The Union Wage Premium in the US and the UK

Transcription

1 The Union Wage Premium in the US and the UK David G. Blanchflower and Alex Bryson Paper for the Industrial Relations Research Association 56 th Annual Meeting, San Diego, 3-5 January 2004 David Blanchflower Department of Economics Dartmouth College Hanover, NH Internet: Alex Bryson Policy Studies Institute, 100 Park Village East London NW1 3SR We wish to thank the Economic and Social Research Council for their financial assistance (grant R ). We thank the BSAS team particularly Katarina Thomson at the National Centre for Social Research for providing the BSAS data. We acknowledge the Department of Trade and Industry, the Economic and Social Research Council, the Advisory, Conciliation and Arbitration Service and the Policy Studies Institute as the originators of the 1998 Workplace Employee Relations Survey data, and the Data Archive at the University of Essex as the distributor of the WERS data. None of these organizations or individuals bears any responsibility for the authors analysis and interpretations of the data.

2 1 Abstract This paper tracks the union membership wage premium in the US and the UK over the last couple of decades. There is evidence of both counter-cyclical and secular decline in the premium. The premium has fallen for most groups of workers, the main exception being public sector workers in the US. By the end of the 20 th Century the premium remained substantial in the US but there was no premium for many workers in the UK. The paper also traces trends in the premium in the US at the level of industry, state and occupation. In all three cases there is upward as well as downward movement in the premium characterized by regression to the mean. At industry level there is clear evidence of a rise in the premium with higher import penetration and mixed results regarding the impact of COLAs and de-regulation. With linked employeremployee data for Britain we show estimates of the membership premium tend to be upwardly biased where rich employer data are absent and that OLS estimates are higher than those obtained with propensity score matching. Among other things, we find a membership premium among covered workers with OLS estimates which is not apparent with PSM estimates. The final section reflects on what the results tell us about the nature of the union wage premium and draws out implications for unions and public policy.

3 2 1. Introduction The decline in union density in the United States and Britain on both sides of the Pond has prompted some commentators to wonder whether unions matter anymore. This is particularly so in the United States where private sector density is around 9% and the equilibrium rate is estimated at 6% (Farber and Western, 2001). But similar voices are increasingly heard in the UK where there is evidence that the new cohort of employers have turned their back on unions (Millward, Bryson and Forth, 2000; Bryson, Gomez and Willman, 2004) and where a decline in the in-flow to union membership indicates unions are finding it increasingly difficult to reach a new generation of workers (Bryson and Gomez, 2003). In both countries the public sector is accounting for an ever-increasing percentage of all union members. There are also signs that unions economic effects in the UK are increasingly benign, recent evidence suggesting that their negative effects on firms financial performance, so apparent in the 1980s, all but disappeared in the 1990s (Bryson, 2003 reviews the evidence). There has been speculation that the intensification of competition since the 1980s, coupled with a diminution of union bargaining strength, has prevented unions from obtaining the sort of wage premium they achieved in the past. This would be consistent with a diminution of unions economic effects due to a reduced ability to extract rents from employers through their monopoly power over labor. A declining union wage premium could, in turn, account for the fall in union density since, other things equal, it implies a reduction in the net benefits of membership. It is evident that unions are not as central to the economy as they used to be, but their gradual disappearance is important in its own right since it has had a profound impact on wage inequality in the US and the UK (Card, Lemieux and Riddell, 2003). Furthermore, union decline is not apparent everywhere: many

4 3 employers continue to contend with strong unions, raising important questions about union effects in those sectors. This paper draws on our recent research estimating trends in the union wage premium over the last few decades in the UK and the US. We identify which workers benefit from a premium, and how this has changed over time, analyzing trends by worker type and, for the US, by industry, state and occupation. We also address an issue that is of particular importance in the UK, but which has attracted attention recently in the US too, namely the relationship between union coverage and union membership. Using particularly rich linked employer-employee data, we explore the links between the membership and coverage premiums in the late 1990s. We also show how sensitive union premium estimates are to alternative estimation techniques and data richness, comparing OLS estimates with those obtained with propensity score matching. We conclude by reflecting on what our empirical results tell us about the nature of the union wage premium and by speculating on the implications of our analyses for unions and public policy. The remainder of the paper is set out as follows. The next section discusses the estimation of the union wage premium in the empirical literature. Section Three briefly reviews the findings from that literature in the US and the UK. Section Four presents our empirical findings for the US, with sub-sections on trends in the premium, and changes in the premium by worker type, state and industry. Section Five presents empirical evidence for the UK over time. Section Six presents cross-sectional estimates of the union wage premium in Britain using linked employer-employee data, comparing OLS estimates with PSM estimates and assessing the relationship between coverage and membership. Section Seven concludes.

5 4 2. Empirical estimation of the union membership wage premium There are two ways unions can affect wages in the economy (Farber, 2001). The first is the direct effect on the wages of workers in jobs where wages are set through collective bargaining. This may affect non-members and members wages. The second is the impact that the presence of unions has in the economy: this can change the level and distribution of wages generally. In theory, these general equilibrium effects may both raise and reduce the level of aggregate wages in the economy. Since it is not possible to observe the counterfactual (wages in the absence of unions) this union effect is not easily estimable. The union-non-union wage differential (the wage premium, or gap), defined as W W W u n =, (1) n is estimable because we observe the wages of members (W u ) and non-members (W n ). Provided differentials are small, this expression is usefully approximated by u n, (2) which says that the measured union wage gap is roughly equal to the difference in the proportional effects of unions on the union and non-union wage. The union wage premium in equation (1) can be usefully approximated by the difference in log wages, implying that ln( W ) ln( W ). (3) u n The union wage gap may reflect the direct effect of unions on the wages of unionized workers, and the offsetting effects on non-union workers. Of course, there may be endogenous selection into union status arising for two reasons. First, there is worker choice in which workers only choose membership if the union wage is greater than the wage available to the individual outside the union. It is often assumed that workers with a lower underlying earning

6 5 capacity have more to gain from membership than higher quality workers, in which case this selection process will understate the union wage premium. The second selection process arises through queuing, since not all workers desiring union employment can find union jobs. Under this model, union employers may choose the best of the workers among those desirous of a union job. This employer selection implies a positive bias in the union premium but, a priori, it is not clear whether this bias is greater or less than the negative bias implied by worker selection. Either way, if there is endogenous selection the membership mark up estimated using standard cross-sectional regression techniques can be interpreted as the average difference in wages between union and non-union workers, but it can not be interpreted as the effect of union membership on the wage of a particular worker (Farber, 2001: 11). Causal inference is problematic because, where workers who become members differ systematically from those who do not become members in ways which might affect their earnings, independent of membership, we cannot infer the non-union wage for union members simply by comparing union members wages with those of non-members. The problem of selection bias is usually tackled by modeling union status determination simultaneously with earnings to account for the simultaneity. This usually involves a Heckman estimator where the earnings function and union status determination function are assumed to have errors that are jointly normal. This technique relies on exclusion restrictions whereby variables assumed to affect union status have no direct effect on earnings. In his review of the literature, H. Gregg Lewis (1986) concluded that, because of these arbitrary functional form assumptions and untestable exclusion restrictions, results from these studies were unreliable. Estimates of the union wage gap using simultaneous equation methods tend to produce large and unstable estimates. Panel estimates, which involve making use of data on the same individuals over time

7 6 and observing how wages change as individuals change their union status have problems of misclassification and measurement error which tend to result in estimates of the impact of unions that are downward biased. Lewis (1986) takes the view that the most appropriate way to estimate the impact of unions on wages is using OLS. He suggests OLS may produce an upper bound estimate of the true impact of unions because such estimates suffer from upward bias resulting from the omission of control variables correlated with the union status variable (Lewis, 1986: 9). The assumption is that some of the wage gap attributed to union membership is, in fact, attributable in part to the characteristics of members, their jobs and their employers which would give them higher wages than non-members in any case. In practice, bias in crosssectional OLS estimation due to unobserved heterogeneity may both upwardly or downwardly bias the true impact (Farber, 2001; Robinson, 1989). For most of this paper, our primary concern is with changes in the union wage premium over time. We do not seek to control for the potential endogeneity of union membership. Rather, we adopt the standard approach to estimation of the union-non-union wage gap using individual-level data and estimating by OLS. That is, lnw it = X it ß + δu it + ε it, (4) where subscript it indexes individuals over time, X it is a vector of worker, job and workplace characteristics, U it is a dummy variable indicating union membership, and ε it is a random component. The parameter δ represents the average proportional difference in wages between union and non-union workers adjusted for worker and workplace characteristics, and it is the regression-adjusted analogue of. In our work, we assume that any bias in our estimates of the δ over time arising through unobserved heterogeneity remains constant over time. However, in Section Six we capitalize on very rich linked employer-employee data to compare OLS estimates

8 7 to those derived using propensity score matching, a technique that has not been used in the literature before, and will be explained later. 3. Empirical findings in the US and UK literatures The definitive empirical works in this area are by H. Gregg Lewis, (1963, 1986) 1. Lewis (1986) found that the overall impact of unions in the US economy was approximately 15 per cent and showed relatively little variation across years varying between 12 per cent and 19 per cent between 1967 and Subsequent work confirmed constancy of the differential until the 1990s. For example, Hirsch and his co-authors have produced a series of papers estimating changes in the differential over time and concluded there has been some decline in the premium in recent years (e.g. Hirsch, Macpherson and Schumacher, 2002; Hirsch and Schumacher, 2002; Hirsch and Macpherson, 2002). Bratsberg and Ragan (2002) examine the trend in the private sector union wage differential in the US, , and conclude that dispersion in the wage premium across industries has substantially declined as the US economy has become more competitive but that there has been only a modest decline in the average premium. Bratsberg and Ragan confirmed the stability of the premium over time, as noted in Linneman, Wachter and Carter (1990), but did observe some evidence of a decline in the premium at the end of the 1990s. There are reasons to believe that the union wage gap might vary with the business cycle. If the union premium comes from employers sharing rents, it is plausible that the premium will be higher when those rents are higher, in which case the wage gap would be pro-cyclical. Alternatively, unions may insulate their members from the downward wage pressures workers in general face in more difficult times, in which case the wage gap may be counter-cyclical

9 8 (Freeman and Medoff, 1984). Grant (2001) used panel data on individuals from the CPS from 1975 to 1993 to examine the cyclicality of union and non-union wages over time. He found that the union coefficients in the non-union sector for the two periods and were always procyclical and generally similar in the two periods. In contrast in the union sector Grant found strong procyclicality in the first period, confirming earlier evidence in Moore and Raisian (1980), but weak or no procyclicality in the union sector in the second period. Wunnava and Okunade (1996) used data for men from the Panel Survey of Income Dynamics and found an overall union wage premium of about 12 per cent for the 1980s, which is a good deal lower than reported in most other studies and possibly driven by measurement error in the union status variable. In response to fluctuations in local labor market conditions, proxied by the local unemployment rate, they found a much more flexible wage-setting process in the non-union sector relative to the union sector. The long-term effect of unemployment on non-union real wages suggested an approximate 0.6 per cent decline for every one percentage point increase in unemployment, but the long-term effect of unemployment on real wages of union members was negligible. Wunnava and Okunade s estimates of the union wage premium ranged between 11.6 to 12.3 per cent for the sample period. Union wages were found to be insensitive to short-run fluctuations in local labor market conditions, and counter-cyclical in nature. We come back to this issue later since we find evidence of a counter-cyclical wage gap in the US and UK in the 1990s. In the UK there have been over thirty studies, some based on establishment data and others on individual data (including some using linked employer-employee data). Industrial relations are rather more complex in the UK than they are in the United States. For example, in the UK many more non-members work in workplaces that are covered by union agreements and, conversely,

10 9 more union members are employed in workplaces where unions are not engaged in pay bargaining than is true for the US. Using CPS data for the period , Budd and Na (2000: 783) estimate that 10 percent of private sector employees are covered by collective bargaining but are not union members. Coverage and membership are less highly correlated in Britain than in the United States because there is less pressure on employees to become members where there is a coverage agreement (Hildreth, 2000: ). In the private sector in 1998, 64% of union members belong to a covered occupation, while 13% of non-members are covered. Around one-third (35%) of covered employees are non-members (Bryson, 2002a). There is, correspondingly, a multi-faceted literature in the UK which has investigated the free rider problem (see Booth and, Bryan, 2001; Hildreth, 2000) as well as the importance of union recognition (Blanchflower, 1984; 1986), multiple unionism (Machin, Stewart and van Reenen, 1993) and closed shops (Stewart, 1987; Blanchflower, Oswald and Garrett 1989; Metcalf and Stewart, 1992). There are one or two papers in the US also on the role of coverage, including Blakemore et al. 1986, Budd and Na, 2000, and Schumacher, We shall return to this issue in Section Six in the UK context. But throughout most of the paper our main focus is a comparative one involving the benefits of union membership on wages. The recent spate of studies that have looked at the impact of union membership on wages in the UK has been occasioned by a growing belief that the union wage premium may be falling. Some argue that a decline in the average union premium is consistent with diminishing union influence over pay setting. There is certainly evidence pointing in that direction. First, case studies suggest the scope of bargaining has narrowed substantially in companies that continue to bargain with unions (Brown et al., 1998). Second, pay settlements in the private sector by the end of the 1990s were no greater where trade unions were involved than in their absence (Forth and

11 10 Millward, 2000b). Third, even where managers say employees have their pay set through workplace-level or organization-level collective bargaining, union representatives and officials are either not involved or are only consulted in a substantial minority of cases (Millward, Forth and Bryson, 2001). But there is also evidence to the contrary. For example, unions continue to have a substantial effect on pay structures, bringing up the wages of the lowest paid and thus narrowing pay differentials across gender, ethnicity, health and occupation (Metcalf, Hansen, and Charlwood, 2001). These studies, which indicate union effects despite substantial declines in union density, might suggest that those unions that have survived are the stronger and, as such, better able to command a wage premium (thus raising the batting average of unions). Here we briefly review what studies to date have told us about the size of the union wage premium over time and across workers. The consensus in the earlier literature is that the mean union wage gap was approximately 10 per cent (Blanchflower, 1999). Despite the rapid decline in union density experienced in the UK since 1979, the gap remained roughly constant from 1970 the year for which the earliest estimate is available (Shah, 1984) to 1995 (see Blanchflower, 1999). However, there is some dispute on this question, with some studies pointing to trends in either direction. For instance, establishment-level analyses indicated that the union wage premium in the early 1980s was most evident where unions were strong, as indicated by the presence of a closed shop (Stewart, 1987). This premium seems to have declined in the second half of the 1980s, a trend which has been attributed to a decline in the incidence and impact of the closed shop, coupled with unions inability to establish differentials in new workplaces (Stewart, 1995). On the other hand, Andrews, Bell and Upward (1998) find the bargaining coverage differential over the period moved counter-cyclically, with an underlying upward trend which they attribute to

12 11 the decentralization of pay bargaining. In addition to cross-sectional estimates, there has been a series of papers producing estimates for this period based on longitudinal data for Britain using the British Household Panel Survey (Blanchflower, 1999; Hildreth, 1999; Swaffield, 2001; Machin, 2001). As noted earlier, and as both Lewis (1986) and Freeman (1984) pointed out, these estimates tend to be below the estimates obtained by OLS because of a downward bias induced by measurement error in the classification of union status. As also noted earlier, OLS estimates may be upwardly biased if unobserved heterogeneity accounts for some of wage variation attributed to union membership. Thus, for example, in Blanchflower (1999) the OLS estimate for the years was 10.6 per cent compared with 3.7 per cent when a full set of people fixed effects were included. 2 As Freeman (1984: 24) has suggested, it may well be that the cross-section and fixed effect or panel estimates of the impact of unions on wages bound the true impact of unionism. Studies using individual pay data covering the first half of the 1990s also suggested that, while the union effect was persisting, the premium declined for some workers (Blanchflower, 1999; Hildreth, 1999). For example, Hildreth (1999: 7) argues that stability in the union premium for blue-collar male workers in compared with a declining premium for their white-collar counterparts may reflect their respective abilities to maintain their bargaining power. The picture emerging from research through to 1998/99 is suggestive of a more widespread decline in the premium. Machin s (2001) analysis of longitudinal data from the British Household Panel Survey indicates that, although there was a wage gain for people moving into union jobs in the early 1990s, this had disappeared by the late 1990s. Booth and Bryan (2001) using linked employer-employee data for 1998 also found no significant wage premium. Forth

13 12 and Millward (2000a) find the premium was confined to workers in workplaces with high bargaining coverage or multiple unions. It would be hasty to assert, on the basis of this evidence alone, that unions ability to secure better than market rates for their workers has declined since the 1980s because methodological and data differences across studies make comparisons extremely difficult (Andrews et al., 1998). It is even more difficult to establish what has happened to the trend over time. As Lanot and Walker (1998: 343) note: the existing literature says little about how the differential has changed over time there are so few studies it is difficult to take a view of whether there is any systematic movement over time. For instance, using standard regression techniques deployed in most studies, Booth and Bryan (2001: 12) identify a membership premium of roughly 10 per cent. However, the authors lay emphasis on the results they obtain through the use of instrumental variables techniques. The disaggregated pattern of results reported by Lewis (1986) for the US are broadly repeated for the UK. The main exception is that the wage gap in the UK appears to be larger for females than it is for males (Blanchflower, 1999; Main, 1996). We explore this issue in more detail below. Before moving to estimating union wage gaps, it is appropriate to place these results in the wider context of the changes in the labor market experience of the two countries over the last couple of decades; specifically, in terms of unemployment and employment; wage inequality, real wage growth and union density. This allows for some appreciation of the climate in which unions have been operating. 1. Unemployment was generally higher in the US than it was in the UK from 1965 to The picture reversed itself in the later period, In 2000 and 2001 the

14 13 unemployment rate in the UK was below that of the United States, averaging 3.4 per cent and 4.4 per cent, respectively. Both employment and the size of the labor force increased rapidly over the period in the US. Over this period, employment in the US increased by 14 per cent while the labor force increased by 12 per cent 3. The UK experienced smaller growth along both of these dimensions, with respective growth rates of 7 per cent and 3 per cent Levels of earnings and wage inequality are high in the US and the UK compared with most other countries, and especially so in comparison with most European countries (Blanchflower, 2000). There was substantial growth in earnings inequality in the 1970s and 1980s in the US. Since the early 1970s earnings in the US have become much more unequal between more-skilled and less-skilled workers as well as between workers with high and low levels of education and those with many years of labor market experience compared to those with few. Earnings inequality declined in the UK in the 1970s but increased in the 1980s. Only Britain and the United States have continued to experience a rapid rise in inequality into the 1990s, albeit at a slower rate than had occurred in the 1980s. There is much less evidence of rising wage inequality in other countries (see the various papers in Freeman and Katz, 1995). 3. In the United States real wage growth has been much greater at the top of the earnings distribution than at the bottom. In the hundred years to 1973, real average hourly earnings rose by 1.9 per cent a year. Between 1973 and 1997, CPI-deflated real wages have fallen by about 0.4 per cent a year. The combination of flat average wages and rising inequality means that large numbers of American workers have experienced stagnation or even absolute declines in their real earnings in recent

15 14 decades. And workers at the low end of the earnings distribution have suffered the most, particularly those in the lowest decile. For example, the real hourly earnings of high-school-educated males fell by 20 per cent from 1979 to In contrast, there has been considerable growth in real earnings at the top of the earnings distribution. Senior managers and executives have experienced large increases in real earnings over the last couple of decades, and especially so when total compensation including stock options are included. In contrast to the United States, in most OECD countries (including the UK) there has been strong real earnings growth across the wage distribution 6. The low-paid in most industrial countries have experienced real earnings growth over the last two decades Union density rates declined steadily in the US from In Britain density increased in the 1970s and then declined dramatically (Figure 1). Since 1991, there has been a decrease in union membership of 1.3 million, a fall over the ten-year period of 15 per cent. The fall in union membership has been steeper for men than for women over the past decade: union density for men was 42 per cent in 1991 and 29 per cent in 2001, whereas that for women was 32 per cent in 1991 and 28 per cent in (Figure 1 near here) 4. Trends in the union wage premium in the United States Table 1 presents estimates of the wage gap using separate log hourly earnings equations for each of the years from 1973 to 1981 using the National Bureau of Economic Research s (NBER) May Earnings Supplements to the Current Population Survey (CPS) 8 and for the years since then using data from the NBER s Matched Outgoing Rotation Group (MORG) files of the CPS 9. For

16 15 both the May and the MORG files a broadly similar, but not identical, list of control variables is used, including a union status dummy, age and its square, a gender dummy, education, race and hours controls plus state and industry dummies. (Table 1 near here) The first column of Table 1 reports time-consistent estimates of the union wage premium for the union coefficient in log hourly earnings equations for the total sample whereas the second and third columns report them for the private sector. Results obtained by Hirsch and Schumacher (2002) with a somewhat different set of controls are reported in the final column of the table. For a discussion of the reason for these differences, see Blanchflower and Bryson (2003a). The time series properties of all three of the series are essentially the same. The wage gap averages between 17 and 18 percent over the period, and is similar in size in the private sector as it is in the economy as a whole. Table 1 confirms Freeman and Medoff s comment (1984: 53) that the late 1970s appear to have been a period of substantial increase in the union wage premium. What is notable is the high differential in the early-to-mid 1980s and a slight decline thereafter, which gathers pace after 1995, with the series picking up again as the economy started to turn down in (Figure 2 near here) Figure 2 plots the point estimates of the US private sector union wage premium, taken from the second column of Table 1, against unemployment for The premium moves counter-cyclically. There are three main factors likely influencing the degree of counter-cyclical movement in the wage gap. The first, cited by Freeman and Medoff (1984: 52-53) as the reason for the widening wage gap during the Depression of the 1920s and 1930s, is the greater capacity for unionized workers to fight employer efforts to reduce wages when market conditions are

17 16 unfavorable. Conversely, when demand for labor is strong, employees rely less on unions to bargain for better wages because market rates rise anyway. The second factor is that union contracts are more long-term than non-union ones and, as such, less responsive to the economic cycle, so union wages respond to economic conditions with a lag. When inflation is higher than expected, a greater contraction in the premium can occur because non-union wages respond more to higher inflation. However, the third factor, which should reduce the cyclical sensitivity of the union wage premium, is the cost-of-livingadjustment (COLA) clauses in union contracts that increase union wages in response to increases in the consumer price level. According to Freeman and Medoff (1984: 54) the percentage of union workers covered by these agreements rose dramatically in the 1970s, from 25 percent at the beginning of the decade to 60 percent at the end of the decade. However, their estimates for manufacturing suggest that COLA provisions contributed only a modest amount to the rising union advantage in the 1970s. Bratsberg and Ragan (2002) revisit this issue and find the increased sensitivity of the premium to the cycle is due in part to reduced COLA coverage from the late 1980s, but we find no such evidence (see below). Commenting on the growth of the union wage premium during the 1970s, Freeman and Medoff (1984:54) suggested that at least in several major sectors the union/nonunion differential reached levels inconsistent with the survival of many union jobs. They were right. In the 1970s and early 1980s, the wage gap in the private sector rose while union density fell, as predicted in the standard textbook model of how employment responds to wages where the union has monopoly power over labor supply. In the classic monopoly model, demand for labor is given, so a rise in the union premium results in a decline in union membership since the premium hits employment. The fact that unions pushed for, and got, an increasing wage premium over

18 17 this period, implies that they were willing to sustain membership losses to maintain real wages, or that unions were simply unaware of the consequences of their actions. From the mid-1990s, the continued decline in union density was accompanied by a falling union wage premium because demand for union labor fell as a result of two pressures. The first was increasing competitiveness throughout the U.S. economy: Increasing price competition in markets generally meant employers were less able to pass the costs of the premium onto the consumer, so that pressures for wages to conform to the market rate grew. Secondly, unionized companies faced greater nonunion competition. Declining union density, by increasing employers opportunities to substitute nonunion products for union products, fueled this process. So too did rising import penetration: If imports are nonunion goods, regardless of U.S. union density, they increase the opportunity for nonunion substitution. These same pressures also increased the employment price of any union wage gap (the elasticity of demand for unionized labor). It is possible that some of the decline in the private sector wage premium is driven by differential selection into union membership over time. We compared estimates of the regression-adjusted premium since 1983 from column 2 of Table 1 with the unadjusted premium where the wage equation contains only the membership dummy 10. We find the unadjusted wage gap is higher than the adjusted wage gap throughout, implying a positive association between union membership and wage enhancing employee or employer characteristics. However, the unadjusted gap has declined more rapidly than the regression-adjusted estimate. In 1983 the unadjusted estimate was 128 percent higher than the adjusted estimate. By 2002, the difference had fallen to 91.5 percent higher. This might imply greater selection into union membership among employees with lower underlying earnings potential along dimensions that we do not

19 18 observe. Alternatively, the fact that observable characteristics account for an increasing percentage of the union wage gap might imply a decline in the real causal impact of unions in obtaining a premium for their members. Hirsch, Macpherson and Schumacher (2002), in a formal decomposition of the private sector wage gap between 1986 and 2001, find almost half (46 percent) of the decline in the union-non-union log wage gap is accounted for by a decline in the regression-adjusted wage gap, 16 percent arises through changes in worker characteristics and payoffs to those characteristics, and 38 percent is due to sectoral shifts and payoffs to the occupational sectors of workers. (Table 2 near here) Private sector by worker type. Table 2 presents union wage gaps obtained from estimating a series of equations for sub-groups of private sector employees since the mid-1970s. To ensure large sample sizes we pooled together six successive May CPS files from and compare those to wage gaps estimated for the years using data from the Matched Outgoing Rotation Group (MORG) files of the CPS. Two points stand out from these analyses. First, no group of workers in the private sector sample has experienced a substantial increase in their union premium. Indeed, the only group recording any increase at all is those aged whose premium rose from 13 percent to 14 percent. Clearly, unions have found it harder to maintain a wage gap over time. Second, with the exception of the manual/nonmanual gap, those with the highest premiums in the 1970s saw the biggest falls, so there has been some convergence in the wage gaps. This trend may be due to an increasingly competitive U.S. economy, where workers commanding wages well above the market rate are subject to intense competition from nonunion workers. Nevertheless, with the exception of the most highly educated and non-manual workers, the wage premium remains around 10 percent or more.

20 19 Public Sector by worker type. The size of the public sector grew (from 15.6 million to 19.1 million or 22.4 percent) between 1983 and 2001, but as a proportion of total employment it fell from 18.0 percent to 16.1 percent. Union membership in the public sector grew even more rapidly (from 5.7 million to 7.1 million or by 24.6 percent). Furthermore, by 2001 public sector unions accounted for 44 percent of all union members compared with 32.5 percent in (Table 3 near here) Table 3 is comparable to Table 2 for the private sector in that it presents disaggregated union wage gap estimates but, due to data constraints, the base period is from the early 1980s. Because sample sizes in the public sector are small using the May CPS files we use data from the ORG files of the CPS for the years for comparison purposes with the data. Data for the years could not be used, as no union data are available. A further advantage of the data is that information is available on individuals whose earnings were allocated who were then excluded from the analysis. The main findings are as follows. 1) The private sector union wage gap has fallen over the two periods (21.5 percent to 17.0 percent) whereas a slight increase was observed in the public sector (13.3 percent to 14.5 percent, respectively). 2) The majority of the worker groups in Table 3 experienced increases in their union wage premium over the two periods, but wage gaps declined markedly for those under 25 and with less than a high school education. 3) There was little change in public sector union wage gaps for men or women. In marked contrast to the private sector where men had higher differentials than women, wage gaps in both periods in the public sector were higher for women than for men.

21 20 4) Unions benefit workers most in local government and least in the federal government, although the differential for federal workers increased over time. 11 5) Just as for the private sector, the wage benefits of union membership are greatest for manual workers, the young, and the least educated. 6) There are only small differences in union wage gaps for nonwhites compared to whites in both the public and the private sectors. 7) In contrast to the private sector where wage differentials were greatest in the South and West, in the public sector exactly the opposite is found. Differentials are higher in the public sector in New England and the Central region in both time periods whereas the reverse was the case in the private sector. 8) Wage gaps increased over time for teachers, lawyers, firefighters and police. Industry, Occupation and State-Level Wage Premia 12. So far, we have focused primarily on union wage effects at the level of the individual and the whole economy. However, the literature on the origins of the union wage premium focuses largely on firms and industries because the conventional assumption is that unions can procure a wage premium by capturing quasi-rents from the employer (Blanchflower et al., 1996). If this is so, there must be rents available to the firm arising from its position in the market place, and unions must have the ability to capture some of these rents through their ability to monopolize the firm s labor supply. Individual-level data can tell us little about these processes. Instead, the literature has concentrated on industrylevel wage gaps. In this section we model the change in the union wage premium at three different units of observation-industry, state, and occupation. Industries. We used our data to estimate separate results for 44 two digit industries for and We chose these years as it was possible to define industries identically

22 21 using the 1980 industry classification. We found considerable variation in the size of the wage gap by industry, but there is less variation in the later period than in the earlier period. Only 3 industries - construction (41 percent), transport (36 percent) and repair services (37 percent) - had a differential of over 35 percent in the second period, compared with 6 in the earlier period which includes the same three -- construction (52 percent); transport (44 percent) and repair services (37 percent) plus agricultural services (41 percent); other agriculture (56 percent) and entertainment (47 percent). Where is the union wage premium rising, and where is it falling? In contrast to the analysis by worker characteristics, which reveal near universal decline in the premium at least in the private sector we found that the wage gap rose in 17 industries and declined in 27. The gap rose by more than ten percentage points in autos (+12 percent) and leather (+19 percent). It declined by more than twenty percentage points in other agriculture (-33 percent) retail trade (-20 percent) and private households (-29 percent). Many of the industries experiencing a rise in the union premium between 1983 and 2001 would have been subject to intensifying international trade (machinery, electrical equipment, paper, rubber and plastics, leather) but this was equally true for those experiencing declining premiums (such as textiles, apparel and furniture). There is a negative correlation between change in union density and change in the premium (correlation coefficient 0.39). Some of the biggest declines in the premium have been concentrated in the three industries with more than a 10 percent share in private sector union membership in In construction and transport, which both make up an increasing proportion of all private sector union members, the premium fell by around 10 percentage points. In retail trade, where the share of private sector union membership has remained roughly constant at 10 percent, the premium fell 20 percentage points. The decline in the wage gap for

23 22 the whole economy, presented earlier, is due to the fact that the industries experiencing a decline in their wage gap make up a higher percentage of all employees than those experiencing a widening gap. The results are similar to those presented by Bratsberg and Ragan (2002) who found that, over the period , the regression-adjusted wage gap closed in 16 industries and increased in 16 others. (Table 4 near here) To explore these changes in the private sector industry union wage premium over time we ran panel fixed effects estimates. Our first step was to estimate separate first-stage regressions for each of our industries in each year from with the dependent variable the log hourly wage along with controls for union membership, age, age squared, male, 4 race dummies, the log of hours, and 50 state dummies. Three sectors with very small sample sizes (toys, tobacco, and forestry and fisheries) were deleted. We extracted the coefficient on the union variable, giving us 19 years * 42 industries or 798 observations in all. The coefficient on the union variable was then turned into a wage gap taking anti-logs, deducting 1 and multiplying by 100 to turn the figure into a percentage. We used the ORG files to estimate the proportion of workers in the industry who were union members both in the private sector and overall and mapped that onto the file. Unemployment rates at the level of the economy are used as industryspecific rates are not meaningful: workers move a great deal between industries and considerably more than they do between states. Regression results, reported in columns 1 and 2 of Table 4, estimate the impact of the lagged premium, lagged unemployment, and a time trend on the level of the industry-level wage premium. The number of observations is 756 as we lose 42 observations in generating the lag on the wage premium and the union density variables.

24 23 In the unweighted equation in column (1) the lagged premium is positively and significantly associated with the level of the premium the following year indicating regression to the mean. Unemployment and the time trend are not significant. However, once the regression is weighted by the number of observations in the industry in the first-stage regression, (column (2)) lagged unemployment is positive and significant, indicating counter-cyclical movement in the premium, while the negative time trend indicates secular decline in the premium. Bratsberg and Ragan (2002) reported that the industry-level premium was influenced by a number of other variables. 13 In particular they found that COLA clauses reduced the cyclicality of the union premium and that increases in import penetration were strongly associated with rising union premiums. They also found some evidence that industry deregulation had mixed effects. Their main equations (their Table 2) did not include a lagged dependent variable. Table 5 reports results using their data for the years using their method and computer programs that they kindly provided to us. Column 1 of Table 5 reports the results they reported in column 2 of their Table 2. Column 2 reports our attempt to replicate their findings. We are unable to do so exactly but there are several similarities we find import penetration both in durables and nondurables, COLA clauses, deregulations in communications, and the unemployment rate all have positive and significant effects. We also found, as they did, that deregulation in finance lowered the premium. In contrast to Bratsberg and Ragan, however, the inflation rate and the two interaction terms with the unemployment rate were insignificant. The model is rerun in column 3, but without the insignificant interaction term. A linear time trend is added in column 4: this is negative and significant, and eliminates the COLA effect and the negative effect of deregulation in the finance sector. Column 5 adds the lagged union wage premium, which is positive and significant. Its introduction makes inflation positive and

25 24 significant. In columns (6) to (8) models are run without the four insignificant deregulation dummies. Column (6) indicates that using an unweighted regression, the size of the lagged premium effect drops markedly and the time trend and inflation lose significance, showing these results are sensitive to the weighting of the regression. The smaller coefficient on the lagged dependent variable is unsurprising given that there is much less likely to be variation in the union wage gap estimates in industries with large sample sizes that have higher weights in the former case. We are able to confirm Bratsberg and Ragan s finding that the unemployment rate, deregulation in communications, and import penetration in both durables and nondurables have positive impacts on the premium but not the findings on COLA, inflation, or any of the other deregulations identified. That import penetration in durable and nondurable goods sectors increases the premium suggests that union wages are more resilient than nonunion wages to foreign competition. Import penetration is likely correlated with unmeasured industry characteristics that depress the premium inducing a negative bias that is removed once industry characteristics are controlled for. Import penetration has likely reduced demand for union and nonunion labor, with union wages holding up better than nonunion wages, but at the expense of reduced union employment. There are theoretical and empirical reasons as to why this might occur. For instance, since union wages tend to be less responsive to market conditions generally, union wages may be sluggish in responding to increased import competition. Alternatively, industries characterized by endgame bargaining may witness perverse union responses to shifts in product demand as the union tries to extract maximum rents in declining industries (Lawrence and Lawrence, 1985). Another possibility is that increased import penetration reduces the share of union employment in labor - intensive firms and increasing it in capital-intensive firms. Greater capital intensity reduces

26 25 elasticity of demand for union labor, allowing rent maximizing unions to raise the premium (Staiger, 1988). It isn t obvious that weights should be used if we regard each industry as a separate observation. In cross-country comparisons which, say, contrasted outcomes for Switzerland, the UK and the U.S., it wouldn t make a lot of sense to weight by population and thereby make the observation from the U.S times more important than that of the UK and 39.3 times more important than Switzerland. 14 Columns (1) to (6) are GLS estimates accounting for potential correlation in error terms. Column (7) switches to a weighted OLS and shows that results are not sensitive to the switch. The unweighted OLS in column (8) gives broadly the same results as the unweighted GLS in column (6). Taking off the weights has a much bigger effect than switching from GLS to OLS. Furthermore, the industries defined by Bratsberg and Ragan are very different in size. Some industries are very broadly defined for example industry 32 Services covers SIC codes whereas tobacco, for example, covers one SIC code (#130). Retail trade averaged 19,075 observations. Column 9 of Table 5 illustrates the sensitivity of the results to industry exclusions. It is exactly equivalent in all respects to column 5 of Table 5 except that it drops the 32 observations from retail trade. The lagged dependent variable falls dramatically from.60 to.32. The COLA variable is now significantly positive while the inflation variable moves from being significantly positive to insignificant. The unweighted results (not reported) are little changed. Bratsberg and Ragan s results appear to be sensitive to both the use of weights and the sample of industries used. States. A similar procedure was adopted to estimate state-level premia over time for the 50 states plus Washington D.C. We compare results using merged samples of the CPS s MORG

27 26 for and files. 15 The correlation between changes in state density and state premia is negative but small (-0.10). The variation in the union wage premium is much less by state than it is by industry. (Full results by state are presented in Blanchflower and Bryson, 2003b). Only two states in the earlier period had gaps of at least 35 percent - North Dakota (35 percent) and Nebraska (37 percent) and none in the later period. There has been a downward shift in the premium generally. The mean state union wage gap was 23.4 percent between 1983 and 1988, falling to 17.2 percent in The premium fell in all but five states, with South Dakota recording the biggest decline (16.8 percentage points). In four of the five states where the premium rose, it only increased by a percentage point or two (Vermont, Massachusetts, Wyoming, and Hawaii). The premium only rose markedly in Maine, where it increased 9 percentage points (from 7 percent to 16.1 percent). Since the early 1980s, union density fell by an average of 5.7 percentage points, with Pennsylvania (-10.6 percent) and West Virginia experiencing the biggest decline (-11 percentage points). The premium appears to have declined more in smaller states than it has in bigger states. The five biggest states of California, Texas, Florida, New York, and Illinois had small changes in their wage gaps (-1.4 percent; -6.7 percent; percent; -0.6 percent and -4.5 percent, respectively). The five smallest states measured by employment tended to have big declines in the differentials: New Mexico ( percent); Alabama ( percent); Nebraska ( percent); Arkansas ( percent); South Dakota ( percent). 16 We then ran 969 separate first-stage regressions, one for each state in each year from with the dependent variable the log hourly wage along with controls for union membership, age, age squared, male, 4 race dummies, the log of hours, and 44 industry dummies. The sample was restricted to the private sector. We extracted the coefficient on the