Zoltan J. Acs, University of Maryland
Felix R. FitzRoy, University of St Andrews
Ian Smith, University of St Andrews
Using 4 years of data from 37 American cities and 6 high technology groupings we present the first estimates of University R&D spillover effects on employment at this level of disaggregation, while controlling for wages, prior innovations and state fixed effects. Wages and employment are strongly positively related, which can be explained in various ways. Consistent with studies showing R&D spillover effects on innovation at the state level, we find robust evidence that university R&D is a statistically significant determinant of city high technology employment and some evidence for employment effects of innovation.
Localized clusters of high technology firms, such as California's famous Silicon Valley or Boston's Route 128, are of significant interest to policy makers and economists alike. For in addition to favorable effects on international competitiveness, such clusters generate considerable regional benefits in terms of jobs, growth and economic development. Understanding the determinants of the spatial distribution of high technology activities is therefore important for both regional and industrial policy.
Traditional explanations of locational choice refer to transportation costs and proximity of raw material, fuel or labour inputs as major factors in the agglomerative process. Another response which has received serious attention recently points to the role of geographically-bounded knowledge spillovers from universities and federal laboratories in the location decision of firms. Informal evidence suggests there is a close association between high technology clusters and major research universities in the United States (Acs, 1993). Typically cited are the links between Stanford University and Silicon Valley or MIT and Route 128. Such a nexus has scarcely emerged in Europe except in the form of a few fledgling research parks such as Cambridge, England (Lumme et al, 1993). Formal tests conducted by Jaffee (1989) provide econometric evidence for the real effects of academic research in terms of its spillover to corporate patenting activity. In addition, papers by Jaffee, Trajtenberg, and Henderson (1993), and Almeida and Kogut (1994) demonstrate the significant degree of localization of these knowledge externalities with respect to patent citations. However, spillovers from university research to commercial innovation are not the only effects of relevance to theory and policy. The ultimate economic interest lies chiefly in the product markets and jobs that are generated by R&D. The aim of this paper is to test for the existence of such spillovers from university R&D to local high technology employment. This is a question of considerable policy importance (Business Week, 1994) which has only been discussed systematically to date by Beeson and Montgomery (1993) who, in contrast to our results, find no statistically significant effect of university research and development expenditures on high technology employment shares.
The discussion is organized into five sections. The first outlines discursively the theoretical background. The second section provides a preliminary analysis of the data. We use unique annual data for six high technology sectors in 37 American Standard Metropolitan Statistical Areas (SMSAs) for the period from 1988 to 1991 to investigate the relationship between university R&D expenditure and employment. Jaffee (1989) used American states as his geographical unit of analysis. This has drawbacks in those cases where state borders cut through economic areas or where states contain several large cities. Our use of SMSA data should clearly subject the theoretical argument for spillovers based on spatial proximity to a more precise test. The model is specified in section three and, in the fourth section, the econometric results are reported and discussed. Given the inclusion of real wages in the employment function, we apply simultaneous estimation techniques to study university R&D spillovers on high technology employment over time and across cities, while controlling for common macroeconomic effects. A final section concludes the paper.
There are two related hypotheses explaining the development of high technology clusters in the vicinity of major university R&D activity.
The first explanation argues that university research is a source of significant innovation-generating knowledge which diffuses initially through personal contacts to adjacent firms, especially those based in a science park. Since both basic and applied university research may benefit private enterprise in various ways it induces firms to located nearby. Lund (1986) in a survey of industrial R&D managers confirms the proximity of university R&D as a factor in the location decision due to the initial spillover from neighboring university research to commercial innovation. Of course, as research results are used and disseminated, the learning advantage created by close geographic proximity between local high technology activity and the university would fade but these learning lags may be long. Information flows locally therefore, through a variety of channels discussed below, more easily and efficiently than over greater distances.
There is a growing body of evidence which supports this hypothesis, especially in the United States1. Spillovers form university R&D to patent activity in the same state have been identified econometrically by Jaffee (1989). Acs, Audretsch and Feldman (1992, 1994) reinforce this result with, instead of patents, a more direct measure of economically useful knowledge production, namely the number of innovations recorded in 1982 by the US Small Business Administration from the leading technology, engineering and trade journals. Likewise, Nelson (1986), using surveys of research managers finds university research to be a key source of innovation in some industries, especially those related to the biological science where he finds some degree of corporate funding of university projects. University research spillovers may be a factor which explains how small, and often new, firms are able to generate innovations while undertaking generally negligible amounts of R&D themselves2. There is econometric evidence for this result based on data from both the United States (Acs, Audretsch and Feldman, 1994) and Italy (Audretsch and Vivarelli, 1994).
Despite the presumed advantages of geographical proximity for receiving spillovers, the mechanisms by which knowledge is transferred are not well understood. Information flow are usually attributed to the use of faculty as technical consultants and post-graduate students as research assistants, the use of university facilities, informal communication between individuals at trade shows, industry conferences, seminars, talks and social activities, or joint participation in commercial ventures by university and corporate scientists through contracted research projects. The latter has grown in importance since the late 1970s as the universities established formal Offices of Technology Transfer (or Licensing) to foster interaction with industry and the commercialization of research results. This partly reflects pressure applied by US government agencies to universities, for economic growth reasons, to hasten technology transfer from their laboratories to the private sector (Parker and Zilberman, 1993). Federal Acts passed in the early 1980s also promote knowledge spillovers. The Stevenson-Wydler Technology Innovation Act of 1980, for example, encourages cooperative research and technology transfer and the 1981 Economic Recovery Tax Act gives tax discounts to firms that provide research equipment to universities. Some universities have created industry consortia to help fund research. Firms pay membership fees to join these consortia and in return benefit from access to the research output and have some voice in the research agenda. Such channels would be expected to flourish given that universities as public institutions do not face the same incentives as private corporation to keep research results secret. In both the San Francisco Bay and Boston areas, for example, the introduction and growth of the biotechnology industry is a direct result of university R&D spillovers. Presumably, the chief benefits of geographical proximity to the spillovers source consist in a reduction in both the transactions costs of knowledge transfer and in the costs of commercial research and product development. As a caveat, it ought to noted that we do not argue that proximity is a necessary condition for spillovers to occur, only that it offers advantages in capturing them.
The Labour Market
The second university based explanation of clustering highlights the provision of a pool of trained and highly qualified science and engineering graduates. The high level of human capital embodied in their general and specific skills is another mechanism by which knowledge is transmitted (Beeson and Montgomery, 1993). To the extent that they do not migrate, such graduates may provide a supply of labour to local firms or else a supply of entrepreneurs for new start-ups in the high technology sector (Link and Rees, 1990). Some evidence for this latter link is provided by Bania, Eberts and Fogarty (1987, 1993) who, using cross-section data, find a significant effect of university research expenditure on new firm start-ups. University scientists themselves, of course, may provide the entrepreneurial input, working part-time as directors of their own start-up companies, or even leaving academia to take a position in a high technology firm. Parker and Zilberman (1993, p.97) report, for example, that MIT has incubated about 40 biotechnology firms since the late 1980s. Lumme et al (1993) in their study of academic entrepreneurship in Cambridge (England) identified 62 high technology companies whose business idea was based on the exploitation of knowledge developed or acquired in either a university of a research institute. However, even if university research is either negligible or irrelevant to industry, university training of new industrial scientists alone may be sufficient to generate local labour market spillovers. Nelson (1986, p.187), for example, motes that industrial interest in academic departments of physics is confined mainly to their output of potential industrial scientists rather than to their research results.
A university and its associated science park may also play an important signaling role in locational choice (Shachar and Felsenstein, 1992) in the sense that they signal the presence of local technological capacity. Thus firms may be attracted even if the university spillovers are not in fact that great.
PRELIMINARY DATA ANALYSIS
The first step3 is to identify the high technology sectors. We proceeded by selecting those with a relatively high ratio of R&D to industry sales. Thirty two three-digit Standard Industrial Classification (SIC) industries were identified in this way, and then grouped into the six sectors.
Next we selected the 22 most important SMSAs for these industries, most of which also have major university R&D activity. For comparison and sample variation we also include 15 additional SMSAs with only minor university research.
The relationship between university R&D and high technology employment can be analyzed in a preliminary fashion using scatter diagrams. Figure 1 plots aggregate high technology employment in 1989 against university research expenditure in 1985 for all 37 SMSAs. Both variables display great variation across metropolitan areas though there is a clear positive association between them. The simple correlation coefficient is 0.60. University research expenditure and high technology employment are both high in the major cities of Los Angeles, Boston, New York and Baltimore.
Figure 2 plots a scatter diagram of high technology employment against the number of scientists and engineers per 100 workers by SMSA. The motivation is that the stock of university science graduates with good general and specific skills influences the location of a high technology cluster. Empirically the association with high technology employment is not that strong. The simple correlation coefficient is 0.26. The plot illustrates that Austin, San Jose, Seattle and Raleigh have a high share of engineers and scientists relative to the level of high technology employment. In other words, the labour quality is relatively high in these SMSAs. In contrast, the large number of employees in Los Angeles appear to be concentrated in low skilled occupations.
In spite of data limitations, we estimate a parsimonious structural labour market model. The main missing variable is a proxy for product demand faced by high technology firms, such as sales, which is not available for individual SMSAs. Our initial specification for the employment equation is written down in natural logarithms as:
(1)EMPmit = a0 + a1Wmit + a2RDmi + a3POPmt + a4HKm + a5INNOVm + aX + u1mit
where 'm' indexes SMSA, 'i' indexes industry, and 't' indexes time: m=1,...,37; i=1,...,6; and t=1988,...,1991. EMPmit refers to high technology employment and Wmit is the corresponding annual real wage per employee, defined as nominal wages deflated by the appropriate industry producer price index. Since the panel includes only four years of annual data, cross-sectional variability dominates4. For this reason, attempts to estimate equations specified in terms of employment growth rates proved fruitless.
For reasons of data availability, RDmi university R&D, is specified for only a single year, 1985. Given the time span of our data set, it seems reasonable the use of R&D inputs dated in 1985 provides and appropriate lag for the knowledge externality to be transmitted into commercial products and employment. Edwards and Gordon (1984), for example, find that innovations made in 1982 resulted from inventions made on average 4.2 years earlier. The R&D data include industry funded university research, a component which rarely exceeds 10% of the total and is usually considerably less. Notice that RDmi varies by both SMSA and industry. Total university R&D spending in each city is desegregated by broad science department and allocated to each of the six industries. This is appropriate given substantial differences in the commercial applicability of university research across academic departments. Thus employment data by industry sector are linked to the relevant component of university research expenditure. The assignment of university department to industrial sector is listed in Table 1. This is close to Jaffee (1989) but it is doubtless not the only plausible allocation.
POPmt refers to city population and controls for local market size. Of course, the market extends beyond SMSA boundaries but we do not have a more appropriate measure of demand. The number of scientists and engineers as a proportion of the labour force of each SMSA represents the potential human capital, or quality of the labour force, available for employers, HKm. Data are only available for a single year, 1989. INNOVm is a simple count of the number of innovations by MSA in 1982, the year for which this variable has been collected. It attempts to control for the effect of pre-existing commercial innovation, that leads to product development and marketing with substantial time lags, on subsequent employment levels. Finally X represents a vector of industry, state and annual time dummies. These control for effects specific to each which may not have been captured by the continuous variables.
Since employment and real wages are jointly determined in the labour market, equation (1) should be estimated by a simultaneous method. The corresponding real wage level equation is given by:
(2)Wmit = b0 + b1Wmi,t-1 + b2EMPmit + b3HKm + b4CWmit + bX + u2mit
This includes the average SMSA hourly wage CWmit and a lagged dependent variable in addition to human capital. Rank and order conditions indicate that both the wage and employment equations are overidentified. Two stage least squares (2SLS) is therefore adopted as the method of estimation.
Table 2 presents the main summary statistics by variable for the 37 MSAs in aggregate.
Table 3 reports the 2SLS estimates of equations (1) and (2). OLS estimates are listed for comparative purposes, though it will be noted that the coefficients do not differ much from their 2SLS counterparts. Student t-ratios are in parentheses. The coefficients on the fixed effects are not tabulated but for the employment equation, their joint significance cannot be rejected by and F-test. Taking the fixed effect groups separately, none fail a variable deletion test5.
Notes: (i) t-statistics are in parentheses; (ii) all variables are in natural logarithms; (iii) R2 is the adjusted multiple correlation coefficient, s is the estimated standard error of the regression, and n is the number of observations; (iv) unreported dummy variables for industry, time and state are also included in each of these regressions.
Each 2SLS equation was estimated in natural logarithms over the three year period 1989 to 1991 using 666 observations (3 years x 6 sectors x 37 cities). The coefficients should therefore be interpreted as elasticities. Statistically the wage and employment equations are satisfactory, with one notable exception, have the expected signs though not all are significant at conventional levels.
The central result is a positive and statistically significant coefficient on the R&D variable in the employment equation. Although the magnitude of the employment elasticity is small (0.08), this is evidence of a direct spillover of university research on the thigh technology employment. In unreported regressions, we found the same result when the dependent variable is specified in terms of high technology employment share, in striking contrast to the statistically insignificant coefficient reported by Beeson and Montgomery(1993)6.
A further result is that real wages and employment are positively related ceteris paribus. This is counter to our theoretical priors based on the perfectly competitive model. Dropping the wage variable did not markedly affect the signs and significance of the remaining regressors so this outcome does not vitiate our spillover story. Neither reestimating as single equations using OLS nor as random effects models produced major differences in the results. So the estimates are robust with respect to estimation technique.
At first blush, this result is quite surprising. However it is quite consistent with two important features of high technology industries. First, output markets with continual product innovation and imperfect information are far from the traditional model of perfect competition. It follows that some proxy for product demand should be included in the employment equation but such a variable was not available. Thus we are estimating a reduced form rather than a true structural demand model. Second, specialized skills are often required in high technology sectors. Locational advantages that attract high technology firms may also generate shortages of skilled workers that lead to higher wages. Other wages typically follow to maintain differentials. The positive correlation between high technology employment and wages thus probably reflects the crucial shortages and imperfect mobility of skilled labour that has been the subject of much policy discussion and concern.
Equally plausible, and without relying on market imperfections, it may simply be the demand for products produced by the most skilled and highly paid workers that has grown most rapidly.
The University based labour market spillover story has weaker support. The proportion of engineers and scientists in a city, the human capital variable, is statistically insignificant in both the employment and wage equations. Our population and technical innovation variables, however, are both well determined.
We have also estimated employment functions separately for each of the six high technology industries. The mean SMSA employment, wage and R&D in each sector is listed in Table 4 and the equation estimates are provided in Table 5. The R&D coefficients are positive and statistically significant (or near significant) in all but the Energy and Chemicals industry. A plausible explanation is the Energy and Chemicals represents a rather traditional industry dependent on both raw materials and products that are much more costly to transport than the inputs of other sectors. Access to port facilities, and other transport infrastructure, is thus likely to be much more important in the location decision, weakening considerably the role of the R&D variable.
Although the results for labour quality and prior innovation are mixed in terms of both sign and statistical significance, it is the university R&D effect, our key variable of interest, which is most consistent. The coefficients are largest in the Defense and Aerospace and High Technology Research sectors.
Notes: (i) t-statistics are in parentheses; (ii) all variables are in natural logarithms; (iii) R2 is the adjusted multiple correlation coefficient, s is the estimated standard error of the regression, and n is the number of observations; (iv) unreported dummy variables for time and state are also included in each of these regressions.
Key: EC: Energy and Chemicals, DA: Defense and Aerospace, ITS: Information and Technology Services, HTR: High Technology Research, BB: Biology and Biomedical, HTM: High Technology Machinery and Instruments
Previous empirical work on R&D spillovers has focused on their relationship with innovation and patent counts at the level of individual U.S. states. With new data for 37 American Standard Metropolitan Statistical Areas including the main university R&D centres we have found a statistically significant and robust spillover to employment in five high technology sector, after controlling for state fixed effects. This confirms the popular view of high technology clusters and provides the first quantitative evidence that academic research has a positive local high technology employment spillover at the city level. A further result is that innovation was also strongly related to high technology industry employment after a long time lag, again a plausible but hitherto untested proposition.
These results are clearly of relevance for regional policy. They provide support for the importance of high technology clusters in the U.S. and possible lessons for Europe and Japan where such clusters are much less well-developed and where there is no evidence of the localization of knowledge spillover, at least in the semi-conductor industry (Almeida and Kogut, 1994). In spite of dramatic declines in the costs of information transmission, local spillovers underline the importance of personal contacts and face-to-face communication in transferring scientific progress into jobs and products. Clearly more research is required on the nature of the transmission process as well as on the skill composition of high technology employment and the relationship of training and skills to wages and employment in local labour markets. Another significant unexplored issue is the role of rent sharing in an industry where human capital is particularly important. Our short panel precluded any dynamic analysis but longer time series could throw light on the determinants of high technology employment growth that have generated the distribution and composition of existing clusters.
The second and initially somewhat surprising result is the strong positive correlation between wages and employment in high technology industries. This association is apparent in all equations and appears to be quite robust. Our view is that the positive partial correlation in the employment equation could arise from two sources which are not mutually exclusive omitted variables and skill aggregation.
The most important omitted variable is the absence of a demand or sales variable in the employment function. This equation therefore essentially captures a labour supply relationship. For example, if the demand for products produced by the most highly skilled, and paid, labour grows fastest due to innovation, government procurement or whatever, employment and real wages would be positively related in a regression which did not control for demand effects. Likewise, if there are shortages or bottlenecks of key skilled personnel, their wages would rise with demand for their products or services to include a scarcity rent.
A complementary explanation relates to the heterogeneity of skill categories in the high technology sector. It is obviously rather crude to estimate employment demand implicitly assuming a single, homogenous category of labour. There exists both skilled and relatively unskilled employment in the high technology sector and the demand for these categories may move in different directions.
Naturally, university knowledge spillovers are not the only reason for high technology clusters. Other forces for localization are quite strong. They would include the development of specialized intermediate goods industries, economies of scale and scope, and network externalities. With respect to the latter, innovations by different producers may be complementary, yielding related new products or processes when combined. On these questions too, further research is called for.
Acs, Z. J., Audretsch, D. B. & Feldman, F. (1994) R&D Spillovers and Recipient Firm Size, Review of Economics and Statistics,100(1).
Acs, Z. J., Audretsch, D. B. & Feldman, F. (1992) Real Effect of Academic Research: Comment, American Economic Review, 82(1), 363-67.
Acs, Z. J. (1993) US High Technology Clusters, Department of Economics, University of St. Andrews, Discussion Paper, No. 9315.
Almeida, P., Kogut, B. (1994) Technology and Geography: The Localization of Knowledge and the Mobility of Patent Holders, Department of Management, Wharton School, University of Pennsylvania.
Audretsch, D. B., Feldman, M. P. (1994) Knowledge Spillovers and the Geography of Innovation and Production, Center for Economic Policy Research Discussion Paper, No. 953.
Audretsch, D. B., Vivarelli, M. (1994) Small Firms and R&D Spillovers: Evidence for Italy, Centre for Economic Policy Research Discussion Paper, No. 927.
Bania, N., Eberts, R. & Fogarty, M. (1987) The Role of Technical Capital in Regional Growth, Presented at the Western Economic Association Meetings.
Bania, N., Eberts, R. & Fogarty, M. (1993) Universities and the Start-Up of New Companies: Can We Generalize from Route 128 and Silicon Valley? Review of Economics and Statistics, 75,761-66.
Beeson, P., Montgomery, E. (1993) The Effects of Colleges and Universities on Local Labor Markets, Review of Economics and Statistics, 75, 753-61.
Business Week (1994) Why Are We So Afraid of Growth?, 62-72, May 16th.
Jaffee, A. B. (1989) Real Effects of Academic Research, American Economic Review, 79(5), 957-70.
Edwards, K. L., Gordon, T. J. (1984) Characterization of Innovations Introduced in the US Market in 1982, The Futures Group.
Jaffee, A. B., Trajtenberg, M. & Henderson, R. (1993) Geographic Localization of Knowledge Spillovers as Evidenced by Patent Citations, Quarterly Journal of Economics, 577-98.
Krugman, P. (1991) Increasing Returns and Economic Geography, Journal of Political Economy, 99(3), 483-99.
Link, A. N., Rees, J. (1990) Firm Size, University Based Research, and the Returns to R&D, Small Business Economics, 2, 25-31.
Lund, L. (1986) Locating Corporate R&D Facilities, Conference Board Report No. 892, Conference Board, New York.
Lumme, A., Kauranen, L., Autio, E. & Kaila, M. M. (1993) New Technology Based Companies in Cambridge in an International Perspective, Working Paper No. 35, Small Business Research Centre, University of Cambridge.
Nelson, R. R., (1986) Institutions Supporting Technical Advance in Industry, American Economic Review, Papers and Proceedings, 76(2), 186-89.
Parker, D. D., Zilberman, D. (1993) University Technology Transfers: Impacts on Local and U.S. Economies, Contemporary Policy Issues, XI(2), 87-99.
Shachar, A., Felsenstein, D. (1992) Urban Economic Development and High Technology Industry, Urban Studies, 29(6), 839-55.
1) Shachar and Felsenstein (1992) report evidence from studies conducted in Europe and Japan which show very few benefits arising from the close physical proximity of high technology firms to a local university.
2) It should be noted that R&D is not a good measure of small firm inputs into knowledge production since such inputs often arise informally without the support of an R&D laboratory.
3) See Acs (1993) for a detailed description.
4) Inclusion of a lagged dependent variable in (1) yielded a coefficient of almost unity and impaired the explanatory power of most other variables, suggesting a relative lack of movement in employment over time.
5) All equations were also estimated with coarser regional instead of state fixed effects, yielding very similar results, though the R&D employment elasticity was smaller in this model. White standard errors to control for heteroscedasticity differed little from the reported results.
6) Note that Montgomery and Beeson omit real wages and prior innovations from their employment equation. They do, however, include several variables to control for the effects of other area attributes on local labour market conditions, which we capture using state dummies.
|Return to Babson College
Main Home Page
©1996 Babson College. All rights reserved.
Last updated November 14, 1996 by Cheryl Ann Lopez