Is mobility of technical personnel a source of R&D spillovers? ¤ by Jarle M¿en z Draft, July 10, 2000 Abstract: Labor mobility is often considered to be an important source of knowl- edge externalities, making it di±cult for ¯rms to appropriate returns to R&D in- vestments. In this paper, I argue that inter-¯rm transfers of knowledge embodied in people should be analyzed within a human capital framework. Testing such a framework using a matched employer-employee data set, I ¯nd that the technical sta® in R&D-intensive ¯rms pays for the knowledge they accumulate on the job through lower wages in the beginning of their career. Later they earn a return on these implicit investments through higher wages. This suggests that the potential externalities associated with labor mobility, at least to some extent, are internalized in the labor market. JEL classi¯cation: J24, J31, J62, O32 Keywords: Labor mobility, Compensating di®erentials, Human capital, R&D-capital, R&D spillovers, Matched employer-employee data ¤ Paper prepared for the NBER Summer Institute 2000. I am grateful to Zvi Griliches for suggesting this topic to me, and to Tor Jakob Klette for his continuous advice and encouragement. I have also received useful comments from Pierre Azoulay, Torbj¿rn H½geland, Oddbj¿rn Raaum, Kjell G. Salvanes, Scott Stern and seminar participants in Oslo and Cambridge. Remaining errors are the sole responsibility of the author. The project is ¯nanced by the Research Council of Norway. z Norwegian School of Economics and Business Administration, Department of Economics, Hellevn. 30, 5045 Bergen, Norway; and Statistics Norway, Microeconometric Division. E-mail: jarle.moen@nhh.no. http://www.nhh.no/sam/cv/moen-jarle.html Tel: + 47 55 95 95 49. Fax: + 47 55 95 95 43. \Don't let your employees do to you what you did to your former boss." The golden rule of protecting trade secrets, as de¯ned by Intel general counsel Roger Borovoy (Jackson; 1997) 1 Introduction Labor mobility is likely to be a very important source of knowledge di®usion. Sur- veying one hundred founders of companies on the 1989 Inc. `500' list of the fastest growing companies in the United States, Bhide (1994) ¯nds that 71 percent \repli- cated or modi¯ed an idea encountered through previous employment." With respect to technical employees, Almeida and Kogut (1996), demonstrate by an analysis of patent data from the semiconductor industry that ideas are spread through mobil- ity of key engineers. Evidence of this kind, however, does not justify the common proposition that labor mobility is an important source of knowledge spillovers, i.e. an externality causing underinvestment in research among private ¯rms. The aim of this paper is to examine this proposition empirically, and I will argue that there are market mechanisms that, at least partially, solve the problem. The link between labor mobility and knowledge spillovers dates back to Arrow's (1962) article on the public good aspect of knowledge. Arrow writes that \no amount of legal protection can make a thoroughly appropriable commodity of something as intangible as information" and adds that \[m]obility of personnel among ¯rms provides a way of spreading information" (p. 615). Following Arrow's seminal work, a large literature on R&D spillovers has evolved, and economists working in the ¯eld have continued to consider labor mobility an important spillover channel. Geroski (1995) expresses what appears to be a common view1 , writing that \[l]ast but not least, spillovers occur when a researcher paid by one ¯rm to generate new knowledge transfers to another ¯rm (or creates a spin-o® ¯rm) without compensating his/her former employer for the full inventory of ideas that travels with him or her." That workers do not make such compensations seems obvious since they already possess their employers' knowledge when they decide to leave. The timing of events that Geroski implicitly suggest, however, is misleading. To the extent that research 1 Similar statements are found e.g. in Ja®e (1996), writing that \[k]nowledge spillovers also occur when researchers leave a ¯rm and take a job at another ¯rm", in Stephan (1996) writing that \[f]uture work should also focus on the role mobility within the industrial sector plays in facilitating spillovers" and in Gersbach and Schmutzler (1997) writing that \[s]pillovers arise because employees who change jobs take with them all their knowledge, some of which is not speci¯c to their original ¯rm." 1 work has a general training element, workers may pay for knowledge as it is accu- mulated. To what extent labor mobility actually reduces appropriability and R&D investments, therefore, is an empirical question. The approach taken in this paper is to test key implications of models that as- sume perfect markets. Even though there may exist numerous market imperfections, an analysis of R&D spillovers should in my view have perfect markets as its point of departure. If using standard methodologies2 without ¯rst considering such a `bench- mark' case, the results of ordinary market exchange may mistakenly be interpreted as R&D spillovers, and public policy may be misguided3 . The basic implications of labor mobility follows from classical human capital theory, cf. Mincer (1958) and Becker (1962, 1964). To the extent that workers in R&D-intensive ¯rms get access to valuable knowledge on the job, they will expect higher wages in the future. When holding jobs that give access to such knowledge, they should therefore be willing to pay for what they learn by accepting wages below their alternative wage. This hypothesis can be tested by using extended Mincer (1974) wage regressions, which is the standard approach in the training literature. Utilizing a large matched employer-employee data set from the Norwegian ma- chinery and equipment industry, I ¯nd that the technical sta® in R&D-intensive ¯rms pay for the knowledge they accumulate on the job through lower wages in the beginning of their career, and that they later earn a return on these implicit investments through higher wages. If choosing an `R&D intensive' career scientists and engineers have to accept a wage discount in the order of six percent in their ¯rst year after graduation. This should be considered a conservative estimate, due to a likely ability bias. Towards the end of their career, they receive a wage pre- mium in the order of seven percent. Similar results apply for workers with secondary technical education. When estimating the price paid for learning separately from the return to research experience4 , I ¯nd that having work experience from R&D intensive ¯rms is associated with higher wages, while the employers current R&D intensity reduce wages for workers with less than 20 years experience. Furthermore, as predicted by human capital theory, the youngest workers appear to invest most heavily in on-the-job learning. These ¯ndings suggest that the potential externali- ties associated with labor mobility, at least to some extent, are internalized in the labor market5 . 2 Cf. e.g. Ja®e (1986) and Ja®e, Trajtenberg and Henderson (1993). 3 Zucker, Darby and Armstrong (1998), Klette and M¿en (1999) and Klette, M¿en and Griliches (2000) elaborate on this point. 4 I will use `R&D experience' as a short term for experience in R&D intensive ¯rms. 5 This does not guarantee optimal R&D investments, however, as credit restrictions or risk averse preferences may reduce worker's willingness to `co-¯nance' R&D. I will return to this in the concluding section. 2 With respect to mobility patterns, I ¯nd that excess labor turnover is less in R&D intensive ¯rms. This e®ect is particularly pronounced for workers with secondary technical education. If changing employer, workers tend to move to a ¯rm with an R&D intensity similar to their former employer. Consistent with the lower turnover in R&D intensive ¯rms, research experience from the current employer appears to be more valued than research experience from previous employers. 2 R&D investments and human capital theory Research is a learning process, and R&D investments, therefore, may not only in- crease a ¯rms' stock of innovations, but also increase the human capital of research workers. In the literature, however, R&D capital (Griliches; 1973), and human cap- ital (e.g. Becker; 1964) are rarely discussed together. To clarify the relationship between these concepts I ¯nd it useful to classify knowledge along two main dimen- sions; ¯rst to what extent it can be codi¯ed, and second to what extent it is di®used throughout the economy. R&D capital is knowledge that is not di®used throughout the economy. As long as knowledge is exclusive, those who possess it can earn a monopoly rent and this rent is what motivates investments in R&D. When knowledge is di®used and have become available in the public domain, it will earn a normal rate of return for those who undertake the cost necessary to make it part of their human capital. A research project whose results are perfectly codi¯able and not di®used, can be considered pure R&D capital. It can in principle be protected by patents or other intellectual property right instruments. Hence, labor mobility is not a concern to ¯rms when it comes to appropriating returns in this case. However, many innovations cannot be properly codi¯ed or protected by patents. Appropriability of rents then hinges on the ¯rms' ability to avoid that knowledge leaks out to competitors. Innovations of this type make R&D spillovers and the problem of underinvestment an important issue. If the intellectual property rights cannot be protected by patents, the R&D capital of ¯rms is to a large extent embod- ied in their employees. Such knowledge resembles what Zucker, Darby and Brewer (1998) have called intellectual human capital. A ¯rm that has ¯nanced R&D creat- ing intellectual human capital, cannot prevent its workers from leaving and taking the knowledge with them. Under these circumstances labor mobility is potentially a threat to the ¯rms, and Pakes and Nitzan (1983) analyze the investment incen- tives of entrepreneurs facing such a situation. Pakes and Nitzan conclude that it is possible to design labor contracts which solve the problem, and that labor mobility therefore does not reduce the appropriability of R&D investments. I will return to their analysis below. 3 Over time, as knowledge di®uses, intellectual human capital may either become more codi¯able and eventually be learned in schools, or it may become human cap- ital acquired through on-the-job training. The latter kind of human capital also has relevance for an analysis of labor mobility and R&D investments, as there may be more to learn in ¯rms conducting research because such ¯rms are likely to use the most up-to date technology and frequently change its products and production processes. This training may be valuable to other ¯rms. Furthermore, the distinc- tion between intellectual human capital and on-the-job training does not constitute a clear dichotomy. Many innovations are incremental product and process improve- ments made at the factory °oor, and in the limit they may as well be considered excellent craftsmanship as innovations. A case were di®erent ¯rms o®er di®erent opportunities for on-the-job training is analyzed by Rosen (1972). The rest of this section will outline the theoretical models of Pakes and Nitzan (1983) and Rosen (1972). The main predictions of these models will be discussed and tested in the empirical part of the paper. The Pakes-Nitzan model The point of departure in Pakes and Nitzan (1983) is Arrow's (1962) reference to labor mobility as a source of R&D spillover. They argue that even though mobility of scienti¯c personnel will spread knowledge produced in industrial laboratories, it need not be a mechanism which reduces the pro¯tability of research projects and employment in such projects. Both scientists and ¯rms are aware of the fact that working on a research project gives access to valuable information6 . Once such information is disclosed or developed, scientists, if they are to stay with the ¯rm, will have to receive a wage increase re°ecting their new market value. Thus, scientists expect that accepting a research position implies a future wage increase, and consequently they accept an initial wage below their alternative wage. Next, Pakes and Nitzan notice that if the innovation makes the ¯rm a true monopolist, it will never be pro¯table for the ¯rm and the scientist to split, since the sum of rents in a duopolistic market will be less than the monopoly rent7 . Mobility, therefore, should only be observed when it increases the joint pro¯t of the ¯rm and the scientist. This may happen if the ¯rm cannot avoid that other ¯rms 6 Pakes and Nitzan (1983) explicitly model the uncertainty involved in research. This feature of the model does not alter the simple intuition given here, however, because they assume that utility functions are linear in income. Discussing this assumption, they acknowledge that both risk aversion and a lower bound on wages will a®ect R&D investments. 7 Pakes and Nitzan (1983) model only a situation with one entrepreneur and one scientists. If several scientists have equal access to the same critical information, this will complicate the analysis because of potential strategic interaction among the scientists. 4 get access to valuable information and enter the market8 . The scientist, by setting up a rival, will then break into pro¯ts which otherwise accrue to third parties, and since this pro¯t will be part of the scientists alternative wage in `period two', it is possible for the ¯rm to extract this rent when setting the `period one' wage. Another situation which may induce the scientist to join or set up a rival is when the research project create `spin-o®s', some of which are better exploited in a separate ¯rm due to coordination costs. Summarizing the insight of their model, Pakes and Nitzan writes that mobility of scienti¯c personnel is not, in itself, a source of concern to entrepreneurs. ... [A]n optimizing entrepreneur who is free to choose among alternative contracts will always choose one which only induces the scientist to leave and join a rival if the sum of the bene¯ts to the two agents increases as a result of the scientist's leaving. Contracts which specify labor payment in the form of a °at wage and stock option (or other pro¯t sharing agreement) ought to be able to induce close approximation to this behavior. Balkin and Gomez-Mejia (1985) provide empirical evidence in support of Pakes and Nitzan's prediction. Surveying 105 companies in the Route 128 region around Boston, they ¯nd that incentive pay programs are far more common in high-tech ¯rms than in other ¯rms, and that such programs are used for broad levels of technical employees. In addition, key scientists and engineers who help form the companies at an early stage, are given long term stock options. Rosen's 1972 model The Pakes and Nitzan (1983) model is a two period model of scientists and entrepreneurs, where scientists get access to valuable information, but don't increase their generic productivity. Rosen (1972) models on-the-job learn- ing in a more general human capital context, although one where ¯rms do not have market power. He uses a compensating di®erential framework, and turns it into \an economic theory of occupational mobility". Rosen thinks of jobs as tied packages of work and learning. Workers sell the services of their skills and simultaneously purchase an opportunity to augment those skills. Some jobs provide more learning opportunities than others. The di®erence between the maximum market rental of a worker's existing skills and the wage that he or she receives in a given job, is the 8 Note that spillovers at this point enter the story, but mobility will be a consequence of spillovers, not a source of spillovers. Information can leak out to third parties by reverse engineering, inspec- tion of patent documents, independent research on the same technological problem, etc. Cf. Levin, Klevoric, Nelson and Winter (1987) for a survey of the importance of various information channels. Labor mobility receives a middle score in their study. 5 implicit price the worker pays for learning. Basic human capital theory suggests that a worker's incentive to accumulate human capital is largest at young age. As the worker grows older he or she will have fewer years to collect returns on a given invest- ment, and obviously workers have no incentives to pay for increasing their human capital in the last year before retirement. This imply that the \optimal human cap- ital investment program is implemented by a sequence of job assignments in which workers systematically move and are promoted across jobs that o®er successively smaller learning opportunities" (Rosen; 1986). The point of departure in Rosen's model is a net wage equation y = !H ¡ P (k) (1) where y is income, ! is the unit rental price of human capital and k is an index measuring potential learning-by-experience on the job, k 2 [0; k]. P (k) is an implicit or shadow price function giving the market equalizing wage di®erential between a job with no learning potential and a job with learning potential k. Assume that the actual amount of learning by individual i is proportional to k and depends on individual i's ability, ®i 2 [0; 1] such that : Hit = ®i k: (2) The workers problem is then to choose a sequence of jobs, kt , over his or her lifetime, T , to maximize the present value of income, i.e. Z T maxV = [!Ht ¡ P (kt )] e¡rt dt (3) kt 0 : subject to an initial stock of human capital, H0 and Hit = ®i k. Optimization requires that at any time, t 2 [0; T ], P 0 (kt ) !£ ¤ = 1 ¡ e¡r(T ¡t) : (4) ®i r The expression on the left hand side is the marginal cost of investing in human capital, and the expression on the right hand side is the discounted marginal return. It seems reasonable to assume that P 0 (k) > 0 and P 00 (k) > 0; i.e. that the marginal cost of learning is positive and increasing. Given this, optimality requires kt to be largest at the time of entry into the labor market and then to decrease monotonically over time. Note that the marginal cost of a given real investment in human capital de- creases with ability. Hence, workers with higher ability will, all else equal, ¯nd it pro¯table to choose jobs with greater learning potential. In the words of Rosen (1972): \Economic incentives induce more `able' workers to learn more and to ac- cumulate knowledge more rapidly than the less `able'." This will give rise to a potential selection problem (ability bias) in the empirical application of the model. 6 3 Data issues The data used in this study comes from three main sources: Governmental admin- istrative records prepared by Statistics Norway9 , the annual manufacturing census of Statistics Norway10 , and the biannual R&D survey of Statistics Norway11 sup- plemented with other surveys of immaterial investments and innovations done by the same agency12 . The Norwegian data are extraordinary in the sense that the entire working population can be followed over a number of years, and in the sense that extremely rich information is available both about the workers and about their employers. When analyzing labor mobility, the extensive coverage o®ered by the Norwegian data is a great advantage. I have chosen to focus on the technical sta®13 in the machinery and equipment industries as these industries have many high-tech ¯rms and have a fairly complete coverage in the R&D surveys. The matched employer-employee data set covers the years 1986 to 1995, and I have only included men employed full time in the analysis below. Women do not constitute a large share of the labor stock in these industries, and they are known to have di®erent career patterns and preferences than men. Even though the data set is rich, I do not have complete information about the workers' careers. Small ¯rms are not necessarily included in the R&D surveys14 , and the matching between the di®erent data sources is not perfect. Due to these prob- lems R&D information is missing for approximately 20 percent of the worker-year observations, and some of the R&D information present is imputed from previous or later ¯rm observations15 . On the positive side, however, I do have some information 9 Cf. Barth and Dale-Olsen (1999), appendix 2, for some details on the various registers in- cluded in the data base. I have taken great care to improve the data quality by checking for consistency across years and across related variables for the same individual, and by ¯lling in missing information where possible. 10 The census is documented in the series Manufacturing statistics, O±cial Statistics of Norway NOS C36, Statistics Norway, Oslo. Microdata are available annually from 1972. 11 Microdata is available for 1970, biannually 1975-81, annually 1981-85 and biannually 1985-95. Prior to 1991 the R&D surveys were conducted by the Royal Norwegian Council for Scienti¯c and Industrial Research. The 1970 survey has been linked to the 1972 manufacturing census. 12 A survey of immaterial investments was conducted in 1988 covering the years 1986-88 and in 1990 covering the years 1988-90. An innovation survey was conducted in 1993 for the year 1992. 13 I de¯ne the technical sta® as workers with secondary technical education and workers with higher technical or scienti¯c education. I refer to the latter group as scientists and engineers. 14 The R&D surveys have close to full coverage for ¯rms with more than 20 employees. 15 I have used the following procedure when constructing the R&D database: First I have linked the R&D surveys to the manufacturing census. Next, for ¯rms and years not included in the R&D surveys I have used R&D information from the surveys of immaterial investments, and from the innovation survey. For ¯rms and years were R&D information is still missing, I have used survey information about planned R&D one and two years ahead, and information about previous R&D. In 7 about workers' careers prior to 198616 , the ¯rst year included in the matched data set. I know when the workers started the job they held in 1986, and this can be combined with information about the employers' R&D investments, in some cases dating as far back as 1972. Both the (normalized) length of the highest attained education, and the type of education, is recorded in the data. Occupation, however, is not available. Hence, it is not possible to look speci¯cally at researchers, and workers' learning will be proxied by the employers' R&D intensity. I measure R&D intensity as R&D man-years per employee at the three-digit line of business level within ¯rms17 . I have censored this variable at 0.8 in order to reduce the in°uence of outliers18 . If all workers within a ¯rm participate equally in the ¯rm's R&D e®orts, R&D man-years per employee will measure the share of time that each worker uses to perform R&D. Since R&D work obviously is not shared equally among the employees, R&D intensity is a noisy proxy for what we want to capture. Measurement errors in the R&D variable, cf. footnote 15, add to this noise. Earnings is measured as taxable labor income19 . The experience measure is real work experience for the youngest cohorts, years since graduation for older cohorts, and potential work experience for cohorts graduating before November 197020 . Both experience and tenure are measured in years completed at the beginning of the calendar year. Information about trimming procedures is given in table A1. In the mobility analysis, all observations with complete information are used, whereas observations with unreliable earnings measures have been excluded from the wage regressions. Trimming based on earnings reduces the sample by 8 percent. Table A2-A6 describe the ¯nal stage, missing R&D variables were imputed by linear interpolation, and by extrapolating the ¯rst observed R&D intensity backwards in time and the last observed R&D intensity forward in time, ¯rm by ¯rm. Firms R&D investments are known to be stable over time. Imputing missing information when possible, therefore, seems preferable to deleting the observations. 80 percent of the worker-year R&D variables are from surveys, 5 percent are imputed by interpolation and 15 percent are imputed by extrapolation. 65 percent of the imputed R&D intensities are zero. 16 Note, however, that for 16 percent of the observations, the starting date is censored at April 30th 1978. I have used a dummy variable to resolve this problem in regressions where tenure is included. 17 This means that R&D intensity is measured at a level `in between' the ¯rm and the plant. I will use the term ¯rm level R&D intensity in what follows. If R&D man-years were not reported, this variable has been imputed based on the ¯rms' R&D spending. 18 This a®ects 0.4 percent of the observations with positive R&D intensity. 19 I will often refer to this as the workers' wage. 20 Potential work experience is age minus schooling minus seven. Real work experience is mea- sured as years since graduation adjusted for pre graduation work experience, part time employment and unemployment within the sample years 1986-1995. When dummy variables are used, they are based on the integer of experience. 8 the sample and the main variables. 4 R&D investments and labor mobility. At ¯rst sight, both Rosen (1972) and Pakes and Nitzan (1983) seem to have speci¯c predictions with respect to mobility patterns. A main prediction of Rosen's model is that workers consistently move to jobs with less learning opportunities. In my context, that may imply that workers move from more to less R&D intensive ¯rms, but as pointed out by Rosen himself, there is heterogeneity with respect to the learning content of jobs not only across, but also within ¯rms. Hence, a clear prediction cannot be deduced. Pakes and Nitzan (1983) predict that R&D ¯rms are able to avoid turnover, and thereby spillovers, by sharing the monopoly rent at stake with the workers. In the presence of spin-o® innovations or sources of spillovers other than labor mobility, however, they show that mobility actually can be a way of appropriating returns. The model, therefore, like Rosen's, fails to give clear predictions with respect to worker mobility between ¯rms. Furthermore, Pakes and Nitzan (1983) do not con- sider ¯rm speci¯c knowledge. If ¯rms with di®erent levels of R&D intensities di®er with respect to ¯rm speci¯c human capital, this will also in°uence the relationship between turnover and R&D investments21 . In lack of strong predictions, empirical mobility patterns cannot be used as a test of the theories. A descriptive analysis of mobility patterns still has interest, however, as it will give insight into the outcome of the di®erent forces at play. R&D investments and worker °ows Based on Rosen's model, despite the lack of an absolute prediction, one would expect a tendency for workers to move from more to less R&D intensive ¯rms as a way of reducing their learning in accordance with an optimal human capital investment plan. To investigate this I have calculated a transition matrix of job changes for technical employees between plants with known R&D intensities. The matrix is reported in table 1. The most striking result is that workers tend to move between ¯rms with similar levels of R&D intensity. 65.5 percent of workers leaving a ¯rm that does not conduct R&D (within the plant's line of business) move to another ¯rm that does not conduct R&D, even though jobs in such ¯rms account only for 34.6 percent of all jobs. 64.0 percent of workers leaving a ¯rm with R&D intensity above 0.2 move to another ¯rm with R&D intensity above 21 In the training literature, the e®ect of training on turnover propensities has been used to assess whether the human capital built up is general or ¯rm speci¯c, cf. e.g. Loewenstein and Spletzer (1999) and Parent (1999). 9 0.2, although such ¯rms only account for 5.9 percent of all jobs. The pattern is the same for workers leaving ¯rms with intermediate levels of R&D intensity. One explanation for the observed stability in R&D intensity across jobs may be that there is some speci¯city associated with a given technology level within the industry. As workers grow older, they will then prefer jobs with less learning, within ¯rms at the same level of R&D intensity as those they have previously worked for. Another explanation may be that workers have preference for work at a given technology level22 . The discovered stability is worth investigating further because this feature of the data may help impute missing career information when later estimating the e®ect of R&D on wages. Table 2 reports the results of regressing past R&D intensities on current R&D intensity to see how well current R&D intensity performs as a predictor. The regressions are run separately for workers who stay with the same employer and for workers who change employer. Not surprisingly, there is a high degree of stability for stayers, but the results for workers who have changed employer con¯rm that there is also a fairly high degree of stability in R&D intensity across jobs. R&D investments and labor turnover As explained above, Pakes and Nitzan (1983) investigate the relationship between R&D and labor turnover theoretically without reaching a ¯rm conclusion. Table 3 reports labor turnover for technical employees in ¯rms with di®erent levels of R&D intensity in my sample. What seems most relevant to explore is how R&D investments a®ect `churning', i.e. hires and quits over and above the level necessary to accomplish changes in the number of employees. Excess turnover, a measure of churning23 , seems to decrease with R&D intensity both for workers with secondary technical education and for workers with higher technical or scienti¯c education. A descriptive analysis of excess turnover is not su±cient, however, as a closer inspection of the data reveal that there are signi¯cant di®erences between ¯rms having di®erent levels of R&D intensity, with respect to other characteristics known to in°uence turnover such as workers' experi- ence. In order to isolate the e®ect of R&D on excess turnover, therefore, a regression framework is called for. Table 4 reports regression results for both a tobit and maximum likelihood 22 The work of Almeida and Kogut (1996), Stern (1999) and others suggests that scientists and technical personnel have preferences regarding the technological environment that they work in. 23 Cf. Burgess, Lane and Stevens (1996) and Barth and Dale-Olsen (1999). The excess turnover rate is half the churning rate. I have calculated the excess turnover rate as separations out of jobs that continue divided by the number of continuing jobs. 10 grouped logit estimator24 . The estimated relationship is excess turnover it = f (R&D-int. ¤ Dsec. edu. ; R&D-int. ¤ Dhigher edu. ; X) (5) The unit of observation is educational groups within plants. Control variables, X, include a quadratic in the educational group number of workers, a quadratic in their average experience, a quadratic in plant age and year dummies25 . In the tobit regression I have followed Barth and Dale-Olsen (1999) by excluding small units, limiting the sample to educational groups that consist of at least ¯ve workers, cf. footnote 24. Both the tobit and the grouped logit speci¯cation show that excess turnover is lower in R&D intensive ¯rms. The e®ect is, however, particularly evident for workers with secondary technical education. One possible explanation is that human capital accumulated by workers with secondary technical education is more ¯rm speci¯c than human capital accumulated by workers with higher technical or scienti¯c education. It may also indicate that the mechanisms related to spin- o® innovations and other spillover channels, modelled in Pakes and Nitzan (1983), are relevant in the industries investigated. Workers with higher education would probably be most a®ected by these mechanisms which increase turnover. The results in table 4 is consistent with other ¯ndings in the empirical liter- ature. Pacelli, Rapiti and Revelli (1998) who estimate the probability of worker ¯rm separations in Italy, ¯nd that \more innovative ¯rms cultivate more durable employer-employee relationships", and Greenhalgh and Mavrotas (1996) analyzing 24 Barth and Dale-Olsen (1999) estimate the e®ect of employers' wage policies on excess turnover, and treat the excess turnover rate as a characteristic of the ¯rm. This leads them to use a tobit estimator. Within such a framework, the observed excess turnover rate must be considered an estimate of a target rate implicit in the ¯rms' personnel policy, and Barth and Dale-Olsen (1999) think in terms of a latent variable censored from below at zero. (One might add to this that the excess turnover rate is also censored from above at one.) As an estimate for the target rate, however, the observed rate is heteroscedastic with a variance proportional to the inverse of the number of employees. Barth and Dale-Olsen (1999) do not explicitly discuss this, but alleviate the problem by limiting the analysis to large ¯rms. One of the reasons they give for this is that the excess turnover rate is measured with \large uncertainty" for very small ¯rms. Grouped logit eliminates this heteroscedasticity problem. Thinking of the data in this way also changes the perspective from the ¯rm unilaterally deciding an excess turnover rate to individual employer-employee relationships which may or may not continue, depending on both ¯rm and worker characteristics. I ¯nd this to be a preferable perspective, as individual employer-employee relationships is the true unit of observation, and it makes sense conceptually to divide observed quits into two groups, those who are replaced, and those who are not replaced. The ¯rst type of quits constitute excess turnover while the second type of quits are due to job destruction. If we knew which of the workers who separate that belong to which group, we would no doubt use logit or probit. When we only know the proportion of workers belonging to each group, we can apply grouped logit or probit, cf. Greene (1997, chapter 19.4.3). 25 A complete list is given in the subtext to table 4. 11 the British labor market, ¯nd that sectoral R&D is associated with lower mobility. They attribute this only to the presence of ¯rm speci¯c human capital, however, claiming that \the skills acquired [in R&D intensive sectors] are rather more speci¯c than average". 5 The e®ect of R&D investments on wages Pakes and Nitzan (1983) predict lower starting wages and higher wage growth for workers choosing a research career, and Rosen (1972) predicts the same pattern more generally for workers having jobs with a high learning potential. A key assumption behind both models is that workers mainly acquire general human capital on the job. Testing these models, i.e. testing to what extent di®erent ¯rms o®er di®erent learning opportunities, and to what extent workers pay for their knowledge accumu- lation, we would like to estimate equation (1) which is Rosen's point of departure. In principle this is possible. Human capital, H, can be decomposed and the price or relative weight of its various components can be estimated using a standard log- linear hedonic wage regression. Furthermore, potential learning-by-experience on the job, k, may be proxied by the employer's R&D intensity as it seems reasonable to assume that workers in `high-tech', R&D intensive ¯rms learn more than work- ers in `low-tech' ¯rms. However, some problems are immediately evident. Work experience needs to be decomposed according to the training or research content of the jobs that workers have had at di®erent stages of their career, but as explained earlier, complete information about the worker's career histories is not available. Furthermore, it is far from obvious how one can summarize what is known about the workers' experience from di®erent ¯rms into a good measure of human capital. In what follows, I will suggest several solutions to these problems. A ¯rst look at the e®ect of R&D on the earnings pro¯le One way to get around the missing career data, is to assume that workers career trajectories are such that the R&D intensity is constant over their career. Tables 1 and 2 show that this assumption is valid as an approximation. We can then utilize the structural relationship between k and H, given in equation (2) together with the optimal time path for learning investments implicit in (4). Under this assumption the R&D intensity will at each point of time reveal information both about k and about the component of H representing accumulated R&D experience. More speci¯cally, the estimated joint e®ect will give the returns to R&D experience minus the cost of learning. Working for a highly R&D intensive employer should cause a large negative wage premium early in the career, re°ecting the implicit price paid for the R&D experience. At the same time, this experience has not had much time to a®ect 12 the stock of human capital. As time goes by, workers' willingness to pay for human capital accumulation decrease and approaches zero, but di®erences in previous R&D experience will translate into di®erences in human capital. Workers who are in R&D intensive ¯rms and have a long R&D intensive career behind them, will therefore have a large positive wage premium re°ecting the human capital accumulated. Table 5 reports the results of simple OLS wage regressions where cross-terms between experience and current R&D intensity are added to test the hypothesis that employees with a career in R&D intensive ¯rms have a steeper experience- earnings pro¯le than other workers. Additional control variables included are years of schooling, seven experience dummies26 , a quadratic in plant number of employees and year dummies. In column 1 and 3 the experience dummies are interacted directly with R&D intensity while column 2 and 4 report the results of interacting the experience dummies with a dummy which is one if the R&D intensity is above 0.227 . An R&D intensity of 0.2 represents the 97th percentile for workers with secondary technical education, and the 82th percentile for workers with higher technical or scienti¯c education. The dummy approach is used as an easy way to assess the magnitude of the e®ect of R&D intensity on wages. An alternative illustration is given in ¯gure 1, where earnings-experience pro¯les for workers in ¯rms with no R&D and in ¯rms with R&D intensity 0.2 is graphed, based on a speci¯cation with a quartic in experience interacted with a quadratic in R&D intensity. The results support the main theoretical prediction of Pakes and Nitzan (1983) and Rosen (1972). Early in the career both workers with secondary technical edu- cation and scientists and engineers accept a signi¯cant wage discount when working for R&D-intensive ¯rms, but over time this discount is changed into a signi¯cant wage premium. The pattern strongly suggests that R&D-investments of ¯rms trans- late into general human capital, and that workers both pay and get paid for the knowledge they accumulate. It is evident from table 5, columns 2 and 4, using the dummy variable approach, that the discounts as well as the premia are of economic signi¯cance. Scientists and engineers working in ¯rms with an R&D intensity above 0.2, have on average 6.1 percent lower wages in their ¯rst year than scientists and engineers in ¯rms 26 I have chosen to use experience dummies rather than a higher order polynomial in the main speci¯cation because the tabulation of cross terms between R&D intensity and a higher order polynomial is di±cult to interpret. A polynomial in experience interacted with R&D intensity also imply a stronger restriction on the e®ect of R&D over the career. 27 In these regressions, workers in ¯rms with medium R&D intensity have been excluded. Medium R&D intensity is de¯ned as an R&D intensity between 0.05 and 0.2. The exclusion is done to facilitate a sharper comparison between workers in ¯rms with high and low R&D intensity. The results are robust to including workers in ¯rms with medium R&D intensity, and to using the 90th percentile for each group as a cuto® point instead of 0.2 R&D intensity. 13 with R&D intensity below 0.05. Scientists and engineers with more than 35 year experience and working in a ¯rm with R&D intensity above 0.2, have wages that on average are 6.8 percent above the wages of scientists and engineers with similar experience in ¯rms with R&D intensity below 0.05. The magnitudes of the discounts and premia are similar for workers with secondary technical education in R&D intensive ¯rms. They have a 5.5 percent wage discount in the beginning of their career, and an 8.6 percent premium in the end of their career. One way to check the plausibility of the coe±cients is to calculate the internal rate of return to choosing an R&D intensive career. For a worker with secondary technical education, the internal rate of return is 5.7 percent, and for workers with higher technical or scienti¯c education it is 3.6 percent28 . These numbers should be considered rough estimates, but they are in a reasonable range. Estimates based on earnings growth One major obstacle to identifying com- pensating di®erentials, whether associated with training or other job amenities, has been the potential correlation between job amenities and unobserved individual char- acteristics. In Rosen's model, an ability bias arises because highly talented workers have a lower cost of learning, and absorb more knowledge in a job with a large po- tential for learning, than less talented workers29 . This imply a tendency for talented workers to self-select into R&D intensive ¯rms, causing the wage discount in the beginning of the career to be underestimated, and the wage premium in the end of the career to be overestimated30 . Another potential bias in table 5 arises from workers switching between employ- ment in `high-tech' and `low-tech' ¯rms. Although table 1 indicates that this kind of behavior is not very common, it clearly does happen. A bias then arises because the regressions in table 5 assume that we can compare experienced workers in R&D intensive ¯rms to experienced workers in less R&D intensive ¯rms, and learn how much more human capital is accumulated in R&D intensive ¯rms. Workers who transfer out of R&D intensive ¯rms, however, will increase the wage level of the `comparison group' in the less R&D intensive ¯rms, and cause a downward bias on 28 The calculation is based on the regressions in table 5, column 2 and 4. I assume that the workers are employed in a ¯rm with 100 employees, and that the business cycle is as it were in 1995. Workers with secondary education are assumed to have 12 years of schooling and work for 45 years. Workers with higher education are assumed to have 15 years of schooling and work for 42 years. 29 Cf. Autor (2000) for a model with the same feature. 30 It is in this respect interesting to note that the estimated coe±cients on R&D-intensity become smaller (more negative) if the share of scientists with post graduate degrees at the plant is included in the regression, despite this variable being strongly correlated with R&D intensity. One possible explanation is that the share of post graduate scientists also is correlated with unobserved worker ability. This would be consistent with the `O-ring theory' of Kremer (1993). 14 the estimated gain from working in R&D intensive ¯rms. In the same way, workers who transfer from ¯rms that do not invest much in R&D to ¯rms that do, have less human capital than those who have been in R&D intensive ¯rms for their entire career. Hence, they will reduce the average wage level in R&D intensive ¯rms and add to the bias. The result is that the wage premia associated with the last periods of a `high tech career' is underestimated, i.e. we will underestimate the steepness of the experience earnings pro¯le. A simple way to avoid the potential ability and `switching' bias, is to estimate the wage equation in ¯rst di®erences, i.e. investigate how ¯rms' R&D intensity a®ect wage growth directly. This is done in table 6. The drawback of this speci¯cation is that we do not learn about the e®ect of R&D on the wage level. Given that ability is expected to bias results against ¯nding support for the hypothesis that workers pay for R&D experience, however, this is not a serious problem. The broad picture emerging from the upper part of table 6 is that workers with technical or scienti¯c education in R&D-intensive ¯rms who do not change employer, have higher wage growth throughout their career31 . This is consistent with the previous ¯nding that R&D translates into human capital that workers earn a return on32 . Wage growth also appears to level o® towards the end of the career, consistent with workers having less incentive to accumulate human capital when getting closer to the retirement age. Since the correlation between ¯rms' R&D intensity and workers' learning invest- ments is expected to be strongest for young workers, it should be possible to observe changes in `payment' associated with transitions between ¯rms with di®erent R&D intensities. Moving from an R&D-intensive ¯rm to a less R&D-intensive ¯rm early in the career should induce a wage increase, and transitions the opposite way should induce a wage decrease. Both types of moves will contribute to a negative relation- ship between wage growth and change in R&D intensity. For old workers, a change in R&D intensity should not a®ect wages as much, since they are not expected to invest much in human capital. The estimated coe±cients do not fully con¯rm these hypotheses. For old workers, the coe±cients are small and not very signi¯cant as expected, and for young workers with secondary technical education the coe±cient 31 Note, however, that wage growth for workers with secondary technical education is negatively correlated with the employers' R&D intensity in the ¯rst two years of the career. This is also evident in table 5, column 1. It may re°ect that it takes some time to `absorb' the complexity of R&D intensive ¯rms, or that workers due to imperfect information about the quality of the training, are unwilling to pay the full cost of the training at once, but that ¯rms are able to extract more through lower wage growth during the very ¯rst years of the workers career. 32 Cash °ow before wage payments per worker, is included to control for the rent sharing e®ect of successful innovations found by van Reenen (1996). Such a rent sharing e®ect is present in the data, but it does not dominate the e®ect of R&D. 15 is negative and highly signi¯cant, but for young scientists and engineers the coef- ¯cient is positive and signi¯cant. A problem with the estimates, however, is that mobility cannot be considered exogenous with respect to wages. Estimating the price of learning and the return to R&D experience sep- arately Table 5 utilize cross sectional information only, and estimates in one co- e±cient the return to previous R&D experience minus the price paid for current learning opportunities. Utilizing the longitudinal dimension of the data set it is possible to specify these two components separately. The learning opportunity that a worker faces depends only on current R&D intensity, while average R&D inten- sity in previous years reveal information about the workers' R&D experience. Note, however, that the stability in R&D intensity over the workers careers, evident in table 1 and 2, makes current and previous R&D intensities somewhat collinear. A high level of precision can therefore not be expected when including both variables. Table 7, column 2 and 4, reports the results of interacting current R&D intensity and the average of previously observed R&D intensities separately with experience dummies. The ¯rst thing to notice is that the coe±cients on the average of previously observed R&D intensities, i.e. the return to R&D experience, are mostly positive, while the coe±cients on current R&D intensity, i.e. the implicit price paid for learning opportunities, are mostly negative. Note also that current R&D intensity has a more negative impact when previous R&D experience is included, cf. column 1 and 3. The price paid for learning decreases over time as predicted by theory, but the data do not bring out the expected wage increase over time that should be associated with R&D experience. Furthermore, the coe±cients on current R&D, i.e. learning, does not go to zero, but becomes positive late in the career. These two features seem connected. The employer's current R&D intensity appears to be a better proxy for old workers' human capital than the average of previously observed R&D intensities. This could be due to some selection process where workers whose technological experience has become obsolete, move out of or are displaced from R&D intensive ¯rms. In order to asses the importance of learning for the industry on an aggregate level, I have summarized the estimated wage discount for all R&D ¯rms. This sum amounts to 0.7 percent of the total wage bill for technical personnel in all R&D performing ¯rms and 2.6 percent of industry R&D investments. Looking only at ¯rms with R&D intensity above 0.2, the wage discount represent 3.0 percent of their total wage bill and 2.5 percent of their R&D investments. These numbers are not very big, but nor are they negligible. 16 The value of R&D experience from current employer vs. previous em- ployers Lengermann (1996) and Loewenstein and Spletzer (1998, 1999) who study the e®ect of formal on-the-job training, ¯nd that the return to training received from previous employers exceed the return to training received from the current employer. Loewenstein and Spletzer argue that this may re°ect that employers extract some returns to general training, and that workers do not realize the full returns until they change jobs. If something similar applies to the value of experience from R&D intensive ¯rms, it would imply that the potential R&D spillovers involved when workers in R&D intensive ¯rms change employers, is only partially internalized in the labor market. In order to investigate this possibility, I have for each employee where su±cient career information is present, calculated the average observed R&D intensity in previous years when working for the current employer and the average observed R&D intensity in years working for previous employers. With a smaller sample size and three R&D measures, an extension of the spec- i¯cation with experience dummies interacted with R&D-intensities, used in table 5 and 7, is not feasible. It is necessary to put more restrictions on the speci¯cation and I have chosen to approximate the price paid for learning opportunities with cur- rent R&D intensity and its interaction with years of overall work experience. R&D experience built up with the current employer is proxied with the average observed R&D intensity in previous years working for this employer times years of tenure with this employer. R&D experience built up with previous employers is proxied with the average observed R&D intensity while working for previous employers times years of experience prior to the current employment relationship. These measures, resembling sums of R&D intensities, are consistent with equation (2). Table 8, column 1 and 3 reports the results. Column 2 and 4 report a slightly less restrictive speci¯cation where non-linear interactions with experience and tenure is allowed. All regressions con¯rm the previous ¯nding that current R&D intensity have a signi¯cantly negative impact on wages early in the career. The positive cross-term with experience also con¯rm that this negative impact, interpreted to be the price paid for learning opportunities, diminishes over time. With respect to the R&D experience built up over the career, both R&D experience from the current employer and R&D experience from previous employers have a positive and signi¯cant impact on wages. R&D experience from the current employer, however, seems to be more highly valued. Unfortunately, this result is more suggestive than conclusive. In order to construct the variables needed, all years working with the current employer must be included in the sample, while information about previous employers can be less complete. Hence, the average R&D intensity in years working for previous employers is measured with greater error than average R&D intensity in years working for the current employer, and coe±cients on variables involving 17 the former measure will therefore be more biased towards zero33 . In addition, the coe±cient on R&D experience with the current employer could be upward biased. This would happen if recent R&D experience show that knowledge accumulated earlier in the career has not become obsolete. The results for old workers in table 7 indicate that this may be the case. Robustness and econometric issues A number of alternative speci¯cations have been tried to asses the robustness of the results34 . In one speci¯cation, more than 30 additional control variables were included, such as proxies for hours worked35 , the capital to labor ratio, the Her¯ndal index, the market share of the ¯rm, the union density36 and, four digit industry dummies. This did not change the quantitative results. The results are also robust to including ¯rm size, years of education and union density in interaction with experience. Dividing the sample into di®erent time periods, however, reveals that the e®ect of R&D on the wage-experience pro¯le is more pronounced in the 1980s than in the 1990. This may be related to the severe re- cession in the Norwegian economy starting in the late 1980s, causing a restructuring 33 If the sample is restricted to workers whose complete career is known, the return to R&D experience from previous employers appears to be above the return to R&D experience from the current employer for workers with higher education, while both coe±cients become insigni¯cant for workers with secondary education. For these workers the coe±cient on previous R&D experience even has a negative sign. 34 In addition to trying out di®erent speci¯cations within the sample of workers with technical education, I have also run the basic regressions on workers with non-technical education. The e®ect of R&D experience on workers with non-technical secondary and higher education resembles the e®ect on workers with technical education in that they seem to have a steeper experience-earnings pro¯le if working in R&D-intensive ¯rms. The results are fairly strong for workers with secondary non-technical education, but less evident for workers with higher non-technical education. It is not clear why these workers should be a®ected by the R&D-intensity of their employers, but several explanations are possible. First, R&D intensive ¯rms may be advanced along many dimensions, and hence o®er valuable work experience also to the non-technical sta®. Second, R&D intensive ¯rms also appear to be intensive in formal training. In years where the dataset includes measures of both R&D investments and formal training, these measures are signi¯cantly, positively corre- lated. Third, it is possible that not only the technical sta®, but also administrative managers in R&D intensive ¯rms have access to sensitive technological information. Then the Pakes and Nitzan (1983) model applies to this group as well as to the technical employees, and it is in any case con- ceivable that R&D intensive ¯rms to a larger extent than other ¯rms use stock options and similar compensation schemes for their managers, e.g. due to cash constraints. Finally, the Norwegian economy is strongly unionized. Unions often demand similar earnings plans for all employees in a ¯rm. 35 The following measures are available: Average hours per week worked at the plant, number of part time jobs and number of months unemployed. 36 The union density is only available after 1990. In 1990 and before, I have used the 1991 value, since union density as a ¯rm characteristic is fairly stable over time. 18 of, and a decline in, some of the most innovative subindustries. All regressions reported in tables 5, 7 and 8 allow for correlated error terms across observations of the same individual in di®erent years. However, one could also argue that error terms for workers belonging to the same ¯rm may be correlated. Allowing for such correlations when computing the standard error of the estimated coe±cients, reduce their signi¯cance, but the qualitative results are even robust to including ¯rm speci¯c ¯xed e®ects in the regressions. A comparison with the training literature Before concluding it may be worth- while to compare the overall results to similar analyses of `on-the-job training' in the labor literature. Although this paper, as far as I know, is the ¯rst to look at the e®ect of R&D on wages, there exists a large literature on the e®ect of formal training. In this literature, a number of authors have found training to be correlated with wage growth, but ¯nding support for a negative e®ect on starting wages such as human capital theory predicts, is unusual, cf. e.g. Barron, Black and Loewenstein (1989), Lynch (1992) or Barron, Berger and Black (1999)37 . Common interpreta- tions are that workers do not pay for general training, or that the implicit price is masked by a positive ability bias. In this perspective, the strong negative e®ect of R&D on starting wages present in this sample, is remarkable. It suggests that ¯rms' technology levels are more important to wages than formal on-the-job training. One explanation for this could be that while most formal training is short term, working in a technologically challenging environment a®ects human capital accumulation for the entire duration of a job. 6 Concluding remarks Labor mobility is often considered to be an important source of knowledge exter- nalities, making it di±cult for ¯rms to appropriate returns to R&D investments. Pakes and Nitzan (1983), however, analyze the problem formally, and ¯nd that la- bor turnover should not be a problem for R&D ¯rms. Both scientists and ¯rms are aware of the fact that working on a research project gives access to valuable information. Once such information is disclosed or developed, scientists, if they are to stay with the ¯rm, will have to receive a wage increase re°ecting their new market value. Thus, scientists expect that accepting a research position implies a 37 One exception is Autor (2000). Studying temporary help ¯rms, he ¯nds that \[w]ages are lower at ¯rms o®ering training by a modest, but statistically signi¯cant magnitude". Lynch (1992) ¯nd a negative e®ect of uncompleted training for workers with less than high school education, but not for workers with a high school degree or some college education. 19 future wage increase, and consequently they can accept an initial wage below their alternative wage, without experiencing a welfare loss. Research ¯rms are likely to use the most up-to date technology and frequently change its products and production processes. Because of this, one would think that even workers who don't have direct access to the results of the R&D projects, learn more in these ¯rms. Rosen (1972) provides a model where di®erent ¯rms o®er di®erent opportunities for on-the-job learning, and derive implications with respect to wages that closely resemble those of Pakes and Nitzan (1983). Rosen thinks of jobs as tied packages of work and learning. Workers sell the services of their skills and simultaneously purchase an opportunity to augment those skills. I have argued in this paper that inter-¯rm transfers of R&D-results embodied in people, should be analyzed within a human capital framework similar to the models of Pakes and Nitzan (1983) and Rosen (1972). Testing such a framework using matched employer-employee data from the Norwegian machinery and equipment industry, I ¯nd that the technical sta® in R&D-intensive ¯rms indeed pays for the knowledge they accumulate on the job through lower wages in the beginning of their career. Later in the career they earn a return on these implicit investments through higher wages. This suggests that potential externalities associated with labor mobility, at least to some extent, are internalized in the labor market38 . Determining whether workers pay for the full value of the knowledge they ac- cumulate in R&D intensive ¯rms, however, is di±cult, and beyond the scope of this paper. Information asymmetries and other barriers to mobility may certainly produce spillover mechanisms39 , and I hope to return to this question in the future. 38 The only related ¯nding I know of in the literature is Zucker, Darby and Armstrong (1998) who in an academic setting, claim that \competitive university salaries are lower, other things equal, in areas where faculty expect the possibility of receiving substantial outside income or wealth as a result of skills developed doing research at the university." 39 Cf. Acemoglu and Pischke (1999) although these authors do not write with reference to R&D investments. A particularly important kind of imperfection may be distortions in the wage structure which makes a wedge between wages and marginal productivity increase with the workers' human capital. Firms then have an incentive to invest in R&D producing general human capital because they get a share of the return. A simple mechanism which could cause such wage compression, is that ¯rms receive a fraction of the productivity of the workers as pro¯t due to matching, search costs or other sorts of labor market friction. If the employer receives a fraction of the workers' productivity, the employers level of pro¯t will increase with the workers' productivity and therefore with their human capital. Another possible mechanism is complementarity between ¯rm speci¯c and general human capital. In this case, the alternative wage for the scientist will increase less than his or her productivity as he or she receives training. Mechanisms which induce employers to pay for general human capital accumulation, create a positive externality to the worker's future employer if the worker decides to quit or if the ¯rm goes out of business. If labor mobility does create some knowledge spillovers, multiple equilibria may be possible. One mechanism that may produce such a result is complementarity between ¯rms' own research investments and the faster 20 Furthermore, even if workers pay for all the knowledge they accumulate, this `so- lution' to the spillover problem does not guarantee optimal R&D investments. If workers co-¯nance R&D through lower wages, and if the value of the knowledge they accumulate depend on the outcome of the R&D project, they become exposed to the risk associated with the project. Risk aversion among workers may then become a new source of distortion40 . Liquidity constraints making workers unwilling to trade o® current wage for future wage on a large scale, may also create problems. References Acemoglu, D. and Pischke, J.-S. (1999). Beyond Becker: Training in imperfect labor markets, Economic Journal 109: F112{F142. Almeida, P. and Kogut, B. (1996). Thechnology and geography: The localization of knowledge and the mobility of patent holders. Working Paper, The Huntsman Center for Global Competition and Innovation, Wharton School, University of Pennsylvania. Arrow, K. J. (1962). Economic welfare and the allocation of resources for invention, in R. R. Nelson (ed.), The Rate and Direction of Inventive Activity: Economic and Social Factors, Vol. 13 of NBER Special Conference Series, Princeton University Press, New Jersey, pp. 609{625. Autor, D. H. (2000). Why do temporary help ¯rms provide free general skills train- ing. NBER Working Paper No. 7637. Balkin, D. B. and Gomez-Mejia, L. R. (1985). Compensation practices in high- technology industries, Personnel Administrator (June): 111{123. Barron, J. M., Berger, M. C. and Black, D. A. (1999). Do workers pay for on-the-job training?, Journal of Human Resources 34(2): 235{252. Barron, J. M., Black, D. A. and Loewenstein, M. A. (1989). Job matching and on-the-job training, Journal of Labor Economics 7(1): 1{19. knowledge di®usion and improved matching associated with a high-mobility equilibrium. Cf. e.g. Cohen and Levinthal (1989) for an analysis which stress the importance of complementarities between spillovers and ¯rm's own R&D investments. A tolerant attitude towards movement of personnel, reported to be present e.g. in Silicon Valley (Saxenian; 1994), could then promote R&D investments and growth, whereas a situation where ¯rms make labor mobility costly could produce an equilibrium with low R&D investments and growth. The latter result would be due to low quality matches combined with a low in°ow of new ideas, and despite the fact that lower probability of quits, all else equal, improves the incentives to invest in R&D. 40 Cf. footnote 6. 21 Barth, E. and Dale-Olsen, H. (1999). The employer's wage policy and worker turnover, in J. C. Haltiwanger, J. I. Lane, J. R. Spletzer, J. J. M. Theeuwes and K. R. Troske (eds), The Creation and Analysis of Employer-Employee Data, Contributions to Economic Analysis 241, North Holland, Amsterdam, pp. 285{ 312. Becker, G. S. (1962). Investment in human beings, Journal of Political Economy pp. 9{49. Becker, G. S. (1964). Human Capital, Columbia University Press, New York. [Third Edition by the University of Chicago Pess 1993]. Bhide, A. (1994). How entrepreneurs craft strategies that work, Harvard Business Review 72: 150{161. Burgess, S., Lane, J. and Stevens, D. (1996). Job °ows and workers °ows: Issues and evidence from a panel of US ¯rms, in R. Schettkat (ed.), The Flow Analysis of Labour Markets, Vol. 3 of Studies in the Modern World Economy, Routledge, London and New York, pp. 96{114. Cohen, W. M. and Levinthal, D. A. (1989). Innovation and learning: The two faces of R&D, Economic Journal 99: 569{596. Geroski, P. A. (1995). Do spillovers undermine the insentive to innovate?, in S. Dowrick (ed.), Economic Approaches to Innovation, Edwar Elgar, Aldershot, UK, pp. 76{97. Gersbach, H. and Schmutzler, A. (1997). Endogenous spillovers, the market for hu- man capital, and incentives for innovation. Discussion Papers No. 248, UniversitÄt a Heidelberg. Greene, W. H. (1997). Econometric Analysis, Third Edition, Prentice Hall, New Jersey. Greenhalgh, C. and Mavrotas, G. (1996). Job training, new technology and labour turnover, British Journal of Industrial Relations 34(1): 131{150. Griliches, Z. (1973). Research expenditures and growth accounting, in B. R. Williams (ed.), Science and Technology in Economic Growth, MacMillian, Lon- don, pp. 59{95. Jackson, T. (1997). Inside Intel: Andrew Grove and the Rise of the World's Most Powerful Chip Company, Dutton Books. Ja®e, A. (1986). Technological opportunity and spillovers from R&D, American Economic Review 76: 984{1001. Ja®e, A. (1996). Economic analysis of research spillovers: Implications for the 22 advanced technology program. National Institute of Standards and Technology, Report GCR 97-708. Ja®e, A. B., Trajtenberg, M. and Henderson, R. (1993). Geographic localization of knowledge spillovers as evidenced by patent citations, Quarterly Journal of Economics 108: 577{598. Klette, T. J. and M¿en, J. (1999). From growth theory to technology policy - co- ordination problems in theory and practice, Nordic Journal of Political Economy 25(1): 53{74. Klette, T. J., M¿en, J. and Griliches, Z. (2000). Do subsidies to commercial R&D re- duce market failure? Microeconometric evaluation studies, Research Policy 29(4- 5): 471{495. Kremer, M. (1993). The O-ring theory of economic development, Quarterly Journal of Economics 108: 551{575. Lengermann, P. A. (1996). How long do the bene¯ts of training last? Evidence of long term e®ects across current and previous employers, education levels, test scores, and occupations. Working Paper 96-18, CAHRS, Cornell University. Levin, R. C., Klevoric, A. K., Nelson, R. R. and Winter, S. G. (1987). Appropri- ating the returns from industrial research and development, Brookings Papers on Economic Activity, Microeconomics pp. 783{831. Loewenstein, M. A. and Spletzer, J. R. (1998). Dividing the costs and returns to general training, Journal of Labor Economics 16(1): 142{171. Loewenstein, M. A. and Spletzer, J. R. (1999). General and speci¯c training, evi- dence and implications, Journal of Human Resources 34(4): 710{733. Lynch, L. M. (1992). Private sector training and the earnings of young workers, American Economic Review 82(1): 299{312. Mincer, J. (1958). Investment in human capital and personal income distribution, Journal of Political Economy 66: 281{302. Mincer, J. (1974). Schooling, Experience and Earnings, Columbia University Press, New York. Pacelli, L., Rapiti, F. and Revelli, R. (1998). Employment and mobility of workers in industries with di®erent intensity of innovastion: Evidence on Italy from a panel of workers and ¯rms, Economics of Innovation and New Technologies 5: 273{300. Pakes, A. and Nitzan, S. (1983). Optimum contracts for research personnel, re- search employment, and the establishment of `rival' enterprises, Journal of Labor Economics 1: 345{365. 23 Parent, D. (1999). Wages and mobility: The impact of employer-provided training, Journal of Labor Economics 17(2): 298{317. Rosen, S. (1972). Learning and experience in the labor market, Journal of Human Resources 7: 326{342. Rosen, S. (1986). The theory of equalizing di®erences, in O. Ashenfelter and R. La- yard (eds), Handbook of Labor Economics, Vol. I, North-Holland, Amsterdam, chapter 12, pp. 641{692. Saxenian, A. (1994). Regional Advantage: Culture Competition Between Route 128 and Silicon Valley, Harvard University Press, Cambridge, Mass. Stephan, P. (1996). The economics of science, Journal of Economic Literature 34(3): 1199{1235. Stern, S. (1999). Do scientists pay to be scientists? NBER Working Paper No. 7410. van Reenen, J. (1996). The creation and capture of rents: Wages and innovation in a panel of UK companies, Quarterly Journal of Economics 111: 195{226. Zucker, L. G., Darby, M. R. and Armstrong, J. (1998). Geographically localized knowledge: Spillovers or markets?, Economic Inquirey 36: 65{86. Zucker, L. G., Darby, M. R. and Brewer, M. B. (1998). Intellectual human cap- ital and the birth of US biotechnology enterprises, American Economic Review 88: 290{306. 24 Table 1. Worker mobility between plants with known R&D intensity with with with with Total no R&D R&D- R&D- R&D- number intensity intensity intensity of sepa- ∈ 〈0,.05] ∈ 〈.05,.2] > .2 rations left a non R&D-plant and joined a plant 65.5 27.3 5.7 1.5 3 168 left a plant with R&D-intensity ∈ 〈0,.05] and joined a plant 27.8 61.1 8.9 2.2 3 330 left a plant with R&D-intensity ∈ 〈.05,.2] and joined a plant 11.9 42.6 40.4 5.1 2 841 left a plant with R&D-intensity > .2 and joined a plant 12.9 9.3 13.9 64.0 497 percentage of jobs in plants 34.6 42.3 17.2 5.9 The numbers are percentage shares of total separations from each category of plants and sum to 100 horizontally. The sample consists of men with technical or scientific education employed full time in the machinery and equipment industry in Norway 1986-1995 at a plant where the R&D-intensity is known. Transitions out of the sample have been excluded. R&D intensity is measured as R&D man-years per employee at the three-digit line of business level within firms. i Table 2. Stability in R&D intensity over the career A: Previous R&D intensity predicted for workers who have not changed employer R&D- R&D- R&D- R&D- R&D- R&D- intensity at intensity at intensity at intensity at intensity at intensity at t-1 t-1 t-5 t-5 t-9 t-9 R&D-intensity at t .845*** 1.015*** .674*** .960*** .495*** .785*** (.003) (.007) (.008) (.011) (.011) (.013) Estimator OLS 2SLS OLS 2SLS OLS 2SLS Sample size 266 009 266 009 81 772 81 772 10 133 10 133 R-squared 0.79 0.57 0.60 0.58 0.73 0.66 B: Previous R&D intensity predicted for workers who have changed employer R&D- R&D- R&D- R&D- R&D- R&D- intensity at intensity at intensity at intensity at intensity at intensity at t-1 t-1 t-5 t-5 t-9 t-9 R&D-intensity at t .279*** .625*** .294*** .658*** .298*** .486*** (.014) (.013) (.012) (.012) (.020) (.017) Estimator OLS 2SLS OLS 2SLS OLS 2SLS Sample size 16 233 16 233 27 533 27 533 6 294 6 294 R-squared 0.36 0.44 0.35 0.45 0.39 0.45 Instruments in the 2SLS-regression are a quadratic in the share of postgraduate scientists at the plant, a quadratic in the number of employees at the plant and four digit industry dummies. Instruments are used to correct for measurement error in the R&D variable. Year dummies are included in all regressions, but not reported. R&D intensity is measured as R&D man- years per employee at the three-digit line of business level within firms. Standard errors, adjusted for heteroscedastisity and correlated error terms within individuals, are given in parentheses. The regressions in part A of the table may be considered a weighted firm level regression where the weights are the number of employees. The rationale for the weighting would be that the R&D intensity is a measure of average R&D man-years per employee, and therefore is more precisely measured in large firms. The sample consists of men with technical or scientific education employed full time in the machinery and equipment industry in Norway 1986-1995. *** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level ii Table 3. Labor turnover by education and R&D intensity Turnover Excess Average Number of rate turnover years of job-year rate experience observations Secondary technically educated in a plant with no R&D .194 .095 16.4 110 091 with R&D-intensity ∈〈0, .05] .210 .091 17.6 84 886 with R&D-intensity ∈〈.05, .2] .211 .072 16.3 37 280 with R&D-intensity >0.2 .208 .059 14.8 7 246 Higher technically or scientifically educated in a plant with no R&D .191 .074 19.8 14 806 with R&D-intensity ∈〈0, .05] .210 .075 19.5 20 444 with R&D-intensity ∈〈.05, .2] .211 .071 17.8 20 782 with R&D-intensity >0.2 .216 .065 14.1 11 838 The turnover rate is separations in year t as a share of employment in year t. The excess turnover rate is separations out of jobs that continue as a share of continuing jobs. R&D intensity is measured as R&D man-years per employee at the three- digit line of business level within firms. The sample consists of men with technical or scientific education employed full time in the machinery and equipment industry in Norway 1986-1995. Table 4. The effect of R&D intensity on excess turnover R&D-intensity * secondary technical education -.119*** -2.041*** (.030) (.416) * higher technical or scientific education -.042 -.721* (.030) (.376) Estimator Tobit Grouped logit Sample size 6 904 266 173 Pseudo R-squared 0.42 0.01 The dependent variable is the excess turnover rate within plant educational groups. The excess turnover rate is separations out of jobs that continue as a share of continuing jobs. Control variables included in the regressions, but not reported are a dummy for higher technical or scientific education, plant job destruction rate, plant job creation rate, a quadratic in the educational group number of workers, a quadratic in their average experience, a quadratic in plant age and year dummies. The sample sizes in the tobit regressions refer to the number of within plant educational groups. Educational groups with less than five workers have been excluded from the tobit regressions due to the turnover estimates being uncertain when based on few workers. The sample sizes in the grouped logit regressions refer to the number of workers. Standard errors are given in parentheses. In the grouped logit regressions, the standard errors are adjusted for heteroscedastisity and correlated error terms within plants. R&D intensity is measured as R&D man-years per employee at the three-digit line of business level within firms. The sample consists of men with technical or scientific education employed full time in the machinery and equipment industry in Norway 1986-1995. *** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level iii Table 5. The effect of R&D on the experience-earnings profile (1) (2) (3) (4) Secondary technical Higher technical or scientific education education R&D * less than one year experience -0.207*** -.055** -.132*** -0.061*** (.049) (.024) (.044) (.014) * 1-2 year experience -0.297*** -.048** -.097*** -0.044*** (.052) (.019) (.025) (.009) * 3-5 year experience -0.163*** -.029*** -.049** -0.025*** (.032) (.011) (.021) (.007) * 6-10 year experience -0.169*** -.025*** -.012 -0.018*** (.026) (.009) (.022) (.007) * 11-15 year experience -0.083*** .001 .008 -0.009 (.032) (.010) (.026) (.009) * 16-20 year experience -0.065* .023** .025 0.001 (.035) (.011) (.032) (.009) * 21-35 year experience 0.088*** .045*** .101*** 0.031*** (.029) (.009) (.029) (.009) * more than 35 year experience 0.222*** .086*** .229*** 0.068*** (.048) (.016) (.055) (.019) R&D measure intensity dummy intensity dummy Sample size 244 657 207 776 71 372 50 216 R-squared 0.20 0.20 0.28 0.26 The dependent variable is ln (real annual earnings). Control variables included in the regression, but not reported are seven experience dummies, years of schooling, a quadratic in plant number of employees and year dummies. The coefficients are estimated using ordinary least squares. Standard errors, adjusted for heteroscedasticity and correlated error terms within individuals, are given in parentheses. R&D intensity is measured as R&D man-years per employee at the three-digit line of business level within firms. The R&D dummy is one if the R&D intensity is above 0.2. Observations with R&D-intensity between 0.05 and 0.2 are excluded from the regressions in column (2) and (4). The sample consists of men with technical or scientific education employed full time in the machinery and equipment industry in Norway 1986-1995. *** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level iv Figure 1. Estimated earnings-experience profiles Workers with higher technical or scientific education 400 earnings in 1000 1995 NOK 350 in firms with no R&D in firms with R&D intensity 0.2 300 Workers with secondary techncial education 250 in firms with no R&D 200 in firms with R&D intensity 0.2 150 0 5 10 15 20 25 30 35 40 45 experience in years The graphs are based on regressions similar to those in table 5, column (1) and (3) except that experience dummies interacted with R&D intensity is exchanged with a quartic in experience interacted with a quadratic in R&D intensity. R&D intensity is measured as R&D man-years per employee at the three-digit line of business level within firms. The sample consists of men with technical or scientific education employed full time in the machinery and equipment industry in Norway 1986-1995. The graphs for workers with higher education are based on 15 years of education and a firm with 100 employees. The graphs for workers with secondary education are based on 12 years of education and a firm with 100 employees. Business cycle conditions are assumed to be like 1995. v Table 6. The effect of R&D on biannual earnings growth Secondary technical Higher technical or education scientific education Stays with same employer from year t-2 to year t * 2 year experience * R&D-intensity -.203** .021 (.082) (.040) * 3-5 year experience * R&D-intensity .077*** .065*** (.027) (.017) * 6-10 year experience * R&D-intensity .079*** .048*** (.016) (.012) * 11-15 year experience * R&D-intensity .101*** .039*** (.016) (.014) * 16-20 year experience * R&D-intensity .093*** .049*** (.018) (.015) * 21-35 year experience * R&D-intensity .063*** .007 (.011) (.011) * above 35 year experience * R&D-intensity .047** .008 (.022) (.021) Separates in year t-1 * 2-10 year experience * ∆R&D-intensity -.232*** .097*** (.057) (.037) * 11-20 year experience * ∆R&D-intensity .044 .037 (.045) (.033) * above 21 year experience * ∆R&D-intensity -.076* -.021 (.045) (.040) Sample size 139 108 42 466 R-squared 0.09 0.13 The dependent variable is the first difference of ln (real annual earnings) between year t and year t-2. Control variables included in the regression, but not reported are cash flow before wage payments per employee, seven experience dummies, a dummy for being a separator in year t-1interacted with dummies for the three levels of experience used for separators, a quadratic in the change in plant size measured by number of employees and year dummies. The coefficients are estimated using ordinary least squares. Standard errors are given in parentheses. R&D intensity is measured as R&D man-years per employee at the three-digit line of business level within firms. R&D intensity for stayers is the average over year t, t-1, and t-2. The sample consists of men with technical or scientific education employed full time in the machinery and equipment industry in Norway 1986-1995. *** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level vi Table 7. The effect of current R&D and previous R&D experience on earnings (1) (2) (3) (4) Secondary technical Higher technical or scientific education education Current R&D-intensity * less than one year experience -.204*** -.204*** -.134*** -.130*** (.050) (.050) (.044) (.044) * 1-2 year experience -.314*** -.342*** -.131*** -.266*** (.062) (.073) (.030) (.038) * 3-5 year experience -.190*** -.233*** -.058** -.123*** (.031) (.039) (.023) (.027) * 6-10 year experience -.177*** -.208*** .010 -.097*** (.028) (.032) (.023) (.025) * 11-15 year experience -.080** -.084** .014 -.062** (.032) (-.037) (.027) (.069) * 16-20 year experience -.069* -.079* .031 -.022 (.036) (.041) (.034) (.038) * 21-35 year experience .089*** .038 .094*** .039 (.029) (.030) (.030) (.032) * more than 35 year experience .234*** .285*** .222*** .265*** (.048) (.053) (.056) (.062) Average R&D-intensity over previous career * 1-2 year experience .055 .261*** (.093) (.051) * 3-5 year experience .094 .142*** (.057) (.038) * 6-10 year experience .070 .186*** (.055) (.036) * 11-15 year experience .010 .167*** (.062) (.038) * 16-20 year experience .028 .134*** (.059) (.052) * 21-35 year experience .171*** .165*** (.060) (.053) * more than 35 year experience -.158** -0.149 (.072) (.103) Sample size 227 418 227 418 65 422 65 422 R-squared 0.19 0.19 0.28 0.28 The dependent variable is ln (real annual earnings). Control variables included in the regression, but not reported, are seven experience dummies, years of schooling, a quadratic in plant number of employees and year dummies. The coefficients are estimated using ordinary least squares. Standard errors, adjusted for heteroscedasticity and correlated error terms within individuals, are given in parentheses. R&D intensity is measured as R&D man-years per employee at the three-digit line of business level within firms. The sample consists of men with technical or scientific education employed full time in the machinery and equipment industry in Norway 1986-1995. Workers, for whom no R&D information from previous years is available, have been excluded. *** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level vii Table 8. The effect of current R&D, R&D experience from the current employer and R&D experience from previous employers (1) (2) (3) (4) Secondary technical Higher technical or scientific education education current R&D intensity -.316*** -.480*** -.104*** -.157*** (.042) (.063) (.034) (.052) * experience .017*** .038*** .003 .008 (.002) (.008) (.002) (.007) * experience 2 -.001*** -.0001 (.0002) (.0002) mean R&D intensity in previous years with current employer * tenure .028*** .124*** .032*** .089*** (.008) (.018) (.007) (.014) * tenure 2 -.014*** -.008*** (.002) (.002) mean R&D intensity in years with previous employer(s) * (experience – tenure) .008** .005 .005** .013** (.003) (.008) (.002) (.006) * (experience – tenure) 2 .0001 -.0004 (.0003) (.0003) Sample size 62 243 62 243 17 675 17 675 R-squared 0.23 0.24 0.33 0.33 The dependent variable is ln (real annual earnings). Control variables included in the regression, but not reported, are years of schooling, a quadratic in plant number of employees, a quartic in experience, a quadratic in tenure, a dummy for having changed employer at least once, year dummies and a dummy variable for job relationships whose starting date is censored at April 30th 1978 together with its interactions with all tenure variables. The coefficients are estimated using ordinary least squares. Standard errors, adjusted for heteroscedastisity and correlated error terms within individuals, are given in parentheses. R&D intensity is measured as R&D man-years per employee at the three-digit line of business level within firms. Mean R&D intensity is calculated over the years where information about the R&D intensity is available. The sample consists of men with technical or scientific education employed full time in the machinery and equipment industry in Norway 1986-1995. Workers in firms where R&D information is not available in the sample year and in at least one prior year, and workers who have had previous employment without R&D intensity being known in at least one year, have been excluded. *** Significant at the 1% level ** Significant at the 5% level * Significant at the 10% level viii Table A1. Sample size and trimming procedures Total number of observations in the machinery and equipment industries 1986-1995 810 559 – Women 125 111 – Part time workers 11 314 – Workers with unknown education 8 968 – Workers with primary education 141 216 – Workers with secondary or higher non-technical/non-scientific education 94 325 Total number of observations of full time working male technical staff 429 625 – Workers in firms that cannot be matched to the time series files of the manufacturing statistics 39 527 – Workers in firms where R&D information is not available 46 744 Total number of observations of full time working male technical staff in the matched sample 343 354 – Workers not working for the whole year because they are entering the labor force 9 982 – Workers not working for the whole year because they are leaving the labor force 14 044 – Workers with secondary technical education and earnings below NOK 75.000 (1995 value) 2 723 – Workers with higher technical or scientific education and earnings below NOK 150.000 (1995 value) 566 Main sample (trimmed) 316 029 Each entry refers to the number of observations deleted among the observations left after the deletions in the rows above have been conducted. Workers with secondary technical education and earnings below NOK 75.000 (1995 value), and workers with higher technical or scientific education and earnings below NOK 150.000 (1995 value) have been excluded because such low earnings suggest that they have not worked full time for an entire year. Tables 1-4 are based on all observations of full time working male technical staff in the matched sample, whereas the wage regressions in tables 5-8 are based on the trimmed ‘main sample’. ix Table A2. Observations in main sample by year and education Number of Secondary technical Higher technical or workers education scientific education 1986 29 256 75.0 % 25.0 % 1987 30 329 75.8 % 24.2 % 1988 29 450 76.0 % 24.0 % 1989 29 952 76.2 % 23.8 % 1990 31 576 77.1 % 22.9 % 1991 31 482 79.6 % 20.4 % 1992 33 857 79.2 % 20.8 % 1993 33 261 78.8 % 21.2 % 1994 35 315 78.3 % 21.7 % 1995 31 551 77.4 % 22.6 % Observations 316 029 77.4 % 22.6 % Table A3. Observations in the main sample by experience and R&D intensity Observations No R&D R&D-intensity R&D-intensity R&D-intensity ∈ 〈0,.05] ∈ 〈.05,.2] >.2 Less than one year experience 7017 36.6 % 42.9 % 16.5 % 4.0 % 1-2 year experience 18446 36.8 % 40.3 % 16.9 % 6.0 % 3-5 year experience 36167 37.7 % 37.7 % 17.6 % 7.0 % 6-10 year experience 57802 42.1 % 32.8 % 17.5 % 7.5 % 11-15 year experience 44545 41.0 % 32.3 % 19.3 % 7.4 % 16-20 year experience 37563 42.9 % 32.0 % 18.6 % 6.4 % 21-35 year experience 86846 40.6 % 33.9 % 19.3 % 6.2 % More than 35 year experience 27643 41.4 % 35.7 % 17.8 % 5.1 % 316 029 40.6 % 34.4 % 18.4 % 6.6 % R&D intensity is measured as R&D man-years per employee at the three-digit line of business level within firms. x Table A4. Worker characteristics by education Secondary technical Higher technical or education scientific education Years of education mean 11.2 14.5 st.dev. (0.9) (1.7) 10th percentile 10.0 13.0 90th percentile 12.0 17.0 Years of experience mean 16.8 17.4 st.dev. (11.9) (11.6) 10th percentile 3 3 90th percentile 34 35 Years of tenure‡ mean 6.3 6.0 st.dev. (5.6) (5.1) 10th percentile 0.9 0.9 90th percentile 13.2 12.5 Wage in 1995 NOK mean 245 400 353 500 st.dev. (71 000) (125 900) 10th percentile 176 500 240 200 90th percentile 336 100 479 700 Union membership share 44% 27% Working at R&D performing plant share 54% 78% R&D-intensity if at R&D performing plant mean 0.057 0.125 st.dev. (0.085) (0.134) 10th percentile 0.002 0.006 90th percentile 0.152 0.278 The numbers are based on all worker-year observations in the machinery and equipment industry included in the main sample, cf. table A1. An R&D plant is a plant belonging to a firm that conducts some R&D within the plant’s three-digit ISIC industry. R&D intensity is measured as R&D man-years per employee at the three-digit line of business level within firms. Wage in 1995 NOK is rounded to the nearest 100. ‡ 16 percent of the observations have the job starting date censored at April 30th 1978. xi Table A5. Plant characteristics by plant size Number of employees: Less than 50 50-200 more than 200 All plants Number of employees mean 18.3 96.2 483.3 83.3 st.dev. (12.9) (38.6) (347.7) (169.9) 10th percentile 4 54 221 6 90th percentile 39 158 906 185 Average experience of technical staff mean 16.9 17.1 17.2 17.0 st.dev. (6.6) (4.4) (4.4) (5.9) 10th percentile 9.3 12.0 12.2 10.2 90th percentile 25.3 23.0 22.5 24.1 Average tenure of technical staff‡ mean 5.7 6.0 6.1 5.8 st.dev. (3.3) (3.0) (3.3) (3.2) 10th percentile 1.9 2.3 2.1 2.0 90th percentile 10.0 9.8 10.7 10.0 Average education of technical staff mean 11.7 11.7 12.0 11.8 st.dev. (1.1) (0.8) 1.0 (1.1) 10th percentile 10.7 10.8 11.2 10.8 90th percentile 13.0 12.9 13.4 13.0 Share of work force with higher technical or scientific education mean 10% 10% 13% 10% st.dev. (16) (11) (12) (14) 10th percentile 0 1% 2% 0 90th percentile 32% 23% 31% 29% R&D performing firm share 34% 54% 63% 42% R&D man-years per employee if R&D performing firm mean 0.14 0.07 0.08 0.10 st.dev. (0.18) (0.09) (0.10) (0.15) 10th percentile 0.007 0.005 0.003 0.006 90th percentile 0.39 0.17 0.21 0.26 Capital per employee in 1995 NOK mean 805 800 831 700 977 100 828 900 st.dev. (1 137 600) (682 300) (746 400) (995 100) 10th percentile 245 700 268 300 233 700 262 000 90th percentile 1 359 600 1 575 200 1 838 500 1 472 800 Union density among technical staff mean 23% 33% 41% 27% st.dev. (36) (39) (43) (38) Market share mean 2% 6% 19% 5% st.dev. (8) (12) (25) (13) Part of multi-plant firm share 37% 47% 64% 42% Plants founded before 1972 share 44% 64% 66% 52% Number of plants 4 728 2 195 697 7 620 The numbers are based on all plant-year observations in the machinery and equipment industries included in the main sample, cf. table A1. An R&D plant is a plant belonging to a firm that conducts some R&D within the plants three-digit ISIC industry. R&D man-years per employee and R&D sales ratio are measured at the three-digit line of business level within firms. Market share is measured at the five-digit line of business level for the firm that the plant belongs to. Capital per employee is rounded to the nearest 100. ‡ 16 percent of the underlying employee observations have the job starting date censored at April 30th 1978. xii Table A6. Aggregate growth from 1986 to 1995 and R&D intensity by sub-industries Number of plants Number of observations R&D ISIC 1986 1995 ∆ 1986 1995 ∆ intensity 38210 Engines and turbines 9 11 2 939 856 -9 % 0.03 38220 Agricultural machinery 52 33 -19 514 597 16 % 0.04 38230 Metal and wood-working machinery 37 26 -11 200 183 -9 % 0.04 38241 Oil and gas well machinery and tools 92 104 12 4709 8270 76 % 0.02 32249 Other industrial machinery 73 104 31 607 878 45 % 0.06 38250 Computers and office machinery 50 28 -22 1052 334 -68 % 0.26 38291 Household machinery 11 8 -3 95 92 -3 % 0.04 38292 Repair of machinery 709 458 -251 480 594 24 % 0.11 38299 Other machinery 339 351 12 4132 3197 -23 % 0.09 38310 Electric motors and eq. for el. production 139 153 14 2428 2010 -17 % 0.07 38320 Radio, TV and communication apparatus 190 135 -55 3335 2858 -14 % 0.17 38330 Electrical household appliances 32 20 12 251 187 -25 % 0.12 38391 Insulated cables and wires 12 17 5 689 627 -9 % 0.12 38399 Other electrical apparatus and equipment 124 100 -24 596 282 -53 % 0.04 38411 Building of ships 163 188 -25 3738 4350 16 % 0.01 38412 Building of boats 438 232 -206 535 400 -25 % 0.04 38413 Ship and boat engines and motors 36 29 -7 557 353 -37 % 0.04 38414 Components and fixtures for ships/boats 53 55 2 590 981 66 % 0.02 38421 Railway and tramway equipment 1 1 0 136 178 31 % - 38422 Repair of railway and tramway eq. 18 8 -10 1258 1015 -19 % - 38430 Motor vehicles 174 80 -94 740 1207 63 % 0.06 38440 Motor cycles and bicycles 1 2 1 114 77 -32 % - 38450 Aircraft 28 20 -8 1167 1173 1% 0.01 38490 Other transport equipment 6 12 6 11 52 373 % 0.02 38510 Professional and scientific instruments 57 109 52 306 749 145 % 0.11 38520 Photographic and optical goods 10 8 -2 77 51 -34 % 0.22 382-385 All machinery and equipment industries 2 854 2 292 -562 25 863 30 698 19% 0.10 The number of plants is taken from the manufacturing census. The number of observations refers to the technical staff in the main sample, cf. table A1. The growth in the technical staff does not imply that there has been employment growth in these industries, but is a result of old workers with primary education not included in the sample, gradually being replaced by workers with secondary education. R&D-intensity is the weighted average R&D man-years per employee, measured at the three-digit line of business level within firms, for the plants in the sample over the years 1986-1995. xiii