemented regional innovation policies based on the presence of universities. The UK government has emphasized interaction between research institutions and business, and the role of geographic innovative clusters in improving innovation performance (HM Treasury, 2004), and many countries provide financial incentives to foster collaboration between firms and universities.

In this context, we investigate two related research questions. First, whether firms locate their research and development (R&D) facilities near to university research departments, and second, conditional on location, whether R&D-doing firms situated closer to university research are more likely to engage with universities, either through formal collaborative research agreements or more informal knowledge sourcing. We combine novel data on the location of firms’ R&D in Great Britain and information on R&D-doing firms’ interactions with universities, with measures of the presence and quality of university research in relevant subject areas at a high level of geographic dis-aggregation. This allows us to examine the role of geographic proximity using continuous spatial measures, enabling a better understanding of how physical distance affects firm–university interactions.

We find evidence that pharmaceutical firms locate R&D near (within 10 km) to world-class rated chemistry research departments, consistent with the importance of accessing localized knowledge. But in this industry and others the location of R&D also appears to be linked to the presence of science parks. Many UK R&D-intensive and science-based start-ups including university spin-outs locate in science parks, which aim to support and promote technology transfer among innovative organizations. Science parks are frequently linked to local universities, which often played a role in their establishment.

In the chemicals industry we find that while firms locate their R&D in relation to university research on a wider spatial scale (up to 80 km from materials science departments), those R&D-doing firms that are geographically close (within 10 km) tend to engage more with the research base, in line with geographic proximity facilitating firm–university interaction.

In other industries we find less, or no, evidence of close co-location with university research, and little evidence that geographic proximity is associated with greater engagement. In the vehicles and machinery industries we find that R&D facilities tend to be located in areas with a higher number of materials science departments within 10–50 km, in areas with higher manufacturing employment and in areas which are relatively specialized in the respective industry. This suggests that the co-location of R&D with production, both within and external to the firm, might be more important in these industries. In electrical machinery, TV and radio equipment, aerospace and precision instruments we find no consistent evidence that the location of business R&D is related to the presence of university research.

The article contributes to the empirical literature on the existence of geographically mediated knowledge spillovers, and in particular the effects of university research on regional innovation activity (Section 2). We extend this literature by providing descriptive evidence on the location of firms’ R&D with respect to the presence of individual departments rather than a measure of total university activity. The research builds on Abramovsky et al. (2007) and extends it by using continuous measures of spatial proximity and by considering firm–university interactions directly. As such the article also contributes to the literature on the role of geographic proximity in business–university co-operation.

The article is structured as follows. Section 2 discusses firms’ R&D location decisions, firms’ incentives to co-operate in R&D with universities and the importance of geographic proximity and social networks. Section 3 describes our data. Section 4 examines the co-location of R&D labs and university research and Section 5 investigates the role of proximity in firm–university interactions. Section 6 concludes.

2. R&D location decisions and firm–university interactions

A firm’s choice of R&D location is a long-term investment decision, which will be affected by factors including local labour market conditions, site availability and planning regulations, and access to infrastructure such as road and rail networks. Firms may benefit from locating R&D with other activities within the firm, or near to other firms or R&D labs in the same industry in order to benefit from localization externalities, arising for example through a pooled labour market for specialized skills or knowledge spillovers (Marshall, 1890). Urbanization externalities arising through knowledge spillovers from diverse industries (Jacobs, 1969) may attract firms to areas with a high density of economic activity, in addition to agglomeration economies arising from improved labour market matching and ease of market access (Duranton and Puga, 2004; Rosenthal and Strange, 2004 provide surveys). Firms may also benefit from locating R&D near to related university research departments, either to facilitate direct connections, e.g. in the form of research joint ventures, to draw on the expertise of university staff or research students through consultancy or direct recruitment, or to access knowledge more informally. Individual academics and research departments may also directly affect the amount of private-sector R&D in a location by commercializing their research through the formation of spin-out companies.

The extent to which firms can increase their research productivity through successful interaction with universities may depend on establishing and sustaining long-term working relationships with university scientists—a potentially valuable asset to both partners. These relationships may be of greater importance for formal, market-based interaction such as collaborative R&D, compared to more ad-hoc, non-market-based knowledge sourcing. Geographic proximity, enabling frequent face-to-face social interactions, may play an important role in developing and maintaining links between researchers, and may be crucial if the primary mechanism through which knowledge is transferred is direct personal interactions within social networks. However, geographic proximity may be less relevant if knowledge is codified or if tacit knowledge is transferred through well established links, such as long-term research collaborations or alumni connections, which endure despite researcher mobility (Ponds et al., 2010).

Capturing knowledge spillovers is a key motivation for engaging in co-operative R&D—generating new knowledge, and internalizing it within the partnership (Katz, 1986; D’Aspremont and Jacquemin, 1988; Kamien et al., 1992; Katsoulacos and Ulph, 1998). Other motivations include spreading the cost and risk associated with innovation, particularly in relation to basic research. Firms’ incentives to co-operate will also depend on their ability to assimilate knowledge from university research, i.e. their absorptive capacity (Cohen and Levinthal, 1989). Many empirical studies examine firms’ propensity to engage in collaborative R&D with universities, (Cassiman and Veugelers, 2002; Abramovsky et al., 2009), but do not consider the role of geographic proximity.

Evidence suggests that academic knowledge benefits firms and that the mechanisms through which non-market spillovers occur may be localized (Jaffe, 1989; Jaffe et al., 1993).¹Breschi and Lissoni (2009) highlight the spatial scale of co-invention networks in determining the extent to which knowledge spillovers are localized. They argue that in large part spillovers are found to be geographically concentrated because co-invention networks are themselves localized due to limited researcher mobility. Knowledge spillovers are inherently difficult to measure and a common approach is to use information from patent citations to track knowledge flows (Jaffe et al., 1993; Griffith et al., 2007; Breschi and Lissoni, 2009). Here we exploit survey data where firms are asked directly how important information from universities is for their innovative activity (Cassiman and Veugelers, 2002 use equivalent data), and also about collaborative innovation agreements, which will capture market-based knowledge flows.

Our article contributes to the empirical literature on geographic knowledge spillovers and the effects of university research on regional innovation activity.² A closely related article is Woodward et al. (2006) who find a positive but small impact of proximity to university research (measured by total university R&D expenditures in science and engineering) on numbers of high-tech start-ups. Instead we examine the location of firms’ R&D with respect to the presence of individual science and engineering departments. Our research builds on previous work (Abramovsky et al., 2007) by using continuous distance measures rather than analysing activity within discrete geographic units. While Woodward et al. (2006) find evidence consistent with geographic knowledge spillovers from university research across many high-tech industries, our findings suggest considerable heterogeneity, with evidence of close co-location with university research only apparent for pharmaceuticals, R&D services and chemicals R&D establishments. Moreover, we find that these relationships vary with respect to specific scientific fields of university research, and indicators of research quality.

The article also contributes to the literature on geographic proximity and business–university connections. Rosa and Mohnen (2008) find that greater distance decreases the proportion of total R&D expenditure that firms paid to universities. Ponds et al. (2007) analyse co-publications and find that proximity is more relevant for collaboration between organizations with different institutional backgrounds, such as firms and universities, than for purely academic sector collaboration, although firm–university collaborations are also prevalent at a national scale.³ Our analysis using survey data also illustrates the difficulty of drawing general conclusions. In that, for only one industry—chemicals—out of the four we investigate, do we find any consistent evidence that geographic proximity is related to greater interaction between firms and universities.

3. Data

In this section we outline each of the datasets we use.

3.1. Business sector R&D activity

We use the UK Office for National Statistics (ONS) establishment-level Business Enterprise Research and Development (BERD) data for the period 2000–2003, which provide information on the population of establishments performing intramural R&D in Great Britain. Each establishment’s full postcode, industry, total number of employees and ownership status are registered.⁴ The ONS collects more detailed information about R&D expenditure by surveying a sample of establishments from this population. This includes a census of large establishments and a stratified sample of the remainder of the population. Because not all establishments are surveyed, in our analysis we rely mainly on the basic information available for the whole population.⁵ We use one piece of information that is not reported for all establishments—the product group for which R&D is being conducted, discussed below.

Our measure of the presence of business R&D activity is constructed at the postcode district level, defined by the first part of the full postcode, e.g. ‘OX1’.⁶ We use postcode districts as the unit of observation to tie in with our second analysis of firm–university interactions using the Community Innovation Survey where this is the finest level of geographic information available. Postcode districts vary in geographic size according to whether an area is rural or urban, so in our empirical analysis we pay careful attention to controlling for a range of other factors that may determine the concentration of R&D establishments and universities in a particular postcode district.

Our measure is a count of the average number of establishments carrying out intramural R&D between 2000 and 2003, in a product group. Product group information is not collected for smaller, sampled establishments and non-sampled establishments.⁷ However, each company’s Standard Industrial Classification (SIC) code is known and the ONS assumes that R&D expenditure is for the product group corresponding to that SIC category. For example, R&D carried out in a small firm that operates in the chemicals industry will be assigned to the chemicals product group. This results in a large number of small establishments being classified as R&D services providers, whereas it is very likely that they do R&D for other product groups. This means that we may not be capturing the locations of all R&D activity devoted to the product groups we consider, (although we will be capturing the locations of establishments accounting for the vast majority of expenditure).

We focus on eight product groups that account for 69% of total intramural R&D in 2003: pharmaceuticals, chemicals, machinery, electrical machinery, TV and radio equipment, vehicles, precision instruments and aerospace. Table 1 breaks down total intramural R&D expenditure and counts of R&D establishments by product group. R&D expenditure is highly concentrated: expenditure in pharmaceutical products accounts for a quarter of the total, followed by aerospace (12%) and vehicles (9%). The distribution of establishments is much less concentrated across products; the same eight product groups account for around 30% of total R&D establishments, implying that a small number of establishments account for a large share of total expenditure, (large firms account for ∼75% of total R&D performed in UK businesses, National Statistics, 2005). As an additional exercise we also look at the location of small R&D services labs operating in natural sciences and engineering. These comprise a further 15% of establishments.

3.2. Firm–university interactions

To obtain information on firm–university linkages we use data from the Community Innovation Survey (CIS), which is conducted every 4 years by EU member states to collect information on firms’ innovative activity. To tie in with the analysis of R&D location decisions we focus on R&D-doing firms only. The firms included in the CIS are a sample drawn from the ONS Inter-Departmental Business Register (IDBR).⁸ The IDBR also encompasses the population of establishments in the BERD data. We combine CIS3 data for 1998–2000 and CIS4 data for 2002–2004 for Great Britain to derive indicators of business–university R&D interactions. We use two pieces of information. First, whether a firm co-operates with higher education institutions (HEIs) in the local area.⁹ This variable equals one if a firm reports that it co-operates with any HEI within ∼50 miles (80 km) in the CIS3 (question 13) and within 100 miles (160 km) in the CIS4 (question 18). Second, whether a firm sources knowledge for its innovative activities from HEIs. Firms are asked how important different information sources have been for their innovation activities (question 12, CIS3; question 16, CIS4). Our variable takes the value one if the firm placed low, medium or high importance on information from universities, and zero if information from HEIs was not used. The data allow us to locate R&D-doing firms within postcode districts. We conduct our analysis for four of the main product groups above (chemicals excluding pharmaceuticals, machinery, vehicles and precision instruments) for which sample sizes are large enough.¹⁰ These account for a substantial share of UK R&D (Table 1).

The CIS contains information on firm characteristics which might affect whether firms engage with HEIs, which we use as control variables. These are a measure of size (log employees); the percentage of the firm’s employees qualified to degree level or above in science and engineering and R&D intensity (intramural R&D expenditure over turnover for the year 2000 for CIS3 and 2004 for CIS4)—both of which should capture absorptive capacity; and whether a firm received financial public support for innovation activities. We provide further detail in Table 8 (Section 5).

3.3. Measuring university research presence and quality

We use the results of the 2001 Research Assessment Exercise (RAE) to map the presence and quality of research carried out by universities, and their specific research departments in Great Britain. The Higher Education Funding Council for England carries out the RAE to produce research quality ratings used to allocate the main grant for research use among universities. Each university submitted research activity carried out in the 5 years to the end of 2000 for assessment, on some fraction of the research staff in departments of their choice. There were 2598 submissions by 173 UK universities on 68 research areas.¹¹ Each department submission was rated on a scale of 1, 2, 3, 4, 5 and 5*; a higher number indicating higher rated research. Top rated departments received a funding weight over three times higher than lower rated departments (HEFCE, 2005).¹² In some cases departments from the same university chose to send more than one submission, and we use the maximum rating achieved.

Our measures of university presence are at the postcode district level and consist of count measures within a certain radius of the centre of the postcode district. We construct count measures of universities, the total number of relevant research departments, the number of relevant research departments rated 5 and 5*, and those rated 1–4, and the log of the total number of research students in all departments.¹³ We focus on distances within an inner ring of 10 km and an outer ring between 10 and 50 km. We also experiment with expanding the inner ring to 25 km, and the outer ring to 80 km to tie in with the narrower definition of local/regional HEIs in the CIS.

We construct the distance between the centre of a postcode district and a university or research department using National Grid references. We calculate the central point of each postcode district by taking the mean of the Eastings and Northings of all the postcodes within each postcode district from the National Statistics Postcode Directory (NSPD). We then link each university’s full postcode with the NSPD to obtain Eastings and Northings and use Pythagoras’ theorem to calculate the Euclidean distance between the centre of each postcode district and each university.¹⁴

To define relevant research we use the 1994 Carnegie Mellon Survey (CMS) that reports the importance of ten research fields (biology, chemistry, physics, computer science, materials science, medical and health science, chemical engineering, electrical engineering, mechanical engineering and mathematics) for different manufacturing industries.¹⁵ We consider a field as relevant for a product group if it was rated moderately or very important (a score of at least three on a four-point scale) for the corresponding industry by over 50% of respondents. We assign each of the RAE departments to the ten CMS fields (mapping available on request). Descriptive statistics for these measures are provided in Sections 4 and 5.

3.4. Further data

We use a number of control variables in our analysis. We include a measure of economic density at the postcode district level to capture urbanization externalities and market potential and the fact that more populated postcode districts are likely to be physically smaller than rural ones. This is measured by the log of the number of full postcode addresses at the postcode district level, from the 2006 NSPD. We also include controls at the ‘postcode area’ level, where a postcode area is the first two letters of the postcode, e.g. ‘OX’ (see footnote 6). We include a measure of the skill composition of the workforce which may affect the type of firms operating in the area and also contribute to the innovation process and to knowledge spillovers (Bartel and Lichtenberg, 1987). We use the percentage of the economically active population in the postcode area qualified to degree equivalent or above (level 4), constructed from official labour market statistics for local areas (NOMIS, Labour Force Survey data).

In our co-location analysis we include the log of total manufacturing employment in the postcode area to further control for the scale of each area and potential agglomeration externalities arising from co-location with production activity. We include the percentage of total manufacturing employment in the postcode area that is in the relevant industry (i.e. the industry corresponding to each product group), to control for potential industry localization externalities and the co-location of R&D facilities with related production (Audretsch and Feldman, 1996). These are constructed using the ONS plant-level ABI-ARD population data for 2000. We also include a measure of the presence of science parks located within 10 km of the centre of each postcode district from the UK Science Park Association (UKSPA). Firms may locate in science parks due to a range of factors: access to university research or other specialized infrastructure; localization benefits; or lower costs of establishing an R&D facility at this type of site. To the extent that science parks are located near to specific university research departments this is likely to make it more difficult to separately identify relationships with proximity to relevant research.

4. Do firms locate their R&D near universities?

We provide some descriptive statistics and then outline our empirical approach and findings.

4.1. Descriptive statistics

Table 2 provides information on our measures of university presence across postcode districts. The first row shows that postcode districts have on average three universities located within 10 km of their centre, but there is variation with over half having no university located in their immediate proximity and with postcode districts in inner London having over 30 universities located within a 10 km radius.¹⁶ Not surprisingly postcode districts have a higher average number of universities located between 10 and 50 km, around 12, and the number of postcode districts with zero universities located in that band is only 98.

The table also covers the specific departments which we use in the analysis. These are often concentrated in the immediate geographic proximity of a small number of postcode districts. Medicine departments are present within 10 km in the largest number of districts,¹⁷ while materials science departments are present near the fewest, with 1850 having none within 10 km. Most districts have a relevant department between 10 and 50 km. Again materials science departments are relatively rare and medicine and computer science departments the most common. Looking at the presence of departments rated 4 or below and rated 5 or 5*, there are few striking differences in the average number within 10 km by rating. As would be expected there are some very high correlations between our main department-level variables, meaning that there may not be enough variation to separately identify relationships with individual departments.¹⁸

Descriptive statistics for the other control variables are shown in the final section of the table and descriptive statistics on the number of R&D labs by product group, our dependent variables, are displayed with the regression results in the bottom row of each table in Section 4.3.

4.2. Empirical approach

Our empirical approach needs to encompass more than one underlying model of firm behaviour since our data cover a heterogeneous set of R&D establishments. Some, in particular those owned by multinational firms, are likely to be highly geographically mobile. For these firms the relevant decision is whether locating in close proximity to a university, all else equal, is likely to increase their R&D productivity relative to an alternative location. Evidence of co-location with university research departments could then indicate that geographic proximity is important to capitalize on beneficial knowledge spillovers. Other establishments will be small R&D start-ups. Here the relevant decision might be whether an individual chooses to set up a new business in the area where he lives or works, rather than where to set up a new business. Hence, in this case a positive association with particular research departments is potentially in line with individuals in those departments having a higher propensity to set up their own commercial ventures. Given this potential heterogeneity in the underlying decision process we estimate a negative binomial count data model to describe the general pattern of location outcomes in our data, rather than specify a more precise model for such a heterogeneous set of firms.^19,20 However, we also estimate separate models for foreign-owned and small establishments to examine whether location patterns vary by firm type.

While we attempt to control for other factors that may affect the location of business R&D the results from the above specification should be interpreted as correlations rather than causal relationships. For example, common unobserved factors may determine both the presence and quality of research departments and the location of R&D, or there may be reverse causation from the location of business R&D to the quality of research departments. In general these effects are likely to bias the results towards finding evidence for the co-location of business R&D and university research. Moreover, many of the location choices we observe in our data are the result of long-term historical decisions, hence our empirical results only present a descriptive snap-shot of current location patterns.

We experiment with alternative measures of geographic proximity and conduct further robustness checks which we report for the pharmaceuticals industry. We look at foreign-owned and small establishments separately to investigate heterogeneity in location patterns. We experiment with excluding postcode districts located in London, and also with including regional dummies to try to address common unobserved heterogeneity. We also look at specialized R&D labs, many of which are likely to conduct pharmaceuticals R&D.

We present results in the form of incidence rate ratios (IRRs). An IRR greater (lower) than one corresponds to a positive (negative) coefficient. For ease of exposition the tables display incidence rate ratios minus one. Hence an IRR of 1.3 is displayed as 0.3 and means that for a one-unit increase in the explanatory variable there is a 30% increase in the expected number of R&D establishments in a postcode district. We report z-statistics in parentheses and indicate statistical significance at the 1, 5 and 10% levels. Each table includes information on the dependent variable’s mean and standard deviation and the number of postcode districts where the dependent variable is zero.

4.3. Results

Table 3 shows results for the location of R&D establishments with respect to the presence of research departments located within 10 km. Given the number of hypotheses being tested we would expect to find some significant effects. However, 3 out of the 21 estimated effects of the presence of specific departments are significant at the 5% level, indicating that the results are significantly stronger than would be expected by chance. We find that pharmaceuticals establishments are more likely to be located in postcode districts with chemistry departments within 10 km. The results in column 1 suggest that an additional chemistry department within 10 km is associated with a 65% increase in the expected number of R&D establishments. On average a postcode district has around 0.10 pharmaceuticals R&D establishments, although the distribution is very skewed. Note that an increase of one department is a large change—as shown in Table 2, the average number of chemistry departments within 10 km, across all postcode districts is only 0.5.²¹ Along similar lines, an additional materials science department within 10 km of a postcode district is associated with an 18% increase in the expected number of R&D establishments in chemicals.²²

We also find a significant negative relationship between the presence of materials science departments and the number of R&D establishments in aerospace. The aerospace industry is characterized by considerable economies of scale and Table 1 shows that there are very few R&D establishments in this industry in our data. As shown in Table 2 the distribution of materials science departments is also strongly geographically concentrated—there are few postcode districts with departments within 10 km, and these tend to be in major cities such as London and Manchester. It is likely that the nature of the aerospace industry, which requires very large business premises and access to infrastructure such as runways prohibits location in close proximity to the centres of large cities. The analyses for machinery, electrical machinery, TV and radio equipment, vehicles and precision instruments also provide no evidence that R&D facilities are located in proximity to university research.

The control variables are also of interest. Conditional on the presence of relevant departments, the number of R&D establishments does not appear to be correlated with university presence in general or with the number of research students. The coefficient on the log of manufacturing employment in the postcode area is positive and highly significant in chemicals, machinery and vehicles. In all cases the coefficient on the measure of postcode area industry specialization is also positive and significant, indicating that R&D establishments are likely to be located close to centres of manufacturing activity in their own industries, possibly indicative of localization externalities. The coefficient on the number of postal addresses is positive and highly significant in all cases suggesting the presence of more general agglomeration economies. The percentage of the population educated to degree-level or above (L4+) enters positively and significantly in some sectors with the strongest relationship in pharmaceuticals.

Table 4 adds measures of research departments located at a distance of 10–50 km. Only the coefficients for the relevant research fields are displayed. For pharmaceuticals the coefficient on chemistry departments within 10 km remains positive and significant, and is substantially higher than that on chemistry departments located between 10 and 50 km. This pattern of results points towards potential knowledge spillovers from chemistry departments to private-sector pharmaceuticals R&D and suggests that these may be increasing in geographic proximity.

For chemicals the coefficient on materials science departments located between 10 and 50 km is positive and significant and of a similar magnitude to that on departments within 10 km. R&D establishments in machinery and in vehicles are also more likely to be located in postcode districts with a higher number of materials science departments between 10 and 50 km distance, and R&D in machinery is more likely to be located in districts with a higher number of mechanical engineering departments 10–50 km from the centre. Although indicative of potential spillovers from relevant university research this pattern could be driven by firms in these industries locating their R&D outside immediate urban areas, possibly alongside production activity. As the results in Table 3 indicate, R&D activity in chemicals, machinery and vehicles is likely to be located in areas with a greater concentration of manufacturing employment, potentially due to planning requirements for large production facilities. We find no evidence of co-location with university research in the remaining industries.

In Table 5 we investigate whether research quality matters, distinguishing between 5 and 5* departments deemed to be carrying out frontier research and those rated 4 and below. The results imply a very strong tendency for pharmaceuticals R&D to be co-located with 5 and 5* rated chemistry departments. The finding for chemicals R&D and materials science departments is no longer significant when we split departments into higher and lower rated. Relative to Table 3 we find a mixed pattern of results for precision instruments, an industry that comprises a heterogeneous set of products from medical equipment to measurement instruments. R&D establishments tend to locate in postcode districts with a larger number of 5 and 5* rated medicine departments in close proximity. However we find negative coefficients on medicine departments rated 4 or below (this is also the case for pharmaceuticals) and electrical engineering departments rated 5 or 5*.²³ It may be that this pattern is driven by the location of establishments carrying out R&D in medical equipment.

In Table 6 we replicate the regressions in Table 3 also controlling for the number of science parks within 10 km. For half of the product groups we find a positive and significant relationship between the presence of science parks and the number of R&D establishments in a postcode district. The relationship is strongest for pharmaceuticals where an additional science park is associated with a 90% increase in R&D establishments. Interestingly, the coefficient on the number of chemistry departments becomes much smaller and insignificant. However, this does not necessarily imply that (highly-rated) chemistry departments are not a relevant factor driving these location decisions. As discussed above the location of science parks is itself endogenous, for example, if science parks arise because of a demand for space in proximity to frontier chemistry departments. Instead these findings suggest that science parks located close to chemistry departments are positively correlated with the location of pharmaceuticals R&D labs.²⁴ Co-location with science parks may also capture access to specialized infrastructure, or knowledge spillovers from other technology-intensive businesses. For the other product groups the coefficients on the research departments generally remain similar to those in Table 3.