Wage inequality, including the college/high school education premium, has increased substantially in the United States. A key part of the most widely accepted explanation for this is that skill-biased technological change accelerated during this time. This article suggests that the impact of skill-biased technological change was closer to constant in the second half of the twentieth century. This leaves a large unexplained decrease in the college/high school education premium in the 1940s and a large unexplained increase in the 1980s. The current article provides evidence that the upsurge and decline in union power during those respective periods provide a good explanation for these unexplained wage inequality changes.
Organisation for Economic Co-operation and Development (OECD)90/10 data are from OECD, Deciles Summary Database (Paris, France: OECD, no date).
2.
George E. Johnson , “Changes in Earnings Inequality: The Role of Demand Shifts,” Journal of Economic Perspectives11, no. 2 ( 1997): 41-54; John Bound and George Johnson, “ Changes in the Structure of Wages in the 1980's: An Evaluation of Alternative Explanations,” The American Economic Review82 (1992): 371-92; Kevin M. Murphy and Finis Welch, “The Structure of Wages,” Quarterly Journal of Economics107 (1992): 285-326; Kevin M. Murphy and Finis Welch, “Occupational Change and the Demand for Skill, 1940—1990,” The American Economic Review83, no. 2 (1993): 122-26; and Lawrence F. Katz and Kevin M. Murphy, “Changes in Relative Wages, 1963—1987: Supply and Demand Factors,” Quarterly Journal of Economics107 ( 1992): 35-78.
3.
Richard B. Freeman, The Overeducated American (San Diego, CA: Academic Press, 1976), 218.
4.
Adrian Wood , “How Trade Hurt Unskilled Workers,” Journal of Economic Perspectives9 ( 1995): 57-80; Nicole M. Fortin and Thomas Lemieux, “ Institutional Change and Rising Wage Inequality: Is There a Linkage?” Journal of Economic Perspectives11, no. 2 (1997): 75-96; Robert C. Feenstra, “Integration of Trade and Disintegration of Production in the Global Economy,” Journal of Economic Perspectives12 ( 1998): 31-50; and John DiNardo, Nicole M. Fortin, and Thomas Lemieux, “ Labor Market Institutions and the Distribution of Wages, 1973—1992: A Semiparametric Approach,” Econometrica64 (1996): 1001-44.
5.
This mismeasurement of the premium to remain nonunion would then, in turn, lead to a mismeasurement of the impact of unions on the wages of unionized workers. This is because the impact of unions on wages is measured in current mainstream analysis as the difference between union wages and nonunion wages. Nonunion wages are incorrectly assumed to be the unconstrained market-clearing rate.
6.
David H. Autor , Lawrence F. Katz, and Allan B. Krueger, “Computing Inequality: Have Computers Changed the Labor Market,” The Quarterly Journal of Economics13 (1998): 1169-1213; Claudia Goldin and Lawrence F. Katz, “The Returns to Skill across the Twentieth Century United States,” NBER Working Paper no. 7126 ( 1999); and Murphy and Welch, “ Occupational Change and the Demand for Skill,” 122-26.
7.
Robert H. Topel , “Factor Proportions and Relative Wages: The Supply-side Determinants of Wage Inequality,” Journal of Economic Perspectives11, no. 2 (1997): 55-74.
8.
Freeman, Overeducated American, 218.
9.
The elasticity of substitution between high school—diploma workers and college-degree workers is assumed to be 1.5. I discuss this issue later in the text.
10.
Of course, since the real world is not ceteris paribus, the decline in returns to education would have eventually reduced incentives for workers to go to college and led to a smaller increase in the relative supply of college-degree workers. Fewer workers would have gone on to college if the education premium had not been maintained.
11.
I follow Johnson, who reports that estimates of the elasticity of substitution between college-educated workers and high school—educated workers are generally in the neighborhood of 1.5; Johnson, “Changes in Earnings Inequality,” 41-54. Psacharopoulos and Hinchliffe report an estimate of 2.2; George Psacharopoulos and Keith Hinchliffe, “Further Evidence on the Elasticity of Substitution among Different Types of Educated Labor,” The Journal of Political Economy 80, no. 4 (1972): 786-92. In this framework, changes due to technology are treated as being entirely incorporated into change in the relative demand measure, ΔA/A. The elasticity of substitution between college labor and high school labor is treated as constant. In theory, technological change might also manifest itself as a change in the elasticity of substitution. However, the change-in-relative-demand story is more intuitively appealing than a change-in-elasticity-of-substitution story. For example, do we use more engineers per automobile now than in the past primarily because we demand more technological sophistication in cars now or because it is harder now than in the past for less educated workers to do work that makes engineers unnecessary? There is probably some of both but more of the former than the latter. In any event, if technological change primarily works by changing the elasticity of substitution, or even a combination of elasticity and relative demand, change would have had to accelerate both in the 1950s and in the 1980s. Again, there is no obvious large technological change that would explain why elasticity or the combination would accelerate in the 1950s. The math in the rest of the article would also change, but the substantive results would not.
12.
Bound and Johnson, “ Changes in the Structure of Wages,” 371-92; and Jacob Mincer, “Human Capital, Technology, and the Wage Structure: What Do Time Series Show,” NBER Working Papers Series, no. 3581 (1991).
13.
Alan B. Krueger , “How Computers Have Changed the Wage Structure: Evidence from Microdata, 1984—1989,” The Quarterly Journal of Economics108, no. 1 (1993): 33-60.
14.
A second alternative explanation would be that all technological change before and after the 1940s/1950s was in favor of the more educated, but in the 1940s/1950s, this skill bias suddenly disappeared for one decade, then returned to the high levels of bias in favor of the more educated starting in the 1950s and persisted for decades after that. There does not seem to be a distinctive set of technological changes introduced in the 1940s/1950s that, compared to surrounding decades, was completely different in impact in that it favored neither the more nor less educated.
15.
James K. Galbraith, Created Unequal: The Crisis in American Pay (New York: Free Press, 1998), 350.
16.
Galbraith's response to the problems he sees is to suggest that technology monopoly rents explain the increasing returns to skill. This solves the problem of the mismatch between increasing returns to skill and sectors/periods of high productivity growth. However, it does have the same problem as skill-biased technological change in that its presence has to be inferred through the absence of other factors. Also, because it delinks wages from marginal productivity, it would leave even less direct evidence of its effects than the skill-biased technological change story. Thus, it is even harder to study, although this does not mean it is not present. Its elusive nature means there is little evidence to disprove it but also little evidence to support it. Ibid.
17.
Lawrence Mishel and Jared Bernstein, Technology and the Wage Structure: Has Technology's Impact Accelerated Since the 1970s (Washington, DC: Economic Policy Institute, 1996 ).
18.
Autor et al., “Computing Inequality,” 1169-213; and Michael J. Handel, “Trends in Direct Measures of Job Skill Requirements,” Jerome Levy Economics Institute Working Paper Series, no. 301 (2000).
19.
Autor et al., “Computing Inequality: Have Computers Changed the Labor Market,” 1201, columns 7 and 8, row 3. Capital intensity, row 1, also shows no increase from the 1970s to the 1980s, but rather, a constant effect.
20.
Because this is a very small sample, the central-limit theorem does not apply. If the error term is non-normally distributed, t-scores are not necessarily distributed as Student's t distribution. Because the t-score distribution may not be distributed precisely as Student's t distribution, t-tests are not always precisely valid (William H. Greene, Econometric Analysis [New York: Macmillan, 1993], 301). However, Monte Carlo tests show that t-tests are relatively robust to deviations from normality (H. R. Neave and C. W. J. Granger, “A Monte Carlo Study Comparing Various Two-sample Tests for Differences in Mean,” Technometrics 10, no. 3 [1968]: 509-22; Hisashi Tanizaki, “Power Comparison of Non-parametric Test: Small-sample Properties from Monte Carlo Experiments,” Journal of Applied Statistics 24, no. 5 [1997]: 603-32; and author's own Monte Carlo simulations). Why this is so can be seen by exploring the implicit z-score distribution of the t-test. The implicit z-score distribution is the distribution of t-scores adjusted using the Student's t distribution with the correct degrees of freedom, that is, implicit z score = inversenormal(t-tail[df, t-score]). Based on the author's own Monte Carlo studies, even with non-normality—even high levels of non-normality—of the error term, the implicit z-score distributions are produced with standard deviations very close to 1. Deviations in the implicit z-score distribution are concentrated in the non-normality of the distribution and not the standard deviation of the distribution. A consequence of this is the relative robustness of the t-test; only very large deviations from normality greatly weaken the accuracy of t-tests. If the repressors are taken as fixed and the error term is very highly non-normal, then the t-test will be merely overly conservative and confidence intervals extra wide. If the regressors are taken as stochastic and both the error distribution and the independent variable(s) distribution(s) are highly non-normal, then the t-score and confidence intervals may be overly conservative or overly optimistic. Very high levels of non-normality would be a kurtosis well above 4.5. Even in a case in which the t-test with stochastic regressors is weakened by very high levels of non-normality, a p value of .01 might be wrong by a factor of 2 or 2.5, not 5 or 10, that is, actual p value of .025 or .004. A p value of .05 would be even less biased and might represent an actual p value of .063 or .037. Such a level of bias in the t-test would not invalidate the conclusions here when the p values are so low. Again, this robustness is created because the standard deviation of the implicit z-score distribution is close to one, and only the normality of the distribution is affected. Moreover, smaller samples—especially extremely small samples—with stochastic regressors, such as the six cases with four degrees of freedom here, are actually more robust to large deviations from normality than larger, moderately sized data sets. Four degrees of freedom is more robust than six. Six degrees of freedom is more robust than ten.
21.
“Union Trends and Data: Union Statistics” (New York: Labor Research Association, online database, www.laborresearch.org,cited2006).
22.
I have also added 0.12 as an offset to bring the 3.3 percent line down over the union density lines. This means the 2.95 percent constant skill-biased technological change line and the 3.3 percent line have differing intercepts.
23.
Richard Freeman , “How Much Has De-unionization Contributed to Earnings Inequality?” in Uneven Tides: Rising Inequality in America, ed. Sheldon Danziger and Peter Gottschalk (New York : Russell Sage Foundation, 1993); Bound and Johnson, “Changes in the Structure of Wages ,” 371-92; and DiNardo et al., “Labor Market Institutions and the Distribution of Wages ,” 1001-44.
24.
Bound and Johnson, “Changes in the Structure of Wages,” 371-92.
25.
Note that I have corrected the Bound and Johnson estimate of the effects of de-unionization. Bound and Johnson subtract the change in union density of college-educated, male workers from the change in the union density of male high school workers. Freeman shows that changes in the male college union density have very little effect on the male college/high school education premium, so I use only the change in the male high school union density to compute the Bound and Johnson estimate. Ibid.; and Freeman, “How Much Has De-unionization Contributed to Earnings Inequality?”
26.
I have adjusted Freeman's results upward due to an apparent typo: union differential × change in union density = .16*(—.12) = —.019, not —.014; ibid.
27.
Michael Goldfield, The Decline of Organized Labor in the United States (Chicago : University of Chicago Press, 1987 ), 23.
28.
Ibid., 23.
29.
Ibid., 90-91.
30.
The exact definition of the detailed industry codes varies some by year. Using two-digit industry codes produces essentially the same pattern of results. Demographic/ compositional changes and changes in relative prices related to experience, race, and industry not explicitly modeled do have an impact on the final overall education premium, but these are filtered out by including them as control variables in the wage equation. While these demographic, compositional, and relative price factors are important, the role of union power versus technological change is the more novel aspect of the change and justifies the focus on these factors. These demographic, compositional, and relative price factors are important and can explain a highly significant 5.1 percent of the total 13.7 percent increase in the education premium from 1978 to 1989. Changes in industry composition and relative wage across industry explain 2.7 percent of the increase. Change in experience and race distribution and changes in relative price across experience and race explain 2.4 percent of the increase, excluding the differences in relative price between older and younger workers as a group that is explicitly modeled and measured—that is, only changes in relative prices according to experience within older and younger worker groups. The changes in relative wages across industries could be due to international trade factors, but international trade as a factor is not a focus of this article. The most likely international trade effect is the changing relative prices between tradable and nontradable sectors due to change in the exchange rate, with college workers more concentrated in nontradable industries. The alternative explanation is international trade's affecting whole sectors including both college and high school workers together, as per the Ricardo-Viner sectoral model, with more college workers in sectors benefiting from sectoral effects. What the evidence does not support are general factor effects in the Stolper-Samuelson framework affecting college and high school workers uniformly as groups.
31.
Freeman finds the union premium for college workers to be “insignificant”; Freeman, “How Much Has De-unionization Contributed to Earnings Inequality?” I also find that adding a dummy for unionized college workers has no significant effect on the results.
32.
An alternative explanation that obviously would make little sense would be to attribute it to technological change that is biased against nonunion, older workers but not biased against unionized, older workers. Another unlikely alternative is to suggest that unions were getting stronger during the 1980s but only benefited older, unionized workers; this explains the increase in union-nonunion relative wages among older workers.
33.
Richard B. Freeman, “Are Your Wages Set inBeijing?” Journal of Economic Perspectives9 (1995): 15-32; George Borjas, Richard Freeman, and Lawrence Katz, “On the Labor Market Effects of Immigration and Trade,” in Immigration and the Work Force, ed. George Borjas and Richard Freeman (Chicago : University of Chicago and NBER, 1992), 213-44; and Wood, How Trade Hurt Unskilled Workers.
34.
Borjas et al., “ On the Labor Market Effects of Immigration and Trade,” 213-44; and David S. Lee, “Wage Inequality in the United States during the 1980s: Rising Dispersion or Falling Minimum Wage?” The Quarterly Journal of Economics114, no. 3 (1999): 977-1023.
35.
35. The specific education premium referred to is the College No Grad School/High School premium. Comparison of the regression estimate of the effect to data that do not use controls is appropriate because the regression analysis measures the relationship between changes in union power and the education premium using the long-term data that do not include controls.
36.
36. The 1970s is an exception to relative balance. In this period, the relative supply changes were much larger than the estimated effects of constant, skill-biased technological change.
37.
37. For the 1999 data, the measurement of groups by education level changes slightly because the questions about education on the current population survey changed. The college group now consists of those with a bachelor's degree and those without a bachelor's degree but who have completed four or more years of college, using two different survey questions to identify this group. Previously, the survey identified those who had completed a fourth year of college.
