Unions, Norms and the Rise in American Earnings Inequality

Transcription

1 Unions, Norms and the Rise in American Earnings Inequality Bruce Western 1 Harvard University Jake Rosenfeld University of Washington, Seattle November, Department of Sociology, Harvard University, 33 Kirkland Street, Cambridge MA western@wjh.harvard.edu. This research was supported by a grant from the Russell Sage Foundation. We gratefully acknowledge Jim Alt and seminar participants at Princeton University who commented on an earlier version of this paper.

2 Abstract From 1972 to 2006, union membership among U.S. men working full-time in private-sector jobs declined from 33 to less than 10 percent. Inequality in men s hourly wages earnings increased by 40 percent in this same period. We study the effect deunionization on rising inequality, with a variance decomposition that accounts for how unions raise average earnings and also reduce inequality in earnings among union workers. Going beyond earlier research, we also argue, that unions reduce inequality by establishing norms for fair wages, reducing the dispersion of wages in the nonunion sector. Accounting for the effect of unions on the wages of union members and the effect of unions on nonunion wage norms suggests that union decline explains one third of the growth in inequality an effect equal to the growing stratification of earnings by education.

3 The decline of organized labor in the United States over the last 30 years coincided with a large increase in wage inequality. Among men working full-time in the private sector, union membership declined from 33 percent to less than 10 percent, while wage inequality increased by 40 percent. Economic research typically attribute only a small fraction of the growth in inequality to deunionization, with a larger share due to the rising wages of highly educated workers (Autor, Katz, and Kearney 2008; Acemoglu 2002). This evidence is often interpreted to support a market, rather than institutional, account of the trend in inequality. We revisit the effects of deunionization on the growth of earnings inequality offering two extensions to earlier research. First, we study the effects of union decline while controlling for the effects of schooling. Estimating the effects of schooling and unions allows a comparison of the effects of market forces (the growing demand for college-educated workers) and changing institutions (union decline) on rising inequality. Second, unlike earlier research we consider the effects of unions on nonunion firms. In particular we consider whether wages are higher and inequality lower among nonunion workers in regions and industries that are highly unionized. The effect of unions on nonunion workers can be motivated in several ways. Research on union threat effects argues that nonunion employers raise wages to head off a possible organizing drive. More fundamentally, unions help establish norms for fair pay and working conditions, even for nonunion workers. In the early 1970s, when 1 in 3 workers were union members, union leaders were often prominent voices for higher and more equitable pay and better working conditions for workers generally. Workers were more likely to spend time in union jobs, so the benefits of 1

4 unionism were experienced by a large fraction of the workforce. Labor legislation institutionalized the role of organized labor as a model for fair pay and conditions in the labor market as a whole. From this perspective, the decline of unions weakened norms of equity in the labor market. With these weakened norms, today s workers expect less and are less able to press for pay raises. To study the effects of unions on rising wage inequality, we analyze thirty-four years of data from the Current Population Survey from 1972 to Our analysis decomposes the change in the variance of log hourly wages among male full-time private sector workers. The decomposition shows that about a third of the growth in inequality is related to the increasing stratification of earnings by education. The effect of deunionization is half as large if we consider only wages of union members. If we add the effects of unions on wages in the nonunion sector, deunionization contributes as much to rising inequality as wage growth among college graduates. Trends in Earnings Inequality and Unionization A large research literature traces the trend in US economic inequality since the 1970s (for recent reviews, see Gottschalk and Danziger 2005; Autor, Katz, and Kearney 2008; Lemieux 2008). The research shows that inequality increased more among men than women, and that inequality has increased for a variety of income measures, including hourly wages, weekly and annual earnings, and family incomes. Economic inequality is also higher among, and has increased more rapidly, in the private sector. The private sector accounts for about 85 percent of all men s employment and we focus on men s private sector wages in studying the effects of union 2

5 decline on inequality. The trend in wage inequality in the private sector is shown in Figure 1. Inequality is measured by the variance of log hourly wages for men working full-time in private sector jobs from 1972 to The variance of log wages increased by 40 percent, from.25 to.35, with most of the increase unfolding from 1978 to The total variance in log wages is divided into within-group and between-group components. These components were obtained from a regression on hourly wages with age, education, region, union membership and industry unionization rates as predictors. The within-group component is given by the residual variance, and this describes the inequality in wages among workers of the same age, education, region, and union status. The between-group component is given by the variance of predicted wages and describes inequality across groups. Figure 1 shows that rising inequality between and within groups increased total wage inequality in the 1980s. Rising inequality in the 1990s was due only to rising within-group inequality. Rising wage inequality accompanied falling unionization. Figure 2 shows that unionization among private sector full-time male workers declined steadily since 1972, from about 35 percent to under 10 percent by The time series on unionization is taken from the May and the Merged Outgoing Rotation Group files of the Current Population Survey. Employer surveys show that CPS survey respondents mistakenly report their union status 2 to 3 percent of the time (Card 1996). Figure 2 reports the observed proportion of unionized workers and an adjusted series that corrects for the misclassification. 1 Though unionization is continuously falling from the early 1970s, two large declines can be seen from Measurement error in union status is discussed in greater detail below. 3

6 Between Group Within Group Variance of Log Wages Figure 1. Inequality in hourly wages among private sector men, working fulltime, Inequality is measured by the variance of log hourly wages. Between-group and within-group inequality are shown separately, where groups are defined by age, education, race, union status, region, and the unionization rates of regional industries. 4

7 Proportion Unionized Observed Unionization Adjusted for Misclassification Figure 2. Union membership rates among private sector men, working full-time, The observed series is compared to an alternative series that adjusted for misreporting of union membership in the Current Population Survey. to 1975 and from 1981 to These declines in unionization among full-time men follow trends in the private labor market as a whole. In the public sector unionization rates are significantly higher and have remained relatively unchanged since the early 1980s. Explanations of the decline in US union membership usually appeal to structural changes in the labor market or intensified political conflict in the workplace (Farber and Western 2001; Freeman and Medoff 1984; Goldfield 1987). Employment has grown fastest in industries with low unionization rates and there has been virtually no net increase in the num- 5

8 ber of jobs in unionized firms. In addition to the effects of structural change, employers have became more resistant to unionization through the 1970s, and new organizing at least through the union election process plunged through the 1980s (Tope and Jacobs 2009). As a result, unionization rates have fallen even in highly organized industries. Unions and Inequality Labor unions might influence inequality in the labor market in two main ways: Collective bargaining affects the wage distribution for union members, and unions may also affect the wages of nonunion workers. Nearly all research on unions and inequality focus on the wages of union members. The Wage Premium, Threat, and Spillover Unions try to raise the wages of their members through collective bargaining. The effect of unions on wages can be defined as the difference in wages between union and nonunion workers with the same productivity. The earliest studies using this framework compared wages in industries that varied in their unionization rates. In the decade after World War Two, unions were estimated to raise wages by 10 to 15 percent (Lewis 1963, 5). The next generation of research, emerging in the 1970s, used crosssectional survey data to regress wages on measures of schooling, work experience, and union membership (see Lewis 1986 for a review). Recent cross-sectional estimates for the United States in the 1990s and 2000s suggest that union membership raises wages by 15 to 20 percent (Hirsch 2003). Cross-sectional estimates only controlled for observed characteristics, like education and experience, and union wage effects may have also reflected unobserved characteristics of workers. Panel data estimates that 6

9 adjust for positive selection into union membership also indicate an average union wage premium in the United States for males of roughly 15% (Card 1996). In contrast to these studies of survey data, recent analysis of firm-level wages suggests that unions in newly-organized workplaces in the 1980s and 1990s barely raised wages at all (DiNardo and Lee 2004). The effect of unions on wage inequality depends on who is being organized. Milton Friedman (1962, 124) argued that unions were strongest among highly-paid workers, in skilled crafts. For Friedman, unions increased inequality by raising pay among those already at the top of the wage distribution and by causing unemployment among nonunion workers. Empirical research since the 1970s finds that unions equalize the wage distribution. This equalizing effect happens in two main ways. First, unions raise wages among low-skill workers, closing the skill gap in wages. Second, unions reduce inequality among union members. Within-sector inequality is reduced by unions that standardize wages across firms in an industry. Union contracts also tend to compress wages within the workplace. In several important studies, Richard Freeman (1980, 1993) found that wage compression among union members had a large effect on wage inequality. In light of these findings, researchers began to study the association between declining rates of unionization since the 1970s and rising wage inequality. From the late 1970s through the late 1980s, declining unionization was estimated to explain 10 to 20 percent of the growth in men s wage inequality (Freeman 1993; Card 1992; DiNardo, Fortin, and Lemieux 1996). Research on unions and inequality focuses on the wages of union members. A key hypothesis for our analysis is that unions influence nonunion 7

10 wages. Studies of nonunion wages have compared the spillover and threat effects of unions. When unions raise wages for their members, employers may cut union employment forcing unemployed workers to look for jobs in the nonunion sector. The spillover of workers into the nonunion labor market causes nonunion wages to fall. The threat effect, on the other hand, results from nonunion employers raising wages to the union level to avert an organizing drive. To test for the effects of spillover or threat, researchers examined whether nonunion wages were in highly unionized industries and localities. Some reported a positive relationship between industry unionization and nonunion wages, and nonunion wages were also higher in highly-organized metro areas (Freeman and Medoff 1981; Neumark and Wachter 1995). Other research also supported the threat effect. Nonunion wages fell faster than union wages as the threat of organization receded following deregulation of the airline and telephone industries (Farber 2005). Within firms, nonunion workers are paid more in the presence of a unionized workforce. Thus, Leicht (1989) finds that for every dollar increase in union wages, nonunion wages in the same plants increase by $.43 to $3.89 (1042; see also Freeman and Medoff 1984). Though the designs vary, the research on threat and spillover share a focus on the effects of unions on the average level of nonunion wages. If threat effects dominate, we expect unions to reduce between-group inequality by reducing the wage gap between union and nonunion workers with similar characteristics. In sum, previous research indicates three main effects of unions on inequality. Unions reduce inequality between groups between college and non-college workers, say. Unions also reduce within-group inequality by compressing wages among unions workers. Finally, if the threat of 8

11 union organization reduces the union wage effect, we also expect we expect unions to reduce between-group inequality by reducing the wage gap between union and nonunion workers with similar characteristics. These effects of unions on inequality suggest deunionization has added significantly to rising between-group and within-group inequality. The Moral Economy of the Labor Market Besides these effects, we argue that unions help establish norms for fair pay and conditions that shape not just the level but also the inequality of nonunion wages. In the sociology of labor, unions are often seen to promote an ethos of solidarity that sparks collective action (Dixon, Roscigno, and Hodson 2004). We also claim a solidaristic influence for unions, though we see these effects on wages not strikes. From our perspective, wages are not only set by the forces of supply and demand. The labor market is also embedded in a moral economy in which norms of equity and fairness help regulate employment relations. We often think of the moral economy in historical context determining, for example, fair prices for bread and flour under the British Corn Laws (Thompson 1971) or the relative rank and standing of English workers in the nineteenth century (Polanyi 1957, 153). Though the customs and institutions that sustain the moral economy are perhaps most vivid in historical context, modern industrial relations also embody norms of equity (Swenson 1989, 11 19). In Western Europe, pay norms among nonunion workers are often institutionalized by collective bargaining. Centralized collective bargaining systems regularly extend union agreements to nonunion workers, increasing wage equality and reducing the union wage premium (Traxler 1996; Pencavel 2005). 9

12 US unions also promoted norms of equity that narrowed pay disparities in the nonunion sector. The normative influence of unions has cultural, political and institutional dimensions. Culturally, industrial unions in particular often promoted a public language of social solidarity. Speaking of unions role as an advocate for fairer working and living conditions, a former Secretary-Treasurer of the AFL-CIO once put it: We are the peoples lobby (Kotz 1977: 146). Under Walter Reuther s leadership, the United Auto Workers (UAW) pressed the Kennedy and Johnson Administrations for quick passage of Civil Rights legislation. Reuther spoke at the March on Washington for Jobs and Freedom, urging the government to fulfill its promise of full citizenship rights and equal rights for all Americans (Los Angeles Times 1963). Politically, American unions have been frequent advocates for redistributive public policy. Highly unionized states have been found to have higher minimum wage rates and congressmen from such states are more likely to support minimum wage increases (Cox and Oaxaca 1982; Kau and Rubin 1978). Union political pressure extended beyond wages to social legislation. For example, major unions regularly backed proposals for universal health-care and supported the creation of Medicare in the mid-1960s. In the 1970s, 53 unions joined numerous states in a lawsuit challenging cutbacks in the federal food stamp program, and threatened to sue again in 1980 to keep the program solvent. Perhaps more important than unions cultural or political role, industrial relations institutions operated, at least informally, to extend union conditions throughout the labor market. Often acting alongside government and employer representatives, unions historically influenced industrywide wage standards. For example, since World War II, tripartite boards comprised of leaders from unions, industry, and government helped set 10

13 wages and standards in an effort to ward off runaway inflation in the midst of wartime production efforts. A similar body, the Wage Standardization Board (WSB) was recreated during the Korean War. The WSB monitored abnormal wage increases among industry leaders and aimed to narrow inter-firm wage differentials, reducing wage inequality in the wider economy (Ross and Rothbaum 1954). Centralized wage setting in the U.S. was not limited to wartime, either. Both the Kennedy and Johnson Administrations instituted wage and price guideposts with the explicit goals of not only keeping prices stable during periods of near full-employment, but also of ensuring distributional equity of wages (Ulman 1998, 170). A key objective of the guideposts was to peg labor cost increases to rises in productivity: government officials urged wage restraint on union leaders and heads of major firms, promising in return healthy wage and profit levels so long as productivity continued to rise. The Nixon Administration s Pay Board repeated the tripartite wagesetting arrangement common during wartime. Concluding that No program works without labor cooperation, (Matusow 1998, 160), Nixon s board was comprised of five union heads, five industry leaders, and another five public appointees. Board initiatives again extended well beyond urging wage restraint in major contract negotiations, focusing also on executive pay levels, boosting pay floors for low-wage workers, and monitoring merit pay increases (Mitchell and Weber 1998). Governmentled efforts to guide wage-setting continued in the 1970s, when President Carter created the Tripartite Pay Advisory Committee, tasked with advising the Council on Wage and Price Stability on voluntary wage and price guidelines. Besides these efforts at centralized wage-setting, unions also helped to 11

14 establish pay norms throughout local labor markets. In some industries, union influence was extended by law. The federal Davis-Bacon Act and its state-level variants require public construction projects to pay at least the locally prevailing wages and fringe benefits paid on similar projects in the area. Studies from the 1970s show that Davis-Bacon raised construction wages in nonunion firms and closed the gap between union and nonunion wages (Goldfarb and Morrall 1981). Beyond the sphere of federal contractors, major union agreements often set the pattern for industry-wide wage negotiations. Employers of the 1960s and 1970s commonly used industry wage surveys to reduce the wage differentials between union and nonunion industry leaders (Dunlop 1977, 15; Foulkes 1980). To be sure, the adoption of union standards in nonunion firms was significantly motivated by the threat of unionization. Still, unions also helped set standards for equitable and legitimate treatment, particularly in large nonunion firms. Foulkes (1980, 156) reports that in the 1970s, even in lightly unionized industries where the threat of unionization was relatively remote, nonunion companies closely monitored union contracts. Predominantly nonunion firms with small union workforces abandoned merit pay, and norms of equal treatment governed the distribution of fringe benefits. Summarizing his survey of pay practices in large nonunion firms, Foulkes (1980, 153) writes: In many environments, providing and demonstrating equity generally means that a company s pay rates favorably compare with those of unionized companies. It would be acceptable to say that the activities of many unions in the United States are benefiting many nonmembers; in other words, unions are doing much good for people who do not pay them any dues. 12

15 The long decline of unionization rates through the postwar period accompanied the erosion of the labor market s moral economy. The cultural, political, and institutional influence of unions on equitable pay norms declined with union membership. Voss and Sherman (2000, 311) see the air traffic controller s strike in 1981 as a turning point when corporate leaders stopped playing by the rules. Levy and Temin (2007) divide the postwar history of the American labor market into the eras of the Treaty of Detroit until 1980, and the Washington Consensus which continues today. The era of the Treaty of Detroit was named for the landmark wage agreement in 1948 between General Motors and the United Autoworkers that provided an annual wage increase of two percent plus cost-ofliving. Wages across manufacturing industries moved broadly according to this formula until The Treaty of Detroit was succeeded by the Washington Consensus, an era of deregulation and eroded pay norms in which earnings inequality increased with the compensation of managers and professionals. Because unionization and its decline varies across regions and industries we view the weakening influence of unions not as a discrete turning point, but as a gradual process unfolding unevenly across labor market through the 1980s and 1990s. Our research design exploits this variation to study how deunionization is associated rising wage inequality among nonunion workers in regional labor markets. Because of the influence of unions on norms of equity in the labor market, we expect more wage equality among nonunion workers more within group equality in highly unionized sectors where normative influence is strongest. From this perspective, deunionization signals the erosion of norms of fair and equitable pay and is associated with rising within-group inequality among 13

16 all workers in the labor market. Data and Methods To assess the effects of union decline on earnings inequality we decompose shifts in wage inequality with a variance function regression (Western and Bloome 2009). Unlike the usual regression model, the variance function regression lets the mean and the variance of an outcome depend on independent variables, providing a model for between-group and withingroup inequality. The variance decomposition is based on a regression on the log hourly wage, y i, for respondent i for a given year of the Current Population Survey. Covariates include schooling, age, race, and ethnicity, and region. Our key predictors are a dummy variable for union membership, u i, and a continuous variable, ū i, that records for each respondent the unionization rate for the industry and region in which they work. The variance function model has two equations, one for the conditional mean of log wages, ŷ i = x i α 1 + u i α 2 + ū i α 3, and one for the conditional variance, logσi 2 = x i β 1 + u i β 2 + ū i β 3, where all the covariates are collected in the vector, x i. The model can be fit with a weighted least squares regression on y i and a gamma regression on the log squared residuals. We expect union membership and industryregion unionization to have positive effects on average wages, but negative effects on the variance of wages, our measure of wage inequality. 14

17 The variance function model provides the conditional means and variances for decomposing the trend in wage inequality. Between group inequality is measured by the variance of mean wages across groups defined by the independent variables, B = N w i (ŷ i ȳ) 2, (1) i=1 where w i is the sample weight for respondent i, normed to sum to 1, and ȳ is the grand mean of log wages. The within-group variance is given by: W = N w i σi 2. (2) i=1 The idea that each respondent has a variance, σi 2, may not be intuitive, but it is simply estimated by the squared residual from the regression on y i in the variance function model. Because unions mostly affect withingroup inequality, we expect deunionization to have large effects on the increase in the within-group variance, W. The overall level of inequality is measured by the total variance which is the sum of the between-group and within-group components: V = B + W. We decompose the change in variance by calculating adjusted variances that fix in a baseline year either the effects of some predictors, or the distribution of some predictors. Our analysis is based on three adjusted variances. First, to study the effect of individual union membership, we ask what level of inequality would we observe if unionization rates had remained at their 1972 level. This compositional effect is estimated by reweighting the data to preserve the 1972 unionization rate 15

18 across all years, from 1972 to The adjusted weights are then used in equations (1) and (2) to calculate adjusted variances for all years, based on the 1972 unionization rate. Second, to study the effects of union threat and wage norms on nonunion workers we ask what level of inequality would we observe if industry-region unionization and its effects were held constant at the 1972 level. For these adjusted variances, the industry-region unionization rates and their coefficients, α 3 and β 3, are fixed at the 1972 level yielding adjusted values of ŷ i and σi 2 that are used in equations (1) and (2). Finally, we compare the magnitude of the union effects by asking what level of wage inequality would we observe if the effect of education on average wages had remained fixed at the 1972 level. In this case, we construct new values of ŷ i for each year based on the 1972 education coefficients, and an adjusted variance is obtained by plugging these conditional means into equation (1). Calculations for the adjusted variances are reported in the Appendix. A key challenge for this analysis is the measurement error in union membership. Card (1996), using data from a validation survey of employers, showed that survey respondents mis-report their union status about 2.5 percent of the time. If unionization were 50 percent, the number of nonunion workers mis-reporting union status would be equal to the number of union members who were mis-reporting. With the observed unionization rate at 10 percent, a relatively large number of nonunion workers (2.5 percent of 90 percent) report they are union members. Because of the measurement error, union effects will be biased in both the mean and the variance equations. The bias will tend to grow over time as unionization declines. If the misclassification rate is λ, and the observed unionization rate is ū, then the true unionization rate is (ū λ)/(1 2λ). With observed 16

19 unionization at.10 (10 percent) and a misclassification rate of.025 (2.5 percent), the true unionization rate is.079, a measurement error of about 20 percent. A variety of approaches have been taken to adjust for measurement error in union status (Card 1996; Hirsch 2003). We augment the likelihood function for the wage data with an extra term for the probability that union status is misclassified. We also introduce a Bayesian prior distribution to describe uncertainty in the misclassification rate. Although this method has not been used before we find that it significantly reduces bias in the estimation of union effects. (See the appendix for details.) Bias in the regression results will also result from omitted variables correlated with unionization and wages. If union members are positively selected, wage effects may be upwardly biased because of the higher productivity of union workers. The effects of industry-region unionization will also be confounded with factors correlated with local patterns of unionization and wage inequality. Industry concentration, firm size, and technology, for example, might all be correlated with industry unionization rates and wage inequality. Industry and region fixed effects might reduce these biases, though fixed effects account for 80 to 90 percent of the variation in industry-region unionization, leaving little residual variance for the effects of interest. We try to reduce bias with region fixed effects. We prefer this simple specification because our decomposition analysis aims to summarize and interpret changes in wage inequality, rather than offer a causal analysis of wages determination. In addition, because we analyze the change in inequality rather than its level at a point in time, biases of constant magnitude will not affect our analysis of the trend. Still, union effects will certainly reflect the influence of correlated omitted variables if 17

20 the degree of bias changes over time. Our data are compiled from May files of the Current Population Survey from 1973 to 1981, and the Merged Outgoing Rotation Group files of the Current Population Survey from 1982 to Descriptive statistics for key variables are reported in Table 1. These surveys yield labor market data for the period 1972 to We analyze data for male workers, working full-time (30 hours a week or more), in the private sector. This accounts for about 80 percent of full-time male workers. The dependent variable is log hourly wages, adjusted for inflation to 2001 dollars. Non-response to wage and income questions increases over time, and by 2007, about a third of CPS workers are not reporting wages. The CPS imputes wages to non-respondents, though imputed wages may have lower variance than observed wages. We restrict our analysis to observations with observed wages. Top-coding of hourly wages has changed over time and the proportion of CPS respondents who are top-coded also varies. We take the top 2 percent of hourly wages in each year and impute wages from a Pareto distribution. We experimented with a variety of methods for top-codes and they all yielded the same substantive results as those reported here. Descriptive statistics show that the average hourly wages for union and nonunion workers are about the same in as they were two decades earlier (Table 1). Wage inequality has increased in the union and nonunion sectors, but the increase has been faster among union members. The schooling of male workers has also shifted significantly from 1972 to The proportion of high school dropouts in workforce dropped from around 1 in 4 workers in the 1970s to fewer than 1 in 10, by the 2000s. Fewer than 20 percent of male private sector workers were college 18

21 Table 1. Descriptive statistics for analysis of unionization and hourly wage inequality among private sector men, working full-time, N U N U N U Log Hourly Wages Mean Standard deviation Level of Schooling (proportion) Less than high school High school graduate Some college BA or more Demographic characteristics (proportion) White Black Other Age (years) Note: N=nonunion workers, U=union workers. graduates in the 1970s and more than a quarter had college degrees three decades later. The education gap between union and nonunion workers also declined. Nonunion workers were about five times more likely than union members to hold college degrees in the 1970s, compared to 2.6 times in the decade. Usually studies of union effects on nonunion workers relate local unionization rates (often measured at the industry level) to average nonunion wages (e.g., Freeman and Medoff 1981; Neumark and Wachter 1995). We extend the analysis by also studying the relationship between wage inequality and industry-level unionization in four regions. To study this regional unionization, we measure unionization in 18 industries across four census regions (the northeast, the south, the midwest, and the west). Figure 3 shows the average level of unionization for the period 1972 to 2006 for the 72 industry-regions. Unionization rates were highest (gen- 19

22 erally over 20 percent) in utilities, transportation, and communications, particularly in the northeast and the midwest. Unionization rates were extremely low in agriculture, finance, and retail trade. We can see preliminary evidence for the relationship between unionization and wages in Figure 4. Here, unionization for the 72 industryregions is plotted against the mean level of nonunion wages for full-time private sector male workers, each year from 1972 to In each year, the relationship is positive, so where unions are strong regions, average wages of nonunion workers are higher. Of course, higher wages may reflect the skills of workers in different industries, or other factors like patterns of industrial concentration. Additional support for the normative influence of unions on nonunion wages is given by Figure 5, relating local unionization rates to wage inequality, measured by the variance of log wages. In each year, from 1972 to 2006, nonunion wages are more compressed in local labor markets that are highly unionized. Results To study the effects of declining union membership on wage inequality, we fix the unionization rate at 1972 levels, and allow the union effects and the effects of schooling and demographic variables to vary as observed. Figure 6 shows that the largest effect of declining union membership is on withingroup inequality. The observed within-group variance increases by.051 log points, but the adjusted variance with the fixed unionization rate increases by only.034 log points. This implies that a third of the increase in within-group inequality is associated with declining union membership. Between-group inequality increases by.046 log points and the adjusted 20

23 Utilities and Sanitation Transportation Manufacturing Communications Transportation Mining Retail Grocery Construction. Non durable Manufacturing Durable Manufacturing Printing and Publishing Entertainment and Recreation Wholesale Trade Professional Services Repair Services Business Services Retail Trade Finance, Insurance, and Real Estate Agric., Forestries, Fisheries Northeast Midwest South West Proportion Unionized Figure 3. Average unionization rates for male private sector workers, in 18 industries, and 4 regions,

24 Unionization Rate Mean Log Wage Figure 4. Unionization rates and mean log hourly wages for nonunion workers, by 72 industry-regions, Points go from light gray to black, from

25 Unionization Rate Variance of Log Wage Figure 5. Unionization rates and the variance of log hourly wages for nonunion workers, by 72 industry-regions, Points go from light gray to black, from

26 Observed Adjusted Variance of Log Wage Within Group Between Group Figure 6. Observed and adjusted within and between group variances of log hourly wages, full-time private sector men, Adjusted variances fix union membership at the 1972 level. series increases by.042, indicating that union decline explains about 10 percent of the rise in between group inequality. Summing the between group and within group effects together, the decline in unionization from 33 to 8 percent explains about 21 percent of the rise in inequality in hourly wages among full-time, private sector men. We obtain larger effects of deunionization on inequality if we also account for the effects of unions on nonunion wages (Figure 7). Again, the largest effects of deunionization are on the growth in within-group inequality. If we fix at the 1972 level, the unionization rate and the indus- 24

27 try unionization rate of regional labor markets, within-group inequality grows by just.013 log points, and not the.051 log points we actually observe. Considering the effects of local unionization rates, union decline explains nearly 75 percent of the increase in within-group inequality. The between group effect runs in the opposite direction. If local unionization rates had remained fixed, and overall union membership levels had remained at 1972 levels, between group inequality by 2006 would be slightly higher than observed. The between group effect, reflects the large wage advantage of nonunion workers in highly unionized industries in The combined between and within group effects, allowing for the union threat and normative effects indicate that union decline accounts for about 21 percent of the rise in inequality similar to the results obtained by adjusting just for falling unionization rates. An alternative approach fixes not just the industry unionization rate, but also the effects of industry unionization on the mean and variance of log hourly wages. Fixing the unionization coefficients, reflects the weakening influence of union strength on nonunion wages in regional labor markets. The coefficients for industry unionization rates themselves offer some evidence of these declining effects. The wage advantage of nonunion workers in highly unionized industries grows through the 1980s, but then declines from the late 1980s (Figure 8). Wage inequality among nonunion workers in highly unionized industries is relatively low in the early 1970s, but this inequality rises steadily. This indicates that, not only does wage inequality increase, but the pattern of inequality across regions and industries becomes more homogeneous as the variance coefficients move to zero. The growing resemblance of wage distributions in the union and nonunion 25

28 Observed Adjusted Variance of Log Wage Within Group Between Group Figure 7. Observed and adjusted within and between group variances of log hourly wages, full-time, private sector men, Adjusted variances fix union membership and regional industry unionization at the 1972 level. 26

29 (a) Effect on Mean Wages Local Unionization Effects (b) Effect on the Variance of Wages Local Unionization Effects Figure 8. Mean and variance coefficients of the local unionization rate in a regression on log hourly wages, full-time, private sector men,

30 sectors is illustrated in Figure 9. In this case we fix the level of union membership, industry unionization rates and the effects of industry unionization. These normative and threat effects are especially large for withingroup inequality. Nearly all the rise in within-group inequality is associated with declining unionization and the effects of industry unionization on within-group inequality among nonunion workers in regional labor markets. The between-group effect is again disequalizing, reflecting the high average wages of nonunion workers in highly unionized industries. All the union effects taken together can explain a third of rise in the total variance of men s hourly wage inequality. This accounting, that includes the effects of unions on nonunion wages, is 50 to 100 percent larger than the union effects on inequality reported in other decompositions (cf. Di- Nardo, Fortin, and Lemieux 1996; Card 2001). The size of the union effects on rising wage inequality can be compared to the increasing educational gradient in earnings, a focus of much of the research on men s earnings inequality. To study the education effects on inequality, we calculate an adjusted variance that fixes at 1972, the education coefficients for the mean and variance of log wages. The effect of deunionization, and the weakening threat and normative effects of unions on the nonunion sector account for about a third of the rise in men s earnings inequality. The rising education gradient in wages and the rising inequality of wages among highly educated workers explain a similar share of the rise in wage inequality (Figure 10). Finally, to assess the combined effects of union decline and the rising effects of schooling on wages we sum the two effects together, and compare these to the observed trend in earnings inequality (Figure 11). The observed variance of log wages increased by 40 percent. The adjusted vari- 28

31 Observed Adjusted Variance of Log Wage Within Group Between Group Figure 9. Observed and adjusted within and between group variances of log hourly wages, full-time, private sector men, Adjusted variances fix at the 1972 level: union membership, regional industry unionization rates, and the regional industry unionization effects. 29

32 Variance of Log Wage Observed variance Adjusting for education effects Adjusting for union effects Figure 10. Total variance of log hourly wages, full-time, private sector men, The adjusted variance fixes at the 1972 level, education effects, union membership, and regional industry unionization. Variances have been standardized to equal one in

33 Variance of Log Wage Observed variance Adjusted variance Figure 11. Total variance of log hourly wages, full-time, private sector men, The adjusted variance fixes at the 1972 level, education effects, union membership, and local industry unionization. Variances have been standardized to equal one in ance, fixing union and education effects, increased only 15 percent. Nearly two-thirds of the rise in men s hourly wage inequality is associated with declining unionization, the declining effects of unions on union wages, and the increasing effects of education on wages. These results of the variance decomposition are summarized in Table 2. Once we account for the effects of unions on nonunion wages, nearly all the rise in the within-group variance is related to deunionization. About half the rise in between group inequality is related to the increasing inequality in wages by levels of schooling. Adding all the effects, 63 percent 31

34 Table 2. Summary of the decomposition of the change in the variance of log hourly wages, private sector, full-time men, Between Within Total Observed change Change in variance fixing: Union membership Local unionization rate Local unionization effects Education effects Percentage of change explained: Union membership 10.4% 33.4% 22.4% + Local unionization rate Local unionization effects Education effects Note: Adjusted variances are obtained by successively adding the listed effects. For example, adjusted variances in line 2 are obtained by fixing union membership, adjusted variances in line 3 are obtained fixing union membership and local unionization rates, and so on. of the total rise in wage inequality from 1972 to 2006 is associated with union and education effects. Discussion This paper revisited the effect of declining union membership on rising earnings inequality in the United States. Studying wage inequality among full-time men working in the private sector, we argued that unions were important not only for raising the wages of union members but also for nonunion workers by helping to establish norms for fair pay. Like similar research on union threat effects, we found strong evidence that unionization rates in detailed industries for geographic regions was positively associated with the earnings of nonunion workers. As unionization rates fell in the national labor market, local unionization rates also declined, and 32

35 unions gradually receded from the economic scene as a voice for higher pay and better working conditions. In our variance decomposition analysis we estimated both the effect of union membership decline and the normative effect of declining regional industry unionization rates. When individual union membership is considered, union decline accounts for 22 percent of the large growth in men s earnings inequality. Adding the normative effects of unions on nonunion pay, increases the union decline effect to 33 percent. By this measure, the decline of the American labor movement has added as much to rising inequality as the increasing pay of college graduates. While the increasing stratification of earnings by education affected inequality between groups, deunionization added mostly to inequality within groups. More generally, the slow disappearance of labor unions from private sector workplaces incurred more than a loss of voice for formerly unionized workers. Unions were a normative presence that helped sustain the labor market as a social institution, in which norms of equity helped guide the allocation of wages. Of course, not all unions of the early 1970s were in the vanguard of egalitarianism. In skilled trades and construction, unions often reinforced racial and ethnic inequalities. Neither were all nonunion workplaces dominated by norms of fairness and limited markets. Still, unions offered a striking alternative to an unbridled market logic, and this institutional alternative employed over a third of all male private sector workers. The social experience of organized labor bled into the nonunion sectors, contributing to greater equality overall. As unions declined, not only did the logic of the market increasingly encroach on what had been the union sector, the logic of the market deepened in the nonunion sector too, contributing greatly to a large rise in wage inequality. 33

36 Appendix 1. Calculating the Adjusted Variances For year t, the between-group and within-group variances can be written in matrix form: W = w t(ŷ t ȳ t ) 2 and, B = w tσ 2 t where w t is a vector of survey weights, y t is a vector of conditional means of log wages, ȳ t is the grand mean, and σ 2 t is the vector of residual variances. The conditional means and variances are given by the variance function regressions (now, in matrix notation for the full sample in year t): and, ŷ t = X t α 1,t + u tα 2,t + ū tα 3,t σ 2 t = X t β 1t + u tβ 2,t + ū tβ 3,t. The first adjusted variance is based on the reweighted data: { w ti = w ti ū b /ū t if u ti = 1 w ti (1 ū b )/(1 ū t ) otherwise, where b is a baseline year set here to, b = The adjusted weights, w t, are then plugged into the between-group and within-group equations above. The second adjusted variance modifies the conditional means and variances by fixing the industry unionization terms at their 1972 values: and, ŷ t = X t α 1t + u tα 2t + ū b α 3b σ 2 t = X t β 1t + u tβ 2t + ū b β 3b. The adjusted means and variances are then plugged into the equations for the between and within group variance. The third adjusted variance fixes the education effects in α 1t and β it at their 1972 values. The modified coefficient vectors yields and adjusted set of conditional means, which are plugged into the between-group and within-group equations. 34

37 Appendix 2. Correcting for the Misclassification of Union Status The variance decomposition analysis largely follows the approach described in Western and Bloome (2009), though an adjustment must be made for the misclassification of union membership status in the Current Population Survey. Card (1996) has described this misclassification in detail, showing that validation data from an employer survey indicates that the misclassification rate of true union status is about 2.5 percent (λ =.025). If 10 percent of respondents report union membership, 2.5 percent of the 10 percent are in fact not union members, and 2.5 percent of the 90 who report being nonmembers are in fact members. More formally, if an indicator for true union status, u = 1, then p(u i = 0 ui = 1) = λ and p(u i = 1 ui = 0) = λ. If we write the observed unionization rate, ū, and the true unionization rate, ū, then ū = (1 λ)ū + λ(1 ū ). Rearranging terms yields an adjustment to the observed rate, providing the true rate of unionization: ū = ū λ 1 2λ. Card s (1996) evidence for misclassification of union status is based on survey data from 1977 January Current Population Survey. Misclassification in other years, may be larger or smaller. To accommodate this uncertainty, we d ideally like to place a prior distribution on λ. True union status, ui is also unobserved, and uncertainty about union status should be incorporated in the final results. To describe this uncertainty, we can write a probability distribution for ui, given the observed union status. The misclassification rate, λ, gives the probability u i given ui. Rearranging terms we can find the probability distribution for true union status, ui, p(u u u = 0,λ) = π 0 (1 π 0 ) u 1, where, and π 0 = λū 1 ū p(u u = 1,λ) = π 1 u (1 π 1 ) u 1, 35

38 where, (1 λ)ū π 1 =. ū To incorporate uncertainty about the misclassification parameter, we write p(u u) = p(u u,λ)p(λ)dλ, where p(λ) is a prior distribution. In the paper, we specify this distribution to be uniform on the interval, [.02,.03]. If log wages, y i, are normal, the variance function model, conditional on true union status, ui, can be written, where and y i N(ŷ i,σ 2 i ), ŷ i = x i α 1 + u i α 2 + ū i α 3 logσ 2 i = x i β 1 + u i β 2 + ū i β 3 Because y i and ui are independent, conditional on u i, we can write the likelihood, p(y i θ,u i ) = p(y i ui,θ) p(u i u i), u where all the model parameters are collected in the vector θ. A full Bayesian analysis is obtained by supplying proper priors to θ. 36