Technology and the Wage Structure

This paper reports direct evidence on how recent changes in technology are related to changes in wage gaps by schooling, experience, and gender. Wage gaps by schooling in the full-year 1979 and 1989 Current Population Surveys increased the most in industries with rising R&D intensity and accelerating growth in the capital-labor ratio. Estimates of the impact of hightech capital are mixed, with wage gaps increasing the most in industries that were high-techcapital intensive in 1979 and decreasing in the industries with the greatest increase in high-techcapital intensity. Contrary to popular notions that technological change harms older workers, wage growth of experienced workers is much greater in R&D-intensive industries than in industries with little R&D activity. The gender gap narrowed more in industries that most intensively used high-tech capital in 1979, especially among younger women with a high school degree or some college. The impact of technological change on the skill mix of the labor force is one of the oldest questions in the social sciences. It is now receiving considerable attention because wage differentials by skill have widened significantly since 1980 and because there is much indirect evidence indicating that technological change is responsible (Katz and Murphy (1992); Juhn, Murphy, and Pierce (1993), Bound and Johnson (1992)). Berman, Bound, and Griliches (1994) find large within-industry increases in the share of nonproduction workers in U.S. manufacturing despite relative price shifts encouraging the opposite pattern. Their results also show that increases in the share of nonproduction workers are associated with R&D and computer investment. Berman, Bound, and Machin (1998) extend this work over a sample of OECD countries and find the largest increases in nonproduction shares tend to occur in the same industries. Autor, Katz, and Krueger (1998) find in the U.S. that the increased demand for skilled workers has been taking place over the last five decades and accelerated in the 1980s in manufacturing. They also find that the most skill upgrading took place in the industries that are most computer-intensive. These results support the hypothesis that skill-biased technological change (SBTC) has taken place, but do not deal directly with the question of whether SBTC is associated with changes in relative wages. There is very little evidence relating technology indicators to changes in the wage structure. Bartel and Sicherman (1999) carefully examine the relationship between technology variables and wage levels using the 1979-93 National Longitudinal Surveys, but their work does not consider how such variables relate to changes in the wage structure over time. Mincer (1991) showed that relative earnings of college to high school graduates with 6-10 years of experience increased with R&D intensity in aggregate, annual data for 1963 to 1987. Krueger (1993) found that workers using computers at the workplace earn 15 percent more than those who do not use computers. Because college graduates are more likely to use computers at work than workers with less schooling, Krueger contends that increased use of computers 2 accounts for one-third to one-half of the observed increase in returns to schooling between 1984 and 1989. This earnings premium also could reflect unobserved ability, if more able workers are assigned to jobs using computers (DiNardo and Pischke (1997)). Card, Kramarz, and Lemieux (1996) find wage growth in the 1980s was associated with computer use in 1989 for both men and women, using data aggregated by age and schooling. Studies using establishment data in manufacturing to examine the relationship between wage growth and the use of a wide range of advanced technologies at the end of the sample period (Chennells and Van Reenen (1997), Entorf and Kramarz (1997), and Doms, Dunne and Troske (1997)) have found no relationship. However, these studies lack controls for the use of advanced technology at the beginning of the sample period. Despite traditional concerns that technological change can make the skills of older, experienced workers obsolete, the question of how technology has been affecting such workers has not received much attention. Wage differentials by experience have been widening, which could reflect greater training needed to cope with an increased pace of change. However, as noted by Bartel and Sicherman (1993), an unexpected change in the rate of technological change could lead to flatter pay profiles and earlier retirement. The goal of this paper is to provide direct, broadly based evidence on how changes in technology are associated with changes in the wage structure in the 1980s. It focuses on wage changes between 1979, when wage gaps by schooling reached a trough and the business cycle was at a peak, and 1989, the next peak in the business cycle. The study will examine all industries, not just manufacturing. By including the early 1980s, this study can assess the role of computers when the use of PCs in the workplace was relatively limited. The data on hightech capital (KHT) also include scientific and engineering instruments and telecommunications equipment, the impact of which has not been assessed. 3 The main innovation of this study is to exploit variation across and within industries to determine how measures of technological change relate to changes in the wage structure, using the full-year Current Population Survey (CPS). The use of industry data on technology variables is dictated by the lack of micro data linking workplace technology to wages that covers this time period (and covering industries outside of manufacturing). It is linked to cells broken down by schooling, experience, and gender categories in the CPS so that the analysis can be done in a fixed effects framework. This study examines a wide range of indicators of technological change: research and development intensity, usage of various forms of KHT, growth in the capital-labor ratio, growth in total factor productivity, and recentness of capital. Complementarity between capital and skilled labor is a well-established result in the literature on labor demand; this paper extends this concept by examining whether some types of capital are more complementary with skilled labor than others. To sum up, the features that distinguish this paper from others in the literature on wages and technology are its focus on wage growth (as opposed to skill upgrading and wage levels); the breakdowns of wage growth by schooling, experience, and gender within all industries (not just one dimension of wage change, not just manufacturing); the use of a wide range of technology indicators (not just computers); the measurement of technology indicators at the beginning and end of the sample period (not just one or the other); and the use of a fixedeffects framework to control for unobserved heterogeneity. The paper is organized in the following way. Section 1 outlines a theoretical framework within which the approach used here has strong microfoundations and identifies critical underlying assumptions. Section 2 discusses the technology parameters used in the empirical analysis. Any study examining changes in the interindustry wage structure would seem quixotic, 4 because it is so stable over time. This is true in the sense that wage levels across industries are autocorrelated over very long time periods, but it also is misleading because, as shown in Section 3 below, there are sizable differences in important parameters -the intercept, wage differentials by schooling and experience, and the gender gap -over time within industries. The main results of this paper, reported in Section 4, are that (1) increases in R&D and acceleration in the growth of the capital-labor ratio (K/L) coincide with increased wage gaps by schooling within industries and (2) increases in R&D are associated with wider wage gaps by experience. The results for KHT-intensity are mixed; wage gaps increase with KHT-intensity in 1979, but decrease with the log change in KHT-intensity. Section 5 assesses how much of the observed change in the wage structure can be explained in terms of observable indicators of technological change. I. THEORETICAL FRAMEWORK The idea that human capital is complementary with physical capital, whereas both are substitutes for raw (unskilled) labor dates back at least to Griliches (1969) and appears to have widespread empirical support in the factor demand literature, as shown in Hamermesh (1993). The link between human capital and technological change is based on the argument that educated individuals are better able to adjust to changing economic conditions (Nelson and Phelps (1966), Welch (1970), Schulz (1975), Bartel and Lichtenberg (1987)). For technological change to result in changes in wage differentials that vary by sector, a necessary condition is that the labor market must consist of two or more sectors and the supply curve of skilled labor (S) in each sector must be upward sloping. If the supply of S (and U=unskilled labor) is infinitely elastic to all sectors at prices that produce equalizing differentials (i.e. cover the cost of training and education), then all technology-induced changes in the relative 5 wage structure of a sector would be short-lived. The magnitude of any change in wage differentials would depend on the magnitude of the S-biased technological shock, along with the absolute value of the elasticities of supply and demand in that sector. For instance, suppose that there are two sectors that are equally Sintensive, but one where new technology leads to a large shock in relative factor demand (e.g., continuous process industries) whereas the new technology has a negligible impact in the other sector (e.g., traditional manufacturing). There would be greater excess demand for S in the sector where the new technology can be used most effectively. Turning from these abstractions to the data, it is useful to consider three groups of workers: experienced workers, inexperienced workers with post-secondary education, and inexperienced workers with no more than a secondary education. Among experienced workers, consider a model where (1) S has a positive supply elasticity in each sector and (2) the supply of U is more elastic than the supply of S, conceivably even infinitely elastic. Here the change in ws/wu by sector increases with the technology-induced demand shock. This framework seems applicable for experienced workers for two reasons: industry-specific investments in training and greater payoffs to such training in sectors undergoing technological change. Because of industry-specific investments in training, experienced workers face significant wage losses if they change industries. A large share of on-the-job training is firm-specific and much general training is tied to a particular product market or technology. Empirical indications of the importance of industry-specific human capital are the low overall rate of inter-industry mobility among experienced workers (Murphy and Topel (1987)) and the wage losses faced by displaced workers who change industries (Neal (1995)). Suppose further that the returns to training vary across industries, leading to variation in the wages of workers with the same levels of schooling and ability. Lillard and Tan (1986) have 6 argued that the payoff to training is greater in sectors undergoing rapid technological change, mainly because very few firms will be at the cutting edge of a particular technology, limiting the supply of trained workers. In business school parlance, it becomes cheaper to make than buy trained help. The training of inexperienced workers with post-secondary education also frequently has an industry-specific component that makes such workers less than perfect substitutes. Some academic units focus on industries such as agriculture, education, journalism, law, medicine, and public administration, to name a few. Many degrees in engineering also have a large degree of industry-specificity. The responsiveness of pay in the various markets for college graduates to supply and demand factors has been well-documented (Freeman (1986)), so as long as the impact of technological change varies across different industries, one should not be surprised to see a correlation between technological change and wage growth for this group that would persist for a decade or more if the pace of change is accelerating. Johnson and Stafford (1996) have pointed out that the economy-wide consequences of S-biased technological change for unskilled workers depend on whether the change increases the ability of S to do tasks traditionally done by S (which they call skill-intensive technological change) or the ability of S to perform tasks traditionally done by U (skill-extensive technological change). In the former case, U benefits in absolute terms, but in the latter case U is worse off. It is more difficult to argue that technological change should have an impact on interindustry wage differentials for inexperienced workers with relatively little schooling. Such workers have little to no industry specific human capital; their schooling also is unlikely to provide industry-specific skills. In summary, S-biased technology shocks lead to wider wage gaps by skill in a multisector model with industry-specific investments in human capital. This can be examined both by 7 looking at wage differentials within each sector (which should widen with technology shocks) and by studying how inter-industry wage differentials change for each skill group (they should widen with technology for S, but not for U). II. MEASURES OF TECHNOLOGICAL CHANGE The theoretical literature provides relatively little guidance for empirical work. Taken literally, the term "technological change" can mean two very different things. The standard economic definition is the ability to produce more output with the same amount of inputs, usually the consequence of better knowledge or organization. The appropriate measure of this type of technological change is total factor productivity (TFP) growth, the growth in output that cannot be explained in terms of changes in the quantity or quality of inputs. This measure is always problematic in empirical work because it is by definition a residual and questions about whether the data have been completely purged of all changes in input quantity and quality never can be completely resolved. Alternatively, technological change can mean a change in equipment and job requirements. The substitution of personal computers, software, and printers for typewriters would qualify as a change in technology under this definition, but not necessarily under the former. The distinction is important because recent changes in the wage structure may very well be attributable to the adoption of certain types of equipment that are highly complementary with skilled labor. To measure changes in equipment, this study examines R&D intensity, KHTintensity, K/L growth, and the recentness of capital. This study examines changes in wage equation parameters across 39 industries at the one or two-digit level of aggregation. Because only 19 industries are in manufacturing, a paramount consideration in the choice of right-hand side variables is the availability of data for 8 nonmanufacturing industries. This is straightforward for TFP growth and the growth in K/L, because Jorgenson and Fraumeni have constructed data (described in Jorgenson (1990)) for 35 industries using definitions that are close to the ones used for this study. The growth rates for 1970-79 are matched with the 1979 CPS, whereas the growth rates for 1980-85 are matched with the 1989 CPS. R&D intensity is widely used (e.g., BLS, OECD) to indicate which industries qualify for "high tech" status. The nature of R&D work varies in fundamental ways from other forms of work -the tasks are nonroutine and undefined; the output is hard to measure; decisions have to be based on relatively little hard data and end up being based on intuition and politics (Pasmore (1997)). High wages are dictated by the need for continuous learning. In addition some reward for discoveries that can generate huge streams of future revenues is dictated by equity concerns. While these often take the form of bonuses and stock options, pay adjustments beyond a standard merit raise (of 1.5 percent on the margin) are made for key performers. In R&D shops, the dictate is "price the person, not the job (Gomez-Mejia, Balkin, and Milkovich (1990))." The R&D intensity measures published by the National Science Foundation are limited to manufacturing industries. A further limitation of these data is that they pertain to the industry where an innovation originates, not the industry where the innovation is actually used. An alternate measure was developed to incorporate the nonmanufacturing sector and possible spillovers of innovation across industry lines -the percentage of employees who were scientists and engineers for each industry in the full-year CPS. Unlike the corresponding measure published by NSF, this is not restricted to persons engaged in R&D activity, a potentially desirable property for industries that are heavy R&D consumers, but engage in very little R&D themselves. 9 The ratio of scientific and engineering employment to total employment is highly correlated with the measures published by NSF for manufacturing industries aggregated at the two digit level. The correlations for 1989 are: CPS ratio and employment share of R&D scientists and engineers 0.960 CPS ratio and company's own R&D funds as a percent of sales 0.868 With such high correlations, the employment share of scientists and engineers should be a reasonably good indicator of innovative activity by industry. Use of this measure in analyzing wage structure changes is problematic in one regard. Scientists and engineers are more highly paid than other college graduates. Because their employment share has risen, there can be little doubt that this occupational shift is driving the aggregate trends to some extent and that this is part of the explanation of changing wage structure that this paper seeks to develop. The danger is that the sampling frame of the CPS was not designed to measure the distribution of scientific personnel across one-and-two-digit industries. In industries where scientific personnel end up being overor undersampled, average wages for college graduates will be overor underestimated. As a result, a model linking contemporaneous values of wage differentials and a CPS-based proxy for R&D will overstate the wage impact of the technological changes resulting from increased innovative activity. To prevent this from happening, values of the R&D proxy from the 1980 and 1990 CPS are used in the regressions reported here. (The 1979 and 1989 values will be used when the discussion focuses on means by industry and aggregate trends.) Most of the economic literature on the introduction of KHT, including computers, into the workplace has found a positive correlation with wages and skill levels. In practice the impact of KHT on skills and compensation depends not just on the knowledge and abilities needed to operate the new equipment, but also the change in work processes. Zuboff (1988) points out that information technology (IT) can be applied in two different ways. It always automates certain 10 tasks, thereby making work more routine and lowering skill requirements. In many cases it also can informate (a term coined by Zuboff), which means that the technology generates more information about the underlying work processes. In the latter situation, work becomes more complex and analytical because of the lower costs associated with obtaining information and communicating with others about what it means. Data on the stocks of various types of KHT for 1979 and 1989 were obtained from the Bureau of Economic Analysis' (BEA) RENPR files. Data on investment by industry were broken down by BEA into investment by industry and type of asset, based on historical patterns. BEA used the 1982 input/output table for 1989 and interpolated the values for 1979 from the 1977 and 1982 tables. The allocations of capital expenditures by type of equipment within an industry are based largely on the occupational distribution. In other words, more computer capital gets allocated to industries with more programmers and systems analysts. At a minimum, this creates measurement error. Conceivably, it could create a built-in correlation between the capital measure and wage growth provided that those employed in computer-related occupations experience higher (or lower) wage growth relative to other occupations within a given educational group. BEA then used age-efficiency functions to aggregate the annual investment data into capital stocks. KHT consists, as in Berndt, Morrison and Rosenblum (1992), of four asset codes in the BEA data set: office, computing, and accounting machinery (14), communications equipment (16), scientific and engineering instruments (25), and photocopy and related equipment (26). This study will focus primarily on two measures using these data: a capital-labor ratio based on all four types of capital (called high-tech and office capital or HTOK) and one based solely on assets 14 and 25 (called high-tech capital or HTK). The reasons for examining the latter measure are that (1) much communications and photocopy equipment is not very high-tech; (2) such equipment is related to white collar employment in a 11 quasi-Leontief relationship; (3) communications capital accounts for more than half of HTOK; and (4) 83 percent of communications capital is used in a single industry. The ratio of net to gross capital stock in 1979 and 1989 from Fixed Reproducible Tangible Wealth in the United States was used to measure the recentness of the technology. This measure has two obvious limitations: it makes no distinction between plant and equipment, even though the latter is a much better indicator of recentness of technology than the former, and it pertains to private capital only, even though some of the industries include public sector workers. Some important characteristics of these technology measures stand out in Table 1. First and most importantly, the technology indicators vary widely across industries. In most cases the standard deviation is well above the mean. It is highly unlikely that technological change and any shocks to relative labor demand caused by such change are uniform. Second, the technology indicator that seems to have undergone the largest change is the growth in KHT, with a mean change (in logs) of 1.6 for both measures. The change in R&D intensity is a modest 0.5 percentage points in absolute terms, but a notable 18 percent in proportional terms. Third, the technology indicators tend to be highly autocorrelated; the industries using the most advanced technologies in 1979 tend to be the ones using the most advanced technologies in 1989. The only measure that is not autocorrelated is the TFP measure; a good property for a residual, but perhaps not-so-good a property for an empirical indicator of the use of advanced technologies or the pace of change. Fourth, the technology variables show no consistent relationship between initial levels and 1979-1989 changes. The sectors with the largest increase in R&D intensity were those with the highest initial levels. However, and this will be important in interpreting the empirical results reported below, the growth in HTK/L was the greatest in those sectors which had the lowest 12 initial levels. Fifth, for the most part the indicators of technological change are largely independent. R&D intensity is highest in those industries with the greatest HTK-intensity, suggesting some overlap between these two measures. Lastly, how do the highest ranking industries compare to the lowest ranking ones and is this consistent with the "popular wisdom" about the rate of change in the workplace? R&D intensity and HTK-intensity come out relatively well in this regard. Chemicals and petroleum refining have the largest R&D intensity in 1979; retail trade and personal services have the lowest values. Communications and utilities join chemicals and petroleum refining as the industries with the highest HTK/L ratios in both 1979 and 1989; agriculture, apparel, and retail trade trail. Some sectors with the highest levels of TFP growth (apparel in the 1970s, lumber in the 1980s) do not seem to be leaders in technological change. Similar problems arise with K/L growth (high values for agriculture and leather, low values for financial services). III. CHANGES IN THE INTERINDUSTRY WAGE STRUCTURE The full-year CPS for 1979 and 1989 are used to estimate changes in the wage structure across industries. The main advantage of these data sets is the very large sample size. The sample periods are selected to coincide with the upswing in returns to schooling in the 1980s. Fortuitously, the business cycle was nearing the end of a lengthy expansion in each of these years as well. The sample is restricted to persons between the age of 18 and 64 for whom the CPS reports average weekly earnings, usual weekly hours, race, sex, age, years of schooling, SMSA status and industry. In cases where usual weekly earnings was top coded at $999, a value of $1450 was assigned. Observations with average hourly earnings above $125 were deleted from the sample; in most cases these persons were employed in occupations where 13 such a high wage is implausible. No exclusions were made with respect to the minimum wage because it arbitrarily truncates the sample and the resulting bias could vary tremendously across industries. Industries were defined with two criteria in mind: (1) adequate sample size in both periods and (2) consistency with industry definitions used for other variables in the analysis (e.g., TFP growth). The 39 industrial categories are listed in the first column of Table 2. They are highly similar to the categories used in the CPS "detailed industry" classification system. The sample size in many cases was substantial enough to permit a more disaggregated analysis, but there was rarely more detail available in the data sources on other industry characteristics. The wage variable is the log of the ratio of usual weekly earnings to usual weekly hours for salaried workers and the log of the wage rate for hourly workers, schooling equals years of schooling completed, and experience is min(age-schooling-6, age-18). Estimates of the returns to schooling (from a Mincer model where schooling enters continuously), the wage gap (in logs) between workers with college and high school degrees (from a model where there are four schooling categories: less than 12 years, 12 years, 13 to 15 years, and 16 or more years) and the wage gap between workers with 0 and 30 years of experience for each industry (from the aforementioned Mincer model with experience as a quadratic) are reported in Table 2. There is considerable dispersion across industries at any point in time in the wage equation parameters. Returns to schooling in 1979 averaged 5.7 percent, with a standard deviation of 1.4 percent and a range between 3.0 percent in eating and drinking places and 9.4 percent in business services. Returns to schooling in 1989 averaged 8.0 percent in 1989, with a standard deviation of 2.1 percent and a range between 3.3 percent in eating and drinking places and 11.2 percent in medical services. The standard deviation increased to 2.1 percent. Estimates of other wage function parameters are similarly dispersed. Industries with high 14 returns to schooling also have greater returns to experience, with correlations of 0.342 in 1979 and 0.268 in 1989. There also is considerable flexibility in the interindustry wage structure over time. In the average industry, the rate of return to schooling increased by 2.3 percent between 1979 and 1989. Yet returns to schooling barely changed or actually fell in three industries (lumber, restaurants, and entertainment) and rose by four percentage points or more in four others (nonelectrical machinery, miscellaneous manufacturing, petroleum, and welfare and religious). The standard deviation of the increase in returns to schooling across industries was 1.2 percent. The change in the log wage gap between high school and college graduates is also widely dispersed with a standard deviation of 0.082, relative to a mean of 0.127. Does examining wage patterns by industry add any "new information" beyond what is already known from studies of changes in the skill mix? The three industries with the largest increases in the share of nonproduction workers are electrical machinery, nonelectrical equipment, and printing and publishing (Berman, Bound, and Machin). The former two have larger than average increases in wage differentials by schooling, whereas the latter has a smaller than average increase. Among the industries with larger than average increases in wage gaps by schooling, there are three (petroleum, rubber, miscellaneous manufacturing) with negligible changes in the share of nonproduction workers. The correlation between the change in returns to schooling and the change in the employment share of college graduates is 0.207, indicating that examining changes in the wage structure will produce more than a rehash of existing studies. The wage equation parameters are positively autocorrelated, as shown below: Returns to schooling 0.856 Log wage gap between college and high school 0.735 Log wage gap between 0 and 30 years of experience 0.840