+ All Categories
Home > Documents > Fernandez M., Roberto - Skill-Biased Technological Change An

Fernandez M., Roberto - Skill-Biased Technological Change An

Date post: 04-Apr-2018
Category:
Upload: davide-esposito-bonito
View: 226 times
Download: 0 times
Share this document with a friend

of 48

Transcript
  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    1/48

    AJS Volume 107 Number 2 (September 2001): 273320 273

    2001 by The University of Chicago. All rights reserved.

    0002-9602/2001/10702-0001$10.00

    Skill-Biased Technological Change and Wage

    Inequality: Evidence from a Plant

    Retooling1

    Roberto M. Fernandez

    Massachussetts Institute of Technology

    One of the most popular explanations for the increased wage in-equality that has occurred since the late 1970s is that technologicalchange has resulted in a downward shift in the demand for low-skill workers. This pattern is also alleged to account for the growthin racial inequality in wages over the same period. This article re-ports on a case study of the retooling of a food processing plant. Aunique, longitudinal, multimethod design reveals the nature of thetechnological change, the changes in job requirements, and themechanisms by which the changes affect the wage distribution forhourly production workers. This research finds that, indeed, theretooling resulted in greater wage dispersion and that the changeshave also been associated with greater racial inequality in wages.However, contrary to the claims of advocates of the skill-bias hy-pothesis, organizational and human resources factors strongly me-diated the impact of the changing technology. Absent these highroad organizational choices, this impact on wage distribution would

    have been even more extreme.

    One of the most prominent explanations for the increased wage inequality

    that has occurred since the late 1970s is that technological changes oc-

    curring over the same period have resulted in a downward shift in the

    1 Funding for various phases of this project has been provided by the National Science

    Foundation, the Russell Sage Foundation, the Rockefeller Foundation, the Institute

    for Research on Poverty, the U.S. Department of Health and Human Services, and

    the Stanford Integrated Manufacturing Association. Chris Wellin, Judith Levine, andDavid Harris provided excellent research assistance on various phases of the project.

    I would like to thank David Card, Christopher Jencks, Frank Levy, Paul Osterman,

    and the AJS reviewers for their helpful comments and suggestions. Direct correspon-dence to Roberto M. Fernandez, Sloan School of Management,Massachussetts Institute

    of Technology, 50 Memorial Drive, Cambridge, Massachussetts 02142. E-mail:

    [email protected]

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    2/48

    American Journal of Sociology

    274

    demand for low-skill workers (for reviews, see Morris and Western 1999;Danziger and Gottschalk 1995; Levy and Murnane 1992). Skill-biased

    technological change is also alleged to play an important role in the growth

    in racial inequality in wages that has been observed over the same period.

    While many others factors may also be at work (Moss and Tilly 1993),

    changing technologyespecially in manufacturingis often implicated

    as one of the major factors worsening the economic plight for minorities

    (Wilson 1987, 1996; Kasarda 1988).2

    In this article, I study the impact of technological change on changes

    in the overall wage distribution and on racial differences in wages within

    the context of a longitudinal case study of the retooling of a food processing

    plant. I argue that this plants experience during the retooling should be

    regarded as a natural experiment, and as such, it offers unique advantages

    over extant research in this area. The longitudinal, multimethod designdeveloped here affords rare insight into the nature of the technological

    change, the changes in job requirements, the organizational context of

    the change, and the mechanisms by which the wage distribution is affected

    by the technological changes.

    The natural-experiment design solves the major problem vexing even

    the best extant studies of skill-biased technological change. While all pre-

    vious empirical studies of the phenomenon infer an exogenous demand-

    side shift in the labor market, the workers at this company experienced

    such a shift in a dramatic way. As such, this study provides an excep-

    tionally clean setting in which to observe the key processes alleged to be

    operating in the skill-biased technological change account of growing

    wage inequality. Since this company endeavored to keep all its workers

    through the change in technology, this study also avoids the main threat

    to validity in extant skill-bias studies, that is, the problem of self-selection

    of people into jobs for which their skills complement the technology. Past

    studies have run the risk of attributing observed wage changes to the use

    of the technology rather than to the individual factors that led the person

    to the job in the first place. In contrast, for the production workers in

    the company studied here, there is no issue of self-selection. For them,

    2 The skill-biasing effects of technology are a specific instance of what has been called

    in sociology job-skill mismatch (Kasarda 1988; Morris and Western 1999). Job-skillmismatch processes are conceptualized in broader terms, encompassing phenomena

    like sectoral changes in the economy such as the shift from manufacturing to services.

    While these broader processes may also be at work in the economy (see Morris, Bern-

    hardt, and Handcock [1994] for an innovative article distinguishing the upgradingand polarizing effects of sectoral shifts), they cannot address the apparently within-

    firm nature of increased wage inequality (see below). I focus here on a specific kind

    of mismatchthe skill-biasing effects of technological changebecause the techno-logical changes observed in this firm offer a natural-experiment test of economists

    preferred explanation of growing within-firm wage inequality (see below).

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    3/48

    Skill-Biased Change

    275

    the retooling comes as a truly exogenous, demand-side shock to their labormarket. As such, this study distills the essence of the skill-biased change

    argument and provides a unique opportunity to observe closely workers

    experiences adjusting to a new technology. Moreover, because the plant

    also has a good representation of minority workers, the validity of ar-

    guments attributing growing racial inequality in wages to skill-biased

    technological change can directly be assessed.

    This article begins by briefly reviewing the literature on past approaches

    to skill-biased technological change. It then describes the research setting

    and the details of the research design. It next discusses the changing

    technology and presents the evidence on the changes in job requirements

    that occurred with the retooling. The strategy is to compare data first on

    various dimensions of job characteristics to assess whether there is direct

    evidence of job requirements changing with the introduction of the new

    technology. Then the wage distribution for hourly production workers

    over time is compared to see whether there is evidence of increased wage

    inequality associated with the changing technology. The wage changes in

    the factory are also contrasted against the baseline of the wage changes

    occurring in the labor market at large. Next, the impact of these changes

    on the changing pattern of racial inequality in the plant is documented.

    The article concludes with a discussion of the organizational factors that

    appear to mediate between the changing technology and changes in the

    wage distribution.

    THE EVIDENCE ON SKILL-BIASED TECHNOLOGICAL CHANGE

    Much of the evidence for the skill-biased technological change explanation

    of rising wage inequality has been using data from supply-side surveys

    of employees (e.g., Bound and Johnson 1992; Katz and Murphy 1992; for

    reviews, see Levy and Murnane 1992; Danziger and Gottschalk 1995).

    Such studies have the advantage of broad coverage, often spanning whole

    sectors of the economy. Unfortunately, this breadth has come at the ex-

    pense of depth of information on key variables of interest, that is, tech-

    nology and skill. Most of these studies proceed by attempting to control

    for alternative explanations of changes in the wage distribution (e.g.,

    changes in product market demand, immigration, globalization, etc.).

    Changes in the wage distribution that cannot be attributed to thesesources

    are then inferred to be taking place within firms. Although there are no

    direct measures of technology in these studies, by virtue of its being a

    process that takes place within firms, skill-biased technical change is im-

    plicated as a key factor accounting for growing wage inequality.

    Another set of studies takes a more direct approach to studying changes

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    4/48

    American Journal of Sociology

    276

    in the skill distribution. Work in this tradition has relied upon job analyses(i.e., detailed observations of tasks being performed at work) in order to

    measure multiple dimensions of job tasks. Studies using this approach

    start by looking at the changing characteristics of jobs using data sources

    such as the Dictionary of Occupational Titles (e.g., see Spenner 1990) or

    proprietary data on specific employers (Cappelli 1993). A number of stud-

    ies examine whether changing job requirements have created a mismatch

    between the skills demanded by jobs and skills of the existing workforce

    (Johnston and Packer 1987; Mishel and Teixeira 1991). The results, how-

    ever, have been affected by the specific measures used and the particular

    specification of the model, so there is little agreement across these studies.

    In addition, few studies examine the link between changes in the skill

    distribution and wages. Even if there have been changes in job requi-

    rements that have led to a skill mismatch, this would not demonstrate alink to changes in the wage distribution. The story becomes even more

    complicated when it is taken into account that the measured skills of the

    workforce have been changing as well (Hunt 1995; Murnane and Levy

    1996). How these changes net out and what effect they have on the wage

    distribution is unclear from these studies.

    One prominent study has tried to establish a direct link between tech-

    nology use and wages. Krueger (1993) used Current Population Survey

    (CPS) data from the 1980s to estimate within-job returns to the use of

    personal computers on the job. He concludes that wage returns to com-

    puter use after controlling for education and other human capital factors

    are on the order of 10%15%. More recent research casts doubt on Krue-

    gers interpretation of his results. Several studies have raised the issue of

    whether these estimates of the returns to computer use may be upwardly

    biased. DiNardo and Pischke (1997) suggest the findings are mostly due

    to unobserved human capital factors correlated with computer use. Using

    German data, they show that the use of pencils at work appears to have

    almost as big a wage return as does computer use.3 Because Krueger

    cannot correct for selection of people into jobs that use computers, his

    study winds up attributing wage increases to the use of the technology,

    rather than the individual factors that led the person to the job in the

    first place.

    From my perspective, these studies have presented a less than wholly

    persuasive story of how growing wage inequality is due to technological

    3

    The same is true for use of a telephone on the job. In contrast, the use of a hammeris associated with lower wage returns. Taken together, these patterns suggest that the

    use of these tools has little to do with commanding a wage premium but simply are

    indexing white-collar vs. blue-collar status. Since wages have been rising for white-collar work, and dropping for blue-collar work, the tools pick up the effects of white-

    collar vs. blue-collar.

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    5/48

    Skill-Biased Change

    277

    change. For one, they have tended to black box the changes in tech-nology that are supposed to be driving the greater dispersion in the wage

    distribution. Even when the technology under study is precisely defined

    (e.g., use of personal computers), the precision seems to be misplaced

    because this measurement ignores how the technology is used, that is, in

    what ways job tasks are altered. What is missing here is a description of

    the context in which new technology is being introduced and any sense

    of the subtle interplay between production technology and job require-

    ments. Moreover, these studies are also blind to any attendant changes

    in organizational or human resources practices that might accompany

    instances of technological change (an exception here is Siegel [1999]).

    There is a rich literature on technological change and its effects on jobs

    (for reviews, see Attewell 1987; Spenner 1990). These studies have shown

    that the relationship between technological change and the skill require-

    ments of jobs is indeterminate because the same capital equipment can

    be surrounded with varying job routines, which can have very different

    effects on job requirements (e.g., Flynn 1988). Moreover, changes in work

    routines are often implemented at the same time as changes in capital

    equipment, so it is dangerous to attribute causal weight to the machinery.

    Indeed, the literature on high-performance manufacturing organizations

    (e.g., Commission on the Skills of the American Workforce 1990) is all

    about how the new automation relies on the alignment of work routines,

    human resources practices, and new production machinery to deliver large

    productivity and quality gains. Large-scale, supply-side studies tend to

    be blind to the highly contextual organizational software of these chang-

    ing methods of production.While technology studies attend quite carefully to the nature of the

    technology and the changing job tasks, these studies have only indirectly

    addressed the stratification consequences of these technological changes,

    that is, they have not engaged the literature on the changing wage dis-

    tribution. To the extent that these studies discuss wages at all, they argue

    that firms organizing along the high-performance work organization

    model would invest heavily in worker training and pay high wages when

    relying on new production technology. In contrast, low road firms might

    avoid training investments and seek to implement technical change in a

    wage-minimizing manner. Beyond this simple high road/low road dis-

    tinction, crucial questions remain about the mechanisms by which chang-

    ing technology affects the wage distribution. If high road firms are likely

    to pay better, which features of the high road model leads them to do

    so? That is, what is it that firms are buying with these policies? What

    kinds of skills are being valued? Does high road imply high wages for

    all production workers? What are the race and gender implications of

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    6/48

    American Journal of Sociology

    278

    taking the high road? What role does the local labor market play in thewage-setting process?

    I argue that there is a vital need to link the literature on high-per-

    formance organizational practices to the literature on wage inequality.

    Currently, there is an almost mantra-like invocation of skill-biased tech-

    nological change as an explanation of growing wage inequality in policy

    circles, even though the supporting evidence is virtually always of the

    black box variety. Studies of changing reorganized work practices serve

    to open up the black box, but they do not go far enough in elaborating

    the mechanisms by which the wage distribution is linked to changing

    technology.

    The current study is offered as a means of bridging these literatures. I

    study the wage implications of technological change within the context

    of a longitudinal case study of the retooling of a food processing plant.Even more important, however, the design of this study solves the major

    problem vexing even the best studies of skill-biased technological change.

    While all large-scale studies of the phenomenon infer an exogenous

    demand-side shift in the labor market, the retooling induced precisely

    such a shift for the production workers at the company studied here. Since

    it is the jobs that have changed, this study gets around the thorny problem

    evident in all past skill-bias studies of the self-selection of people into jobs

    for which their skills complement the technology. Without taking account

    of the endogeneity of such choices, these studies will always run the risk

    of attributing wage increases to the use of the technology rather than the

    individual factors that led the person to the job in the first place. In

    contrast, for the production workers in the company studied here, there

    is no issue of self-selection. For them, the retooling comes as an exogenous

    demand-side shock to their labor market. As such, this company serves

    as a natural experiment distilling the key processes alleged to be operating

    in the skill-biased technological change account of growing wage ine-

    quality and provides the unique opportunity to closely observe workers

    experiences adjusting to a new technology.

    Unique among studies of the labor market effects of changing tech-

    nology, I surveyed production workers and conducted participant obser-

    vation research in the plant both before and after the retooling. This

    longitudinal, multimethod design affords rare insight into the nature of

    the technological change, the changes in job requirements, the organi-

    zational context of the change, and the mechanisms by which the wage

    distribution is affected by the technological changes. While this focus onone firm introduces some special methodological issues (see the appendix),

    I agree with Morris, Bernhardt, and Handcock (1994, p. 217) that qual-

    itative work will be important for establishing a causal connection be-

    tween industrial restructuring and inequality. Furthermore, in this project

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    7/48

    Skill-Biased Change

    279

    I have directly confronted Morris et al.s (1994) challenge by indepen-dently measuring skills required by jobs (demand) and skills possessed

    by workers (supply) (p. 217) in a manner that is sensitive to the organ-

    izational context. I argue that such an approach is needed to open up the

    black box and get in-depth information on the relationship between tech-

    nological change and wages. In addition to documenting the firms

    changes over time, I also compare the firms wage changes against the

    baseline of changes that were occurring in the local labor market over

    the period of the study.

    RESEARCH SETTING

    These issues are addressed by studying the retooling of a food processingplant located in the midwestern United States with a substantial number

    of minority workers. This firm is a wholesale supplier of food ingredients

    to other companies that produce finished, retail foods to the consumer.

    The company employed 195 hourly production workers (all unionized) at

    the beginning of the study period: 55% of these workers are racial mi-

    norities (43% black and 12% Hispanic), and 31% are female.

    In early 1989, the firms management received approval to build a new

    plant in order to accomplish a massive upgrade of the companys pro-

    duction equipment. The old facility was located in a cramped, 100-year-

    old, multistory plant in the citys central business district. The company

    invested $92 million in building a new facility located 10 miles from the

    old plant. The ground breaking for the new plant was in early 1992, and

    production started in the new plant in mid-1993. The companys presidentcited competitive pressures as the main reason for the investment: If we

    didnt make this move, within five years we would be out of business.

    By the time this study began late in 1990, the company and the union

    had agreed to cooperate through the move. First, the company gave a

    no-layoff guarantee through the period of the retooling. Second, the

    company pledged that production workers wages in the new plant would

    be no lower than their wages in the old plant, irrespective of the job in

    which workers would land in the retooled plant. In return, the union

    agreed to more flexible work arrangements through the period of the

    move, temporarily suspending seniority considerations in job transfer and

    other work rules.

    I developed a multimethod, before and after research design to track

    the retooling process, studying production workers in both the old and

    new plants. Beginning in the spring of 1991, a team of field-workers did

    participant observation of the production work in both plants. These

    researchers worked as temporary employees in the plants, rotating through

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    8/48

    American Journal of Sociology

    280

    various jobs. While the field-workers could not work in all the jobs, theydid observe and record field notes about all the production jobs at both

    plants. I used the field notes to develop direct measures of changes in

    tasks associated with the retooling. I also surveyed workers at both plants,

    including new workers at the new plant and workers who left the company

    between the first and second waves of the survey. The survey consisted

    of hour-long, face-to-face interviews about their demographic back-

    grounds, job tasks, and job rewards (response rates of 83% at time 1 and

    85% at time 2). Finally, I collected employment records, work documents,

    and other archival materials from company records.

    The Changing Technology

    The plant under study is an industrial food processor. The company takesraw food inputs (such as sugar, flour, lecithin), combines them with other

    key ingredients according to myriad recipes, and then cooks the ma-

    terials. The products are then shipped wholesale to retail food companies

    in large batches (e.g., by the pallet in 50-lb. boxes). While this basic

    description applies to both the old and new plants, the new plant has

    made extensive use of smart machine technology (such as programmable

    logic controllers and computer-controlled pneumatic material transport)

    to select recipes and to speed the flow of products through the two basic

    cooking processes used in the plant. Where the old plant relied on op-

    erators to feed raw ingredients to stand-alone machines and physically

    direct the transfer of the results of each process to the next step, the new

    plant links these refining machines via pneumatically run lines, which

    automatically direct product flows across them in ways that avoid bot-

    tlenecks. Operators in the new plant sit in air-conditioned control rooms,

    directing the process by clicking a mouse on a computer with a 20-inch

    monitor with a graphical display of the entire production process (see

    Zuboff 1988). Changeovers from one product to another used to be very

    time consuming in the old plant; the new plant can accomplish these

    changes much more quickly.4 It would be fair to characterize this company

    as moving from a mass production system to a flexible specialization

    model (Piore and Sabel 1984).

    A glimpse of how management envisioned these changes is shown in

    table 1, which reproduces the overhead slides that management used in

    4 While there are other noteworthy changes between the two plants (e.g., the new plant

    has gone to a just-in-time inventory management system and the introduction of

    statistical process control into most processes), the core of the technological change isthe automation of the two cooking processes and the fact that these processes are

    now linked in a continuous flow.

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    9/48

    Skill-Biased Change

    281

    TABLE 1Top Management Teams Description of Differences

    between Old and New Facilities

    Current Facility Future Facility

    Individual assignments Work group assignments

    Supervised Coached

    Paper system Computer system

    Trained for specific

    function

    Cross-trained for multiple

    functions

    Directed tasks Self-planned tasks

    Data collectors Data interpretation and

    data entry

    Information provider Decisio n maker

    Physical verification Computerized verification

    Samplers TestersQuality (lab controlled) Quality (self-controlled-

    lab audit)

    Old technology New technology

    Problem identifiers Problem solvers

    Bag count system of

    weight control

    Actual weight control

    (load cell/mass meter)

    Physical implementation Computer implementation

    Physical transport of

    materials

    Pneumatic transport of

    materials

    Primarily mechanical

    equipment

    Mechanical and electronic

    equipment

    Reactive Proactive

    Source.Taken from overhead used in 1989 presentation to the

    capital investment committee during the planning phase of retooling.

    a presentation to the parent firms capital allocation committee when

    requesting $92 million to build the new plant. What is interesting about

    these slides is the front-and-center role that is being afforded to human

    resources factors in describing the differences between the plants. Super-

    visors are to be transformed into coaches, and workers are to be moved

    from individual assignments to group assignments. While the machinery

    itself is mentioned occasionally (e.g., Bag Count vs. Load Cell weight

    control), the technology plays a relatively minor role in the description.

    From the earliest phases, then, organizational and human resources fac-

    tors (the software, if you will) were seen as an integral part of the plant

    retooling.

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    10/48

    American Journal of Sociology

    282

    ANALYSIS

    Changes in Job Requirements

    I begin by examining the nature of the changes in production job requi-

    rements associated with the retooling. As described below, the jobs have

    been reorganized so that their relationships to the production process have

    changed as a result of the retooling. Some tasks have been totally elim-

    inated, while others have been combined. Consequently, it is questionable

    that there is enough continuity of job tasks to be able to make sensible

    comparisons at the level of the job title. Instead, I assess the net changes

    in job requirements along various dimensions (see below) for the pro-

    duction department as a whole in the old and retooled plants. Before

    turning to the specific measures of jobs characteristics, it is necessary to

    say a few words about the nature of the work reorganization that wasinstituted with the new technology.

    Work Reorganization

    In the old plant, 195 frontline (i.e., nonsupervisory) workers filled 23

    distinct jobs, and there are 193 workers in 23 jobs corresponding to various

    tasks in the new production process as well. However, this seeming sta-

    bility in the number of job titles masks a substantially altered division

    of labor. The least common change was to have a machine totally replace

    the work of a human being, and only one job was totally automated away

    in this way. The pumpers job was to direct the flow of intermediate

    products such as liquids, oils, and pastes through a labyrinth of tubes,

    sometimes working against gravity, connecting storage tanks and the re-fining machinery. In the new plant, this job is done by a series of rationally

    ordered, dedicated lines with mechanical pigs running through them to

    clean out old products before reusing the lines to transfer new products.

    The more common kind of change is for formerly separate jobs to be

    combined. For example, operators of the stand-alone refining machines,

    which were prevalent in the old plant, were collapsed into the job of

    control room operators. In the new plant, operators sitting in front of

    computer displays can operate the string of refiners by the click of a mouse.

    Operators work in pairs, switching between working the computer and

    walking the production floor, remaining in contact via hand-held two-

    way radios (see Zuboff 1988).

    A third pattern of job change was the exact opposite of this: some job

    responsibilities were more finely elaborated in the new plant than they

    had been in the old plant. The most important examples here are the jobs

    in the new just-in-time warehouse, where formerly general-purpose fork-

    lift drivers have had their tasks become more specialized. Finally, there

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    11/48

    Skill-Biased Change

    283

    are a number of jobs where the role in the division of labor may be quitestable, for example, running one of the finishing machines, which produces

    the material that will be packed and shippedsome of these machines

    were refurbished and literally transferred to the new site. However, in

    these jobs, some auxiliary job tasks might have been added (e.g., there

    is now a requirement to read and interpret a statistical process control

    chart as part of the finishers job). All of the jobs appear to have changed

    in at least one of these ways.5

    In light of the reorganized nature of the jobs, I seek to describe changes

    in job requirements for the plant as a whole. I adopt a triangulation

    strategy to describe the task changes along a number of dimensions by

    using data from a number of sources.

    Dimensions of Job Changes

    Evidence from participant observation.A coding scheme was devel-

    oped in order to summarize the findings from the participant observations

    of the various jobs. These jobs are scored using the procedures followed

    in the Dictionary of Occupational Titles (DOT). The DOT is based on

    observations of jobs in a wide variety of settings across the economy.

    While the DOT is commonly used in studies that attempt to determine

    whether there has been upskilling or deskilling over time (see Spenner

    1990), it has been shown to suffer from a number of problems that limit

    its usefulness in studying changes in job tasks (Cain and Treiman 1981).

    Many of these problems concern patterns of coverage and the fact that

    stability of job titles does not necessarily correspond to stability of tasks.

    Because my design calls for comparisons of tasks over time, I avoid these

    problems.

    For the purposes of summarizing the qualitative field data, the strength

    of the DOT is that it identifies a number of facets of a job based on direct

    observation. I replicated the DOTs procedures and coded seven variables

    for each job in the old and new plants. The seven variables are the extent

    to which jobs involved data, people, or things (DATA, PEOPLE,

    5 Some of the changes have been quite subtle, however. For example, at first glance,the job of feeding raw materials into hoppers at a dump station appears to be virtually

    identical to the task in the old plant, and indeed, workers talk about this job as if it

    is unchanged. Even here, however, my fieldwork uncovered some changes that are

    noteworthy. Where dumpers in the old plant recorded their work on a clipboard locatednext to the workstation, the new plant has eliminated the clipboards and has data

    entry stations with keypads mounted on the wall for recording inventory. Dumpers

    must now log in and record the product code and batch number for the ingredient.This minor change in the job task has large implications for inventory control purposes,

    since the new tasks support the new systems real-time inventory capability.

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    12/48

    American Journal of Sociology

    284

    THINGS), the extent of general educational development required forlanguage, math, and reasoning tasks (GED-L, GED-M, GED-R), and the

    amount of specific vocational preparation required (SVP).

    DATA, PEOPLE, and THINGS are designed to measure job contents

    (i.e., to describe what people do on the job), while GED and SVP are

    supposed to reflect the kinds of skills and knowledge required by the

    average person to achieve an average level of job performance. The scales

    on DATA, PEOPLE, and THINGS are such that lower scores indicate

    more complex tasks. The most complex rating for DATA is a score of 0

    for synthesizing, while comparing is the least complex job activity and

    is coded a 6. PEOPLE varies from mentoring (coded 0) to taking

    instructionshelping (coded 8). The least complex relationship to

    THINGS is handling (coded 7), while the most complex score is for

    setting up (coded 0). The GED variables all range from 1 to 6, withhigher scores indicating greater complexity. For example, a score of 1

    on GED-M corresponds to being able to perform the four basic arithmetic

    operations, while a 6 might involve mathematical statistics or advanced

    calculus. SVP is coded on a nine-point scale where 1 refers to short

    demonstration only and 9 indicates over 10 years (for details on the

    DOT coding scheme, see Cain and Treiman [1981]).

    Table 2 summarizes the average scores on the DOT variables from the

    field observation computed across all production jobs.6 The general pat-

    tern has been one of greater job complexity at time 2 than at time 1 (the

    rows with a indicate where the scores have changed in this direction).

    Given the crude nature of these scales, I caution against reading more

    precision than is warranted into the changes presented in table 2. My

    sense is that the greatest change among the job content measures has

    been along the DATA dimension but that there also has been a general

    intensification of job requirements. This is tracked in table 2, which shows

    that though the greatest change among the job content measures has been

    along the DATA dimension, the averages for all three variables also have

    declined indicating a general intensification of job requirements. A change

    from 4.5 to 2.6 on DATA roughly corresponds to a shift from the average

    job requiring copying and some computing (understood as performing

    calculations) to compiling with some analyzing of the data. Interestingly,

    this observed shift corresponds closely to the job changes prescribed by

    management in table 1 from data collectors to data interpreters and

    data entry.

    6 Note that these scores vary only across job. In order to capture the redistribution of

    personnel that has occurred with the retooling with these measures,each jobis weightedby the number of incumbents when calculating these averages. Thus, these numbers

    represent the job requirements experienced by the average worker at the plant.

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    13/48

    Skill-Biased Change

    285

    TABLE 2Average Changes in Job Requirements Using DOTScores to

    Summarize Field Observation of Jobs for All Production Jobs

    Time 1

    (1991)

    Time 2

    (1994)

    DOT variables:

    Job content:

    Data* . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . 4.5 2.6

    People* . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . 6.8 6.0

    Things* . . . . . . . . . .. . . . . . . . . . . . . . . . . . . 4.4 3.2

    General educational development:

    Language . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 1.8

    Math . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . 1.4 1.7

    Reasoning . .. .. .. .. .. .. .. .. .. .. .. .. .. 2.2 2.6

    Specific vocational preparation . .. 3.7 3.7N of cases . .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. . 195 193

    * Lower numbers indicate greater complexity. Indicates increased complexity at time 2.

    With respect to the skill variables, the averages on all three have in-

    creased, indicating that the required skill levels have risen over the period

    of the study. However, it is important to note that the absolute levels on

    the language and math dimensions remain fairly simple. The changes in

    the math scores capture the observation that the firm has moved from a

    situation where most jobs required only basic arithmetic at time 1 to one

    where many (if not most) jobs at time 2 require workers to be able to

    compute using decimals and to read a graph (e.g., a statistical process

    control graph). Language-related changes also seem modest, but in the

    direction of greater complexity. Direct evidence of the latter changes is

    presented when the reading materials associated with jobs are examined

    below.

    The only variable that has not shown an increase is SVP: at both time

    points, the average training time is between three and six months. There

    has been an increase in all three GED components, so workers in the

    new plant could be expected to be more highly educated than in the old

    plant. However, the stability observed in SVP would imply that on av-

    erage the retooling has not changed the extent to which job skills are

    firm-specific.

    Evidence from the surveys.In addition to the DOT-like measures, I

    also collected survey data to tap job incumbents assessments of the jobskills and training required to do their jobs (see the top half of table 3),

    as well as individuals own human capital characteristics (bottom half of

    table 3). I asked each survey respondent to estimate the number of years

    of formal education needed to perform their job, as well as the number

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    14/48

    TABLE 3

    Average Responses to Survey Items Asking about Education and Training for All Production J

    Tim

    (199

    Job skills and training items:

    Education most people haveHow many years of formal education do most people in jobs like yours

    have?* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

    ( 142

    Formal education neededHow much formal education do you feel is necessary to do your job well?* .. . 9

    ( 150

    How long to trainHow long would it take to train someone to do your job? . . .. . .. . .. . .. . .. . .. . .. . . .. . .. . . 3

    ( 152

    Human capital measures:

    Years of educationHow many years of school did you complete?

    . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11( 152

    Years of experienceAbout how many years have you worked full-time (35 hours or more a week) since

    you were 16 years old?k . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

    ( 152

    Years of tenure# . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

    ( 195

    Note.Ns are given in parentheses.

    * Coded in years. Indicates increases. Response categories: 1 p a few hours, 2 p a few days to a week, 3 p several weeks, 4 p 25 months, 5p 6 months to a year, 6

    7 p five years or more. Responses are recorded in years.k Responses are recorded verbatim, and recoded to years.# Tenure is calculated from the date of hire, which is available for all workers. We calculate the difference in days between the

    administration and date of hire and convert days to years (365.25 days per year). For nonrespondents, the date of survey administrat

    as the midpoint of the survey field period.

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    15/48

    Skill-Biased Change

    287

    of years of education that most people on the job have. I also askedworkers to estimate the length of training period required to learn their

    job. Insofar as the first two items tap sources of skills obtained external

    to the organization, these items are analogous to the general educational

    development concept. The third item refers to on-the-job training and is

    thus meant to correspond to the specific vocational preparation concept.

    I also examine three standard measures of human capital: years of edu-

    cation, years of full-time labor force experience, and years of tenure with

    the current employer.

    Similar to the pattern found in table 2, the only variable in table 3 that

    has not shown an increase is training time (SVP). Workers appear to agree

    with the judgment that the retooling has not changed the extent to which

    job skills are firm-specific. Consistent with fieldwork-based assessments,

    at both time points, workers describe the average training time neededto be between three and six months.

    The data in table 3 are also consistent with those in table 2 in other

    respects. In workers judgments as well, educational requirements have

    gone up over the period of the study. The magnitude of this change is

    about 1.5 years (9.95 vs. 11.49 years), as measured by the formal edu-

    cation needed item. Workers also estimate that the average level of ed-

    ucation for incumbents of their jobs has also risen by about the same

    amount (10.38 vs. 11.85 on the education most people have item).

    The bottom half of table 3 shows that the workforces human capital

    also increased over this period. While average labor force experience in-

    creased only slightly (from 17.26 to 17.37 years), the average years of

    tenure with the company increased more dramatically over this period

    (from 9.55 to 10.85 years). The average education of the workforce also

    rose over time by about one-half of a year (11.43 vs. 11.75).

    It is interesting to compare how the relationship between education

    and the job skill measures changed over time. Respondents described the

    plant at time 1 as employing workers who were overeducated for their

    job tasks by over one year on average (11.43 years of education for in-

    cumbents compared with 9.95 years of education required to do the job).

    At time 1, respondents underestimate the actual years of education of

    those working in their jobs by about a year (10.38 vs. 11.43). However,

    in the new plant, these relationships have changed considerably. First,

    time 2 respondents estimates are much more accurate with respect to the

    actual educational composition of the workforce (11.85 vs. 11.75). Second,

    the difference between workers judgments of education required andactual education have virtually disappeared (11.75 vs. 11.49). Taken to-

    gether, these data suggest that, at least in workers perceptions, the re-

    tooling has served to eliminate their sense of being ahead of the educa-

    tional requirements for their jobs.

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    16/48

    American Journal of Sociology

    288

    Changes in basic skills requirements: survey evidence.In order toaddress the policy debates on the economic role of basic skills (see Car-

    nevale, Gainer, and Meltzer 1990; Hunt 1995; Murnane and Levy 1996),

    data was collected on literacy and numeracy. Respondents were asked

    about their use of reading, writing, addition and subtraction, multipli-

    cation and division, fractions, and percentages. Responses to the survey

    conducted before the retooling then were compared to responses to the

    same items asked in the follow-up survey conducted after the retooling

    (table 4). The results are clear cut: across all items, respondents report an

    increase in the use of basic literacy and numeracy skills on the job over

    time.

    Changes in basic skills requirements: evidence from job documents.In

    addition to the survey measures of basic skills, I collected an innovative

    set of measures of jobs cognitive demands using direct observations ofworkers on the job. During the fieldwork, I discovered that virtually every

    workstation had a set of documents (usually attached to clipboards) that

    were being used on the job. Workers used these documents in a variety

    of ways, sometimes simply in a lookup fashion (e.g., metric conversions),

    or perhaps to record job tasks performed (e.g., recording the number of

    boxes on a pallet). Often, workers might be required to handle multiple

    forms. As I watched workers interact with these documents, I realized

    that I was observing on-the-job literacy acts. This allowed me to bring

    a fresh perspective to the basic skills debates. Rather than speculate on

    what level of education or score on a test is required to do various jobs

    (Hunt 1995), I was in the enviable position of observing literacy requi-

    rements and how these requirements would change with the retooling.

    Consequently, at the end of the fieldwork periods, I collected censuses of

    the various documents that were used in both plants.

    While this is a tremendous opportunity to inform the basic skills lit-

    erature, there was a challenge in coming up with a way to characterize

    the changes in the documents over time. I identified a set of studies by

    Kirsch and his associates (e.g., Kirsch and Mosenthal 1990) that develop

    a flexible grammar for parsing forms in everyday use (e.g., a train sched-

    ule). This system is called document literacy, and it summarizes the cog-

    nitive complexity of a wide variety of visual materials (e.g., tables and

    graphs) into a small number of dimensions, irrespective of mode of pre-

    sentation (e.g., paper or computer screen). Because of its flexibility and

    generality, the document literacy approach is adopted for coding the on-

    the-job documents used in the two factories.Before turning to the cognitive complexity measures, it is important to

    point out that there has been a huge increase in the number of documents

    in the plant. In the old plant, there were 16 distinct job forms, which

    were distributed across the various workstations in different combina-

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    17/48

    Skill-Biased Change

    289

    TABLE 4Average Responses to Survey Items Asking about the on-

    the-Job Use of Basic Skills for All Production Jobs

    Time 1

    (1991)

    Time 2

    (1994)

    Survey items:* ....................

    Reading . .. .. .. .. .. .. .. .. .. .. .. .. 3.48 3.78

    Writing . .. .. .. .. .. .. .. .. .. .. .. .. 3.58 3.80

    Addition and subtraction . .. .. 3.34 3.47

    Multiplication and division . . . 2.77 3.23

    Fractions . .. .. .. .. .. .. .. .. .. .. .. 1.97 2.23

    Percentages . .. .. .. .. .. .. .. .. .. .. 2.24 2.68

    N of cases . . . . . . . . . . . . . . . . . . . . . . . . . 151 154

    * The exact question wording is How much [of the following] do you have

    to do on your job? Would you say . . . (1 p none at all, 5 p a lot). Indicates increases.

    tions. The maximum number any workstations clipboard contained was

    six, and the minimum was zero. In the new plant, 202 distinct forms were

    identified; every workstation had at least one form, and the maximum

    number of forms for any one workstation was 30. While this increase may

    itself seem staggering, the 202 figure may actuallyunderestimate the extent

    to which documentary material has been made part of production work-

    ers job requirements since it does not reflect the reading demands in-

    volved in dealing with the various computer screens that were introduced

    into the new plant. It is perhaps not surprising that the number of paper

    documents has increased with the introduction of computers, since one

    of the things at which computers are very good is document production.

    Much of the increase in paper forms is directly linked to the introduction

    of various computer systems since many of the forms in the new plant

    are either computer-generated (e.g., statistical process control charts) or

    involve directions on the use of computers (e.g., instructions posted next

    to smart keypads).

    The document literacy variables reflect the structure and complexity

    of the document and the way in which the reader uses the document.

    Documents structure and complexity are measured by the number of

    specific pieces of information being referred to in the document (specifics)

    and organizing categories (labels), which serve to classify or summarize

    specifics. The number of specifics is a measure of the length and amount

    of material in a document; as the number of specifics increase, thedifficultyof the document increases. Since labels group together specifics, they sim-

    plify documents and make them easier to understand.

    With respect to measuring the ways in which the document is used,

    the document literacy system defines a hierarchy of complexity on the

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    18/48

    American Journal of Sociology

    290

    basis of what is done with the document. I coded each documents strategyrating in a manner consistent with this hierarchy. (If a particular form

    asked for multiple tasks of varying levels of difficulty, I assigned each

    form the strategy rating for the most difficult task.) Simply locating in-

    formation is the simplest task (strategy rating of 1). The next most

    complex task is a lookup and comparison (coded 2). Next is a lookup

    and comparison that is conditional on textual information literally and

    explicitly contained in the same document, as might be the case when

    directions are clearly provided (coded 3). A code of 4 on strategy

    indicates that the conditional information defining the lookup and com-

    parison is not literally in the document. Directions might be included,

    but those directions would not use literally the same words to guide the

    user. Consequently, level 4 documents require the user to make an in-

    ference about the applicability of the directions. The highest level of thehierarchy (a strategy rating of 5) makes the lookup and comparison

    conditional on special prior knowledge that is not clued, literally or oth-

    erwise, by the document itself. This would be the case if the document

    were to be used conditionally but there was no clarifying information

    provided on when to use it.

    As a further refinement to the strategy rating, I distinguished whether

    the document required that the procedure be repeated. If the document

    called for the most complex procedure to be repeated, I added 0.50 to the

    strategy score. For example, a document that required the user to re-

    peatedly perform unconditional lookup-and-compare (level 2) tasks

    would be given a strategy rating of 2.5. The shift in warehouse procedures

    provides a good example of this change. In the old plant, forklift operators

    would be handed a bill of lading and asked to retrieve the item on the

    form, for example, a pallet of a certain product (a single lookup and

    compare). When the operator was done retrieving the product, he (it was

    always a he) would hand the bill of lading to the woman in the shipping

    office (it was always a woman) who would put the form on the stack of

    forms she had next to the terminal awaiting entry into the system. In the

    new plant, the woman is gone, there is no pile, and when the operator is

    done retrieving the pallet, it is his or her responsibility to enter the in-

    formation from the bill of lading into the centralized inventory control

    system. This would typically involve a repeated lookup and compare, as

    the information on the formthe product number, the client, the desti-

    nation, and so onis cycled through and made to conform to the requi-

    rements of the inventory system.

    7

    7 It is interesting to note that in this example, it is the way the document is used thatchanged, not the document itself (indeed, the typical bill of lading did not change over

    this period). However, adding formerly clerical tasks to the forklift operators job

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    19/48

    Skill-Biased Change

    291

    As a final measure of how the document is used, each document isexamined for the presence or absence of distractors. Distractors pull the

    users attention away from its proper use, inviting confusion. Train sched-

    ules, for example, are everyday documents with many distractors. Because

    train schedules try to report the entire weekly schedule on a single piece

    of paper, they present lots of opportunities for confusion and misuse (e.g.,

    being on the wrong page of the weekend vs. holiday vs. weekday parts

    of the schedule). Forms containing distractors require that users pay closer

    attention in order to use them properly; consequently, they are deemed

    more cognitively complex by the document literacy system.

    For both time points, a measure of the aggregate levels of document

    literacy needed in each of the plants was developed. I scored all the forms

    using the document literacy measures and then matched each form to the

    jobs that require the form. The four document literacy variables werethen averaged across all the forms corresponding to each job. This yields

    job-level measures of the average document literacy required for each

    job. In order to measure the aggregate levels of document literacy required

    in the plant as a whole, and to be consistent with our treatment of the

    DOT-style measures, I weighted the job-level document literacy averages

    by the number of people in each job.

    Table 5 shows the changes in document literacy measures that occurred

    over the course of the study. There has been a nearly fourfold increase

    in the average number of documents that production workers are being

    asked to use on the job (2.63 vs. 10.32). With respect to the cognitive

    complexity measures, the job forms have become more complex on all

    four dimensions over time. The average number of specific pieces of in-

    formation requested has increased dramatically (from 21.88 to 70.75), also

    indicating a huge increase in the volume of information that workers are

    being asked to process. At the same time, the number of labels has gone

    down by almost half (31.84 vs. 60.49). Since labels serve to organize and

    simplify the presentation of information, this too indicates an increase in

    the cognitive complexity demands of the job. The strategy rating has also

    increased (from 2.14 to 2.47), suggesting that the user of the document is

    being asked to do somewhat more demanding tasks with the forms. In

    rough terms, this corresponds to moving from a lookup and compare at

    time 1 to a repeated application of a lookup and compare at time 2. As

    such, this indicates that the density of interaction with documents has

    gone up over time. Finally, the percentage of documents with distractors

    also has increased over time. I observed a 6.7% increase in the prevalenceof documents containing distractors. This suggests that the time 2 doc-

    requirements has increased the document literacy requirement of this job along this

    dimension.

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    20/48

    American Journal of Sociology

    292

    TABLE 5Document Literacy Scores for Production Jobs

    Time 1

    (1991)

    Time 2

    (1994)

    N of documents . . . . . . . . . 2.63 10.32*

    N of specifics . .. .. .. .. .. . 21.88 70.75*

    N of labels . . . . . . . . . . . . . . 60.49 31.84*

    %with distractors . .. .. .. 50.5 57.2*

    Mean strategy rating . .. 2.14 2.47*

    N of cases . . . . . . . . . . . . . . . 195 187

    * Indicates increase in complexity. Lower numbers indicate greater complexity. We did not obtain forms from the quality control lab ( N p 6).

    uments require that workers pay closer attention in using the forms than

    when they were using the time 1 documents. Along all of these dimensions,

    then, the document literacy analysis strongly supports the inference that

    both the volume and complexity of the documents have increased over

    time.

    Changes in use and knowledge of computers.I assessed changes in

    workers use and knowledge of computers by a series of survey items at

    both time points. I asked workers to report on their use of calculators

    and computers during the course of their work. Table 6 shows the changes

    for these items over time. These results quantify what was seen quali-

    tatively in the fieldwork. At time 1, 83.4% of respondents answered that

    they never used a computer on the job; by time 2, the percentage of

    workers reporting no computer use at all had dropped to 9.8%. At the

    other end of the spectrum of computer use, only 5.3% of time 1 workers

    said they always used computers on the job, compared to 29.4% of time

    2 workers. Use of an electronic calculator on the job also grew over this

    time period from 37.7% to 53.2%. Over half (56.3%) of the time 1 work-

    force used neithera calculator nor a computer on the job. By time 2, this

    percentage had dropped to 8.5%. Clearly, the introduction of computing

    equipment into the work process has been widespread over this period.

    This pattern is also consistent with the evidence presented in table 4 that

    the amount of basic math used on the job increased as well.

    Summary of Job Changes

    The data analyzed thus far tell a consistent story. Across all three data

    sourcesparticipation observation, surveys, job documentswhen mea-

    sured at the mean, job skill requirements have increased over time. In

    light of the need to combine the numerous data sources and the special

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    21/48

    Skill-Biased Change

    293

    TABLE 6Survey Results of Workers Use of Calculators and Computers,

    All Production Jobs

    Time 1

    (1991)

    Time 2

    (1994)

    On the job use:

    %using a calculator* . .. .. .. .. .. .. .. .. .. .. 37.7 53.2

    (151) (154)

    %never using computer . . . . . . . . . . . . . . . . . 83.4 9.8

    (151) (153)

    %always using computer . . . . . . . . . . . . . . . . 5.3 29.4

    (151) (153)

    %not using calculator nor computer . . . 56.3 8.5

    (151) (153)

    Note.Ns are given in parentheses.

    * The exact question wording is Do you use an electronic calculator on the job? (1

    p yes, 0 p no). Indicates increases in computer use. The exact question wording is How often do you use a computer in the course of

    your work at [NAME OF COMPANY]? (15 scale Likert scale where 1 p never, 5p

    always). Percentage reporting no to the item on calculators and never to the item on computer

    use.

    nature of the sample, I use a sign test as a conservative statistical test of

    the pattern of results (see the appendix). Of the 28 job skill measures, 26

    agree in the direction of the changes over time, with the only exceptions

    to this pattern being the two measures of SVP. The sign test shows that

    a pattern this extreme is very unlikely to be due to chance. 8 At least in

    this case, the retooling appears to have been implemented in an upskill-ing manner.

    While job requirements have increased across multiple skill dimensions

    when comparing changes at the means of these various measures, the

    effects of the retooling on other points of the distribution within each skill

    dimension have yet to be examined. Changes at the mean could be con-

    sistent with very different scenarios with very different implications for

    wage inequality. On the one hand, it could be that the retooling has mainly

    served to increase the job requirements for jobs in the upper tails of the

    various skill distributions. Such a pattern might disproportionately influ-

    ence high-wage workers. On the other hand, it is possible that the main

    impact of the technical change has been felt mainly on the bottom ends

    of the distributions of job requirements, thereby mainly affecting low-

    8 By using the sign test, we are treating the over-time comparisons of means of theskill dimensions as analogous to flips of an unbiased coin. The chances of getting 26

    heads out of 28 flips are fewer than one in a million (sign test, P ! .0000001).

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    22/48

    American Journal of Sociology

    294

    wage workers. This section on job changes closes by addressing thisquestion.

    Table 7 presents the results of over-time comparisons of various points

    in the percentile distribution for various skill measures. I omit the DOT-

    style variables coded from the qualitative field data since they are not

    fine-grained enough to support analysis of these distributional features. I

    also modify the analyses of the basic skills survey items reported in table

    4. While the individual items adequately capture central tendencies and

    reasonably support inferences about qualitative (positive or negative)

    changes over time, they are measured on five-point Likert scales and

    therefore provide only coarse information on the rest of the distribution. 9

    I combine the basic skills survey items to form a more finely graded scale

    by taking the average of the six items.10 I describe changes in the extremes

    of the distributions (i.e., the fifth and ninety-fifth percentiles), as well aschanges occurring at each decile of the various skill dimensions. For each

    skill measure, I denote with a whether the values marking each

    percentile distribution point are strictly greater at time 2 than at time 1. 11

    The first row of table 7 shows, with few exceptions, increases in the

    self-reported measures of basic skills (sign test, P ! .033). Similarly broad

    changes are evident for two of the document literacy measures. The pat-

    terns for number of documents and number of labels are statistically

    reliable (sign tests, P ! .0005 and P ! .033, respectively).12 The patterns

    for these measures support the inference that upskilling has occurred

    across the board and has not been localized in either the top or bottom

    tails of the various dimensions of skills.

    In marked contrast, the measure of on-the-job training (row 7 of table

    7) shows no evidence of change at all. The stability of the training measure

    9 While the percentage shifts over time are clear for each point on the Likert scale for

    the computer use item (for time 1 vs. time 2, 1: 83.4% vs. 9.8%; 2: 3.3% vs. 9.8%; 3:

    4.0% vs. 17.0%; 4: 4.0% vs. 34.0%; 5: 5.3% vs. 29.4%), these items are not measured

    finely enough to support the percentile point comparisons we present in table 7.10 Factor analyses of the six items showed that the basic skills items combine to form

    one scale, loading on one dimension in very similar ways at each time point. Bothbasic skills scales are highly reliable, as shown by Cronbach alpha reliability coefficients

    of 0.876 and 0.838 for time1 and time 2, respectively.11 While I think it is important to examine the changes occurring across the distribution,one should bear in mind that the sample size increases the risk of slicing the data too

    thinly in this exercise. Depending on the particular measure, each percentile point

    comparison is based on 1518 cases, with the fifth and ninety-fifth percentile com-

    parisons based on even fewer cases. The limits of what this data can tell us areapproached in table 7.12 The changes for the other document literacy measures are more inconsistent and arenot statistically reliable as measured by sign tests (number of specifics P ! .113; dis-

    tractors and strategy rating are P ! .500).

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    23/48

    Skill-Biased Change

    295

    TABLE 7Pattern of Over-Time Changes in Skills for Various Percentile Points

    Percentiles

    5 th 1 0t h 2 0t h 3 0t h 40 th 5 0t h 6 0t h 7 0t h 80 th 9 0t h 9 5t h

    Basic skills measures:

    Survey scale* .. .. .. .. .. .

    Document literacy:

    N of documents .........

    N of specifics ............

    N of l abels . . . . . . . . . . . . . .

    %with distractors .......

    Mean strategy rating . . .

    Education items:

    How long to train .. .. ..

    Education most

    people have . .. .. .. .. ..

    Education needed .. .. ..

    Human capital:

    Years of education .. .. .

    Years of experience . .. ..

    Years of tenure .. .. .. .. .

    Note. indicates increase.

    * Scale based on average of six items in table 5. Cronbachs a p .876 (time 1), .838 (time 2).

    is evident across the entire distribution, not just at the mean (the chance

    of 0 increases out of 11 comparisons is less than 5 in 100,000, sign-test P

    ! .00049). The education and experience measures, however, show a bi-

    furcated pattern, with changes most evident at the bottom and at the

    very top of the distributions. Changes in workers reports of how much

    formal education is needed and how much education most people have

    in their jobs show a gap between the fiftieth and ninetieth percentiles.

    Changes in the actual educational composition of the workforce show an

    even more bifurcated pattern. For this variable, increases are evident at

    the fifth through the twentieth and at the eightieth percentile points. A

    less dramatic change in the measured education level of the workforce

    would be expected than in the more subjective education measures. Work-

    ers reported that they were overeducated at time 1 (see table 3), suggesting

    that there is room for job requirements to rise without a corresponding

    adjustment in the composition of the workforce. To the extent there have

    been changes in the educational composition of the work force, they ap-

    pear to have been concentrated in the extremes of the distribution. Sincevery few workers were engaged in formal education between time 1 and

    time 2, these changes are a result of replacing workers who turned over

    during the period of the study.

    The patterns in table 7 suggest that the retooling has had a far-reaching

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    24/48

    American Journal of Sociology

    296

    impact on skill requirements. Considering the data in table 7 as a whole,the upskilling appears to have occurred across the board. Of the 132

    percentile points compared in table 7, 81 show increases. The chances of

    observing a result this extreme when changes over time are random are

    quite small (sign test, P ! .0057). However, the question remains whether

    the changes have been differentially concentrated in the top or the bottom

    of these skill distributions. Examining above the median (i.e., the sixtieth

    through the ninety-fifth percentiles), 36 of 60 comparisons show increases,

    yielding a P value of 0.077. Below the median, 37 of 60 percentile points

    increase (P ! .046). These results suggest that the upskilling shifts have

    occurred across the board and have not been localized either below or

    above the median. Removing from consideration the on-the-job training

    variable, which shows no changes at all across the entire distribution,further strengthens the inference. After omitting this variable, 36 of 55

    percentile points changed in the direction of upskilling above the median

    (P ! .015); below the median, 37 of 55 percentile points indicated upskilling

    (P ! .007).

    As a final point, the fact that there has been an across-the-board upward

    shift for many of these job skills is quite consistent with managers im-

    pressions and behavior. From the interviews, I learned that management

    had a broad, multidimensional conception of the nature of the job changes

    and that they also conceived of these changes as an overall upgrading

    of the job requirements. Moreover, the companys management saw a

    need to retrain many workers in advance of these changes. While the

    organization and execution of the retraining effort constitutes a study in

    itself, one feature of this effort is particularly noteworthy in this context.

    Management not only talked of the job changes as being of an across-

    the-board nature, they also acted in a manner consistent with this con-

    ception when they designed the retraining program. Most of the retraining

    effort was not targeted on particular individuals or groups of workers.

    Training in basic skills and new broadly used processes (e.g., statistical

    process control) was given to all production workers. (The company also

    tried to offer all production workers general training in computers before

    they ran short of money. About 75% of the production department re-

    ceived this training). While training focusing specifically on the new cap-

    ital equipment was targeted to the operators who were most likely to be

    running the new machinery, the fact that the company invested significant

    training resources across the entire workforce is consistent with our in-

    terpretation of the patterns of table 7. From managements perspective,

    it is clear that significant changes were expected across the entire range

    of job requirements.

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    25/48

    Skill-Biased Change

    297

    Changes in the Wage Distribution

    Although management has conceived of the new plant as having more

    stringent job requirements across the broad spectrum of jobs, contrary to

    the conventional economic model, these increased job requirements were

    not rewarded with higher wages overall: both the mean and median wage

    for hourly workers barely changed in real terms between 1991 and 1994

    (see table 8). There were, however, marked changes in the dispersion of

    wages over this period. The standard deviation of wages increased over

    30% ($1.78 vs. $2.32), and the interquartile range (the difference between

    the seventy-fifth and twenty-fifth percentiles) jumped over 82% ($1.31 vs.

    $2.39) between 1991 and 1994.

    Further examination shows that the increased wage inequality is due

    to a twisting of the wage distribution around the median, where wagesbelow the median fell, while wages above the median increased. Figure

    1 gives a picture of the changes in the wage distribution: it plots each

    percentile of the 1991 wage distribution (along the x-axis) against the 1994

    wage distribution (the y-axis). The 45-degree, dotted line shows the base-

    line of no change in the wage distribution over this period. Although the

    pattern is choppy, the overall pattern is S-shaped, with wages below the

    median dropping below the 45-degree, dotted line and wages above the

    median rising above the line, especially those wages at the very top of

    the distribution (i.e., those above the ninetieth percentile). This S-shaped

    pattern tracks the changes that have been occurring in the economy as

    a whole: in real terms, pay for low-wage workers has dropped, while pay

    for high-wage workers has increased.

    It is important to note that in contrast to most other analyses of thesechanges that look across many firms and sectors, the changes documented

    here have occurred within the same firm. While many studies touting the

    skill-biased technological change explanation of increasing wage ine-

    quality have argued that such changes must be operating within firms,

    to our knowledge, this study is the first to have demonstrated the predicted

    pattern of increasing wage inequality at the level of the firm. Moreover,

    I have found this pattern in a firm that has undergone a massive retooling

    that has dramatically altered job requirements. As shown below, the re-

    lationship between these job and wage changes is not simply coincidental.

    While this study has shown evidence that the wage distribution changed

    over time in precisely the way predicted by advocates of the skill-bias

    interpretation of growing inequality, a number of puzzles remain to be

    resolved before this account of these findings is accepted.

    While a pattern of wage change that is consistent with the skill-bias

    explanation of increasing wage inequality has been found, the question

    arises whether the changes in wages are statistically reliable. Could these

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    26/48

    American Journal of Sociology

    298

    TABLE 8Hourly Wages for Hourly Production Workers (in 1991 Dollars)

    Before Retooling

    (1991)

    After Retooling

    (1994)

    %Increase

    (199194)

    Minimum . . . . . . . . . . . . . 6.96 6.95 .1

    Percentile:

    5th . . . . . . . . . . . . . . . . . . 8.32 8.02 3.6

    10th . . . . . . . . . . . . . . . . . 8.32 8.16 1.9

    20th . . . . . . . . . . . . . . . . . 9.31 8.21 11.8

    30th . . . . . . . . . . . . . . . . . 9.56 8.96 6.3

    40th . . . . . . . . . . . . . . . . . 9.81 9.23 5.9

    50th . . . . . . . . . . . . . . . . . 10.01 10.06 .5

    60th . . . . . . . . . . . . . . . . . 10.17 10.44 2.6

    70th . . . . . . . . . . . . . . . . . 10.47 10.62 1.4

    80th . . . . . . . . . . . . . . . . . 10.91 10.86 .590th . . . . . . . . . . . . . . . . . 13.25 13.77 3.9

    95th . . . . . . . . . . . . . . . . . 15.00 15.88 5.8

    Maximum . . . . . . . . . . . . . 16.48 20.65 25.3

    Mean . . . . . . . . . . . . . . . . . . 10.30 10.23 .6

    Median . . . . . . . . . . . . . . . . 10.01 10.06 .5

    SD . . . . . . . . . . . . . . . . . . . . . 1.78 2.32 30.3

    Interquartile range . . . 1.31 2.39 82.4

    Valid N of cases . . . . . . 195 187

    patterns to be due to chance? As discussed in the appendix, the approach

    taken in this study to assessing the statistical significance is to use boot-

    strap procedures to determine how common it would be to find results

    that contradict the observed pattern in random resamplings of the data

    (Efron and Tibshirani 1986). The first inference tested in this manner is

    whether wage inequality increased between 1991 and 1994. For both

    measures of inequality (the standard deviation and the interquartile

    range), I seek to determine how safe it is to infer that wage inequality

    has increased over this period. I ran 1,000 bootstrap samplings based on

    the original data and found that the interquartile range for 1991 never

    exceeded the corresponding measure for 1994; thus, the P value is less

    than 0.001. With respect to the standard deviation, the 1991 standard

    deviation of wages exceeded the 1994 standard deviation twice in 1,000

    random samples, yielding a P ! .003. This leads to the conclusion that

    the increase in wage inequality observed over this period is not sensitive

    to the random inclusion (or exclusion) of a few cases in the study.

    While it is clear that wage inequality has increased in these data, theresults of the tests just reported do not specify the particular form of the

    growth in inequality. For example, it is quite possible that the changes

    in inequality might have grown differentially at the top, rather than at

    the bottom of the wage distribution. Indeed, visual inspection of figure 1

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    27/48

    Skill-Biased Change

    299

    Fig. 1.Change in wage distribution, 199194

    seems to suggest such a pattern. In order to locate where the increase in

    inequality is coming from, a test of whether the S-shaped form in figure

    1 is statistically reliable needs to be specified.

    I define a test statistic that summarizes the degree to which the changes

    in the wage distribution follow the hypothesized S-shape. I take a ratio

    between 99 percentile points taken from the 1991 and 1994 wage distri-

    butions (i.e., the first to the forty-ninth and the fifty-first to the 100th,

    excluding the fiftieth). In order to assess whether these changes conform

    to the hypothesized S-shaped change, below the median, I compute the

    ratio of the 1991 percentile points to the 1994 percentile points. This

    measures the extent to which 1991 wages exceed 1994 wages below the

    median. Above the median, the comparison is reversed by taking a ratio

    of the 1994 to 1991 percentile points and measuring the degree to which

    1994 wages exceed 1991 wages. I then average across the 99 percentile

    points to get a summary measure of the change. I also compute two

    submeasures of this statistic in order to assess the extent to which thechanges in the wage distribution are localized below the median (the first

    to the forty-ninth) or above the median (the fifty-first to 100th). I then

    test whether the measures are statistically reliable by applying bootstrap

    techniques.

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    28/48

    American Journal of Sociology

    300

    Test statistic for entire distribution:

    W91 W91 W911st 2nd 49thAverage : ( ) ( ) ( )[ W94 W94 W941st 2nd 49thW94 W94 W9451st 52nd 100th , (1)( ) ( ) ( )]W91 W91 W9151st 52nd 100th

    below median:

    W91 W91 W911st 2nd 49thAverage : , (2)( ) ( ) ( )[ ]W94 W94 W941st 2nd 49thand above median:

    W94 W94 W9451st 52nd 100thAverage : , (3)( ) ( ) ( )[ ]W91 W91 W9151st 52nd 100thwhere corresponds to the nth percentile of the 1991 wage distri-W91

    n

    bution, and refers to the nth percentile of the 1994 wage distribution.W94n

    The test statistics are one when there has been no change in the wage

    distribution (i.e., the 45-degree line in fig. 1), greater than one when the

    numerators are greater than the denominators (corresponding to an S-

    shape), and negative when the denominators are greater than the nu-

    merators (a reverse S-shape).

    Several points are worth noting about these measures. First, the mea-

    sures are equal to 1 when there has been no change in the percentile

    points, so they have a natural baseline to compare to as a null model.

    Second, as averages of ratios, they measure average proportionate changes

    over time in the hypothesized direction, down below the median, and up

    above the median. As such, the measures become more positive as the

    data depart further from the 45-degree line in figure 1 in an S-shaped

    pattern.

    Beginning with the test statistic for the entire wage distribution, I find

    that the observed value of the statistic is 1.045, indicating that the per-

    centiles have shifted an average of 4.5% in the direction of an S-shape

    over time (table 9). I tested whether this statistic is reliably different from

    1, the expected value of the statistic under the null model of no change

    over time (i.e., the 45-degree line in fig. 1). In 1,000 bootstrap replications,

    the test statistic was never one or less, suggesting that the observedS

    -shaped change in the wage distribution is very unlikely to be due to

    chance.

    Separate test statistics for the changes occurring above and below the

    median are also computed. Below the median, the observed value of the

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    29/48

    Skill-Biased Change

    301

    TABLE 9Bootstrap Significance Test of the S-Shaped Change in Wage Distribution,

    199194

    Entire

    Distribution (Eq. 1)

    Below

    Median (Eq. 2)

    Above

    Median (Eq. 3)

    Observed test statistic . . . . 1.045 1.059 1.030

    Bootstrap SE . . . . . . . . . . . . . . .010 .016 .020

    Statistical significance* . . . .001 .001 .052

    Note.Data are from 1,000 bootstrap replications. N p 195 (1991), 187 (1994).

    * Proportion of bootstrap samples where the test statistic is less than or equal to one.

    statistic is 1.059, and values of 1 or less never occur in 1,000 bootstrap

    replications. Consequently, the inference that wages below the median

    dropped over the period of the study is robust to random perturbationsof the original data. The pattern observed above the median reveals an

    average 3.1% shift in the direction of increasing wage inequality. The

    bootstrap significance test shows that violations of the hypothesized S-

    shaped change (i.e., statistics of one or less) occurred 51 out of 1,000 times

    yielding a P value of 0.052. While less impressive than the statistical

    significance level computed for the data below the median, this pattern

    is still quite robust considering the small number of cases that are being

    used to estimate changes across a broad span of the wage distribution.

    A closer look at table 8 and figure 1 reveals that the largest wage changes

    occurred at the very top of the wage of distribution, that is, from the

    ninetieth percentile and above. Even with this thin case base, however,

    the magnitudes of the wage increases that occurred above the ninetieth

    percentile are so large that we can be reasonably confident that wages at

    the very top of the distribution went up in real terms over the period of

    our study. At the risk of slicing the data too finely, I recalculated statistics

    based on changes above the ninetieth percentile and above the ninety-

    fifth percentile. At the ninetieth percentile and above, the 1994 wages

    exceed the 1991 wages by an average of 8.5%. The bootstrap tests show

    that the data did not follow the hypothesized pattern 10 times out of 1,000

    bootstrap replications, yielding a P ! .011. Looking at the ninety-fifth

    percentile and above, wages rose an average of 12.2%. The bootstrap test

    shows only two violations of the hypothesis out of 1,000 trials for a P !

    .003. Despite the small number of cases, there does appear to be a sta-

    tistically reliable tendency for the wages at the very top of the distribution

    to have risen over this period.The jobs that account for this patternthose jobs that are the highest

    paid at both time pointsare the maintenance mechanics and mainte-

    nance electricians. These personnel are charged with repairing and main-

    taining the industrial machinery in the plant, and they are the recognized

  • 7/29/2019 Fernandez M., Roberto - Skill-Biased Technological Change An

    30/48

    American Journal of Sociology

    302

    elite of the plants hourly workforce. They are highly trained, possessjourneymen skills, and, as a consequence, are paid off the union scale at

    both time points.

    Thus far evidence has been shown that the wage distribution changed

    over time in precisely the way predicted by advocates of the skill-bias

    interpretation of growing wage inequality. One of the announced advan-

    tages of the case-study approach is the ability to open up the black box

    and observe the mechanisms by which these changes take place. How

    did these changes come about?

    Turnover plays a key role in these wage changes. Of the 195 production

    workers employed at the factory at time 1, 43 (22%) had left the by the

    time of the second wave of the study three years later.13 While this rate

    is in line with the historical pattern of turnover in this company,14 the

    shift in the composition of the company has had important effects on thewage distribution. These changes are seen clearly by comparing the wage

    data for the entire population of workers (table 8) with the wage data for

    those workers who stayed through the transition between 1991 and 1994

    (table 10), which remove from consideration the data for the time 1 leavers

    and the time 2 new entrants.15

    Comparing the time 1 data for both tables, neither measure of central

    tendency (the mean and the median) changes very much. Nor do the

    inequality measures (the standard deviation and the interquartile range),

    indicating that the people who are turning over are not concentrated in

    any particular part of the wage distribution, but instead are leaving from

    across the wage distribution. While the impact of the time 1 leavers on

    the wage distribution has been relatively even, their replacements (the

    13 In keeping with the companys pledge, none of these terminations were due to layoffs;

    9 occurred through retirement, 1 through death, 16 were quits, and 17 were fired for

    cause. In addition to these 43, two cases were transferred to another facility within

    the company, and two cases were promoted out of the blue-collar ranks. These fourcases are not terminatio


Recommended