The March of the Techies:Job Polarization Within and Between Firms

James HarriganUniversity of Virginia, Paris School of Economics, and NBER

Ariell ReshefCNRS, Universite Paris 1 Pantheon-Sorbonne,

Paris School of Economics and CEPII

Farid ToubalENS-Paris Saclay, CREST and CEPII

April 2020*

Abstract

Using administrative employee-firm-level data on the entire private sector from1994 to 2007, we show that the labor market in France has polarized: employmentshares of high and low wage occupations grew, while middle wage occupationsshrank. At the same time, the share of technology-related occupations (“techies”)grew substantially. Aggregate polarization was driven mostly by changes in thecomposition of firms within industries. Within-firm adjustments and changes inindustry composition were much less important. Polarization occured mostly withinurban areas, with roughly equal contributions of men and women. We study the roleof technology adoption in shaping firm-level outcomes using a new measure of thepropensity of a firm to adopt new technology: its employment share of techies. Wefind that techies were an important force driving aggregate polarization in France,as firms with more techies grew faster.

JEL classifications: J2, O3, D3, F1, F16, F66.

Keywords: Job polarization, technological change, STEM skills, techies, offshoring, firmlevel data.

*james.harrigan@virginia.edu, ariell.reshef@psemail.eu, ftoubal@ens-cachan.fr. This research is sup-ported by a public grant overseen by the French National Research Agency (ANR) as part of the ”In-vestissements d’avenir” program (ANR-10-EQPX-17 CASD). We also acknowledge financial supportfrom the iCODE institute and from the ANR agency under the program ANR-16-CE26-0001-01. Wethank Francis Kramarz for guidance with the data, and many seminar audiences in the United Statesand Europe for helpful comments.

1

1 Introduction

Job polarization—growth in the employment shares of high-wage and low-wage jobs atthe expense of middle wage jobs—is one of the most striking phenomena in many ad-vanced economies’ labor markets in the last several decades.1 While job polarizationhas been mostly studied at aggregate levels, much less is known about the underlyingfirm-level mechanics that drive it. The firm-level dimension is important because firm-level decisions—for example, about investments in technological improvements and onglobal engagement—affect both overall firm employment growth and within-firm internaloccupational changes, which together add up to aggregate changes.

We use administrative data for the entire private sector in France from 1994 to 2007to show that France has indeed experienced job polarization: employment shares ofhigh-wage managers and professionals increased; shares of middle-wage office workersand industrial workers fell; and shares of low-wage retail, personal service and unskilledmanual workers increased. However, the picture that emerges is more complex than thissimple relationship between wage ranks and changes in employment shares. For example,middle managers and technicians earn similar middle-income wages, yet employment inmiddle management occupations declined, while technicians increased their employmentshares.

We then study the dimensions along which polarization happened: industries, firms,geography and gender. Using regressions motivated by theory, we also study whethertechnological change is associated with polarization through changing the composition offirms or by altering occupational shares within firms.

The use of matched firm-level data allows us to shed light on the dimensions alongwhich job polarization occurs. The aggregate employment trends are visible across Frenchurban and rural areas, at the industrial and firm level, and across gender. We comparehow aggregate job polarization manifests in compositional changes of firms, industries,and departments—versus changes within these units.2 As Cerina et al. (2017) do forthe U.S, we also study whether polarization is more apparent for men than for women.We show that the main channel through which job polarization occurs is changes withinindustry in the composition of firms’ size, and not changes in occupational employmentshares within firms. In other words, polarization didn’t happen because firms changedtheir employment mix, it happened because firms intensive in middle-wage jobs grewmore slowly than firms intensive in high and low wage occupations. Indeed, even iffirms had not changed their internal occupational composition at all, we would have seenalmost the same polarization at the aggregate level, due to changes in firms’ sizes. Thissuggests that explanations that rely on substitution (either within local labor markets or

1Job polarization has been documented in the United States (Autor et al. (2006), Autor et al. (2008),Firpo et al. (2011)), the United Kingdom (Goos and Manning (2007)), Germany (Spitz-Oener (2006),Dustmann et al. (2007)), and more generally in Europe (Goos et al. (2009) and Oesch (2013)). Po-larization contrasts with earlier labor market developments, where changes in employment shares ofmiddle-wage jobs were more modest, and the growth of high-wage jobs was at the expense of low-wagejobs. For example, in 1980s in the U.S., changes in employment shares are positively related to wages inthe 1980s (Autor et al. (2008)).

2France is divided into 101 departments. Our analysis is restricted to the 94 departments that com-prise mainland France. The median mainland department had a population of just above 500 thousandin 2007.

2

national industries) are missing an important dimension of the mechanics of polarization.Moreover, as Autor (2019) notes for the U.S., we find that job polarization is primarilyan urban phenomenon.

These results lead us to ask what factors explain changes of firms’ sizes and changesin occupational shares. Most explanations for job polarization focus on the “routiniza-tion hypothesis”: new and cheaper information and communication technologies (ICT)perform “routine”, codifiable tasks that would otherwise be performed by middle-wageworkers. At the same time, ICT has been found to be particularly complementary to“non-routine”, non-codifiable tasks (e.g., analysis, decision making) that are typicallyperformed by high-wage workers.

In France, the price of ICT capital dropped by roughly 30% in our sample. We developa simple model which shows how a drop in the cost of ICT capital affects within-firmoccupational shares and relative firm size. In the model, variation in initial levels of ICTintensity across firms leads to different responses to a common drop in the price of ICT,with more ICT-intensive firms seeing their costs drop more. As a result of this cost-reduction effect, firms that are more ICT-intensive grow relative to less ICT-intensivefirms. The effect of falling ICT prices on within-firm occupational polarization dependson how ICT substitutes for or is complementary to different types of labor. We showthat under a broad set of parameters, more ICT-intensive firms will tend to experiencegreater occupational polarization.

We evaluate the predictions of the model, in particular how changes depend on initialICT intensity, by estimating regressions using weighted least squares (WLS) and two-stage least squares (W2SLS). We use the initial share of techies in the firm as a proxy forinitial ICT intensity, as suggested by the model. In addition to the role of routinizationand technological progress, many authors have linked the phenomenon of polarizationwith globalization.3 The data allow us to measure the importer and exporter statusof the firm. We find that export access to new markets increases employment, whilefirm level imports have ambiguous effects. Access to imported intermediate inputs willincrease a firm’s competitiveness, which will tend to raise employment, but may alsodirectly substitute for some tasks formerly done by workers in the firm, thus reducingemployment of some occupations. Concerning the composition of employment withinfirm, we expect a greater effect of international trade on occupations that are moredirectly exposed to it. Imports may explain the decline of mid-level wage jobs becausethey substitute the tasks associated with jobs that can be carried out by a less costlyworkforce abroad. Since these tasks are to be performed at long distance, they are morelikely to be offshored if they require less face-to-face interpersonal interaction (Blinder(2006)). International trade can also favor high-wage workers by increasing demandfor non-routine cognitive tasks, those associated with within-firm internal occupationalchanges, or those related to management and communication between the firm’s affiliatesthat are located in different countries.

A novel contribution of our paper is our focus on the firm-level employment share oftechies: workers with STEM skills and experience. We focus on techies because of theircentral role in planning, installing, and maintaining information and computer technol-ogy (ICT) and other technologies, as well as in training and assisting other workers in

3Important contributions including Autor and Dorn (2013), Autor et al. (2013), Autor et al. (2015),Goos et al. (2014), and Malgouyres (2017) link polarization with trade.

3

the use of technology. These roles make techies the crucial link between economy widetechnological progress and firm level technology adoption. As noted by Tambe and Hitt(2014), “the technical know-how required to implement new IT innovations is primarilyembodied within the IT workforce” in a firm. Brynjolfsson and Hitt (2003) observe that“In recent years, companies have implemented thousands of large and small innovations insoftware applications, work processes, business organization, supply-chain management,and customer relationship management”. Implementation of these innovations was un-doubtedly mediated by techies, who are a good measure of the “organizational capital”that Brynjolfsson and Hitt (2003) argue is crucial to ICT adoption. The richness of ourdata allows us to make a distinction between two different categories of techie workers:highly-educated technical managers and engineers on the one hand, and lower-rankedtechnicians on the other.

Of course, not all technological change comes through ICT investment, and our mea-sure of techies includes engineers, technology managers, and technicians more broadly.For example, Helper and Kuan (2018) survey the auto parts industry and find that in-novation occurs through the efforts of firms’ engineers and technicians, although withoutexplicit R&D expenditures. Our paper is the first to measure the number of these work-ers at the firm level in the entire economy, which allows us to associate techies with firmoutcomes, including internal occupational change and employment growth.

The econometric results suggest strong effects of techies on firm-level employmentgrowth which are mostly driven by engineers rather than technicians. We also find apositive association between techies and the share of managers within firms and a negativeassociation between techies and the share of skilled and services workers.

Our findings on the importance of techies are consistent with their hypothesized roleas a conduit for technological change, whether through ICT adoption or more broadly.Since techie workers have STEM skills and education, an alternative interpretation ofour results is that it is STEM skills themselves that explain our findings, rather thantechnology adoption.

Our paper is organized as follows. Section 2 discusses the relationship to the existingliterature. We then discuss our theoretical framework for firm-level analysis in Section5. In Section 4 we discuss data, document job polarization in the French labor marketand show how polarization has evolved mostly due to firm composition. In section 6, wedevelop the econometric framework for estimating the impact of technology and tradebetween and within firms. In Section 7, we present econometric results showing how firmcharacteristics are associated both with firm employment growth and with within-firmoccupational change. Section 8 concludes.

2 Relationship to existing literature

This is one of the first papers to describe and analyze polarization across and withinfirms. Since employment decisions are made at the firm level, firm-level data is idealfor studying polarization. Other papers have looked at polarization across industries orregions (see Beaudry et al. (2010), Autor and Dorn (2013), Michaels et al. (2014), andGoos et al. (2014). Goos et al. (2014) is the only previous paper to have found evidenceof job polarization in France, although their discussion is limited to a single line in their

4

Table 2, and using survey data that are inferior to our data sources. Caliendo et al.(2015) analyze within-firm internal occupational change in France with the same datathat we use, focusing on hierarchies, but restrict their attention to manufacturing firms(25% of private sector hours) and do not discuss ICT or polarization.

The main explanation for job polarization in the literature is the “routinization hy-pothesis” (Goos and Manning (2007)). As argued in Autor et al. (2003), technologicalprogress in information and communications technology (ICT) allows machines to replacecodifiable cognitive routine tasks that were once performed by humans. These tasks hap-pen to be more common in occupations that are, on average, in the middle of the wagedistribution. Thus, the diffusion of ICT lowers demand for these occupations. At thesame time, ICT complements non-routine cognitive tasks, and demand for occupationsthat are characterized by these tasks—which are higher up in the wage distribution—rises. Occupations at the bottom of the wage distribution are less affected by ICT, andthey absorb the residual supply of labor.4 Our results broadly support the importance ofthe “routinization hypothesis”, but with important nuances.

Our findings on the importance of techies for overall employment growth, while sub-stituting for some occupations, are consistent with Graetz and Michaels (2018), who findthat industries that invested more in robotization reduced relative demand for workerswho are close substitutes to robots, but overall increased labor demand. As they do, weinterpret this finding through a product demand mechanism: the main effect of firm-leveltechnological change is to lower prices and increase competitiveness, which boosts labordemand. Similarly, Goos et al. (2014) and Gregory et al. (2016) also estimate increasesin overall labor demand within economic units (industries or local labor markets) thatundergo differentially more technological change through a similar competitive effect.

The effect of ICT investment on growth and internal occupational change is shownin several studies that typically focus on a particular industry or specific settings. Incontrast, our paper considers the entire French private sector. Lichtenberg (1995) andBrynjolfsson and Hitt (1996), working with a small number of U.S. firms in the late 1980sfind that IT labor has a positive output elasticity. Tambe and Hitt (2012) corroboratethis finding in a newer data source and more sophisticated estimation technique. Autoret al. (2002) study organizational change due to the introduction of digital check imagingwithin one large bank. Bresnahan et al. (2002) discuss the complementarity between IT,decentralized firm organization, and skilled labor. Our results below are consistent withthis; we find that techies (which we also associate with IT) cause an increase in the em-ployment share of top managers and a decrease of middle management employment shareswithin firms. However, our descriptive results imply that within-firm re-organization maynot be the most important margin of adjustment.

Several studies consider the manufacturing sector due to data availability. Our workis closely related to Maurin and Thesmar (2004), who investigate changes in organizationwithin French manufacturing firms in 1984–1995, the period preceding ours, where theyfocus on skill upgrading and do not consider the role of techies. Dunne et al. (2004)find that computer use within U.S. manufacturing plants is associated with greater skillintensity, although not with greater labor productivity. Bartel et al. (2007) show that ICT

4Acemoglu and Autor (2011) provide an analytical framework that suggests how tasks are bundledacross types of workers (differentiated by education level or skill), and how changes in demand for thesetasks affect employment shares of these types.

5

adoption in the U.S. valve manufacturing industry caused reorganization within firms,raised the skill-requirements for machine operators and increased productivity (throughfaster setup times, greater customizability, and better quality control). Similar to ourresults on techies and employment growth, Barth et al. (2017) find a positive associationbetween the science and engineer share of employment with revenue in U.S. manufacturingplants. In contrast, we study the entire private sector and make causal inferences.

The idea that engineers and other technically-trained workers are important for pro-ductivity growth has also found support in the economic history literature. Kelly et al.(2014) and Ben Zeev et al. (2017) highlight the importance of the British apprentice sys-tem during the British Industrial Revolution in supplying the basic skills needed for tech-nology adoption (whether British technology or other). Maloney and Valencia-Caicedo(2017) construct a dataset of engineer intensity for the Americas and for U.S. countiesaround 1880, and show that this intensity helps predict income today. In addition, engi-neers are at the center of modern (endogenous) growth theory, e.g., Romer (1990).

A second force that could help explain job polarization and firm re-organization in gen-eral is offshoring, where domestic labor is replaced by imported inputs (see among manyothers Feenstra and Hanson (1996), Grossman and Rossi-Hansberg (2008), Rodriguez-Clare (2010), and Blinder and Krueger (2013)). Empirically, our results suggest a rel-atively modest role for offshoring in explaining polarization, as have a number of otherstudies including Feenstra and Hanson (1999), Michaels et al. (2014), Oesch (2013) andBiscourp and Kramarz (2007).

Most complementary to our paper are Bockerman et al. (2019) and Heyman (2016),who use samples of matched employee-employer data for Finland and Sweden, respec-tively, and focus on explaining within-firm organization. In contrast, our data cover everyfirm and employee in the French private sector, and we highlight the paramount impor-tance of changes in firm composition and study what drives these changes—in additionto studying changes in within-firm organization.

A key objective of our paper is to identify causal effects of technology and tradeon firms’ occupational composition and size. Our identification strategy relies on ini-tial conditions across firms to explain changes in occupational composition and size.This strategy is similar to that of Beaudry et al. (2010) and Autor and Dorn (2013),who exploit variation across space and use lagged initial conditions as instruments thathelp identify the propensity of local labor markets to respond to technological change.Michaels et al. (2014) estimate long differences specifications and exploit variation acrossindustries (within countries), and instrument for differences in ICT intensity by usinginitial conditions in the United States.

Our results on the importance of firm composition is related to a recent literature thatstudies the relationship between the growth of so-called “superstar firms” and declinesin the labor share. The role of firm composition is illustrated in Autor et al. (2020), whoargue that when superstar firms, which are larger and more capital intensive to begin with,increase market and employment shares, then aggregate labor payments fall. We find apositive covariance between firm employment growth and techie intensity, suggestingthat techie-intensive firms grow much faster than other firms. This is consistent withincreasing competitiveness for these firms as they gain more from reductions in the costand increases in availability of ICT.

6

3 Data

We use detailed panel data on firms in the French private sector economy in 1994–2007. Our empirical analysis proceeds in two steps. First, we describe the pattern ofjob polarization in France, then we provide an econometric analysis of the drivers ofpolarization. These steps require merging detailed firm-level information on employmentby occupation to firm-level information on trade. The matching process is straightforwardbecause firms in France are identified by the same identification number (called SIREN),which can be followed across years in both data sets.5 This section gives an overview ofour data sources that will be use in subsequent analyses.

3.1 Data on workers

Our source for information on workers is the DADS Poste, which is based on mandatoryannual reports filed by all firms with employees, so our data includes all private sectorFrench workers except the self-employed.6 Our unit of analysis is annual hours paid ina firm, by occupation. For each worker, the DADS reports gross and net wages, hourspaid, and the two-digit occupation of each workers.

Every job is categorized by a two-digit PCS occupation code.7 Excluding agriculturaland public sector categories, the PCS has 22 two-digit occupational categories. For ouranalysis, we aggregate these 22 codes into seven broader categories, which are the boldface headings in Table 1. Table 1 lists the constituent 2-digit codes of each heading,the share of hours and the relative wage of occupations in 1994, and growth in eachoccupations shares of hours from 1994 to 2007.

– Insert Table 1 about here –

In order to aggregate occupations, we base our judgment on the routine intensity ofan occupation on the INSEE documentation for occupations coded in the DADS dataset,which provides a full description of each occupation. For example, an office worker“performs a set of activities related to administrative tasks, copyediting, typesetting orinformation transcription, control of administrative operations or linked to the receptionof clients”. One needs to be more cautious when associating a type of task with anoccupation. We do our best to aggregate occupations within broad categories that are(i) meaningful in terms of tasks and jobs that they unite and (ii) are in similar parts ofthe wage distribution in 1994. Some occupations are in different parts of the wage distri-bution in 1994. The deviation is however small enough so that we still find employmentpolarization with 22 occupations as we will see later on.

Techies are central to our research. Techie occupations combine two of the 2-digitoccupations: “technical managers and engineers” and “technicians”. As is clear fromthe detailed descriptions in Table 2, many workers in these categories are closely con-nected with the installation, management, maintenance, and support of information and

5The data is reported at the level of establishments, which are identified by their SIRET. The firstnine digits of each SIRET is the firm-level SIREN, which makes it easy to aggregate across establishmentsfor each firm.

6The DADS Poste is an INSEE database compiled from the mandatory firm-level DADS (DeclarationAnnuelle des Donnees Sociales) reports.

7PCS stands for Professions et Categories Socioprofessionnelles.

7

communications technology (ICT), and even if they do not work with ICT these are jobsthat require STEM and technical training, skill, and experience.8

– Insert Table 2 about here –

As we will outline in the theory section 5 below, techies mediate the effects of newtechnology within firms: they are the ones who plan, purchase, and install new ICTequipment, and who train and support other workers in the use of ICT. Inspection ofTable 2 supports this argument, although the table also makes it clear that not all techieworkers necessarily work primarily with ICT. In a nutshell, if a firm invests in ICT, itneeds techies, and firms with more techies are probably more technologically sophisticatedfirms.

Each two-digit PCS category is an aggregate of as many as 40 four digit subcategories.Although hours data is not available by four-digit category, the descriptions of the fourdigit categories in Table 2 are helpful in understanding the kinds of tasks performedwithin two-digit categories. The subcategories listed in Table 1 also suggest differences inthe susceptibility of jobs to automation and/or offshoring. For example, service workerssuch as restaurant servers, hair stylists, and child care providers do the sort of “non-routine manual” tasks (c.f. Autor et al. (2003)) that require both proximity and humaninteraction. The same can be said for both skilled and unskilled manual laborers, whosejobs include gardening, cooking, repair, building trades, and cleaning. In contrast, mid-level managers and professionals often do routine cognitive tasks that can be done morecheaply by computers or overseas workers. Skilled and unskilled industrial workers doingroutine manual work are unquestionably directly in competition with both robots andimported intermediate goods.

One potential problem with our hypothesis that firm-level techies are an indicator offirm-level technological sophistication is that firms can purchase ICT consulting services.By hiring a consultant, firms can obtain and service new ICT without increasing theirpermanent staff of techies. However, the vast majority of techies work outside the ITconsulting sector (95% in 1994 and 91% in 2007), which means that almost all techieservices are provided by in-house staff rather than hired consultants.9

Techies are far from equally dispersed over all sectors. They are more prevalent in theIT consulting sector—but not exclusively concentrated there. To see this, we computethe following ratio which divides techie employment shares across two-digit sectors byoverall employment shares: (

techies in industry itotal techies

)(employment in industry i

total employment

)This ratio is about 4.4 in the IT consulting sector in 2007, implying that techies are 4.4times more prevalent in this sector compared to the overall representation of the sector

8The detailed data on the 4-digit occupations codes described in Table 2 are available in the DADSdataset from 2009 only. As our sample ends in 2007, we cannot include these numbers.

9The IT consulting sector is a subset of sectors 62 and 63 of the NAF Rev. 2 classification andcomprises of “Consulting in computer systems and software” and “Third party maintenance of computersystems and applications”. These numbers are somewhat smaller if we use the corresponding sector 72in the NAF Rev. 1 classification. The shares of engineers or technicians in the IT consulting sector aresimilar over the sample period.

8

in aggregate employment. The ratio is stable from 1994 to 2007. Outside of the ITconsulting sector the same ratio in 2007 is on average 1.17, with a standard deviation of1.19.10 This implies that overall, techies are widely dispersed across industries, in a waythat is broadly commensurate with the size of the sector in which they are employed.11

3.2 Data on firm-level trade.

The DADS Poste dataset has no information about the firm beyond the firm identifier andindustry and, implicitly, firm-level aggregates related to employment such as total hoursby occupation, average wages, etc. Our source for firm-level trade data is the FrenchCustoms. For each trade flow observation, we use the identity of the French importing orexporting firm and construct an indicator of the firm status into exporting and importing.We merge the customs data into the DADS Poste database, and keep all non-matchedfirms. Firms that are present in the DADS and that are not matched with customs dataare assumed to have zero exports and imports. The customs data indeed covers tradablegoods, and both exporting and importing firms are mostly in manufacturing and retailers.The share of internationally engaged firms, who either export, import or do both, in thesample is about 8.6% in 1994. These firms account for 49.9% of total hours paid.

From 1994 to 2007, roughly 2.9 million private sector firms appear in the DADS Postedata. Our descriptive analysis includes all 2.9 million firms, but in our econometric analy-sis we focus on the subset of firms that were in operation continuously from 1994 to 2007.There are 297,402 of these “permanent” firms, of whom 85% are in non-manufacturing.The permanent firms represent about a quarter of firms and half of hours paid in oursample in each year.

4 Facts

France has experienced rapid structural change across sectors over the last few decades,which can be seen in the sharp increase in the share of nonmanufacturing (mainly services)employment from 72 to 78 percent between 1994 and 2007. The share of hours employedin the non-manufacturing sector (mostly services) within the private sector grew from analready high share of 72 percent in 1994 to 78 percent in 2007. This transition is notthe subject of our paper, but a potential explanation is that as productivity growth inthe manufacturing sector is much faster than in non-manufacturing, while households’demand for manufactured goods relative to services remains relatively constant, rela-tively fewer workers are needed in manufacturing (Baumol (1967)).12 The evolution ofemployment within the non-manufacturing sector accounts for most of the economy andthe employment composition at aggregate level.

10The average can be greater than 1 because it is an unweighted average across industries.11This does not address the issue that the techie “treatment effect” may be different in the IT consulting

sector versus other sectors. In order to address this, we estimated all our regressions without the sectors62-63. The estimates from these regressions are virtually the same as the main regressions. This isbecause we always control for industry fixed effects and weight our regressions by employment. Sectors62-63 account for only 0.43% of firms in our regression sample, or 1.15% of employment.

12This explanation is consistent with findings in Buera and Kaboski (2009) and in Reshef (2013).

9

In this section, we show how the French job market polarized between 1994 and 2007.We show that polarization was accompanied by the March of the Techies : the growingimportance of technologically-oriented occupations that require STEM skills. We alsoshow that polarization occurred both between and within firms, and that the betweencomponent was much larger than the within component.

4.1 Occupational polarization and the March of the Techies

The French occupational structure polarized between 1994 and 2007, with high-wageand low-wage occupation shares growing at the expense of middle-wage occupations.Polarization is shown in Figure 1, which illustrates the changes in each broad occupation’sshare of hours worked that are reported in Table 1. Figure ?? in Appendix visualizes thesame pattern of polarization using 22 occupations confirming the results based on broadcategories.

As in Table 1, occupations are ranked by average wage, and the width of the bars isproportional to each occupation’s employment share in 1994.

– Insert Figure 1 about here –

Figure 1 shows that the share of hours by upper managers and especially engineers andtechnicians grew substantially, while the share worked by blue and white collar workersfell. In sharp contrast, low-wage personal service jobs grew. This pattern is driven bythe nonmanufacturing sector (about 80 percent of employment), as shown in Figure 2.The pattern is somewhat different in manufacturing. As Figure 3 shows, employment inmanufacturing was characterized by skill upgrading, with lower-skilled and office workersreplaced by higher-skilled workers, along with a huge increase in the share of techniciansand engineers.

– Insert Figures 2 and 3 about here –

Particularly striking in Figures 1 through 3 is the strong growth in the techie occu-pations. The share of techie hours increased from 9.04% to 12.21% of total hours from1994 to 2007.13 We call this phenomenon The March of the Techies and illustrate it inFigure 4. Techie growth was steady over this period, with most of the growth since thelate 1990s accounted for by the higher-paid and more-educated Technical Managers andEngineers category within techies.

4.2 Decompositions of aggregate trends

To better understand the aggregate patterns, we decompose changes in hours shares intochanges within economic units and changes in composition. This allows us to evaluate,for example, whether aggregate changes occur due to within-firm adjustments, changes

13The share of techie hours in non-manufacturing increase from 58% to 64% between 1994 and2007. The share of techie hours in manufacturing increase from roughly 12.7% to 20% while in non-manufacturing it increased from roughly 7% to 10%. The rate of increase of techie employment inmanufacturing and non-manufacturing is very similar, but the level is higher in manufacturing. Thisimplies that the relative expansion of non-manufacturing does not account for the increase in techieemployment growth (in fact, it reduces it).

10

in firm size composition, or both. Thus, the decompositions help identify mechanismsthat underlie the aggregate trends.

The change in occupation o’s share of aggregate hours over some period can be de-composed as follows

∆So =∑i

∆λisio︸ ︷︷ ︸+

between

∑i

λi∆sio︸ ︷︷ ︸within

, (1)

where i indexes different partitions of the aggregate economy into distinct units: indus-tries, firms, geographic units (departments or urban/rural distinctions) and gender. Here∆λi is the change in the hours share of unit i in the aggregate, sio is the average share ofoccupation o within unit i, λi is the average hours share of unit i, and ∆sio is the changein the share of occupation o within unit i. The between component tells how much of theaggregate change is due to changes in composition, holding the within-unit occupationalshares fixed. The within component tells how much of the aggregate change is due tochanges in within-unit occupational shares, holding fixed the composition. At the level offirm for instance, the shift-share analyses examine whether firms that always had a highfraction of workers in an occupation grew more (between firm component), or whetherfirms started to hire more workers of that occupation over time (within firm component).

In the case of firms, we take into account net entry and exit in (1) as follows

∆So =∑i∈P

∆λisio︸ ︷︷ ︸+

between, perm

∑i∈P

λi∆sio︸ ︷︷ ︸within, perm

+∑i∈E

λi2si2︸ ︷︷ ︸entry

−∑i∈X

λi1si1︸ ︷︷ ︸exit

=∑i∈P

∆λisio︸ ︷︷ ︸+

between, perm

∑i∈P

λi∆sio︸ ︷︷ ︸within, perm

+ net entry , (2)

where P denotes the set of permanent firms, i.e., those that are present in both the firstand last period of the decomposition, E denotes the set of firms that enter, i.e. are notpresent in the first period but are present in the last, and X denotes the set of firms thatexit, i.e. are present in the first period but are not present in the last.14

We also use (1) in order to evaluate the contribution of gender and urban versus ruralareas to overall change as follows

Contributionoi = ∆λisio + λi∆sio , (3)

where∑

iContributionoi = ∆So. For example, in the case that ∆So > 0 urban areas (i.e.,

all i’s that are urban) can contribute to aggregate changes either by virtue of having arelatively high occupational share (sio) and growing disproportionately (∆λi), or by largeincreases of the share of occupation o within urban areas (∆sio) given their relative size(λi). While equation (3) is feasible with any number of units, considering two specificunits is particularly informative.

14As it is defined, the set P may include firms that report positive hours in 1994 and in 2007, but nohours in one or more of the intervening years. We exclude a tiny number of firms of this type from ourdata. These firms account for a negligible share of aggregate hours.

11

4.3 Industries versus firms

We start by juxtaposing decompositions for industries with decompositions for firms.Table 3 reports the decompositions of ∆So for all seven occupational groups accordingto equation (1) in 1994–2007 for industries, firms, and then further breaking down firmsaccording to equation (2). In the last row we report the weighted average contributionof each component to changes:

∑o λo,1994×betweeno/∆So and

∑o λo,1994×withino/∆So,

where λo,1994 are aggregate hours shares of occupation o in 1994.The first message from Table 3 is that the lion’s share of the change in aggregate

occupational shares occurs within industries, not due to industry composition. In alloccupations but “Other white collar”, the within component is larger than the betweencomponent. The within component is more than all of the aggregate change for Managers,Techies and Service workers, implying that composition works in the opposite directionof the aggregate change. The weighted average contribution of the within component is91%.

The second message from Table 3 is that in contrast to industries, most of the variationin aggregate occupational shares is due to firm composition, rather than changes withinfirms. For six out of seven occupations the between-firm component is much larger thanthe within component. For example, of the overall 3.17 percent point increase in the shareof techies, four-fifths came from between-firm adjustments while only one-fifth came fromincreases in the within-firm share of techies. This finding suggests that firms with above-average techie shares grew much faster than average. The pattern is particularly starkfor office workers: virtually all of the 3.15 percent point drop in office workers’ shareof employment came from slower growth of firms with above-average shares of officeworkers. The only exception to this pattern is in the lowest paid and fast-growing servicesoccupations, where the between and within split is roughly equal: firms with lots of serviceworkers grew faster than average, while at the same time firms were on average increasingtheir employment of these low-wage occupations. The weighted average contribution ofthe between firm component is 125%, implying that, on average, within-firm adjustmentswork in the opposite direction of aggregate changes.

The third part of Table 3 reports within and between components for permanentfirms (those in set P ) in addition to their overall contribution (within, perm + between,perm) as in (2). Thus, the third message from Table 3 is that the permanent firm sampleexhibits similar patterns for decompositions as the overall sample, with few significantdifferences. Overall, the pattern of polarization within permanent firms is similar to theaggregate, with one difference, which is the share of Unskilled workers increases, ratherthan decreases. However, this class of workers is second to last in terms of wage ranks,so polarization is still apparent. The weighted average contribution of permanent firmsis 0.42 of the aggregate changes. Since the set of P firms account for almost half ofemployment in the sample, 0.5 is the relevant reference point for these contributions.The residual 0.58 is accounted for by net entry. As in the aggregate sample, the betweenfirm component accounts for more than the entire change, while within firm changes workin the opposite direction.

We also correlate the overall between and within components with the correspondingcontributions of the permanent sample firms. The correlation coefficient is .99 (both forthe seven aggregate occupation classes, and for the more detailed 22 PCS codes). The

12

regression of the aggregate change on the change for permanent sample firms gives acoefficient of 1.087 with a t-statistic of 21.56. Thus, while there are specific differencesin the evolutions in shares comparing overall and permanent firms, they are not veryimportant on average.

The dominance of the between-firm channel in accounting for occupational changemeans that an explanation for polarization must explain why some firms grew faster thanothers. We return to this in Section 7 below after describing our estimation procedure.

4.4 Geography

We now turn to decompositions of aggregate occupational hours share by geography. Westart by reporting in Table 4 decompositions of ∆So according to (1) in 1994–2007 for94 French departments.15 Here the message is very clear: aggregate changes in occu-pational shares occur almost entirely within departments. Except for Managers, wherecomposition accounts for -43% of the change in the aggregate hours share (i.e., works inthe opposite direction), the contribution of composition is small. The weighted averagecontribution of the within departments components is 106%.

We then aggregate departments into two groups—22 urban departments and 62 ruraldepartments – according to the definition of the French National Institute of Statistics(INSEE). Urban departments are those where the majority of the population lives inan urban area of 200,000 inhabitants or more. The 22 urban departments account for55.1% of aggregate hours in 1994, and this share is remarkably stable, increasing slightlyto 55.7 in 2007.16 Therefore, it is not surprising that, similar to the decomposition bydepartments, here too we find that virtually all of the aggregate changes come from withinurban and rural areas, not from changes in urban-rural composition of employment.

We then turn to evaluate the separate contributions of urban versus rural areas tochanges in aggregate occupational hours shares according to equation (3). On average,urban areas contribute virtually all of the aggregate changes, 95%. However, this averagemasks much heterogeneity across occupations. For example, urban areas contribute morethan the entire increase in the hours share of Managers, whereas rural areas see declinesin employment shares of managers.17 Rural areas contribute roughly three quarters ofthe aggregate changes for Techies, Other white collar, and for Office workers. In contrast,rural areas contribute two-thirds of the decline in Unskilled workers. Overall, job polar-ization is more of an urban phenomenon, as found by Autor (2019) for the U.S., whereasthe pattern of change for occupational shares in rural areas is less clear.

4.5 Gender

Finally, we turn to decompositions of aggregate occupational hours share by gender.Table 5 reports decompositions of ∆So according to (1) and the separate contributions of

15There are 95 departments in “Metropolitan France”, i.e., excluding overseas departments. We dropCorsica because of reporting discrepancies over time.

16This stability masks some increase in urban density within departments. Appendix Table ?? liststhe urban departments, together with their respective hours shares in 1994.

17The increased concentration of managers in cities is reminiscent of the finding of Duranton and Puga(2005).

13

women versus men according to (3). As with the urban-rural split, here too we find thatvirtually all of the changes occur within groups, with relatively little variation. In only oneoccupation, Skilled Workers, is the between-gender component of the change large (abouta quarter of the total change). This is because men are much more prevalent in this typeof occupation (notably in industrial and construction jobs, drivers, and transport jobs)compared to women, and the employment composition shifts slightly towards women,from 34.7% in 1994 to 35.6% in 2007 of total hours worked in France. The weightedaverage contribution of the within gender component is 96%.

The contribution of each sex is, however, relatively even, where women on averageaccount for 47% of aggregate changes and the remaining 53% is accounted for by men.However, the average masks significant variation by occupation. Since the aggregatehours share of women increases only slightly, the relative contributions by gender foreach occupation are driven by differential changes within each sex. Most notably, womenalone account for more than the entire increase in Managers, increasing their employmentshare in managerial jobs by 0.9 percent points, whereas men see their share decline by0.53 percent points. Women also increase their employment share in mid-level managerialjobs, where the overall decline is completely accounted for by men. Men account for mostof the increase in Techies, Office workers and Skilled workers. Overall, the pattern ofpolarization is clearly apparent for women, but less so for men due to the large declinein employment in top managerial occupations.

5 Theoretical Framework

Our hypothesis is that techies are a channel through which falling ICT prices cause polar-ization. In the Online Appendix, we develop a simple model of firm-level outcomes thatillustrates this. The model shows how a drop in the price of ICT leads to greater em-ployment growth in ICT-intensive firms and to polarization of employment within a firm.These results help motivate our between- and within-firm descriptive and econometricanalyses below. Here we describe the economic logic of the model, and refer interestedreaders to the details in the Online Appendix.

Assume that firms use three types of workers (high, medium and low skilled) alongwith ICT capital to produce output. In addition, the firm hires techies to manage andmaintain the stock of ICT capital. Crucially, suppose that skilled workers are comple-ments to ICT, while ICT substitutes for medium skilled workers, and that techies are anecessary input that is required for productively using ICT.

How does technological progress affect firms’ employment? We model technologicalprogress as a drop in the cost of ICT capital, or equivalently in the cost of the productiveservices provided by ICT. This price drop is common across firms, but the effects differdepending on how intensive firms are in their use of ICT. Since ICT and techies arestrongly complementary, ICT-intensive firms are identified by their techie intensity.

There are two implications of such technological progress. First, firms’ overall costswill fall, and the drop in costs will be larger for firms who use more ICT and employmore techies. This competitiveness effect implies that techie-intensive firms will gainmarket share, boosting their employment relative to less techie-intensive firms. Thesecond implication is that firms will increase their employment of skilled relative to both

14

medium and unskilled workers. The negative effect is stronger for medium skilled workers,for whom ICT is a substitute. Thus, cheaper ICT capital causes polarization within thefirm: employment of both the highest and lowest skilled workers increases relative tomedium-skilled workers. This effect will be stronger in firms who are more reliant onICT and have greater techie intensity.

Putting these two effects together implies that technological progress will lead to po-larization of overall labor demand through both a between-firm "competitiveness" channeland a within-firm "substitution" channel. We investigate the importance of both channelsin our empirical analysis in the following sections.

Turning to the effects of globalization, Feenstra and Hanson (1996) were the first toshow how purchases of imported intermediates (offshoring) can affect the skill compositionof employment. More recently, Acemoglu and Autor (2011) show how offshoring cancontribute to firm-level polarization. These and related analyses make the point thatoffshoring has competing effects on total firm-level employment: a direct substitutioneffect of imported intermediates for workers within the firm, and a cost-reducing effectthat can raise demand for workers whose jobs are not offshored.18 We estimate thesepolarization and net employment effects in our econometric analysis below.

A large empirical literature, reviewed by Bernard et al. (2007), finds that exporting isassociated with higher skill intensity in cross-sections of firms. We know of no theoreticalor empirical analysis of the link between exporting and firm-level employment, perhapsbecause the first-order effect is too obvious: firms that also sell abroad will tend to havegreater labor demand than those who sell only at home. In our econometric analysisbelow we look for such firm-level employment effects both within and between Frenchfirms.

6 Econometric Framework

The theoretical framework implies that firms’ responses to economy wide trends in re-ductions in barriers to trade and to reductions in ICT prices will depend on their char-acteristics. In this section, we develop an econometric framework for estimating howfirms’ global engagement and technology intensity cause changes in employment growthand within-firm changes in occupational composition. In particular, how do predeter-mined firm differences in the propensity to trade and adopt technology lead firms tochange their size and employment mix over time? We answer these questions by regress-ing changes in firm-level employment and occupational shares from on initial levels oftrade and techies. In some specifications we address possible endogeneity of the initiallevels by using long lags as instrumental variables. We always control for 2-digit industryspecific effects, which absorb industry-wide trends such as import competition, as wellas industry-wide average levels of control variables. Thus, our identification comes fromcross-firm variation within industries in initial conditions. We report separate regressionsfor manufacturing and for non-manufacturing, since we expect firm level responses to bedifferent across these two sectors of the economy.

18This is related to the wage effects studied in Grossman and Rossi-Hansberg (2008).

15

6.1 Firm employment growth equation

Optimal firm-level employment depends on both demand and cost conditions. We specifyoptimal log employment for firm f in year t, lnhft, as

lnhft = βf +D (Xf ) · t+Wftγ + εft . (4)

Here βf a firm-specific intercept. The firm specific trend D (Xf ) · t is determined by firm-specific values Xf . This trend absorbs how different firms respond to common, aggregatetrends, for example, the drop in the price of ICT. The corresponding element of Xf isour proxy for θ, the firm-specific techie share. In this case D (Xf ) captures how firmswith different θ (techie intensity) respond differently to the drop in ICT prices. Theeffect of time-varying, firm-specific characteristics is given by Wftγ. Here Wft includesfirm characteristics which we cannot measure, such as capital, intermediate inputs, anddemand shocks. First-differencing (4) gives

∆ lnhf = D (Xf ) + uf , (5)

where uf = ∆Wftγ + ∆εf is a composite error term that includes changes in the firmcharacteristics and changes in the error term εf .

We model the firm-specific time trend D (Xf ) as a function of the initial level of thetechie share and trade status at the beginning of the interval over which first differencesare taken. Firms that do not trade at all, and/or that have no techies at all, are likely tobe distinctly different from firms that do trade and/or have techies, so to accommodatethis we include in Xf indicators for positive values of techies and trade. We also entertaingradual and non-linear effects of techies, by constructing indicators for three terciles oftechie intensity over the positive support (where no techies is the omitted category). Insome specifications we also separate techies into engineers (PCS code 38) and technicians(PCS code 47). To control for the well-established fact that firm growth rates declinewith size, we also include the initial level of employment hf0 as a regressor. Finally, weinclude industry i fixed effects βi to absorb industry trends.

Let techiesft be the share of techies in firm f ’s total hours worked in period t andtechposft be an indicator equal to one if techiesft > 0. Let expposft and impposft beindicators equal to one if firm f exports or imports in period t, respectively. The basicequation to be estimated is the empirical implementation of (5)

∆ lnhf = βi + β1techposf0 + β2expposf0 + β3impposf0 + β4hf0 + uf , (6)

where βi +Xf0β summarizes the D (Xf ) function, and Xf = β1techposf0 + β2expposf0 +β3impposf0. This specification captures the theoretical prediction that in the face offalling ICT prices, employment will grow more in firms that have higher initial levels oftechies, as shown in section 5 above. Similarly, a firm that exports final goods or purchasesimported inputs will be more affected by the increased integration of Eastern Europe,China, and India into the world economy than will a firm that does not trade. Thus,equation (6) allows us to estimate the heterogeneous effect of aggregate trends on firmoutcomes, where the heterogeneity is captured by firm characteristics in the initial period.With industry fixed effects βi, the parameters of interest are identified by variation across

16

firms within industries in the levels of techies, trade, and employment.19

We estimate equation (6) by weighted least squares (WLS) on a single 14-year dif-ference from 1994 to 2007, so that for example, techposf0 =techposf,1994. Because (6) isa long differences specification, it is not informative about the speed of adjustment ofemployment or possible lags in the response to drops in economywide ICT prices or thegrowth of globalization. We also estimate versions of (6) on changes in 2002–2007 byweighted two stage least squares (W2SLS), using 1994 values to instrument for initiallevels. Accordingly, when estimating by W2SLS techposf0 =techposf,2002 in (6), and theinstrument is techposf,1994, and similarly for expposf0 and impposf0. The choice to basethe changes for W2SLS in 2002–2007 is done in order to allow a sufficiently long lag forthe instruments to be considered plausibly exogenous and excludable from (6). We weightall regressions using 1994 firm hours worked. This ensures that the estimated regressionscan be interpreted as conditional expectations across the distribution of hours worked,which is what is relevant for our research question. In particular, this allows for strongerinfluence of larger firms on the estimator.

6.2 Within-firm organization equation

In order to analyze the changes in the occupational structure within French firms, wereplace the dependent variable in (6) by changes in the share of hours employed in eachoccupation, where the denominator of the share excludes techies, in order to avoid acompletely mechanical relationship with initial levels of techies. The approach here isvery similar to the approach in Section 6.1. The firm-occupation outcome measure ofinterest is ∆sfo, the change in the ex-techie share of hours of the six large non-techieoccupations listed in Table 1. For each occupation o our estimating equation is thus

∆sfo = βoi + βo

1techposf0 + βo2expposf0 + βo

3impposf0 + uof . (7)

As with (6), we estimate (7) by WLS in 1994–2007 and by W2SLS in 2002–2007 using1994 levels as instruments for levels in 2002.

7 Econometric Results

7.1 Firm-level employment growth regressions

The theoretical framework of Section 5 provides predictions on the effect of firms’ initiallevel of technology intensity and global engagement on their employment growth rate. Weinvestigate these predictions by estimating the empirical model (6). The estimator usesvariation across firms within the same industry. Any industry-level characteristic, includ-ing demand, import competition, sector-specific trends in ICT prices, etc., is absorbedby the inclusion of industry fixed effects.

19We reported above that the share of techies working in the IT consulting sector out of all techiesrises from 4.7% in 1994 to 9% in 2007. This implies that we may mis-measure the techie labor servicesindicator for a non-negligible number of firms that do not employ techies, but outsource these services.This mis-measurement biases the estimator of the effects of techies towards zero, and makes it harder toinfer large impacts (either positive or negative), and our estimates can be thought of as lower bounds.

17

We report the results of the WLS specifications in Table 6. We regress annualizedgrowth rates in 1994-2007 on initial levels in 1994. In those regressions, we use eitherindicator variables for techie employment (overall techies, engineers and technicians), orindicators for terciles of techie employment intensity on the positive part of the support(the omitted category remains no techie employment). We also estimate specificationsthat split techies into engineers and technicians.

– Insert Table 6 about here –

Columns (1) to (3) of Table 6 report the effect of techies, exports and imports on theannualized growth rate of firm-level hours in the manufacturing sector. Columns (4) to(6) report the results for the non-manufacturing sector.20 Columns (1) and (4) show thatfirms with a positive techie share in 1994 saw significantly faster employment growth thanfirms without techies. The presence of techies involves a sizeable acceleration of firm-levelemployment growth, on the order of 3.5 percentage points in manufacturing industries and1.1 percentage points in non-manufacturing sectors per year. This result is consistent withthe theoretical prediction in Section 5, where falling ICT prices raise the competitivenessof firms that employ more techies. For manufacturing firms, exporters grew more rapidlywhile importers grew more slowly, though these effects are not statistically significant.

Columns (2) and (5) display results where we allow for differential effects of techieintensity within firms. The specifications include three indicators for terciles of techieintensity over the positive support. The coefficients on the second and third tercile in-dicators are statistically significant in manufacturing, and the third tercile indicator issomewhat larger. In non-manufacturing industries all tercile coefficients are statisticallysignificant, and the third is clearly larger than the first and second. In both manufac-turing and non-manufacturing samples, the joint tests of equality between terciles rejectthe hypothesis that they are equal. These results suggest that firms that are the mostintensive in techies have a faster employment growth than firms that have smaller techieshares.

The results so far suggest that firm-level employment growth is faster in techie-intensive firms. The effects of techies may be different across different group of techieworkers, highly-educated technical managers and engineers versus lower-ranked techni-cians. In columns (3) and (6), we exploit the distinction between engineers and techniciansin our data to contrast their effects on firm-level employment growth. We find a strongand positive effect of engineers on non-manufacturing firms’ employment growth. Non-manufacturing firms with a positive presence of engineers increased their employmentgrowth by about 1.3 percentage point compared to firms without techies, but there is vir-tually no effect of employment of technicians.21 Given the size of the non-manufacturingsector, the effect is economically large. In manufacturing, the data do not allow us to sep-arately estimate the effect of engineers from that of technicians because their employment

20To address the issue that the techie treatment effect may be different in digital sectors versus others,we estimated all our regressions without these sectors (NAF sectors 62 and 63, Rev. 2 classification).The estimates from these regressions are virtually the same as the main regressions. This is because wealways control for industry fixed effects and weight our regressions by employment. Moreover, sectors62-63 account for only 0.43% of firms in our regression sample, or 1.15% of employment.

21This somewhat surprising result is not because there are no technicians in non-manufacturing: theirshare of hours is roughly the same as those of engineers in 1994.

18

is too strongly correlated across firms, causing the standard errors to be very big. Bothengineers and technician have a positive effect on employment growth but the effects arestatistically insignificant at conventional levels.

In Table 7 below, we report results using just-identified W2SLS regressions of changesin 2007-2007 on levels in 2002, using 1994 lags as instruments. We report the Kleibergen-Paap F tests which yield values larger than 20. This suggest that our estimations areunlikely to suffer from weak instruments bias. In these regressions we do not estimatemodels with nonlinear effects, and focus only on indicators for techies and trade. Wepresent the W2SLS regressions alongside WLS estimate of the same specification to com-pare the results.

– Insert Table 7 about here –

Columns (1-4) report the results for the manufacturing sample while Columns (5-8)concern the non-manufacturing sample. The OLS estimates suggest a strong and positiveimpact of techies on employment growth over the period 2002-2007. We find that thepresence of techies involves an increase of firm-level employment growth of 1.8 percentagepoints in manufacturing industries and 1.2 percentage points in non-manufacturing sec-tors per year. The impact of techies on employment growth is smaller for manufacturingduring the shorter period from 2002 to 2007 compared to the longer period starting in1994, but it is very similar for non-manufacturing sectors. Columns (3) and (7) repro-duce the specifications in Columns (3) and (6) of Table 6 on the shorter sample. Wenow find positive and statistically significant effects of both engineers and technicians onemployment growth of manufacturing firms. The results for non-manufacturing firms aresimilar to the longer sample, with a positive and statistically significant effects of engi-neers on non-manufacturing firms employment growth. Overall, our findings are broadlyconsistent across periods.

The techie estimates from just-identified W2SLS models are presented in Columns(2), (4), (6) and (8). The results confirm that techies have a strong effect on manufac-turing employment growth of 4.6 percent per year. The W2SLS regressions indicate thatengineers are associated with manufacturing firms’ employment growth much more thantechnicians. Concerning the non-manufacturing sector, we find that the main effect oftechies drops slightly, from 1.1 to 0.9 percent per year, but we lose precision because stan-dard errors increase so much that these effects on employment growth are not statisticallysignificant.

7.2 Changes in the occupational structure of firms

We next examine the changes in the occupational structure within French firms, and inparticular the job polarization that was documented in Section 4. Our hypothesis is thatboth the techie intensity and the global engagement of firms are important factors, andthat firms responses to economy wide changes will depend on these characteristics. Wemeasure changes in a firm’s occupational structure by changes in the share of hours ineach of six PCS occupations, excluding the share of techies. As with the employmentgrowth regression, we compute “ex-techie” shares of employment and ask whether theseare associated with the exporter and importer status of the firm as well as the presenceof techies in the firm. Our estimation approach here is very similar to the approach

19

described above. Each regression at occupational level includes a set of 2-digit sectorfixed effects. The regression coefficients are identified by cross-firm variation in changeswithin 2-digit sectors.

We discuss our results in detail in Appendix D, and summarize them briefly here.We find that the presence of techies in 1994 is followed by within-firm skill upgrading,with an increase in the share of top and middle managers and a decrease in the shareof the three lowest paid occupations. Turning to the effects of international trade, bothimporting and exporting in 1994 are associated with a subsequent increase in the topmanager share.

8 Conclusions

In this paper we use administrative employee-firm-level data to show that the labormarket in France polarized between 1994 and 2007: employment shares of high and lowwage occupations have grown, while middle wage occupations have shrunk. This hasprofound implications for inequality. We show that job polarization occurs mainly withinindustries, through changes in firm sizes rather than through within-firm adjustment.The importance of firm size reallocations for polarization implies that simple theories ofsubstitution across workers within firms miss an important margin of adjustment. Thismargin is changes in competitiveness, where more tech-savvy firms gain market share atthe expense of less tech-savvy firms.

We develop a stylized model which illustrates that variation in initial levels of ICTintensity predict differential responses of firms to the drop in the price of ICT capi-tal. Motivated by the fact that technology adoption is mediated by technically qualifiedmanagers and technicians, we develop a novel measure of the propensity to adopt newtechnology: the firm-level employment share of techies. Firm-level techie intensity helpspredict firm-level outcomes that are consistent with the effects of falling ICT prices andthe growing importance of STEM skills generally, in particular through the competitive-ness margin. As techies become more numerous in the economy, we predict that theseeffects will be commensurately more widespread.

We develop an empirical framework that allows us to study the effect of techie inten-sity on both between-firm and within-firm employment changes. We show that firms withmore techies in 1994 saw substantially faster employment growth in 1994–2007. The ef-fect is pervasive across sectors and is mostly driven by engineers rather than technicians.Our empirical analysis is consistent with technological change improving the competi-tiveness of these firms relative to other firms with fewer techies. Concerning the impactof techies on within-firm occupational changes, we find that techies are associated withskill upgrading, and particularly an increase in the employment share of managers.

20

References

Acemoglu, D. and D. Autor (2011). Skills, tasks and technologies: Implications foremployment and earnings. In O. Ashenfelter and D. Card (Eds.), Hanbook of LaborEconomics, Volume 4b. North Holland.

Autor, D. H. (2019). Aea papers and proceedings. AEA Papers and Proceedings 109,1–32.

Autor, D. H. and D. Dorn (2013). The growth of low-skill service jobs and the polarizationof the us labor market. The American Economic Review 103 (5), 1553–1597.

Autor, D. H., D. Dorn, and G. H. Hanson (2013). The china syndrome: Local labormarket effects of import competition in the united states. The American EconomicReview 103 (6), 2121–2168.

Autor, D. H., D. Dorn, and G. H. Hanson (2015). Untangling trade and technology:Evidence from local labour markets. The Economic Journal 125 (584), 621–646.

Autor, D. H., D. Dorn, L. F. Katz, C. Patterson, and J. Van Reenen (2020). The fallof the labor share and the rise of superstar firms. Quarterly Journal of Economics,forthcoming .

Autor, D. H., L. F. Katz, and M. S. Kearney (2006). The polarization of the u.s. labormarket. The American Economic Review, AEA Papers and Proceedings 96 (2), 189–194.

Autor, D. H., L. F. Katz, and M. S. Kearney (2008). Trends in u.s. wage inequality:Revising the revisionists. The Review of Economics and Statistics 90 (2), 300–323.

Autor, D. H., F. Levy, and R. J. Murnane (2002). Upstairs, downstairs: Computersand skills on two floors of a large bank. Industrial and Labor Relations Review 55 (3),432–447.

Autor, D. H., F. Levy, and R. J. Murnane (2003). The skill content of recent technologicalchange: An empirical exploration. The Quarterly Journal of Economics 118(4), 1279–1333.

Bartel, A., C. Ichniowski, and K. Shaw (2007). How does information technology affectproductivity? plant-level comparisons of product innovation, process improvement,and worker skills. The quarterly journal of Economics 122 (4), 1721–1758.

Barth, E., J. C. Davis, R. B. Freeman, and A. J. Wang (2017). The effects of scientistsand engineers on productivity and earnings at the establishment where they work.NBER Working Paper No. 23484.

Baumol, W. J. (1967). Macroeconomics of unbalanced growth: The anatomy of urbancrisis. The American Economic Review 57 (3), 415–426.

Beaudry, P., M. Doms, and E. Lewis (2010). Should the personal computer be considereda technological revolution? evidence from u.s. metropolitan areas. Journal of PoliticalEconomy 118 (5), 988–1036.

Ben Zeev, N., J. Mokyr, and K. Van Der Beek (2017). Flexible supply of apprenticeshipin the british industrial revolution. The Journal of Economic History 77 (1), 208–250.

Bernard, A., B. Jensen, S. Redding, and P. Schott (2007). Firms in international trade.Journal of Economic Perspectives 21 (3), 105–130.

21

Biscourp, P. and F. Kramarz (2007). Employment, skill structure and international trade:Firm-level evidence for france. Journal of International Economics 72 (1), 22–51.

Blinder, A. S. (2006). Offshoring: The next industrial revolution? Foreign Affairs 85 (2),113–28.

Blinder, A. S. and A. B. Krueger (2013). Alternative measures of offshorability: A surveyapproach. Journal of Labor Economics 31 (2), S97–S128.

Bockerman, P., S. Laaksonen, and J. Vainiomaki (2019). Does ict usage erode routineoccupations at the firm level? LABOUR 33 (1), 26–47.

Bresnahan, T. F., E. Brynjolfsson, and L. M. Hitt (2002). Information technology, work-place organization, and the demand for skilled labor: Firm-level evidence. The Quar-terly Journal of Economics 117 (1), 339–376.

Brynjolfsson, E. and L. Hitt (1996). Paradox lost? firm-level evidence on the returns toinformation systems spending. Management science 42 (4), 541–558.

Brynjolfsson, E. and L. M. Hitt (2003). Computing productivity: Firm-level evidence.Review of economics and statistics 85 (4), 793–808.

Buera, F. J. and J. P. Kaboski (2009). Can traditional theories of structural change fitthe data? Journal of the European Economic Association 7 (2-3), 469–477.

Caliendo, L., F. Monte, and E. Rossi-Hansberg (2015). The anatomy of french productionhierarchies. Journal of Political Economy 123 (4).

Cerina, F., A. Moro, and M. Rendall (2017). The role of gender in employment polariza-tion.

Dunne, T., L. Foster, J. Haltiwanger, and K. Troske (2004). Wage and productivitydispersion in United States manufacturing: The role of computer investment. Journalof Labor Economics 22 (2), 397–429.

Duranton, G. and D. Puga (2005). From sectoral to functional urban specialisation.Journal of urban Economics 57 (2), 343–370.

Dustmann, C., J. Ludsteck, and U. Schonberg (2007). Revisiting the german wage struc-ture. The Quarterly Journal of Economics 124 (2), 843–881.

Feenstra, R. C. and G. H. Hanson (1996). Foreign investment, outsourcing, and relativewages. In G. M. Grossman, R. C. Feenstra, and D. R. Irwin (Eds.), The PoliticalEconomy of Trade Policy, pp. 89–127. Oxford: MIT Press.

Feenstra, R. C. and G. H. Hanson (1999). The impact of outsourcing and high-technologycapital on wages: Estimates for the united states, 1979-1990. The Quarterly Journalof Economics 114 (3), 907–940.

Firpo, S., N. M. Fortin, and T. Lemieux (2011). Occupational tasks and changes in thewage structure. Working Paper, University of British Columbia.

Goos, M. and A. Manning (2007). Lousy and lovely jobs: The rising polarization of workin britain. The Review of Economics and Statistics 89 (1), 118–133.

Goos, M., A. Manning, and A. Salomons (2009). Job polarization in europe. The Amer-ican Economic Review, AEA Papers and Proceedings 99 (2), 58–63.

22

Goos, M., A. Manning, and A. Salomons (2014). Explaining job polarization: Routine-biased technological change and offshoring. The American Economic Review 104 (8),2509–2526.

Graetz, G. and G. Michaels (2018). Robots at work. Review of Economics and Statis-tics 100 (5), 753–768.

Gregory, T., A. Salomons, and U. Zierahn (2016, July). Racing with or against themachine? evidence from europe. Technical report, unpublished.

Grossman, G. M. and E. Rossi-Hansberg (2008). Trading tasks: A simple theory ofoffshoring. The American Economic Review 98 (5), 1978–97.

Helper, S. and J. Kuan (2018). What goes on under the hood? how engineers innovatein the automotive supply chain. In US Engineering in the Global Economy. Universityof Chicago Press.

Heyman, F. (2016). Job polarization, job tasks and the role of firms. Economics let-ters 145, 246–251.

Kelly, M., J. Mokyr, and C. O Grada (2014). Precocious albion: a new interpretation ofthe british industrial revolution. Annu. Rev. Econ. 6 (1), 363–389.

Lichtenberg, F. R. (1995). The output contributions of computer equipment and per-sonnel: A firm-level analysis. Economics of innovation and new technology 3 (3-4),201–218.

Malgouyres, C. (2017). The impact of chinese import competition on the local structureof employment and wages: Evidence from france. Journal of Regional Science 57 (3),411–441.

Maloney, W. F. and F. Valencia-Caicedo (2017). Engineering Growth: Innovative Ca-pacity and Development in the Americas. CESIfo Working Paper No. 6339.

Maurin, E. and D. Thesmar (2004). Changes in the functional structure of firms and thedemand for skill. Journal of labor economics 22 (3), 639–664.

Michaels, G., A. Natraj, and J. Van Reenen (2014). Has ict polarized skill demand?evidence from eleven countries over twenty-five years. Review of Economics and Statis-tics 96 (1), 60–77.

Oesch, D. (2013). Occupational Change in Europe. United Kingdom: Oxford UniversityPress.

Reshef, A. (2013). Is technological change biased towards the unskilled in services? anempirical investigation. Review of Economic Dynamics 16, 312–331.

Rodriguez-Clare, A. (2010). Offshoring in a ricardian world. American Economic Journal:Macroeconomics 2 (2), 227–58.

Romer, P. M. (1990). Endogenous technological change. Journal of Political Econ-omy 98 (5), S71–S102.

Spitz-Oener, A. (2006). Technical change, job tasks and rising educational demands:Looking outside the wage structure. Journal of Labor Economics 24 (2), 235–270.

Tambe, P. and L. M. Hitt (2012). The productivity of information technology investments:New evidence from it labor data. Information Systems Research 23 (3-part-1), 599–617.

Tambe, P. and L. M. Hitt (2014). Job hopping, information technology spillovers, andproductivity growth. Management science 60 (2), 338–355.

23

Table 1: Occupational definitions and employment shares

Employment Shareslevel change relative wage

Occupation 1994 1994-2007 1994Business owners, professionals, managers 9.24 0.37 1.96

Craftsmen 0.70 -0.65 1.32Shopkeepers 0.61 -0.55 1.39Heads of small business 0.75 0.00 2.70Scientific professionals 0.48 -0.04 1.54Creative professionals 0.65 -0.07 1.48Managers 6.04 1.68 2.04

Techies 9.04 3.17 1.59Engineers & technical managers 4.51 2.29 2.04Technicians 4.53 0.87 1.13

Other white collar 4.85 -0.26 1.15Teachers 0.32 0.04 1.05Health and social workers 1.23 0.15 0.95Foremen, supervisors 3.30 -0.46 1.19

Office workers 25.35 -3.15 1.00Office workers, middle-level 12.85 -0.68 1.12Office workers, lower-level 12.50 -2.47 0.84

Skilled workers 27.36 -0.62 0.82Skilled industrial 11.35 -1.14 0.87Skilled manual 8.96 -0.41 0.73Drivers 4.58 0.61 0.74Skilled transport & wholesale 2.48 0.32 0.78

Unskilled workers 13.81 -2.56 0.70Unskilled industrial 9.71 -2.89 0.71Unskilled manual 3.96 0.33 0.61

Services workers 10.49 3.04 0.66Private security 0.69 0.36 0.70Retail 6.30 1.47 0.65Personal services 3.51 1.22 0.63

Note to table: “relative wage” is the occupation’s average wage in 1994 divided by the economy’s median

wage in 1994.

24

Table 2: Techies, representative suboccupations

Technical managers and engineersTechnical managers for large companiesEngineers and R&D managersEletrical, mechanical, materials and chemical engineersPurchasing, planning, quality control, and production managersInformation technology R&D engineers and managersInformation technology support engineers and managersTelecommunications engineers and specialists

TechniciansDesigners of electrical, electronic, and mechanical equipmentR&D technicians, general and ITInstallation and maintenance of non-IT equipmentInstallation and maintenance of IT equipmentTelecommunications and computer network techniciansComputer operation, installation and maintenance technicians

25

Figure 1: Change in economywide hours shares (1994-2007)

Services

Unskilled Wkrs.

Skilled Wkrs.

Office Wrks.

Technicians

Other whitecollar

Engineers

Mgmt.

-4-2

02

4

Chan

ge in

hou

r sha

res

(per

cent

age

poin

ts)

0 10 24 52 77 81 86 91 100Occupational rank based on median wages in 1994

Figure 2: Change in hours shares (Non-manufacturing: 1994-

2007)

Services

Unskilled Wkrs.Skilled Wkrs.

Office Wrks.

Technicians

Otherwhite collar

Engineers

Mgmt.

-4-2

02

4

Chan

ge in

hou

r sha

res

(per

cent

age

poin

ts)

0 14 24 48 7780 85 89 100Occupational rank based on median wages in 1994

26

Figure 3: Change in hours shares (Manufacturing: 1994-2007)

Services

Unskilled Wkrs.

Skilled Wkrs.Technicians

Office Wrks.

Engineers

-6-4

-20

24

Chan

ge in

hou

r sha

res

(per

cent

age

poin

ts)

02 25 60 67 84 89 95 100Occupational rank based on median wages in 1994

Mgmt.

Other White collar

Figure 4: The march of the techies

46

810

12

Sha

re o

f tec

hie

hour

s(%

of t

otal

hou

rs)

1995 1998 2001 2004 2007year

Techies (Total) EngineersTechnicians

27

Tab

le3:

Decompositionofaggregatetrends–In

dust

riesversu

sfirms(1994-2007)

Dec

om

posi

tion

Contr

ibu

tion

share

of

Lev

elin

1994

Ch

an

ge

Ind

ust

ries

Fir

ms

per

man

ent

firm

sto

Wit

hin

Bet

wee

nW

ith

inB

etw

een

Ch

an

ge

Wit

hin

Bet

wee

n

Bu

sin

ess

own

ers.

pro

fess

ion

als.

man

ager

s9.2

40.3

70.9

5-0

.58

-0.9

51.3

20.4

1-0

.50

0.9

1

Tec

hie

s9.0

43.1

73.2

8-0

.11

0.6

2.5

71.0

8-0

.18

1.2

6O

ther

wh

ite

coll

ar4.8

5-0

.26

-0.0

7-0

.19

-0.0

5-0

.21

-0.0

70.0

2-0

.08

Offi

cew

orke

rs25.3

5-3

.15

-1.9

2-1

.23

0-3

.15

-1.0

1-0

.06

-0.9

5S

kil

led

wor

ker

s27.3

6-0

.62

-0.3

8-0

.24

0.1

3-0

.75

-0.4

00.0

1-0

.41

Un

skil

led

wor

kers

13.8

1-2

.56

-1.8

2-0

.74

0.5

6-3

.12

0.4

60.4

00.0

6S

ervic

esw

orke

rs10.4

93.0

44.1

6-1

.12

1.5

51.4

91.3

10.6

70.6

4W

gt.

Avg.

shar

ein

Ch

ange

0.9

10.0

9-0

.25

1.2

50.4

2-0

.13

0.5

5

28

Tab

le4:

Decompositionofaggregatetrends–Geography(1994-2007)

Dec

om

posi

tion

Lev

elin

1994

Ch

an

ge

Dep

art

men

tsL

evel

in1994

Urb

an

/R

ura

lC

ontr

ibu

tion

Wit

hin

Bet

wee

nU

rban

Ru

ral

Wit

hin

Bet

wee

nU

rban

Ru

ral

Bu

sin

ess

own

ers.

pro

fess

ion

als.

man

ager

s9.2

40.3

70.5

3-0

.16

11.4

76.4

80.3

20.0

51.0

6-0

.69

Tec

hie

s9.0

43.1

73.1

50.0

210.9

86.6

43.1

40.0

32.2

50.9

2O

ther

wh

ite

coll

ar4.8

5-0

.26

-0.2

80.0

24.8

94.8

0-0

.26

0.0

0-0

.19

-0.0

7O

ffice

wor

kers

25.3

5-3

.15

-3.0

8-0

.07

28.9

720.8

7-3

.19

0.0

4-2

.37

-0.7

8S

kil

led

wor

ker

s27.3

6-0

.62

-0.6

1-0

.01

22.8

132.9

8-0

.57

-0.0

5-0

.65

0.0

3U

nsk

ille

dw

orke

rs13.8

1-2

.56

-2.8

30.2

710.8

817.4

3-2

.53

-0.0

3-0

.82

-1.7

4S

ervic

esw

orke

rs10.4

93.0

42.9

90.0

59.9

911.1

13.0

40.0

01.6

11.4

3W

gt.

Avg.

shar

ein

Ch

ange

1.0

6-0

.06

0.9

70.0

30.9

50.0

5

29

Tab

le5:

Decompositionofaggregatetrends–Gender(1994-2007)

Dec

om

posi

tion

Lev

elin

1994

Ch

an

ge

Lev

elin

1994

Gen

der

Contr

ibu

tion

Wom

enM

enW

ith

inB

etw

een

Wom

enM

en

Bu

sin

ess

own

ers.

pro

fess

ion

als.

man

ager

s9.2

40.3

76.7

010.5

90.3

9-0

.02

0.9

0-0

.53

Tec

hie

s9.0

43.1

73.0

612.2

23.2

5-0

.08

0.7

22.4

5O

ther

wh

ite

coll

ar4.8

5-0

.26

4.0

95.2

5-0

.25

-0.0

10.0

7-0

.33

Offi

cew

orke

rs25.3

5-3

.15

45.0

614.8

6-3

.45

0.3

0-1

.18

-1.9

7S

kil

led

wor

ker

s27.3

6-0

.62

7.0

638.1

6-0

.47

-0.1

5-0

.01

-0.6

1U

nsk

ille

dw

orker

s13.8

1-2

.56

14.9

813.1

9-2

.56

0.0

0-1

.46

-1.1

0S

ervic

esw

orke

rs10.4

93.0

419.0

45.9

42.9

20.1

21.6

51.3

9W

gt.

Avg.

shar

ein

Ch

ange

0.9

60.0

40.4

70.5

3

30

Table 6: Baseline Results – WLS Regressions (1994-2007)

Dep. Variable Annualized employment growth rate, 1994-2007

(1) (2) (3) (4) (5) (6)

Manufacturing Non-Manufacturing

Techies 0.0349* 0.0109***

(0.0208) (0.00280)Techies (1st terc.) 0.0404 0.0117***

(0.0285) (0.00306)Techies (2nd terc.) 0.0266** 0.00584*

(0.0110) (0.00354)Techies (3rd terc.) 0.0322*** 0.0202***

(0.00571) (0.00625)Engineers 0.0269 0.0128***

(0.0174) (0.00346)Technicians 0.0259 0.00121

(0.0163) (0.00316)Exporters 0.0281 0.0285 0.0246 0.00109 0.00115 0.000532

(0.0378) (0.0377) (0.0354) (0.00478) (0.00480) (0.00478)Importers -0.0448 -0.0452 -0.0488 -0.00202 -0.00190 -0.00259

(0.0514) (0.0535) (0.0542) (0.00504) (0.00503) (0.00500)Hours -0.0156*** -0.0153*** -0.0176*** -0.00880*** -0.00889*** -0.00915***

(0.00420) (0.00264) (0.00545) (0.00131) (0.00131) (0.00144)Sector FE Yes Yes Yes Yes Yes Yes

Obs. 43,476 43,476 43,476 253,926 253,926 253,926

R2 0.005 0.005 0.005 0.001 0.001 0.001F -test (p-value)

Equality of terciles 0.052 0.036 βEngineers= βTechnicians 0.793 0.030

Dependent variable is firm-level annualized employment growth rate. Each specification includes 2-digit sector

fixed effects. WLS estimates with robust standard errors. Standard errors are in parentheses. ***, **, *

significantly different from 0 at 1%, 5%, and 10% levels, respectively.

31

Tab

le7:

IVResu

lts–W

LSand

W2SLSRegressions(2002-2007)

Dep

.V

aria

ble

An

nu

ali

zed

emp

loym

ent

gro

wth

rate

,2002-2

007

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Manu

fact

uri

ng

Non

-Manu

fact

uri

ng

WL

SW

2SL

SW

LS

W2S

LS

WL

SW

2S

LS

WL

SW

2S

LS

Tec

hie

s0.

0185

***

0.04

64***

0.0

115***

0.0

0895

(0.0

0271

)(0

.0134)

(0.0

0377)

(0.0

183)

En

gin

eers

0.0

119***

0.0

601**

0.0

122***

-0.0

0849

(0.0

0325)

(0.0

292)

(0.0

0333)

(0.0

464)

Tec

hn

icia

ns

0.0

0874***

-0.0

201

-0.0

0106

-0.0

112

(0.0

0261)

(0.0

319)

(0.0

0421)

(0.0

398)

Exp

orte

rs-0

.005

17*

-0.0

181

-0.0

0576**

-0.0

225

-0.0

144***

-0.0

275

-0.0

146***

-0.0

325

(0.0

0288

)(0

.0152)

(0.0

0293)

(0.0

154)

(0.0

0549)

(0.0

225)

(0.0

0548)

(0.0

228)

Imp

orte

rs0.

0054

3*-0

.00762

0.0

0421

-0.0

0926

-0.0

0229

-0.0

143

-0.0

0238

-0.0

0707

(0.0

0295

)(0

.0182)

(0.0

0304)

(0.0

186)

(0.0

0561)

(0.0

214)

(0.0

0561)

(0.0

218)

Hou

rs-0

.006

17**

*-0

.005

61***

-0.0

0666***

-0.0

0579***

0.0

00354

0.0

0342

0.0

00276

0.0

0651**

(0.0

0106

)(0

.001

21)

(0.0

0111)

(0.0

0138)

(0.0

0130)

(0.0

0245)

(0.0

0137)

(0.0

0319)

Sec

tor

FE

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Ob

serv

atio

ns

43,4

7643

,476

43,4

76

43,4

76

253,9

26

253,9

26

253,9

26

253,9

26

R2

0.01

90.0

19

0.0

09

0.0

09

F-S

tat

(1sta

ge)

23.9

436.9

247.6

439.2

1D

epen

den

tva

riab

leis

firm

-lev

elan

nu

aliz

edem

plo

ym

ent

gro

wth

rate

.E

ach

spec

ifica

tion

incl

ud

es2-d

igit

sect

or

fixed

effec

ts.

WL

S

and

W2S

LS

esti

mat

esw

ith

rob

ust

stan

dard

erro

rs.

Th

eF

-Sta

t(1

sta

ge)

isth

eK

leib

ergen

-Paap

Wald

stati

stic

s.S

tan

dard

erro

rs

are

inp

aren

thes

es.

***,

**,

*si

gnifi

cantl

yd

iffer

ent

from

0at

1%

,5%

,an

d10%

leve

ls,

resp

ecti

vely

.

32

Appendix

A Change in employment shares 1994-2007 for the whole economy.

For ease of exposition, we aggregate the 22 PCS (Profession et Categorie Socioprofessionelle) occupations at the2-digit level into 7 broad categories. The figure immediately below visualizes the polarization pattern using 22occupations. The pattern depicted in the figure immediately above is clear. The two large highly-paid occupationsPCS 37 (Managers) and PCS 38 (Technical Managers) on the right grew, as did low-wage occupations on theleft: PCS 56 (Personal service workers), PCS 68 (Low-skilled manual laborers), PCS 55 (Retail workers), PCS 64(Drivers) and PCS 65 (Skilled transport and wholesale workers). The main middle-wage occupations that shrankover the period include skilled industrial workers and manual laborers (PCS 62 and 63), unskilled industrial workers(PCS 67), and clerical and middle-management workers (PCS 54 and 46). Exceptions to this pattern in the middleof the wage distribution include technicians (PCS 47) and health and social workers (PCS 43). To summarize,polarization and the march of the techies proceeded together from 1994 to 2007.

Figure A: Change in economywide hours shares 1994-2007

56

68

55

5364

65

63

67

54

43

62

31

47

42

4648

3534

38

37

2221

23

-3-2

-10

12

Chan

ge in

hou

r sha

res

(per

cent

age

poin

ts)

0 25 50 75 100Occupational rank based on median wages in 1994

Figure B: *Note: 21–Craftsmen; 22–Shopkeepers; 23–Heads of small businesses; 34–Scientific professionals; 35–Creative professionals; 37–Top

managers and professionals; 38– Engineers and technical managers; 42–Teachers; 43– health and social workers; 46– Office workers,middle-level; 47–Technicians; 48–Supervisors and foremen; 53–Security workers; 54–Office workers, lower-level; 55–Retail workers;

56–Personal service workers; 62–Skilled industrial Workers; 63–Skilled manual laborers ; 64–Drivers; 65–Skilled transport andwholesale workers; 67–Unskilled industrial workers; 68–Unskilled manual laborers.

Source: Authors’ own computation. DADS.

1

B List of urban departments

There are 94 departments in mainland France of which 22 are classified as urban according to the definition of theFrench National Institute of Statistics (INSEE). Urban departments are those where the majority of the populationlives in an urban area of 200,000 inhabitants or more.

Table A: List of urban departmentdepartment Name Capital Employment

(%, 1994)01 Ain Bourg-en-Bresse 0,82

06 Alpes-Maritimes Nice 1,4913 Bouches-du-Rhone Marseille 2,6731 Haute-Garonne Toulouse 1,6533 Gironde Bordeaux 1,7834 Herault Montpellier 0,9835 Ille-et-Vilaine Rennes 1,4138 Isere Grenoble 1,8044 Loire-Atlantique Nantes 1,8259 Nord Lille 4,3367 Bas-Rhin Strasbourg 2,1369 Rhone Lyon 3,4075 Paris Paris 8,5076 Seine-Maritime Rouen 2,2577 Seine-et-Marne Melun 1,8778 Yvelines Versailles 2,7983 Var Toulon 0,90

91 Essonne Evry 1,9892 Hauts-de-Seine Nanterre 5,6993 Seine-St-Denis Bobigny 2,5594 Val-de-Marne Creteil 2,4195 Val-D’Oise Pontoise 1,91

Total 55,1

2

C Model

Our hypothesis is that techies are a channel through which falling ICT prices cause polarization. In this sectionwe illustrate this in a simple model of firm-level outcomes. The model shows how a drop in the price of ICT canlead to (1) greater employment growth in ICT-intensive firms, and (2) polarization of employment within a firm.These results help motivate our within and between-firm econometric analysis in the following sections, and alsothe descriptive analysis above.1 Formal proofs can be found at the end of this section of the Appendix.

A Technology

We begin with a constant return to scale production function, which combines three types of non-techie laborservices, along with ICT, into output Q:

Q =

(L

1− α− β

)1−α−β(M

α

)α(H

β

)β.

In this function, L is hours worked by low-skill workers. The other components of the production function combinehours worked by medium and high-skill workers, M and H, with ICT services C,

M =[θ

1η C

η−1η + (1− θ)

1η M

η−1η

] ηη−1

H =[θ

1σ C

σ−1σ + (1− θ)

1σ H

σ−1σ

] σσ−1

.

M is an aggregate of the tasks performed by medium-skill workers together with ICT services, and H is similarlyan aggregate of tasks produced by high-skill workers together with ICT services. Our assumption that ICT is asubstitute for M and a complement to H is given by η > 1 and 0 < σ < 1. A key parameter is θ, which indexes theintensity of ICT in production.

ICT technology does not affect production unless it is installed, maintained, and managed by technicians andmanagers with the appropriate education, training, and experience. To express this idea in the simplest way possible,we specify ICT services C as a Leontief function of techies’ labor hours T and ICT capital K,

C = min[T,K] .

The three types of workers are paid wL, wM , and wH . Techies are paid wT , and ICT capital is paid a rental rateof r. The unit cost function corresponding to this technology is

b = (wL)1−α−β

(pM )α

(pH)β,

where the price indices of medium and high-skill tasks are

pM =[θp1−ηC + (1− θ)w1−η

M

] 11−η

pH =[θp1−σC + (1− θ)w1−σ

H

] 11−σ ,

and the price of ICT services is

pC = wT + r .

Using Shephard’s Lemma we compute the demand for each type of worker H, M and L. These are functions ofall factor prices. We then compute relative demand for H/L, M/L and M/L. Factor prices affect these relativedemand depending on the level of θ, inter alia.

The share of ICT services in total cost is increasing in θ. This implies that the employment of techies is alsoincreasing in θ. This motivates our focus on techies as a marker of technology intensity in the econometric analysis.

1Goos et. al. (2014) show how a within and between decomposition of the type we conduct above can be embedded in a model ofthe economy—but they apply this at the industry level.

3

B Techies, competitiveness and total employment

We first discuss the between-firm effect of falling ICT prices. While a drop in the price of computers r benefitsall firms that employ ICT services by lowering their unit costs b (∂b/∂r > 0 when θ > 0), firms that are moreICT-intensive (higher θ) benefit more:

∂2b

∂r∂θ> 0 .

This means that, following a drop in r, ICT-intensive firms become relatively more cost competitive, and under anyplausible demand system this will lead to market-share gains for ICT-intensive firms. This competitive effect leadsto greater employment growth in techie-intensive firms, which motivates our firm employment growth regressionsbelow.

C Techies and polarization within firms

We now turn to the effect of falling ICT prices on relative employment within firms. A drop in r leads to apolarization in employment, with H rising relative to M and L, and M falling relative to H and L,

∂

∂r

(H

L

)< 0,

∂

∂r

(M

L

)> 0,

∂

∂r

(H

M

)< 0 .

The intuition is straightforward: since ICT is a complement to H but a substitute for M , a drop in r leads to greateremployment ofH and less ofM . The key question—which helps motivate our within-firm regression analysis below—is whether the polarizing effect of falling r is stronger in firms where ICT is more important. Mathematically, is

the cross derivative ∂2

∂r∂θ

(HM

)negative? Intuition suggests yes, and we show below that ∂2

∂r∂θ

(HM

)is negative for

most of the relevant regions of the parameter space. We also illustrate the forces at work with a numerical example.The reason that the sign of the cross derivative here is, in general, ambiguous is that it depends in a complex wayboth on relative complementarity and substitutability of ICT with different types of labor and on the degree ofsubstitutability of these types of labor one with the other.2

To summarize, a fall in r will lead to an increase in aggregate demand for H relative to M through a within-firmchannel, and possibly through a between-firm channel. The within-firm effect is due to substitution of H for Mwithin firms, and we illustrate that this effect is likely to be stronger in high-θ firms. There are two effects of fallingr on the total demand for labor in high-θ firms relative to low-θ firms. The first effect is the substitution effect ofICT for medium-skilled labor M , and the second is the complementary effect of ICT on high-skilled labor H. Thenet effect of the substitution and complementarity effects on relative employment of high-θ firms is ambiguous, sothe effect of ICT-intensity on relative employment growth is an empirical matter.

In this section of the Appendix we show how relative employment varies with θ, the distributional parameterassociated with ICT services in the functions H and M , and with r, the cost of ICT capital services.

We begin with a constant returns to scale production function, which combines three types of non-techie laborservices, along with ICT, into output Q:

Q =

(L

1− α− β

)1−α−β(M

α

)α(H

β

)β.

In this function, L is hours worked by low-skill workers. The other components of the production function combinehours worked by medium and high-skill workers, M and H, with ICT services C,

M =[θ

1η C

η−1η + (1− θ)

1η M

η−1η

] ηη−1

H =[θ

1σ C

σ−1σ + (1− θ)

1σ H

σ−1σ

] σσ−1

.

M is an aggregate of the tasks performed by medium-skill workers together with ICT services, and H is similarlyan aggregate of tasks produced by high-skill workers together with ICT services. Our assumption that ICT is a

2This is related to the debate about whether declines in the user cost of capital can explain declines in labor shares Karabarbounisand Neiman (2013) versus Oberfield and Raval (2014). The answer depends on whether the elasticity of substitution between capitaland labor is greater or smaller than unity. The literature has provided different answers to this question, depending on data sources,methodologies and levels of aggregation. However, if there are more than one type of labor with capital-skill complementarity, as in oursetting here with ICT capital, then it is generally impossible to predict outcomes, as illustrated in Reshef and Santoni (2020).

4

substitute for M and a complement to H is given by η > 1 and 0 < σ < 1. A key parameter is θ, which indexes theintensity of ICT in production.

ICT technology does not affect production unless it is installed, maintained, and managed by technicians andmanagers with the appropriate education, training, and experience. To express this idea in the simplest way possible,we specify ICT services C as a Leontief function of techies’ labor hours T and ICT capital K,

C = min[T,K] .

The three types of workers are paid wL, wM , and wH . Techies are paid wT , and ICT capital is paid a rental rateof r. The unit cost function corresponding to this technology is

b = (wL)1−α−β

(pM )α

(pH)β,

where the price indices of medium and high-skill tasks are

pM =[θp1−ηC + (1− θ)w1−η

M

] 11−η

pH =[θp1−σC + (1− θ)w1−σ

H

] 11−σ ,

and the price of ICT services is

pC = wT + r .

Using Shephard’s Lemma, the relative employment levels of workers are

H

L=

β

1− α− β

((1− θ) pσ−1

C wL

θwσH + (1− θ) pσ−1C wH

)M

L=

α

1− α− β

((1− θ) pη−1

C wL

θwηM + (1− θ) pη−1C wM

)H

M=β

αpσ−ηC

(θwηM + (1− θ) pη−1

C wM

θwσH + (1− θ) pσ−1C wH

).

D Techies and polarization within firms: details

How does cross-sectional variation in θ affect the composition of employment within firms? We answer this questionby differentiating the relative employment equations with respect to θ,

∂

∂θ

(H

L

)=

−β1− α− β

pσ+1C wσHwL

(θpCwσH + (1− θ) pσCwH)2 < 0

∂

∂θ

(M

L

)=

−α1− α− β

pη+1C wηMwL

(θpCwηM + (1− θ) pηCwM )

2 < 0

For both H and M , higher θ is associated with lower employment relative to L. The reason is that as the importanceof ICT in producing high- and medium-skill tasks rises, the labor that is required to work with ICT capital falls.Since there is no direct effect of θ on the productivity of L, the ratios H/L and M/L decline with θ. The effect ofθ on H

M can not be signed:

∂

∂θ

(H

M

)=β

α

pσ+1C

(pσ−ηC wHw

ηM − 1

)(θpCwσH + (1− θ) pσCwH)

2

The term in parentheses in the numerator is of ambiguous sign, so the derivative is of ambiguous sign. The effectis more likely to be positive the higher is wH or wM , and the lower is pC .

The parameter θ is an indicator of the importance of ICT services in production. The share SICT of ICT inunit cost b, which is also the elasticity of cost with respect to pC , is a fairly complex function,

SICT =pCb

∂b

∂pC=θpCwMwH

[α (1− θ) pσCw

η−1M + β (1− θ) pηCw

σ−1H + (α+ β) θpCw

η−1M wσ−1

H

][θpCwσH + (1− θ) pσCwH ] [θpCw

ηM + (1− θ) pηCwM ]

.

5

This share is increasing in θ:

∂SICT

∂θ=pCb

∂2b

∂pC∂θ=

α (pCwM )η+1

(θpCwηM + (1− θ) pηCwM )

2 +β (pCwH)

σ+1

(θpCwσH + (1− θ) pσCwH)2 > 0

Given that SICT is increasing in θ, it is not surprising that the share of techie workers in total employment,T/(T + L+M +H), can also be shown to be increasing in θ.

E Polarization with falling ICT prices: details

We next turn to the effect of falling ICT prices on relative employment. Since σ − 1 < 0 and η − 1 > 0, we findthat a drop in r leads to a polarization in employment, with H rising relative to M and L, and M falling relativeto H and L,

∂

∂r

(H

L

)=

β

1− α− β(σ − 1) θ (1− θ) pσCwσHwL(θpCwσH + (1− θ) pσCwH)

2 < 0

∂

∂r

(M

L

)=

α

1− α− β(η − 1) θ (1− θ) pηCw

ηMwL

(θpCwηM + (1− θ) pηCwM )

2 > 0

∂

∂r

(H

M

)=β

α

pσ−ηC [− (1− θ) (η − 1) pσCwHwηM − wσH {θ (η − σ) pCw

ηM + (1− θ) (1− σ) pηCwM}]

(θpCwσH + (1− θ) pσCwH)2 < 0

The intuition is straightforward: since ICT is a complement to H but a substitute for M , a drop in r leads togreater employment of H and less of M .

We now turn to a key question which helps motivate our empirical specification below: is the polarizing effect

of falling r stronger within firms where ICT is more important? Mathematically, is the cross derivative ∂2

∂r∂θ

(HM

)negative? The expression for ∂2

∂r∂θ

(HM

)is quite complex:

∂2

∂r∂θ

(H

M

)= rσ−η

β

α

A−B[−θrwσH − (1− θ) rσwH ]

3

where the cubed term in the denominator is negative, and

A = rηwσHwM (σ − 1) [θrwσH − (1− θ) rσwH ]

B = rσwHwηM [θrwσH (2σ − η − 1)− (1− θ) (η − 1) rσwH ]

Given the assumptions η > 1 and 1 > σ > 0, the term B is necessarily negative. If A > 0, then A−B > 0 and thederivative is therefore negative. This is what intuition suggests: for higher levels of θ, the polarizing effect of a fallin r is greater. However, A need not be positive, though A − B > 0 is still possible when A < 0. The conditionA−B > 0 can be analyzed further by writing it out, and dividing both sides by the positive quantity rηwσHwM , toobtain

(σ − 1) [θrwσH − (1− θ) rσwH ] > rσ−ηw1−σH wη−1

M [θrwσH (2σ − η − 1)− (1− θ) (η − 1) rσwH ]

When will this inequality be satisfied? Since the RHS is strictly negative, a sufficient but not necessary conditionis that σ → 1, so that the LHS → 0. An alternative sufficient condition is that [θrwσH − (1− θ) rσwH ] < 0, whichwill hold for small enough values of θ. Without tediously examining various configurations of the parameter space,we conclude that if the importance of ICT in production θ is not too high, and/or if ICT is not too complementary

with high-skilled labor H, then ∂2

∂r∂θ

(HM

)< 0: the polarizing effect of falling prices for ICT is stronger in firms

where ICT is more important.We illustrate the forces at work with a numerical example.We normalize the wage of the least skilled workers

to 1, and set wM = 2 and wH = 3. The elasticities of substitution are η = 2 and σ = 1/2, and the upper-level costshares α and β are equalized at 1/3. We drop the cost of ICT r from 11 to 1, and analyze how the resulting ratioH/M varies as a function of θ ∈ [0, 1]. The figure below, a contour plot of the level of H/M , illustrates what wefind.

The vertical axis measures the cost of computer capital r, while the horizontal axis measures the parameter θ.Lower levels of H/M are at the upper right of the figure, shaded blue, with higher levels of H/M shading towardorange. Moving from the top to the bottom of the figure illustrates our analytical result that a drop in r leadsto an increase in H/M , as ICT services complement H and substitute for M . This increase is steeper for higherlevels of θ: the more important ICT is in the production function, the greater the polarizing effect of a drop in r

6

Figure C: H/M as a function of r (vertical axis) and θ (horizontal axis)

(to see this, note that when more contour lines are crossed for the same vertical drop, then the level of the functionis changing faster). The figure also shows that the effect of higher θ on H/M is ambiguous: for low levels of r (atthe bottom of the figure), higher θ is associated with higher H/M , but for higher values of r the effect is reversed.

Aside from the effects on within-firm relative labor demand, a drop in r can change economy-wide relative labor

demand by reducing costs more rapidly for firms that use ICT more intensively. This effect follows from ∂2b∂pC∂θ

> 0shown above.

D Further econometric results

Table 8 reports estimates of equation (7), where we regress annualized long changes from 1994 to 2007 of firm-levelex-techie occupation shares on initial levels of techies, export status, and import status in 1994. We find a positiveassociation between techies and managers, both top and middle level, and a negative association between techies andskilled and services workers. These results are remarkably similar in both manufacturing and non-manufacturingsamples. Compared to firms that have no techies, firms with techies saw their top managers’ share of hours increaseby 0.34 percent per year in the manufacturing sector and 0.30 percent per year in the non-manufacturing sector.The impact of techies on middle managers is also positive in both sectors over the period 1994 to 2007—but muchsmaller: about 0.05 percent per year.

We find a negative impact of techies on skilled, unskilled and services workers. The effect on unskilled workersis however not significant in the manufacturing sector.

The last columns of Table 8 display the results when we use indicators for engineers and technicians rather thanjust one for techie employment. These regressions indicate that engineers are associated with firm compositionalchanges, much more than technicians. In both sectors, engineers are positively associated with the change in em-ployment of top managers and negatively with the change of employment of services workers. In non-manufacturing,

7

we find a negative impact of engineers on unskilled workers.Overall, our results suggest that the presence of techies in 1994 is followed by within-firm skill upgrading, with

an increase in the share of top and middle managers and a decrease in the share of the three lowest paid occupations,especially the lowest-paid service worker occupation.

Turning to the effects of international trade, both importing and exporting are associated with an increase inthe top manager share in both manufacturing and nonmanufacturing, while among manufacturing firms importingis associated with skill downgrading within the blue collar workforce (a large negative effect of importing on skilledworkers and a large positive effect on unskilled workers).

We reproduce the same exercise using the overall techie indicator reported in Table 8 for changes between2002-2007 that we regress on indicators in 2002. The results in Table 9 are qualitatively similar. We find positiveeffects of techies on the change in employment of top managers and negative effects on the change in employment ofskilled and services workers. The effects on middle managers and unskilled workers are not statistically significantat conventional levels when using the shorter time span.

– Insert Table 9 about here –

The last columns of Table 9 report estimates of just-identified W2SLS models, where we run the regressions forthe sample between 2002 and 2007 and use 1994 lagged techie and trade indicators as instruments for indicators in2002. The Kleibergen-Papp F tests have values larger than 25 in the non-manufacturing sample and around 10 –and much below for the category of office workers, in the manufacturing sample. We find a negative effect of techieson the change in employment of middle managers in manufacturing firms. There is weak evidence of a positiveimpact of techies on skilled workers in the manufacturing sectors and on unskilled workers in services sector. Wefind a negative association of techies with services workers in the non-manufacturing sector.3

We have experimented with different lags for the instruments in these and in the employment growth regressions.Contrary to the results of the growth regressions, the results of the W2SLS regressions at occupational level arenot robust to the choice of the year that we use as lagged instrument. Therefore, we hedge our interpretations ofthe within-firm occupational composition regressions. We stress that the growth regression estimates are robust tochanging the lag with which we use to instrument the level of techies in 2002, which lends to the plausible exclusionof lags in this context.

3In Harrigan et al. 2016, where we used all 1994-1998 lags to instrument levels in 2002, we find a strong negative effect of techies onthe share of middle managers within firms when we instrument the techie dummy variable in 2002 by its lag in 1994.

8

Tab

le8:

Occupa

tionalChanges–W

LSRegressions(1994-2007)

Dep.

Vari

able

Change

inth

eex-t

ech

iesh

are

of

hours

,1994-2

007

Top

Mid

.O

ffice

Skille

dU

nsk

ille

dServ

ice

Top

Mid

.O

ffice

Skille

dU

nsk

ille

dServ

ice

Mgm

t.M

gm

t.W

krs

.W

krs

.W

krs

.W

krs

.M

gm

t.M

gm

t.W

krs

.W

krs

.W

krs

.W

krs

.P

anel

A:

Manufa

ctu

ring

Tech

ies

0.3

44***

0.0

52***

0.0

36

-0.0

44*

-0.0

032

-0.2

88***

(0.0

16)

(0.0

18)

(0.0

29)

(0.0

27)

(0.0

48)

(0.0

43)

Engin

eers

0.3

00***

0.0

275

0.0

74***

-0.0

39

-0.0

15

-0.2

60***

(0.0

19)

(0.0

17)

(0.0

27)

(0.0

25)

(0.0

92)

(0.0

90)

Tech

nic

ians

0.0

98***

0.0

23

-0.0

29

0.0

11

-0.1

29**

0.0

17

(0.0

234)

(0.0

185)

(0.0

297)

(0.0

35)

(0.0

60)

(0.0

58)

Exp

ort

ers

0.2

38***

-0.0

26

-0.0

35

-0.0

25

-0.1

47*

-0.0

01

0.2

05***

-0.0

30

-0.0

39

-0.0

27

-0.1

11

0.0

03

(0.0

19)

(0.0

19)

(0.0

24)

(0.0

29)

(0.0

76)

(0.0

79)

(0.0

18)

(0.0

19)

(0.0

25)

(0.0

30)

(0.0

77)

(0.0

73)

Imp

ort

ers

0.2

47***

-0.0

52**

-0.0

22

-0.0

75***

0.1

66***

-0.2

48***

0.2

11***

-0.0

55**

-0.0

24

-0.0

78***

0.2

12***

-0.2

49***

(0.0

19)

(0.0

22)

(0.0

26)

(0.0

27)

(0.0

62)

(0.0

59)

(0.0

19)

(0.0

23)

(0.0

28)

(0.0

27)

(0.0

63)

(0.0

60)

Secto

rF

EY

es

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Obs.

29,4

63

20,0

70

17,0

43

32,9

88

40,0

33

35,5

67

29,4

63

20,0

70

17,0

43

32,9

88

40,0

33

35,5

67

R2

0.1

25

0.0

28

0.0

10

0.0

29

0.1

00

0.1

19

0.1

29

0.0

28

0.0

11

0.0

29

0.1

00

0.1

19

Panel

B:

Non-M

anufa

ctu

ring

Tech

ies

0.3

00***

0.0

51***

-0.0

27

-0.1

05**

-0.0

54*

-0.1

32***

(0.0

50)

(0.0

17)

(0.0

40)

(0.0

50)

(0.0

28)

(0.0

29)

Engin

eers

0.2

74***

0.0

123

-0.0

24

-0.0

38

-0.1

03***

-0.1

29***

(0.0

38)

(0.0

22)

(0.0

41)

(0.0

49)

(0.0

31)

(0.0

29)

Tech

nic

ians

0.1

05**

0.0

317

0.0

10

-0.0

49

-0.0

44

-0.0

09

(0.0

44)

(0.0

24)

(0.0

42)

(0.0

49)

(0.0

32)

(0.0

30)

Exp

ort

ers

0.4

08***

-0.0

22

-0.0

58

-0.1

66**

-0.1

07*

-0.0

16

0.3

77***

-0.0

23

-0.0

60

-0.1

67**

-0.0

83

-0.0

08

(0.0

70)

(0.0

27)

(0.0

62)

(0.0

67)

(0.0

62)

(0.0

44)

(0.0

70)

(0.0

28)

(0.0

61)

(0.0

67)

(0.0

60)

(0.0

44)

Imp

ort

ers

0.2

76***

-0.0

28

-0.0

80

-0.1

07

0.0

25

-0.0

67

0.2

46***

-0.0

28

-0.0

81

-0.1

09

0.0

50

-0.0

61

(0.0

68)

(0.0

34)

(0.0

53)

(0.0

75)

(0.0

67)

(0.0

48)

(0.0

69)

(0.0

34)

(0.0

54)

(0.0

75)

(0.0

68)

(0.0

48)

Secto

rF

EY

es

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Obs.

146,6

12

71,6

08

118,8

14

185,9

97

146,6

68

137,4

33

146,6

12

71,6

08

118,8

14

185,9

97

146,6

68

137,4

33

R2

0.3

11

0.0

48

0.1

19

0.0

44

0.0

44

0.0

60

0.3

12

0.0

48

0.1

19

0.0

44

0.0

45

0.0

60

Dep

end

ent

vari

ab

leis

the

chan

ge

inth

eex

-tec

hie

share

of

hou

rsof

the

six

larg

en

on

-tec

hie

occ

up

ati

on

s.E

ach

spec

ifica

tion

incl

ud

es2-d

igit

sect

or

fixed

effec

ts.

WL

S

esti

mate

sw

ith

rob

ust

stan

dard

erro

rs.

Sta

nd

ard

erro

rsare

inp

are

nth

eses

.***,

**,

*si

gn

ifica

ntl

yd

iffer

ent

from

0at

1%

,5%

,an

d10%

level

s,re

spec

tivel

y.

8

Tab

le9:

Occupa

tionalChanges–W

LSand

W2SLSRegressions(2002-2007)

Dep.

Vari

able

Change

inth

eex-t

ech

iesh

are

of

hours

,2002-2

007

WL

SW

2SL

ST

op

Mid

.O

ffice

Skille

dU

nsk

ille

dServ

ice

Top

Mid

.O

ffice

Skille

dU

nsk

ille

dServ

ice

Mgm

t.M

gm

t.W

krs

.W

krs

.W

krs

.W

krs

.M

gm

t.M

gm

t.W

krs

.W

krs

.W

krs

.W

krs

.P

anel

A:

Manufa

ctu

ring

Tech

ies

0.3

51***

-0.0

0568

-0.0

643

-0.0

760*

0.0

902

-0.2

40***

0.1

91

-0.5

05**

0.0

593

0.4

24*

-4.5

81

4.6

73

(0.0

340)

(0.1

09)

(0.0

562)

(0.0

451)

(0.1

15)

(0.0

874)

(0.2

48)

(0.2

54)

(0.2

07)

(0.2

47)

(3.0

47)

(3.2

92)

Exp

ort

ers

0.1

78***

-0.0

360

-0.1

64***

0.0

367

-0.0

658

-0.0

246

0.0

978

0.2

26

0.0

468

0.1

74

-8.3

84*

8.0

49

(0.0

367)

(0.0

789)

(0.0

586)

(0.0

761)

(0.1

25)

(0.0

945)

(0.2

55)

(0.5

73)

(0.1

41)

(0.3

01)

(4.9

64)

(5.1

41)

Imp

ort

ers

0.1

39***

-0.0

651

0.0

179

0.0

599

-0.0

528

-0.0

808

0.4

19

-0.2

55

-0.3

51

-0.2

68

11.8

9*

-11.9

6(0

.0369)

(0.0

996)

(0.0

548)

(0.0

695)

(0.1

23)

(0.0

947)

(0.3

79)

(0.5

22)

(0.2

15)

(0.3

82)

(7.2

00)

(7.5

46)

Secto

rF

EY

es

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Obs.

24,8

33

20,1

24

15,7

15

33,0

24

40,6

24

35,3

34

24,8

33

20,1

24

15,7

15

33,0

24

40,6

24

35,3

34

R2

0.0

51

0.0

20

0.0

06

0.0

11

0.0

66

0.0

92

F-S

tat

(1stage)

10.9

39.1

16

3.3

22

10.8

110.3

010.2

0P

anel

B:

Non-M

anufa

ctu

ring

Tech

ies

0.2

58***

0.0

00924

0.0

553

-0.0

957*

-0.0

380

-0.1

65***

0.1

74

-0.1

26

0.0

188

-0.1

04

0.2

88*

-0.3

28**

(0.0

381)

(0.0

266)

(0.0

637)

(0.0

571)

(0.0

471)

(0.0

353)

(0.2

29)

(0.1

17)

(0.1

95)

(0.2

39)

(0.1

60)

(0.1

32)

Exp

ort

ers

0.3

47***

-0.1

00*

-0.0

902

-0.2

55*

-0.0

137

0.1

35

0.3

23

-0.4

33**

0.2

46

-0.5

16

-0.0

0907

0.4

01

(0.1

17)

(0.0

567)

(0.0

694)

(0.1

36)

(0.0

886)

(0.1

03)

(0.2

53)

(0.1

81)

(0.3

15)

(0.4

30)

(0.2

67)

(0.3

27)

Imp

ort

ers

-0.0

0718

-0.0

948

0.0

324

0.0

472

0.1

48

-0.1

02

0.1

56

0.2

69

-0.2

28

0.3

44

-0.1

55

-0.3

09

(0.0

941)

(0.0

611)

(0.0

703)

(0.1

22)

(0.0

912)

(0.0

944)

(0.2

84)

(0.2

01)

(0.3

15)

(0.4

24)

(0.3

13)

(0.3

25)

Secto

rF

EY

es

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Obse

rvati

ons

120,2

08

73,8

26

120,8

09

188,4

47

151,6

00

129,2

63

120,2

08

73,8

26

120,8

09

188,4

47

151,6

00

129,2

63

R2

0.1

10

0.0

23

0.0

39

0.0

25

0.0

25

0.0

34

F-S

tat

(1stage)

36.9

725.2

325.2

439.4

935.9

632.9

9

Dep

end

ent

vari

ab

leis

the

chan

ge

inth

eex

-tec

hie

share

of

hou

rsof

the

six

larg

en

on

-tec

hie

occ

up

ati

on

s.E

ach

spec

ifica

tion

incl

ud

es2-d

igit

sect

or

fixed

effec

ts.

WL

San

d

W2S

LS

esti

mate

sw

ith

rob

ust

stan

dard

erro

rs.

Th

eF

-Sta

t(1

sta

ge)

isth

eK

leib

ergen

-Paap

Wald

stati

stic

s.S

tan

dard

erro

rsare

inp

are

nth

eses

.***,

**,

*si

gn

ifica

ntl

y

diff

eren

tfr

om

0at

1%

,5%

,an

d10%

level

s,re

spec

tivel

y.

9