Contents lists available atScienceDirect
Structural Change and Economic Dynamics
j o u r n a l h o m e p a g e :w w w . e l s e v i e r . c o m / l o c a t e / s c e dHow does technological innovation and diffusion affect
inter-industry workers’ mobility?
Elisabetta Magnani
∗School of Economics, The University of New South Wales, Kensington, NSW 2052, Australia
a r t i c l e i n f o
Article history:Received January 2007
Received in revised form September 2008 Accepted December 2008
Available online 24 December 2008
JEL classification: J5 L2 L6 Keywords: Workers’ mobility
Technology innovation and diffusion Skill transferability
a b s t r a c t
Does technological change amount to accumulation of general, and so transferable, human capital? To approach this question I rely on a theoretical framework in which the “technol-ogy distance” between industries reduces inter-industry transferability of workers’ skill. Empirically, I use US panel data on individual intra-industry and inter-industry mobility decisions between 1982 and 1990, a period of rapid technological change in all manufac-turing sectors, to estimate a mixed logit econometric specification that does not rely on the IIA assumption. I find support to the main idea that technological innovation and diffusion have different effects on workers’ industrial mobility patterns. “Knowledge spillovers”, dif-ferently from “rent spillovers”, indeed boost the chances of workers’ inter-industry mobility. Surprisingly, this is more consistently so in low-tech industries than in high-tech industries. Consistently with the expectations developed in the theoretical framework, in low-tech industries skilled workers respond more sharply to technology diffusion than unskilled workers.
© 2008 Elsevier B.V. All rights reserved.
(. . .) Technological innovations, institutional reforms, and fresh ideas do not affect the aggregate level of eco-nomic activity abruptly: they need to diffuse from region to region, from activity to activity, cross-boundaries and seas, be evaluated, adapted, and refined. Their pro-moters have to dislodge the entrenched, persuade the skeptic, and reassure the fearful.Mokyr (2005), p. 285 1. Introduction
The Industrial Enlightenment was in part about the expansion of useful knowledge. According toMcGee (2004) technological progress often depended on “analogical” thinking, in which inventors, consciously or subcon-sciously, transform an idea they have already seen into something novel. Since most of these ideas travel embodied in humans, workers’ mobility is paramount for technolog-ical progress. So, who are the fearful?
∗ Tel.: +61 2 9385 3370; fax: +61 2 9313 6337. E-mail address:E.Magnani@unsw.edu.au.
In a “cognitive capitalism” it is the rivalrous nature of intellectual human capital that poses to employers’ the problem of skilled and technical workers’ mobility. Stated in other words, if R&D investment translates into human capital or knowledge that workers can transfer and utilize in other firms, workers’ mobility amounts to knowledge diffusion. While this is generally evaluated as an important condition for the spread of new firms and research units (e.g., inZucker et al., 1998), mobility of workers and par-ticularly of highly technical workers is potentially a threat to the firm whenever it’s the firm/industry that bears the cost of R&D activities(Kim and Marschke, 2005). In other words, intellectual human capital is rivalrous because it is characterized by natural excludability in the sense that its utilization in a firm usually excludes its simultaneous utilization in another firm.1This argument is central to
con-1The case of outsourcing, by which technical staff sells their specialized skills to more than one firm, raises the issue of the relationship between this alternative form of employment and technological diffusion. For more on this issue seeMagnani (2003).
0954-349X/$ – see front matter © 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.strueco.2008.12.003
textualize the interest in skilled workers’ mobility and in how the labor market for skilled technical workers works (Rosen, 1972; Pakes and Nitzan, 1983;Song et al., 1996; Fallick et al., 2005). More fundamentally an analysis of workers’ mobility, what prevents and what engenders it, is paramount to understand the contemporary ways in which capitalism attends to its central imperative and challenge: namely, “to immobilize workers, to tie it to the labor rela-tion and to prevent its flight, the breach of contract and the refusal to work”(Moulier-Boutang, 1998).
This paper investigates the effects of industry R&D intensity on workers’ mobility. A so far overlooked aspect in the literature on how technology change impacts upon the labor markets is the effects of technological change that occurs by means of diffusion of existing knowledge. This oversight has persisted in spite of the intensification of phe-nomena of technology diffusion brought by the deepening of input dependence among firms and industries starting from the 1970s (Wolff, 1997)and by a rise in R&D per-formed by non-manufacturing industries(OECD, 1996). In the face of the rapid technology changes driven by both innovation and diffusion, this paper explores the hypothe-sis that the dual aspects of technological change are among the determinants of workers’ mobility.
I structure the discussion as follows. Section2reviews the literature and introduces a simple model, which relies on the notion of “skill distance” developed in Silverberg et al. (1988)to discuss the impact of technological change on inter-industry mobility. Section 3 describes the data and discusses measurement of technology issues. Section4 introduces the econometric strategies, namely multinomial models and mixed logit models of inter-industry mobility. Section5discusses the empirical findings, while Section6 concludes.
2. Technological change, technological distance and skill transferability
This study’s focus on workers’ mobility is motivated by its importance for an economy’s adjustment to struc-tural change. In particular, strucstruc-tural shift in the 1980s and 1990s sparked a number of adjustments strategies at the firm’s level in the US, roughly characterizing “the US firms’ downsizing”, a term that encompasses phenomena like corporate layoffs, re-engineering, restructuring and job displacements (Baumol et al., 2003). Millions of workers have moved across industries in response to a structural change in the US industrial composition from manufactur-ing to trade and services. Usmanufactur-ing data from the Panel Study of Income Dynamics (PSID),Parrado et al. (2007)estimate that industrial mobility declined from 15 to around 10% between the first and the second half of the 1970s to vary between 15 and 20% in the 1981–93 period (Parrado et al., 2007, p. 442).DiPrete and Nonnemaker (1997a) illus-trate the relative size of US workers’ inter-industry mobility flows by 2-digit industry of employment to find substantial variations across industries (seeTable 1, reproduced from DiPrete and Nonnemaker, 1997a, p. 393).
The argument that technology change facilitates work-ers’ mobility is here developed in terms of the technology effect on workers’ skill transferability. The idea that R&D
investment translates into human capital or knowledge that workers can transfer and utilize in other firms has a number of relevant potential consequences. Firstly, firms may under-invest in R&D unless the labor market is able to internalize such externality. For example, Kim and Marschke (2005)develop and test a model of the patenting and R&D decisions of an innovating firm whose scientist-employees sometimes quit to join or start a rival. They show theoretically that the risk of a scientist’s mobility reduced the firm’s R&D expenditure. Another important consequence is that because intellectual human capital is scarce and valuable, accumulation of general human capital requires a change in the wage structure, rather than in the wage level, that is able to prevent workers’ mobility.2To the extent that R&D activities contribute to the accumulation of general human capital, they should steepen the earning profile and particularly so for technical and skilled workers. The importance of assessing the degree of specificity of human capital dates back toBecker (1962) who ini-tially focussed on the dichotomy between firm-specific and general capital.3More recently, the whole issue of the source of the specificity has been examined to argue that an important role in limiting workers’ mobility may be industry-specific capital rather than firm-specific capital (Neal, 1995; Parent, 2000). Furthermore, there is accumu-lating evidence that workers’ mobility has intensified in the new economy. Understanding the sources of human capi-tal specificity is particularly important in labor markets in which flexibility often relies on individual ability to respond to risk of unemployment or displacement(Magnani, 2001). In this context, human capital specificity can reduce its transferability, thus endangering workers’ ability to flexibly responding to work opportunities elsewhere. Interestingly, this is the approach that Hiscox takes in exploring the rela-tionship between factor mobility and technology change across US industries between 1820 and 1990(Hiscox, 2002). To examine differences in the way the two dimensions of technological change, namely innovation and diffusion, impact on human capital transferability it is useful to refer toSilverberg et al. (1988), who formally derive a short-ening of the distance between internal (firm-specific) and external (general) skills as a consequence of technology diffusion. This is the intuition that inspires the following theoretical framework.
2.1. Technology and workers’ mobility: sketching the relationship
To draw the implications of technological innova-tion and diffusion on workers’ mobility I use a simple model, fully described inMagnani (2003)and inspired by Mortensen and Pissarides (1994)and adapted to focus on
2Note that because the general skills accumulated in R&D intensive industries are useful to all employers, all wage offers from other employers will reflect this increased productivity and should theoretically be identi-cal to the current wage, making the worker indifferent between moving and staying in the current employment situation.
3This distinction was important because while workers had a full incen-tive to invest optimally in general human capital, there were potential incentive problems in financing firm-specific human capital.
Table 1
Net job change, gross job flows and mobility by sex, industry, and occupation: PSID, 1981–1989.
Industry/occupation Net change Gross job inflow Gross job outflow Type of mobility Within-employer Between-employer Between-industry Between-occupation Employment exit Men (N= 26, 833 Person-Year) Industry
Manufacturing, mining, utilities −0.001 0.126 0.128 0.090 0.081 0.060 0.115 0.055
Construction 0.024 0.191 0.167 0.038 0.166 0.101 0.116 0.073 Trade/restaurants 0.019 0.243 0.224 0.046 0.160 0.103 0.133 0.059 Transport/communication 0.019 0.119 0.100 0.089 0.102 0.058 0.105 0.055 Finance/business services 0.043 0.196 0.153 0.059 0.141 0.085 0.098 0.055 Social-personal services 0.015 0.143 0.128 0.084 0.083 0.055 0.088 0.046 Occupation
Services class I/II 0.029 0.170 0.141 0.074 0.097 0.060 0.089 0.038
Routine nonmanual 0.023 0.283 0.260 0.105 0.129 0.096 0.185 0.065 Self-employed IV a, b 0.020 0.269 0.250 0.014 0.127 0.083 0.093 0.059 Skilled/manual supervisory 0.007 0.182 0.174 0.072 0.114 0.074 0.102 0.060 Semiskilled/unskilled 0.010 0.255 0.244 0.073 0.134 0.096 0.149 0.085 Women (N= 23, 973 Person-Year) Industry
Manufacturing, mining, utilities 0.001 0.127 0.126 0.100 0.102 0.077 0.143 0.106
Construction 0.028 0.193 0.165 0.037 0.132 0.106 0.110 0.105 Trade/restaurants 0.017 0.242 0.225 0.053 0.159 0.115 0.157 0.176 Transport/communication 0.026 0.122 0.095 0.077 0.064 0.037 0.089 0.075 Finance/business services 0.041 0.194 0.153 0.085 0.160 0.100 0.162 0.098 Social-personal services 0.17 0.144 0.127 0.053 0.104 0.061 0.091 0.085 Occupation
Services class I/II 0.025 0.173 0.148 0.067 0.107 0.064 0.096 0.071
Routine nonmanual 0.011 0.277 0.266 0.073 0.141 0.096 0.150 0.115
Self-employed IV a, b 0.002 0.261 0.259 0.004 0.128 0.107 0.103 0.220
Skilled/manual supervisory 0.008 0.185 0.177 0.090 0.102 0.058 0.103 0.114
Semiskilled/unskilled 0.003 0.251 0.248 0.061 0.103 0.069 0.125 0.143
Inter-industry variations in mobility flows, 1981–1989, Source:DiPrete and Nonnemaker (1997a), p. 393.
inter-industry mobility. The representative firm (firm h) in the i-th industry at time t can produce good Y= F(L; K; i)
where L is skilled labor employed by means of internal labor markets and iis a measure of the industry-specific
current state of technical knowledge. Let Vhitindicate the
(steady state) evaluation by an internal skilled worker of of the value of employment in the h-th firm in industry i at time t
Vhit=
W (L; K; i)
r (1)
where r is the (exogenously given) market interest rate and W is his/her wage. The value of employment in a different industry must be evaluated net of moving costs and thus depends on the transferability of his skill. We assume that skill transferability is inversely proportional to the techno-logical distance between industry i and industry j, formally |j− i|. The value of employment in industry j can be
defined as Vhjt=
w(L; K; j)
r − (|j− i|) (2)
where (|j− i|) is the loss a worker faces due to
the only partial transferability of his/her human capi-tal when he/she moves from industry i to industry j. Thus (.) > 0.
This simple analytical setup allows us to explore the mobility implications of technological innovation and dif-fusion. More generally, in the short/medium run the
worker’s mobility decision (mob= 1) can be formalized as follows:
mob= 1 if Vhjt− Vhit> (|j− i|) (3)
This simple set up is useful in addressing the question that motivate this paper: does technology innovation at the level of the industry of current employment imply work-ers’ accumulation of general or specific human capital? Expression(3) clarifies that the impact of technological innovation on workers’ inter-industry mobility is complex as it depends on the way technological change in the indus-try of current employment, say changes in i, affects the
two terms of the inequality Vhjt− Vhit> (|j− i|). In
particular, if employment in industry i allows a worker to accumulate general skill (this can happen, for example, because industry-specific technology becomes more gen-eral), technological distance|j− i| between industries
and the loss associated with mobility (.) will drop. This implies that mobility should become more likely, unless of course wages in industry i undergo a process of change. In theory, to the extent that R&D activities contribute to the accumulation of general human capital, they should induce a steepening of the wage–tenure profile. However, as pointed out byLoewenstein and Spletzer (1999)and byAcemoglu and Pichke (1999)a number of market fac-tors may prevent this anti-clock-wise tilting of the earning
profile from happening.4Magnani (2006)offers limited evi-dence of an impact of industry R&D on the wage profile. The difficulty of assessing whether human capital is general or specific by looking at the slope of the workers’ earnings pro-file, makes it necessary to infer skill’s generality by looking at whether skill hinders mobility, and if so what sort of mobility. These considerations lead to the following: Proposition 1. If industry technology innovation is specific and involves accumulation of less than perfectly transferable skill by the worker, industry’s technological innovation would have a negative impact on workers’ inter-industry mobility.
Proof. Assume that technology innovation in industry i
amounts to an increase in the technology distance between industry i and industry j|j− i|. By expression(3), skill
specificity varies proportionally with the technological dis-tance (|j− i|). In this case technology innovation entails
a reduction in workers’ inter-industry mobility opportuni-ties.
This negative link between R&D investment and work-ers’ mobility is particularly likely in high-tech industries where indeed investment in R&D amounts to a outward shift of the technological frontier. Different may be the impact of technological change on workers’ mobility in low-tech industries. As strongly argued byTilton (1971); Allen (1977); Mowery (1983)among others, R&D activities not only amount to technological innovation, as they also enable firms to utilize information.5This argument leads to the following:
Proposition 2. If industry-specific technology innovation facilitates the assimilation and diffusion of new technologies developed elsewhere, the technology distance between indus-try i and indusindus-try j shortens and so the skill distance leading to an increase in workers’ inter-industry mobility.
I will perform a test of both these proposition by check-ing the impact of technological innovation on mobility of workers employed in high-tech and low-tech industries.
The argument developed above is useful to analyze the link between technological diffusion, accumulation of skill, and workers’ inter-industry mobility. To the extent that technology diffusion increases the chances of successful transfer of scarce skill accumulated in the current indus-try it will have no negative impact on the inter-indusindus-try mobility chances of workers, particularly those endowed with technical skills. This is expressed in the following:
Proposition 3. According to the technological diffusion
argument, technological diffusion shortens the technology distance between industries (|j− i|). This should make
4As pointed out byLoewenstein and Spletzer (1999)and byAcemoglu and Pichke (1999)the wage profile may not change in front of workers’ liq-uidity constraints, due to contract enforcement considerations and more generally in non-competitive labor markets. Indeed, even models with human capital externalities and agglomeration economies do not make clear predictions about wages (seeAcemoglu, 1997, p. 453).
5This is the theoretical argument developed inBartel et al. (2005), according to which an increase in the pace of technological change increases outsourcing because it allows firms to use services based on leading edge technologies, without incurring the sunk costs of adopting these new technologies.
workers’ skill more general. The implication is that technolog-ical diffusion should enhance workers’ inter-industry mobility chances.
These propositions will guide our empirical analysis. In the next section we describe the data.
3. Data and measurement issues
The data set used in this study comprises 6625 family unit (FU) heads6 drawn from the Panel Study of Income Dynamics (PSID). As a consequence of low attrition rates and the success in following young adults as they form their own families and recontact efforts (of those declining an interview in prior years), the PSID sample size has grown steadily from its start in 1968. This feature is particularly useful in the present context of addressing issues of labor inter-industry mobility.Appendix Adiscusses in details the nature of the PSID study and the representativeness of its sample. The number of FU heads actually comprising my sample is determined by the technology data as detailed below.
Information on workers’ industrial mobility is gathered by means of PSID information on the industry of employ-ment. The PSID used a one-digit occupation code, and later a two-digit, until 1981 when the 3-digit 1970 Census code became standard for the main jobs of employed Heads and Wives. It is only since 1981 the PSID has collected informa-tion on the industry of employment at the 3-digit level of aggregation.7Such detailed information is crucial to iden-tify the typology of mobility decisions occurring between time t and time t+ 1. With this information I draw a sample of 6625 observations of workers employed in US manufac-turing industries who are either employed or temporarily laid-off at time t, t= 1982, 1985, 1990. Multiple jobholders were excluded from the sample.8For these 6625 observa-tions that constitute the main sample I observe, besides the usual socio-demographic characteristics, the industry of employment at time t and at time t+ 1, with t + 1 = 1983, 1986, 1991, respectively.Zucker et al. (1998)provide evidence of the relevance of the period under scrutiny, namely the 1990s, for a study of workers’ inter-industry mobility.
6Within each wave of data, each FU (family unit) has one and only one current Head. Originally, if the family contained a husband-wife pair, the husband was arbitrarily designated the Head to conform with Census Bureau definitions in effect at the time the study began. The Head of the FU must be at least 16 years old and the person with the most financial responsibility for the FU. If this person is female and she has a husband in the FU, then he is designated as Head. If she has a boyfriend with whom she has been living for at least 1 year, then he is Head. However, if the husband or boyfriend is incapacitated and unable to fulfill the functions of Head, then the FU will have a female Head.
7The PSID released the 1968–1980 Retrospective Occupation-Industry Files in 1999. The aim of these files is to provide 3-digit occupation and industry codes from 1968 to 1980.
8The PSID distinguished between “main” and “extra” jobs. Clearly, someone cannot have an extra job unless he/she has a main job during the same time period. The extra job must be held simultaneously with the main job. Those who are only temporarily laid off are still employed at a main job and, therefore, could have an extra job during that time period.
To study how workers’ mobility decisions respond to technological innovation and diffusion I merge a panel of individual observations of intra-sectoral and inter-sectoral mobility with data on sector-specific measures of technology innovation and diffusion that relies on industry-specific R&D intensities. Notoriously, job mobility is produced by structural forces of expansion and con-traction as well as by individual choices that have been particularly dramatic in the US economy.9 The reader should however keep in mind the specificity (and conse-quent limitations) of the empirical analysis to the US.
From the main sample I draw a sub-sample of 3751 workers employed in high-tech industries (HT) and 2874 workers employed in low-tech industries (LT) in the US manufacturing sector.10 In each sub-sample I compare results obtained for skilled and unskilled individuals. Skilled individuals are those with at least one of the follow-ing two characteristics: (i) technical workers (engineerfollow-ing, computer technicians and technicians); (ii) workers whose education is greater or equal to 12 years of schooling). A number of sub-samples are used to check the robustness of our results. All results related to robustness exercises are available upon request.
3.1. Typology of workers’ mobility
Distinguishing mobility between broad industrial sec-tors (defined at the 2-digit level of aggregation) from mobility within industrial sectors promises to provide valuable insights on the effect of technological change on workers’ labor market opportunities. For mobility within broad industrial sectors (between 3-digit indus-tries) (Inter3D mob= 1) I rely on information about 3-digit industry of employment at time t and at time t+ 1. Simi-larly for mobility between broad industrial sector (2-digit industries) (Inter2D mob= 1) I rely on information about 2-digit industry of employment at time t and t+ 1.
A bit more complex is measuring mobility within 3-digit industry of employment because the PSID unit of observation does not have any identifier for the firm of current employer. Thus the identification of between-firms (within 3 digit industries) mobility needs to rely on the reported tenure with the current employer (Tenure). This leaves the issue of the distinction between a change in tenure involving a change of employer and a change in tenure without a change of employer (for example due to career advancement) open. To unambigously identify intra-industry mobility I proceed as follows. For each indi-vidual h I identify Tenureht and Tenureht+1 where t= 1982, 1985 and 1990 making t+ 1 = 1983, 1986, 1991.
9The choice of US workers as the unit of analysis to study the impact of technological change on patterns of inter-industry mobility is further motivated by comparative studies on “mobility regimes”. For example DiPrete et al. (1997b)analyze rates of job mobility in four countries, US, Germany, Netherlands and Sweden to argue that US mobility rates show the greatest sensitivity to structural change (shifts in technology, markets, and the consequent demand for particular forms of labor) and to the labor market resources of individual workers.
10The definition of high-tech and low-tech industries is based on the level of technology innovation as detailed below.
Mobility across firms (Intra3D mob= 1) but within 3-digit industries occurs if two conditions are satisfied: (i) Tenureht+1≤ Tenureht where tenure is defined as tenure
with the current employer and (ii) the 3-digit industry of current employment between time t and time t+ 1 has not changed.11
3.2. Measuring technology
3.2.1. Measuring technology innovation
In the literature on technological innovation and diffu-sion two different approaches with respect to the modelling of the role of R&D can be distinguished. In the flow approach originating withTerleckyj (1974)“own” technol-ogy is treated as a flow and measured by R&D intensities, that is by R&D expenditures over output or value added. As Griliches and Mairesse (1984)demonstrate, this is equiva-lent to setting the depreciation rate for R&D equal to zero. With this method, a proxy for the technology itused by
industry i at time t is provided by Rflow and it depends on its past R&D intensities, where
Rflowit=
R& D expenditurei, Outputi,
(4)
with ≤ t and = 1973, 1977, 1982, 1985, 1990.12 According to the perpetual inventory method (stock approach) used inGriliches and Mairesse (1984)andCoe and Helpman (1995) among others, the industry spe-cific knowledge stock accumulates by means of (real) gross investments in R&D expenditures and depreciates as knowledge gets obsolete. Notoriously there are a number of issues in the measurement of such knowledge stock (e.g., Griliches, 1979, 1990). Firstly, the R&D process takes time. Thus current R&D expenditure does not immediately add to the current knowledge stock. Secondly, past R&D invest-ments depreciate and may become obsolete. With a fixed rate of depreciation d set, to some extent set arbitrarily, at 15% and the GDP deflator used to go from nominal to real R&D expenditures as inLos and Verspagen (2000), the per-petual inventory method can be applied to construct the variable Rstock, a measure of firm i’s technology stock at
11The relevant PSID question (V11668 in the 1985–wave XVIII question-naire) to measure tenure with the current employer is “How many years (in months) altogether have you (HEAD) worked for your present employer?” Note that this question perfectly clarifies that tenure here must be con-ceived as tenure with the current employer as distinguished from tenure in the current job (e.g., PSID variables V11670 and V11671 “In what month and year did you start working in your present position/work situation?” The distinction between a change in tenure involving a change of employer and a change in tenure without a change of employer (for example due to career advancement) is important because only the first case satisfies condition (i) of the intra-industry mobility definition for intra3d mob= 1. 12Thus for example, the flow measures of innovation for industry i at time t= 1982 and t= 1990 will be the Rflowi1982= [(( R& D expenditurei,1973)/(Outputi,1973))]+ [(( R& D expenditurei,1977)/ (Outputi,1977))]+ [(( R& D expenditurei,1982)/(Outputi,1982))], and Rflowi1990= [((R& D expenditurei,1973)/(Outputi,1973))]+
[((R& D expenditurei,1977)/(Outputi,1977))]+ [((R& D expenditurei,1982)/ (Outputi,1982))]+ [((R& D expenditurei,1985)/(Outputi,1985))]+ [((R& D expenditurei,1990)/(Outputi,1990))], respectively.
time t it, as follows
Rstockit=
(1− d)R& D expenditurei,
GDP Deflator
(5) with ≤ t and = 1973, 1977, 1982, 1985, 1990. The use of one or the other of these two different measures of the industry-specific technology is relevant from an economet-ric viewpoint as well. As discussed byGrabowski (1968), when analyzing the R&D behavior of a group of industries in which substantially different scales of operation exist, research intensity, rather than R&D expenditures, may be more appropriate explanatory variable, particularly when heteroskedasticity is suspected and scale effects tend to dominate the regression equations.
3.2.2. Measuring technological diffusion
As clearly stated inGriliches (1979), the level of knowl-edge in any one sector of the economy not only derives from “own” (direct) R&D investments, but is also affected by the knowledge “imported” from other sectors. This is the process of technology diffusion where the distance (| − i|) between firm-specific technology and
economy-wide technology is shortened as knowledge and technical expertise spread and are assimilated throughout the econ-omy(OECD, 1996). The indirect R&D industry i imports from the rest of the economy will depend on the R&D performed by other industries . However, not all industries’ R&D is relevant. Thus industry j’s R&D activities will be more rel-evant to industry i the shorter the technological distance between industry i and industry j. The perpetual inventory (stock) approach and the flow approach differ substantially in terms of the way technological distance is measured.
In the flow approach technology is treated as a flow mea-sured by R&D expenditure over output or sales. Technology diffusion occurs by means of transactions of intermediate and capital inputs. In this framework, embodied technology diffusion is the introduction into production processes of machinery, equipment and components that incorporate new technology.13According to the flow approach, indirect technology flows from one industry to another when the industry originating the R&D sells products (intermediate (INT) or capital goods) embodying its R&D to other indus-tries to be used as inputs in their production processes. Thus
IndirR&D1it= R&D INTit+ R&D CAPit (6)
where R&D INTitis the R&D intensity embodied in
interme-diate goods and R&D CAPitis the R&D intensity embodied in
13To highlight the importance of technology flows of this kind, suffice to say that in advanced economies much new technology is embodied in the capital goods that industries purchase to expand and improve produc-tion. For instance, the OECD documents that for the US, the contribution of direct R&D in the economy-wide technology intensity (the sum of direct and embodied R&D) has been declining from 1972 to 1993 (OECD, 1996, p. 39, graph 2.6). Services account for an increasing share of total business sector R&D expenditure. In recent years, up to 40 percent of all R&D has been performed by the non-manufacturing sector, mainly by service firms (OECD, 1996, p.29 and graph 2.3). At the same time manufacturing indus-tries continued to shift their input structure away from R&D intensive industries towards service industries(Wolff, 1997).
capital goods that flow to industry i at time t (see(B.3) and (B.4)in the appendix for details on the measurements of these two components of indirect R&D).14The technology diffusion measure becomes
IRflowit=
IndirR&D1i, Outputi, (7)with ≤ t and = 1973, 1977, 1982, 1985, 1990. Thus the higher IRflow is, the lower is the technological difference between industry i and the rest of the economy (| − i|).
Appendix Adescribes the formal definitions of R&D diffu-sion as employed inOECD (1996).
According to the perpetual inventory approach, R&D expenditures are accumulated in a knowledge stock that depreciates as knowledge gets obsolete. Knowledge and technical expertise spread and are assimilated through-out the economy through R&D spillovers to other firms and industries the intensity of which is directly propor-tional to their technological proximity to the originating firms and industries. This amounts to measure indirect R&D as a weighted sum of (real) R&D expenditures performed in industries j other than industry i
IndirR&D2it=
j/= i
[ωijR& D expenditurejt] (8)
Following Jaffe (1986) I adopt a “technology perspec-tive”, where the weights ωij are inversely proportional
to the technological distance between industry i and industry j as measured by the fraction of overlapping of the industry-specific distributions of patenting activi-ties over technology fields. Specifically, to construct the industry-specific measure of indirect R&D I adopt the weighting system derived from Verspagen (1997) (see Verspagen, 1997, Table 1, pp. 10–12). He uses European Patent Office (EPO) data to capture technological linkages between different industrial sectors that are general and as such applicable to the US manufacturing data.15Verspagen (1997) thus obtain a matrix of weights that reflect the patent counts (seeVerspagen, 1997, for details on the pro-cedure used to construct this matrix). It is important to
14More precisely, in the indirect component of industry R&D, the OECD distinguishes between embodied and disembodied technological diffu-sion. Disembodied technological diffusion involves the transmission of knowledge, technical expertise or technology in a way that does not imply the purchase of machinery and equipment incorporating new tech-nology. Conversely embodied technology diffusion is the introduction into production processes of machinery, equipment and components that incorporate new technology. In this study we focus on the embodied indi-rect R&D.
15In about 60% of all the patent applications documents the EPO pro-cessed in the period 1979–1994, each patent is assigned to a single “main patent class”, which reflects the main application area of the patent, and multiple “supplementary” patent classes, which assign secondary appli-cation areas. Verspagen then uses a concordance table that maps 4 digits International Patent Classification (IPC) codes onto one or more of the 22 manufacturing industries of the International Standard Industrial Classi-fication (ISIC) system. The assumption here is that the main IPC code into which a patent is classified provides a good proxy for the industry that produces the knowledge, while the invention information classified into supplementary IPC codes gives an indication of knowledge spillovers to other manufacturing industries.
Table 2
Industry averages of technology measures across 1982, 1985 and 1990, low-tech and high-tech manufacturing industries.
SIC Rflow IRflow Rstock IRstock
High tech industries
Primary metal industries 33 3.81 3.19 11.00 7.21
Indus. and commer. mach. and comp. equip. 35 56.44 6.81 97.21 76.33
Stone, clay and glass products 32 3.83 1.84 9.76 16.86
Misc. manuf. industries 39 4.93 2.54 6.42 18.13
Rubber and misc. plastic products 30 4.80 3.65 12.29 6.33
Petroleum and coal products 29 4.21 1.12 28.32 3.61
Instruments and related products 38 23.34 3.70 64.55 58.83
Electronic and other electrical equip. 36 53.39 5.46 190.75 81.56
Industrial machinery and equip. 35 64.91 8.11 136.27 109.84
Chemicals and allied products 28 13.07 1.68 120.71 33.34
Transportation equip. 37 94.89 13.32 360.26 63.45
Low tech industries
Food and kindred products 20 0.72 1.18 9.8 2.69
Tobacco manufactures 21 0.73 1.20 10.1 3.51
Textile mill products 22 0.33 1.58 1.92 5.32
Apparel and related products 23 0.37 1.72 2.01 5.49
Lumber and wood products 24 0.87 1.50 2.15 1.59
Furniture and fixtures 25 0.91 1.58 2.13 1.62
Paper and allied products 26 1.01 1.54 6.91 9.48
Printing and publishing 27 1.02 1.55 6.95 9.55
Leather and leather products 31 0.36 1.70 1.98 5.36
Fabricated metal products 34 1.84 2.09 8.47 31.5
Source: OECD R&D data.
highlight that in calculating indirect R&D flows, the prin-cipal diagonal elements of the weights matrix have been set to zero to avoid multicollinearity with the direct R&D measures.
We then apply the perpetual inventory method to mea-sure the indirect technology stock of industry i at time t IRstockit=
(1− d) IndirR&D2i, GDPDeflator (9)with ≤ t and = 1973, 1977, 1982, 1985, 1990. Again, we can assume that the higher IRstock the lower the technolog-ical distance between industry i and the rest of the economy (| − i|) is.
Inspection of expressions(6) and (8)for indirect R&D should clarify that while IRflow for industry i depends on the other industries’ R&D that is embodied in their inter-mediate and capital goods purchased by industry i, I Rstock depends on the R&D of other industries weighted by means of “patent” weights. As expected there is a relatively high sample correlation between the two measures of own tech-nology (0.92 between Rflow and Rstock). Also not surprising is the lower correlation (0.71) between IRflow and IRstock. In fact, there is little doubt that the two set of techno-logical diffusion measures, namely IRflow and IRstock, are very different from each other as they capture different aspects of the process of technology spillovers.Griliches (1979)distinguishes between “rent spillovers” and “knowl-edge spillovers”. Rent spillovers occur in relation to actual transactions of intermediate and capital goods between firms/industries. Under competitive pressure, the suppliers of these goods are not able to raise prices in proportion to the quality improvements embodied in their products. As the quality/price ratio rises, firms/industries that use such goods benefit from R&D spillovers generated by the
suppli-ers. Knowledge spillovers are more directly related to the knowledge embodied in the innovation.
3.2.3. Technology data and variables
Data on direct and indirect R&D expenditures and inten-sities for 2-digit industries of the US manufacturing sector that are needed to measure Rflow and IRflow, Rstock and IRstock in(4), (7), (5) and (9), respectively, have been made available by OECD researchers and refer to a limited number of years starting from 1973 to 1993 at unequal lags of 3–5 years (1973, 1977, 1982, 1985, 1990, 199316).17The merg-ing of information for industry technology innovation and diffusion with industry of employment from the PSID nec-essarily implies that my panel data involves the years 1982, 1985, 1990. In this way we aim to explain mobility between, for example, 1982 and 1983 by means of technology vari-ables that are specific to the industry of employment in year 1982.
Appendix B clearly illustrates the methodology used by OECD statistical departments to construct measures of indirect R&D as expressed by (6). Table 2 reports industry-specific average technology measures Rflow and IRflow, Rstock and IRstock across the 3 years (1982, 1985, 1990) for 2-digit manufacturing industries grouped in sub-samples of high-tech (HT) and low-tech (LT) industries. Not surprisingly, the lowest ranked industries in terms of technology measures are leather products, lumber and wood products and apparel, three industries that
presum-16The 1993 R&D data are only in provisional form.
17Given the limited number of years in which OECD has computed the direct and indirect R&D components of expressions(4), (7), (5) and (9), by construction industry-specific measures of technology innovation and diffusion vary only in the years in which OECD has collected data on direct and indirect R&D.
ably have relatively mature technologies. By contrast, in the HT industries’ group the highest ranked industries on this index include electronics, industrial machinery, chemicals and transportation (including aircraft). In my econometric specifications for workers’ mobility I will mea-sure TEC INNit by means of expressions (4) and (5)and
TEC DIFFitby means of expressions(7) and (9).
4. Empirical specifications and econometric strategy
How does technological diffusion shape workers’ labor market opportunities? How do the effects of technological diffusion on mobility differ from the ones of technological innovation? To address these questions econometrically I will use two different econometric specifications, namely a multinomial logit specification and a mixed logit specifi-cation as detailed below.
4.1. A multinomial logit model of labor mobility
Like in any other discrete choice framework involv-ing random utilities models (RUM) we assume that at time t a worker faces a choice among various types of mobility between time t and time t+ 1, namely mobil-ity between firms but within 3-digit industries, mobilmobil-ity between 3 digit industries but within broad industrial sec-tor and mobility between broad industrial secsec-tors (defined at the 2-digit level of aggregation). Let indicate these three types of mobility with Intra3D mob, Inter3D mob and Inter2D mob, respectively. The model illustrated by expression(10)below incorporates four mutually exclusive regimes of mobility. In particular
Mob typeht=
⎛
⎜
⎝
0 if no mobility occurs
1 if mobility within 3-digit industries occurs (Intra3D mob)
2 if mobility within 2-digit industries,but between3− digit industries occurs (Inter3D mob) 3 if mobility between 2-digit sectors occurs (Inter2D mob)
⎞
⎟
⎠
(10)The relative utility associated with each alternative mobil-ity option, Mob typeht= J, J = 0, 1, 2, 3, as evaluated by
individual h at time t, is represented by a utility expression Uht,J= F(Xht, TEC INNit, TEC DIFFit)+
hJt (11)where i is the industry of current employment. In expression (11)Xht is a vector of individual-specific
socio-demographic characteristics as discussed below. Expression (11) represents the benefit from choosing option J, J= 0 (no mobility), J = 1 (Intra3D mob), J = 2 (Inter3D mob), J= 3 (Inter2D mob), net of the mobility costs that a worker faces in transferring his/her skill to a different industry. We test the hypothesis that work-ers’ mobility patterns depend on technology variables TEC INNit, as expressed by(4) and (5), and TEC DIFFit, as
expressed by expressions(7) and (9).
In the present context of modelling labor mobility as a discrete, and unordered, categorical variable, we start with a Multinomial Logit (MNL) specification where we assume that the error term that specifies each mobility option is extreme value distributed. In the MNL framework the assumption of independent and identically distributed error terms in the specification of each alternative (IIA)
holds. This econometric strategy enables us to compare the effects of individual and industry level variables on dif-ferent types of move. Note that we might expect that the variables affecting industry mobility also affect inter-firm mobility, possibly with a smaller impact. If this is true, this might affect the results of the multinomial logit esti-mations. In fact, it is not difficult to envisage scenarios in which the stochastic component associated with observ-ing a particular pattern of mobility may in fact be correlated with that of another mobility decision. To test for the failure of the independence-of-irrelevant-alteratives assumption (IIA) I compute and report Hausman test and Small-Hsiao test and propose a substantially different approach that does not rely on the IIA assumption.
4.2. A mixed logit model of workers’ mobility
When we believe that there is a non-zero correlation across (some of) the alternatives in each choice situa-tion and indeed across choice situasitua-tions, a mixed logit model offers an interesting alternative to a multinomial logit model. The random parameter model presented by Train (1998, 2003) starts with the definition of a utility function for individual h in period t, given that s/he chose alternative J, denoted by Uhit,J= ˇhXhJt+
hJtwhere XhJt isa vector of alternative-specific observable factors and
hJtis an identically and independently distributed extreme value error term, independent of ˇh. The vector of
coef-ficients ˇh can be defined as the sum of two components, ˇh= b + hwhere b is the average effect and hrepresents
an individual’s deviation from the average so that h is
drawn from a probability distribution, ∼hg(|ϑ), where ϑ
is a set of parameters defining the distribution g(.). Replac-ing this definition in the utility function above we obtain Uhit,J= bXhJt+ (hXhJt+
hJt), where the term in bracketsis the random component of Uht,J. Given the assumption
for the error term
hJt and conditioning on the randomcomponent h, the probability that individual h chooses
alternative J in period t can be expressed as in the tradi-tional formulation of a conditradi-tional logit model. Note that presented in this fashion, a mixed logit model is a way to capture individual heterogeneity because the impact of each explanatory variable varies among individuals.
A mixed logit model can be used without a random-coefficients interpretation, as simply representing error components that create correlations among the utilities for different alternatives J, J= 0, 1, 2, 3. In fact, we can re-interpret the random component hXhJtof Uhit,Jas an error
component:
hXhJt= ht,J (12)
where ht,J is correlated over alternatives and
het-eroskedastic. Thus, in our mixed logit model, the relative utility associated with each alternatives J, J= 0, 1, 2, 3 as
evaluated by individual h at time t becomes
Uhit,J= bXhJt+ (ht,J+
hJt) (13)Note that for the standard logit model, hXhJtis zero so that
there is no correlation in utility over alternatives. This lack of correlation gives rise to the IIA property and its restric-tive substitution patterns. Various correlation patterns, and hence substitution patterns, can be obtained by appropriate choice of variables to enter as error components. For exam-ple, we can specify a dummy variable for the “mobility” nest that equals 1 for each alternative involving some kind of mobility and zero for the option of no-mobility. In this case a random quantity enters the utility of all alternative belonging to the same “mobility” nest, inducing correlation across these alternatives.
While a mixed logit specification is sufficiently flexible to allow for randomness of each variable’s coefficient as well as correlations among the coefficients, in this study the primary goal is to represent substitution patterns appro-priately through the use of error components. Thus the emphasis here is on specifying variables that can induce correlations over alternatives in a parsimonious fashion so as to provide sufficiently realistic substitution patterns. For this reason here we assume that the error component ht,J is a random (normally distributed) term with zero
mean whose distribution over individuals and alternatives J produces a correlation structure across the set or sub-sets of alternatives involving mobility, Cov(ht,J, ht,J), for
J, J= 1, 2, 3. It is this non-trivial set of correlations across mobility options a worker faces that makes the IIA assump-tion unnecessary.
4.3. Explanatory variables
The non-random observable component in Uht,J
depends on Xht, a vector of individual-specific
character-istics that are likely to impact on the worker’s mobility decisions. Consistently with an abundant literature on the determinants of worker’s mobility workers’ mobility is also heavily influenced by social factors and personal characteristics. These socio-demographic characteristics can interact with technological factors in ways that is important to investigate, as Almeida and Kogut (1999), Kim and Marschke (2005),Song et al. (2003)among others have argued. Age has often been found to significantly impact on a worker’s propensity to move across industries. For this reason we include a quadratic polynomial in age (AGE, AGESQ ). Standard on-the-job search models of labor mobility predict that the intensity of search and the probability of moving will be negatively correlated with the cost of mobility. Education has been consistently identified as an important factor for mobility decisions, particularly to the extent that it proxies general (and so easily transferable) human capital (EDUC). Conversely skill that is specific (either firm-specific or industry-specific) is less likely to be positively correlated with mobility. Con-sistently with an abundant literature we use a quadratic polynomial in tenure as a proxy for firm-specific human capital (TENURE). Mobility may be more difficult in case of married workers (MARRIED). To the extent that there is still asymmetry in the intra-household bargaining power
Table 3a
Variables’ description.
Explanatory variables
Description
age The worker’s age in years agesq The square of age
educ The education level in years of school completed dumarried dummy variable= 0 if respondent is single income a worker’s annual labor income divided by 1,000 tenure tenure with the current employer in months tenuresq the square of tenure
male Dummy variable= 1 if male
ducomp dummy variable= 1 if computer technician dueng dummy variable= 1 if engineer
dutech dummy variable= 1 if technician
Tech inn measure of technology innovation, Eqs.(4) and (5) Tech diff measure of technology diffusion, Eqs.(7) and (9)
and/or males and females show different propensities to move, gender can be found as a significant factor in determining industrial mobility. In a world where the quality of the match between workers and jobs matters, the current labor income (INCOME) is expected to be neg-atively correlated with the probability of moving across industries when other variables such as education, age and tenure are controlled for (Jovanovic and Moffitt, 1990). Table 3a reports full description of the right-hand-side variables used.
Table 3breports summary statistics for the three main samples of US manufacturing workers, manufacturing workers employed in high-tech industries and manufactur-ing workers in low-tech industries. It is instructive to notice that mobility within 3-digit industries is slightly higher in a sample of low tech manufacturing workers (13% as opposed to 10% in high tech industries and 11% in the full sample of manufacturing workers). Conversely mobility between 2-digit industries is more likely among high-tech workers (34%) than in the sample of low tech workers (26%) or in the full sample (31%). The average level of education is 11.4 years of education in the sample of low-tech workers and just above 12.3 in high-tech workers. Labor income levels also differ substantially being much higher for high tech workers than for low-tech workers. Tenure is also signifi-cantly different in the two main sub-samples as the mean number of months with the current employer is 129 in the sample of workers in high-tech industries but only 110 in the sample of workers in low-tech industries. The sam-ple percentages of technical workers (engineers, computer technicians and technicians) are also evidently higher in a sample of workers employed in high-tech industries than in the sample of workers employed in low-tech industries and so are the sample percentages for the technology inno-vation and diffusion measures.18
The right hand side of expressions(11) and (13)in the multinomial logit and mixed logit models, respectively, depends on industry technology variables, TEC INNit and
TEC DIFFitfor technology innovation and diffusion,
respec-18Unfortunately the PSID does not provide information on other deter-minantes of mobility such as the extent of a worker’s social networks and the special localization of knowledge, factors that are proved to be relevant to mobility decisions of patent holders inAlmeida and Kogut (1999).
Table 3b
Summary statistics for the three main samples of all workers in US manufacturing sector, high-tech industries’ workers and low-tech industries’ workers.
Variable Observations Mean Standard deviation Minimum Maximum
Full sample of US manufacturing workers
intra3dmob 6625 0.114 0.317 0 1 inter3dmob 6625 0.086 0.280 0 1 inter2dmob 6625 0.307 0.461 0 1 age 6625 38.625 10.864 18 76 male 6625 0.905 0.293 0 1 married 6625 1.377 0.971 1 5 educ 6625 11.909 2.506 1 17 empten 6625 120.906 108.624 0 630 emptensq 6625 26415.740 42115.730 0 396900 ducomp 6625 0.018 0.131 0 1 dueng 6625 0.056 0.231 0 1 dutech 6625 0.035 0.184 0 1 hightech 6625 0.566 0.496 0 1 du85 6625 0.301 0.459 0 1 du90 6625 0.384 0.486 0 1 income 6625 24.142 14.835 0 250 tech inna 6625 26.843 34.677 0.19 113.05 tech diffa 6625 4.283 4.233 0.64 15.84 tech innb 6625 76.850 98.572 1.279 315.906 tech diffb 6625 29.231 27.642 1.055 94.880
High-tech industries’ workers
intra3dmob 3751 0.101 0.302 0 1 inter3dmob 3751 0.078 0.268 0 1 inter2dmob 3751 0.340 0.474 0 1 age 3751 39.084 10.515 19 72 male 3751 0.918 0.274 0 1 married 3751 1.326 0.914 1 5 educ 3751 12.274 2.461 1 17 empten 3751 129.006 109.239 0 528 emptensq 3751 28572.510 41692.850 0 278784 ducomp 3751 0.025 0.156 0 1 dueng 3751 0.089 0.284 0 1 dutech 3751 0.049 0.215 0 1 hightech 3751 1 0 1 1 du85 3751 0.312 0.463 0 1 du90 3751 0.376 0.484 0 1 Income/1000 3751 26.994 15.030 0 250 tech inna 3751 46.710 34.840 1.78 113.05 tech diffa 3751 6.363 4.626 0.64 15.84 tech innb 3751 131.322 101.544 3.553 315.906 tech diffb 3751 44.680 26.718 1.433 94.880
Low-tech industries’ workersc
intra3dmob 2874 0.130 0.336 0 1 inter3dmob 2874 0.096 0.295 0 1 inter2dmob 2874 0.265 0.441 0 1 age 2874 38.027 11.278 18 76 male 2874 0.888 0.316 0 1 married 2874 1.442 1.038 1 5 educ 2874 11.434 2.486 1 17 empten 2874 110.334 106.917 0 630 emptensq 2874 23600.830 42504.260 0 396900 ducomp 2874 0.008 0.087 0 1 dueng 2874 0.015 0.120 0 1 dutech 2874 0.017 0.129 0 1 hightech 2874 0 0 0 0 du85 2874 0.285 0.452 0 1 du90 2874 0.394 0.489 0 1 Income/1000 2874 20.421 13.715 0 120 tech inna 2874 0.913 0.548 0.19 2.44 tech diffa 2874 1.568 0.597 0.69 2.79 tech innb 2874 5.756 3.645 1.279 12.267 tech diffb 2874 9.069 10.567 1.055 36.815
aFlow measures of technology innovation and diffusion (expressions(4) and (7)in the text). bStock measures of technology innovation and diffusion (expressions(5) and (9)in the text). cHigh-tech and low-tech industries are defined inTable 2.
E. Magnani / Structur al Change and Economic Dynamics 20 (20 09) 1 6–3 7 Table 4a
Small-Hsiao tests of IIA assumption, Full sample in top panel, High-tech industries workers in central panel, Low-tech industries workers in bottom panel. Flow measures and stock measures of technology at the left hand side and right hand side, respectively, are defined by expression(4) and (7)(flow) and by expression(5) and (9)(stock). Ho: Odds (Outcome-J vs. Outcome-K) are independent of other alternatives.
Multinomial logit specification with flow measures of technology Multinomial logit specification with stock measures of technology
Omitted lnL(full) lnL(omit) chi2 df P > chi2 Evidence Omitted lnL(full) lnL(omit) chi2 df P > chi2 Evidence
Full sample of US manufacturing industries
1 −2605.475 −2585.316 40.317 34 0.211 for Ho 1 −2580.490 −2564.110 32.760 34 0.528 for Ho
2 −2844.003 −2826.285 35.436 34 0.400 for Ho 2 −2775.971 −2758.522 34.898 34 0.425 for Ho
3 −1753.667 −1731.356 44.623 34 0.105 for Ho 3 −1807.308 −1792.445 29.727 34 0.677 for Ho
High-tech industries’ workers
1 −1446.470 −1425.866 41.207 32 0.128 for Ho 1 −1416.268 −1404.645 23.245 32 0.870 for Ho
2 −1603.294 −1582.949 40.689 32 0.139 for Ho 2 −1515.816 −1500.138 31.355 32 0.499 for Ho
3 −878.914 −861.686 34.457 32 0.351 for Ho 3 −843.998 −827.638 32.720 32 0.431 for Ho
Low-tech industries’ workers
1 −1081.668 −1064.379 34.579 32 0.346 for Ho 1 −1128.722 −1064.382 128.681 32 0.000 against Ho
2 −1168.480 −1149.637 37.687 32 0.225 for Ho 2 −1190.299 −1134.068 112.460 32 0.000 against Ho
3 −854.628 −840.249 28.759 32 0.631 for Ho 3 −928.669 −829.005 199.329 32 0.000 against Ho
Table 4b
Hausman tests of IIA assumption, full sample in top panel, high-tech industries workers in central panel, low-tech industries workers in bottom panel. Flow measures and stock measures of technology at the left hand side and right hand side, respectively, are defined by expression(4) and (7)(flow) and by expression(5) and (9)(stock). Ho: Odds (Outcome-J vs. Outcome-K) are independent of other alternatives.
Multinomial Logit Specification with flow measures of technology Multinomial Logit Specification with stock measures of technology
Omitted chi2 df P > chi2 Evidence Omitted chi2 df P > chi2 Evidence
Full sample of US manufacturing industries
1 68.855 34 0.000 against Ho 1 66.538 34 0.001 against Ho
2 57.642 34 0.007 against Ho 2 90.130 34 0.000 against Ho
3 125.899 34 0.000 against Ho 3 85.249 34 0.000 against Ho
High-tech industries’ workers
1 56.417 32 0.005 against Ho 1 49.373 32 0.026 against Ho
2 22.142 32 0.904 for Ho 2 63.549 32 0.001 against Ho
3 97.846 32 0.000 against Ho 3 52.430 32 0.013 against Ho
Low-tech industries’ workers
1 35.932 32 0.289 for Ho 1 36.560 32 0.265 for Ho
2 56.601 32 0.005 against Ho 2 54.637 32 0.008 against Ho
Table 5
Correlation structure of the alternative specific constant (ASC), J= 1, 2, 3, with J = 1 for intra3D mob, J = 2 for inter3D mob, J = 3 for inter2D mob. Variables Full sample Workers in high-tech industries Workers in low-tech industries
Coefficient S.E. Coefficient S.E. Coefficient S.E.
Unrestricted Mixed Logit specification with flow measures of technology
Var(ht,1) 0.02 0.083 0.7 0.357 0.002* 0.023 Cov(ht,2, ht,1) −0.248 0.500 1.602*** 0.478 −0.05 0.288 Cov(ht,3, ht,1) −0.248 0.120 0.391 0.239 −0.067 0.382 Var(ht,2) 3.019*** 0.574 4.247*** 1.019 1.616*** 0.555 Cov(ht,3, ht,2) 0.813** 0.331 1.93*** 0.573 0.811 0.682 Var(ht,3) 1.824*** 0.279 2.616*** 0.494 4.508*** 1.036
Unrestricted Mixed Logit specification with stock measures of technology
Var(ht,1) 0.238 0.678 0.352 0.298 0.001 0.018 Cov(ht,2, ht,1) −0.627*** 0.303 1.54 0.700 −0.037 0.280 Cov(ht,3, ht,1) 0.109 0.490 0.36 0.227 −0.052 0.384 Var(ht,2) 3.900*** 0.708 6.937*** 1.558 1.673*** 0.571 Cov(ht,3, ht,2) 0.490 0.4369 1.026 0.654 0.635 0.669 Var(ht,3) 1.811*** 0.349 2.443*** 0.460 4.351*** 0.994
tively as measured by (4) and (7) and by (5) and (9), respectively, depending on the specification. Our working assumption is that to the extent that technology innovation increases the technology distance between the industry of current employment and other industries, TEC INNit
is expected to have a negative impacts on mobility par-ticularly on Inter3D mob and on Inter2D mob as defined above, or b
INN< 0. However, in low-tech industries this
may not be the case because technology innovation may actually contribute to lower the industry’s technology distance. TEC DIFFit is the embodied R&D (technological
diffusion) characterizing the industrial sector (defined at the 2-digit level of aggregation) of current employment. Our hypothesis is that technological diffusion increases skill transferability thus positively impacting upon work-ers’ mobility. In this case the estimated coefficient of the technology diffusion variable, bDIFF> 0.
Finally, dummy variables for years and for broad indus-trial sectors of employment (high tech sector and low tech sector) appear in all full sample regressions to capture any potentially non-linear time trends in workers’ flows and other unmeasured determinants of mobility that are either job/firm-specific or industry-specific.
5. Technological change and workers’ mobility: empirical results
The empirical results are organized in six tables. Tables 4a and 4bprovide some indication of whether pos-sible deviations from the IIA assumption may impact upon the MNL results.Table 4areports the Small-Hsiao IIA test corresponding to all specifications, with flow and stock measures of technology in the left hand side and right hand side, respectively, and for the full samples and sub-samples of workers in HT and LT industries. Although we can only reject the null hypothesis of IIA in one case (specification with stock measures of technology using LT workers) we suspect that IIA may not hold in more cases (e.g., speci-fications with flow measures of technology with the full sample as well as with the HT workers’ sample.Table 4b report the Hausman tests and confirms that we can reject the IIA assumption particularly in the case of the full
sam-ple and HT workers’ samsam-ples. These results lead us to turn to the mixed logit results in Tables 5–7. The tables for the multinomial logit regression results are available upon request, but the reader should bear in mind that in general MNL results and mixed logit results are broadly consis-tent. Before turning to commenting the main variables of interest, namely TEC INNit and TEC DIFFit it is worth to
notice that a dummy variables for employment in high-tech industries is consistently statistically significant in the full sample specifications, both MNL and mixed logit. The statistically significant coefficients for the dummy variable for employment in high-tech industries motivates our deci-sion to look at samples of HT and LT industries’ workers separately. Full samples results are however, available upon request.
Table 5reports the estimated covariance structure (vari-ances and covari(vari-ances) between the option specific error terms ht,J and ht,J, J= 1, 2, 3; J= 1, 2, 3 in expression
(13).Tables 6 and 7report the mixed logit results for the two main samples, namely a sample of HT workers and a sam-ple of LT workers. Finally,Tables 8a and 8breport the mixed logit results obtained by using samples of skilled workers and unskilled workers in LT and HT industries, using stock measures and flow measures of technology, respectively. 5.1. The impact of technology innovation and diffusion when options of mobility are correlated. Mixed logit results
Table 5illustrates the results obtained when we run preliminary unrestricted mixed logit specifications for the three main samples (all workers, workers employed in HT industries, workers employed in LT industries) with the two sets of technology measures. These unrestricted specifica-tions allow us to estimate the variances and covariances across the three main option-specific error terms ht,J for
J= 1, 2, 3, in expression(13). Two set of results emerge fromTable 5. Firstly, there is a specific ranking in terms of size of the variance of the ht,Jerror term in the three
samples. Specifically, Var(ht,2) > Var(ht,3) > Var(ht,1) in
the full sample’s unrestricted regressions. When a sam-ple of HT workers is employed, the resulting ranking of
Table 6
Mixed Logit results for mobility options, workers employed in high-tech industries, 1982, 1985, 1990. Base category= 0 (no mobility), 1 = intra3D mob, 2= inter3D mob and 3 = inter2D mob.a, b
Specification with flow measures of tech Specification with stock measures of tech
Explan. Var. Coefficient Std. Err. Explan. Var. Coefficient Std. Err.
age1 −0.149** 0.050 age1 −0.149** 0.050 age2 0.274** 0.089 age2 0.264** 0.095 age3 −0.014 0.040 age3 −0.012 0.040 agesq1 0.001** 0.0006 agesq1 0.001** 0.0006 agesq2 −0.004*** 0.001 agesq2 −0.004** 0.001 agesq3 0.0004 0.0005 agesq3 0.0004 0.0005 male1 −0.061 0.397 male1 −0.043 0.395 male2 0.886 0.558 male2 0.454 0.604 male3 −0.146 0.287 male3 −0.185 0.288 dumarried1 0.231 0.306 dumarried1 0.223 0.305 dumarried2 0.030 0.395 dumarried2 −0.031 0.434 dumarried3 −0.172 0.226 dumarried3 −0.187 0.226 educ1 0.035 0.029 educ1 0.033 0.029 educ2 0.064 0.046 educ2 0.068 0.050 educ3 0.060** 0.025 educ3 0.051** 0.026 tenure1 0.016*** 0.002 tenure1 0.016*** 0.002 tenure2 −0.0006 0.003 tenure2 0.0008 0.003 tenure3 −0.007*** 0.002 tenure3 −0.007*** 0.002 tenuresq1 −0.00003*** 0.000005 tenuresq1 −0.00003*** 0.000005 tenuresq2 0.00002** 0.000008 tenuresq2 0.00002** 0.000008 tenuresq3 0.000007* 0.000004 tenuresq3 0.000006 0.000004 ducomp1 0.553 0.433 ducomp1 0.549 0.434 ducomp2 0.222 0.631 ducomp2 0.314 0.643 ducomp3 0.853** 0.352 ducomp3 0.599* 0.356 dueng1 0.250 0.232 dueng1 0.254 0.230 dueng2 −0.183 0.358 dueng2 0.149 0.378 dueng3 0.068 0.207 dueng3 0.087 0.207 dutech1 −0.266 0.356 dutech1 −0.268 0.355 dutech2 0.725* 0.372 dutech2 0.989** 0.396 dutech3 0.260 0.248 dutech3 0.258 0.250 du85 1 −0.499** 0.170 du85 1 −0.491** 0.178 du85 2 0.055 0.211 du85 2 −0.384 0.237 du85 3 0.340** 0.129 du85 3 0.267** 0.136 du90 1 0.304* 0.170 du90 1 0.315* 0.165 du90 2 0.526** 0.246 du90 2 −0.564** 0.258 du90 3 0.587*** 0.143 du90 3 0.370** 0.138 income1 −0.011** 0.006 income1 −0.012** 0.006 income2 −0.006 0.008 income2 −0.006 0.009 income3 0.0002 0.004 income3 −0.0003 0.004
Tech inn1 −0.002 0.008 Tech inn1 0.0001 0.0007
Tech inn2 0.093*** 0.014 Tech inn2 −0.011*** 0.002
Tech inn3 0.007 0.006 Tech inn3 −0.005*** 0.0007
Tech diff1 0.019 0.058 Tech diff1 −0.0001 0.003
Tech diff2 −1.129*** 0.125 Tech diff2 0.009* 0.005
Tech diff3 −0.130** 0.048 Tech diff3 0.010*** 0.002
mobASC 0.025 1.064 mobASC 0.057 1.057 inter3dASC −7.219*** 1.998 inter3dASC −7.815*** 2.142 inter2dASC −0.310 1.171 inter2dASC −0.398 1.174 SD SD mobASC 0.797*** 0.159 mobASC 0.764*** 0.164 inter3dASC 1.787*** 0.282 inter3dASC 2.382*** 0.303 inter2dASC 1.377*** 0.165 inter2dASC 1.395*** 0.169 No. of obs. 15,004 15,004 Log likelihood −3892.0 −3958.9 Chi squared(3) 135.7*** 189.8**
a***, **, * indicate statistical significance at the 1%, 5% and 10% level, respectively.
bIn the first column VarJ, J= 1, 2, 3 is the coefficient of variable “var” in the estimation of the relative odd of event J. variances is Var(ht,2) > Var(ht,3) > Var(ht,1). The
rank-ing resultrank-ing from the specifications with a sample of LT workers is Var(ht,3) > Var(ht,2) > Var(ht,1). The second
aspect illustrated inTable 5concerns the correlation struc-ture across alternative mobility options. While in the full sample and in the sample of LT workers there is positive correlation across the decisions involving moving across
3-digit industries (namely inter3D mob and inter2D mob), HT workers’ specifications show a positive correlation across the three options of mobility (namely intra3D mob, inter3D mob and inter2D mob). We use this information to estimate restricted versions of the mixed logit specifications as inTable 6for HT workers and inTable 7for LT work-ers. In particular, alternative specific constants (ASC) such
Table 7
Mixed Logit results for mobility options, workers employed in low-tech industries, 1982, 1985, 1990. Base category= 0 (no mobility), 1 = intra3D mob, 2= inter3D mob and 3 = inter2D mob.a, b
Specification with flow measures of tech. Specification with stock measures of tech.
Explan. Var. Coefficient Std. Err. Explan. Var. Coefficient Std. Err.
age1 0.035 0.044 age1 0.033 0.044 age2 −0.042 0.053 age2 −0.040 0.052 age3 0.083 0.052 age3 0.080 0.051 agesq1 −0.0004 0.0005 agesq1 −0.0004 0.0005 agesq2 0.0004 0.0006 agesq2 0.0004 0.0006 agesq3 −0.001 0.0006 agesq3 −0.001 0.0006 male1 0.195 0.303 male1 0.208 0.299 male2 0.022 0.367 male2 −0.124 0.361 male3 −0.369 0.346 male3 −0.137 0.340 dumarried1 0.243 0.234 dumarried1 0.238 0.234 dumarried2 0.359 0.294 dumarried2 0.384 0.292 dumarried3 −0.136 0.273 dumarried3 −0.148 0.272 educ1 −0.048* 0.026 educ1 −0.047* 0.026 educ2 −0.042 0.034 educ2 −0.041 0.034 educ3 0.020 0.035 educ3 0.018 0.034 tenure1 0.006** 0.002 tenure1 0.006** 0.002 tenure2 0.003 0.002 tenure2 0.004 0.002 tenure3 −0.007** 0.002 tenure3 −0.007*** 0.002 tenuresq1 −0.000003 0.000003 tenuresq1 −0.00003 0.000004 tenuresq2 −0.00001 0.000006 tenuresq2 −0.00001* 0.000006 tenuresq3 0.00001* 0.000005 tenuresq3 0.00001** 0.000005 ducomp1 0.675 0.849 ducomp1 0.650 0.848 ducomp2 2.069** 0.730 ducomp2 2.136** 0.717 ducomp3 0.970 0.921 ducomp3 0.721 0.928 dueng1 0.302 0.687 dueng1 0.307 0.685 dueng2 0.494 0.851 dueng2 0.437 0.852 dueng3 2.069** 0.617 dueng3 2.118** 0.613 dutech1 −1.016 0.764 dutech1 −0.996 0.763 dutech2 0.583 0.578 dutech2 0.563 0.578 dutech3 0.938* 0.560 dutech3 0.951* 0.556 du85 1 0.310 0.226 du85 1 0.387** 0.173 du85 2 −0.574** 0.282 du85 2 0.316 0.194 du85 3 −0.382 0.254 du85 3 −0.398** 0.183 du90 1 0.662* 0.357 du90 1 0.781*** 0.163 du90 2 −1.390** 0.483 du90 2 0.355* 0.185 du90 3 −0.708* 0.421 du90 3 −0.340* 0.178 income1 −0.023*** 0.006 income1 −0.021** 0.006 income2 −0.011 0.008 income2 −0.012 0.008 income3 −0.003 0.007 income3 −0.002 0.007
Tech inn1 0.002 0.177 Tech inn1 −0.012 0.019
Tech inn2 −0.760*** 0.211 Tech inn2 −0.084** 0.026
Tech inn3 1.495*** 0.261 Tech inn3 0.049* 0.025
Tech diff1 0.075 0.296 Tech diff1 −0.0009 0.007
Tech diff2 1.613*** 0.392 Tech diff2 0.019** 0.009
Tech diff3 −0.206 0.393 Tech diff3 0.049*** 0.008
intra3dASC −2.598** 03956 intra3dASC −2.480** 0.927 mobinterASC −1.869* 1.126 mobinterASC −0.554 1.075 inter2dASC −0.860 1.368 inter2dASC −2.085 1.324 SD SD intra3dASC −0.536* 0.279 intra3dASC −0.523* 0.284 mobinterASC 1.155*** 0.212 mobinterASC 1.140*** 0.205 inter2dASC 1.839*** 0.291 inter2dASC 1.828*** 0.275
No. of obs. 11,496 No. of obs. 11,496
Log likelihood −3182.8 Log likelihood −3197.9
Chi squared(3) 87.11*** Chi squared(3) 87.6***
aseeTable 6.
