How unions affect the distribution of income is a subject that has long intrigued social scientists. The publication of What Do Unions Do? and the related papers by Freeman (1980, 1982, 1984) represented a watershed in the evolution of economists’ views on this question. Until the 1970s the dominant view was that unions tended to increase wage inequality (Johnson, 1975). Using micro data on individual workers in the union and nonunion sectors, Freeman (1980) presented results that challenged this view. He showed that the inequality-reducing effects of unions were quantitatively larger than the inequality-increasing effects. The equalizing effect of unions became a key chapter in What Do Unions Do? and an important component of the authors’ overall assessment of the social and economic consequences of unions.
This is a preview of subscription content, log in via an institution to check access.
eBook EUR 128.39 Price includes VAT (France)
Softcover Book EUR 168.79 Price includes VAT (France)
Hardcover Book EUR 168.79 Price includes VAT (France)
Tax calculation will be finalised at checkout
Purchases are for personal use only
This possibility was emphasized by Lewis (1963). The presence of unionized employers may lead to higher wages in the nonunion sector (if nonunion employers raise wages to deter unionization efforts) or to lower wages (if unionization reduces employment in the union sector, increasing labor supply in the nonunion sector).
Equation (6) is only correct if unobserved skills are rewarded equally in the union and nonunion sectors, although it may provide a good first approximation if the rewards for unobserved ability in the union sector are not too much lower than in the nonunion sector. Lemieux (1998) presented a model in which unobserved attributes are rewarded differently in the union and nonunion sectors.
A couple of studies in the late 1970s and early 1980s also pointed to the conclusion that unions lowered wage inequality. Hyclak (1979) analyzed the determinants of inequality in wage and salary income in urban labor markets and found that higher union coverage was associated with lower earnings inequality. Hyclak (1980) found a negative relationship between the state mean of union density and the percentage of families with low earnings. Hirsch (1982) performed a cross-sectional study at the industry level using a simultaneous equations model of earnings, earnings dispersion, and union coverage. He concluded that the equalizing effects of unions on earnings inequality are larger when allowance is made for the joint determination of union coverage and wage dispersion. Metcalf (1982) also looked at the dispersion of wages across industries in the United Kingdom (without controlling for the joint determination of earnings and union coverage) but concluded that union coverage widened the pay structure across industries. Metcalf also showed, however, that the variation of weekly earnings was lower in the union sector and that unions narrowed the pay structure by occupation and race.
See Lewis (1986) for a review of U.S. studies and Meng (1990) and Lemieux (1993) for Canadian evidence.
In the 2001 CLFS, 2.4 percent of male workers and 1.9 percent of female workers were covered by collective bargaining but not members of a union. The two different measures of unionization lead to nearly identical estimates of the union wage premium in a conventional linear regression of wages on union status, education, and experience.
In the United States, for example, we use bins for the log hourly wage of width 0.05. We use smaller bins for our UK and Canadian samples.
The densities are estimated using a bandwidth of 0.05. See DiNardo et al. (1996) for more detail.This is similar to DiNardo et al. (1996) who showed that the minimum wage has a much larger impact on women than on men.
The derivative of the right hand side of Eq. (4′) with respect to the unionization rate is \( _v+\left(1-2U\right)_w^2 \) . This is negative as long as Δv is large relative to \( _w^2 \) .
Taking a more direct approach, Hirsch and Schumacher (1998) examined test-score data and found that union members with high measured skills had relatively low test scores.
DiNardo et al. (1996), Card (2001), and Gosling and Lemieux (2001) all concluded that de-unionization explains very little of the increase in wage inequality among women in the United States or United Kingdom.
US Data: Since 1979, the US Census Bureau has been collecting data on weekly hours, weekly earnings, and hourly earnings (for workers paid by the hour) for all wage and salaried workers in the “outgoing rotation group” (ORG) of the Current Population Survey (CPS). Beginning in 1983, the ORG supplement of the CPS also asks about the union status of workers (and union coverage). Similar variables are also available in the May supplement of the CPS between 1973 and 1978, though only union membership (and not coverage) is available for this period.
In both the May and ORG supplements of the CPS, workers paid by the hour are asked their hourly rate of pay. We use this variable, which is collected in a consistent fashion over time, as our measure of the hourly wage rate for these workers. The May and ORG supplements also provide information on usual weekly earnings for all workers. For workers not paid by the hour, we use average hourly earnings (weekly earnings divided by weekly hours) as our measure of the wage rate.
Note, however, that weekly earnings are not measured consistently over time. From 1973 to 1993, this variable was collected by asking individuals directly about their earnings on a weekly basis. From 1994 to 2001, individuals had the option of reporting their usual earnings on the base period of their choice (weekly, bi-weekly, monthly, or annually). Weekly earnings are then obtained by normalizing the earnings reported by workers on a weekly basis. The available evidence does not suggest, however, that this change in the way earnings are collected had a significant impact on the distribution of wages (see Card and DiNardo (2002) and Gosling and Lemieux (2001) for more detail).
Another potential problem is that weekly earnings are top-coded at different values for different years throughout the sample period. Before 1988, weekly earnings were top-coded at $999. The top-code was later increased to $1923 in 1988 and $2884 in 1998. For an individual working 40 hours a week, the weekly earnings top-code corresponds to an hourly wage ranging from of $42.6 in 1984 ($2001) to $99.6 in 1973 ($2001). To keep the wage samples relatively comparable over time, we trim observations with wages above $63 ($2001). We also trim observations with wages below $2.5 ($2001), which typically corresponds to about half of the minimum wage. The wage deflator used is the Consumer Price Index (CPI-U). All the US wage statistics reported herein are also weighted using the CPS earnings weights.
Questions about educational achievement were changed substantially in the early 1990s. Until 1991, the CPS asked about the highest grade (or years of schooling) completed. Starting in 1992, the CPS moved to questions about the highest degree. We have recoded the post-1992 data in terms of completed years of schooling to have a measure of schooling that is consistent over time. We then use years of schooling to compute the standard measure of years of potential experience (age-schooling-6). Only observations with potential experience larger or equal than zero are kept in the analysis samples.
Finally, in the 1979–2001 ORG supplements of the CPS, wages or earnings of workers who refuse to answer the wage/earnings questions were allocated using a “hot deck” procedure. We exclude observations with allocated wages and earnings for two reasons. First, wages and earnings were not allocated in the May 1973–1978 CPS. We thus need to exclude allocated observations from the 1984, 1993, and 2001 ORG supplement data to maintain a consistent sample over time. Second, union status is not one of the characteristics used to match observations with missing earnings to observations with non-missing earnings in the imputation procedure (hot deck) used by the US Census Bureau. As a result, estimates of union wage effects obtained from a sample with allocation observations included can be severely biased downward (see Hirsch and Schumacher, 2004 for more details).
UK Data: As mentioned in the text, for the UK we use data from the 1983 GHS and the 1993 and 2001 UKLFS. For the sake of consistency, we exclude observations from Northern Ireland since this region was sampled in the UKLFS but not in the GHS. Real wages are obtained by deflating nominal wages with the Consumer Price Index (Retail Price Index). To limit the effect of outliers, we only keep observations with an hourly wage rate between 1.5 and 50 pounds (in 2001 pounds).
In general, we process the UK samples to make them as comparable as possible to the US samples. In both the UKLFS and the GHS, we use observations for wage and salaried workers with non-missing wages and earnings. We also use the sample weights whenever available (there are no sample weights in the GHS). Since education is not consistently measured over time, we recode education into five broad categories that are consistent over time: university graduates, higher level vocational training and A-level qualifications, middle-level vocational training or O-level qualifications, lower level vocational training, and no qualifications or diploma.
Canadian Data: As mentioned in the text, for Canada we use the 2001 Labour Force Survey (CLFS), the 1991 and 1995 Surveys on Work Arrangements (SWA), and the 1984 Survey of Union Membership (SUM). These data sets are all relatively comparable since both the SUM and the SWA were conducted as supplements to the Labour Force Survey. Relative to the US and UK data however, there are some important limitations in the Canadian data. First, as mentioned in the text, it is not possible to distinguish union membership from union coverage in the SWA. For the sake of consistency over time, we thus use union coverage as our measure of unionization in Canada.
A second limitation is that in the 1984 SUM and the 2001 CLFS missing wages and earnings were allocated but no allocation flags are provided. We thus have to include observations with allocated wages and earnings in the analysis which generates an inconsistency relative to the SWA (where missing wages and earnings are not allocated) and the US and UK data. This likely understates the effect of unions on wages in 1984 and 2001, though it is not possible to quantify the extent of the bias. Another limitation is that age is only provided in broad categories, unlike in the US and UK data where age is reported in years. In particular, it is not possible to separate workers aged 15 from those aged 16. This explains why we use all wage and salaried workers aged 15 to 64 in Canada, compared to workers aged 16 to 64 in the two other countries.
A further limitation is that hourly wages are top-coded at a relatively low level in the Canadian data. The top codes are $45 in the 1984 SUM, $50 in the 1991 SWA, $40 in the 1995 SWA, and $100 in the 2001 CLFS. For the sake of consistency, we trim observations with hourly wages above $44 (in $2001). Wages are deflated using the Canadian CPI for all items. We also trim observations with wages below $2.5 in $2001, which represents about half of the minimum wage.
One final limitation is that only five education categories are consistently available over time. These categories are: 0 to 8 years of school, high school (some or completed), some post-secondary education, post-education degree or diploma (less than university), and university degree. As in the CPS and the UKLFS, all statistics for Canada are computed using sample weights.
