Unionization and the pattern of nonunion wage supplements.

Heywood, John S.

I. Introduction

Negotiated standard wage policies reduce dispersion in the log earnings of union members when compared to those of otherwise equal nonunion members |5; 6~. As large as these dispersion differences are, they may underestimate the true influence of union wage policy. Kahn and Curme |14~ have suggested that high degrees of unionization actually result in reduced nonunion wage dispersion. This follows from a particular model of the union threat effect in which nonunion employers tilt induced wage increases toward low-wage workers. Such tilting makes sense because lower paid workers are assumed to be more likely to support unions, ceteris paribus. This assumption fails to emphasize that low wage workers may be the least likely to be retained by a newly unionized employer and that this also influences their ultimate decision to support unionization. In the face of union avoidance strategies that routinely imply the possibility of job loss, low wage employees may recognize that the employer faces greater costs of bringing them up to a standard union wage and thus their chances for retention are lower.(1) Including this realization results in ambiguous predictions about the influence of unionization on nonunion dispersion.

Not only should the ambiguity of the proper theoretical prediction be highlighted, but it should be recognized that the relevant prediction from theoretical models has not yet been tested. Previous tests are inappropriate because the threat induced wage increases could be tilted toward low-wage workers without a necessary reduction in the typical measures of wage dispersion. Similarly, those measures of wage dispersion may actually decline without a general tilt of wages toward low-wage workers. This possibility exists because of the sensitivity of the usual variance measures to outliers and because of aggregation and endogeneity problems affecting previous methodologies.

This paper uses 1983 and 1988 Current Population Survey (CPS) data to directly estimate the impact of unionization on the earnings of low- and high-wage nonunion workers. After controlling for traditional human capital, locational, and occupational variables, unionization increases the log-earnings of low-wage nonunion workers significantly more than those of high-wage nonunion workers. This result is found with the very data Belman and Heywood |2~ used to demonstrate that inionization may not reduce nonunion wage dispersion suggesting the importance of separating the pattern of threat-induced wage increases from specific measures of wage dispersion.

A variety of alternative specifications initially confirm the consistency of this empirical result. Yet, once managers, and then professionals, are removed from the sample, the results vanish. The ultimate conclusion is that evidence on the pattern of nonunion wage supplements is extremely sensitive to the sample composition. The workers least likely to be receiving supplements for institutional reasons (managers) are also disproportionately in the upper half of the nonunion wage distribution. Removing these workers from the sample shows that wage supplements are not tilted toward either side of the distribution and reinforces the theoretical insight that threat effects need not disproportionately increase the earnings of low wage workers.

II. Threat Effects and Nonunion Wages

Economists have previously identified a variety of specific and contradictory influences of unionization on nonunion earnings. If increased union wages cause displaced union workers to increase the supply of nonunion labor, the equilibrium nonunion wage will fall. Alternatively, if the turnover of unionized positions is high enough, the nonunion wage might actually increase as nonunion workers quit to stand in a queue waiting for a union job. Perhaps the most controversial predicted consequence of unionization dates from Rosen |21~ and is that the threat of becoming unionized will appear so costly to nonunion employers that they will increase wages. Such wage increases are conceived as minimizing expected labor costs by reducing the probability of unionization.

Past empirical evidence provides support for a positive correlation between the extent of unionization and nonunion wages. Podgursky |20~, Moore, Newman and Cunningham |17~ and Hirsh and Neufeld |11~ each confirm that nonunion workers in more unionized markets earn higher wages.(2) Such evidence has been taken to indicate that the threat effect, perhaps combined with the queuing effect, dominates any tendency for previously unionized workers to crowd the nonunion labor market.(3)

The tendency of unions to narrow wage dispersion for its own members argues that an optimizing nonunion firm may respond to a union threat by not giving equal wage increases to each of its employees. Within occupations, union standard rate policies raise the wages of low-wage union workers more than those of high-wage workers when compared to typical nonunion pay systems |6~. Similarly, typical union across-the-board raises and cost of living adjustments also have a tendency to increase earnings of low-wage workers more than would be expected in a nonunion workplace. The consequence is that otherwise equal low-wage workers receive the largest wage premium from union membership. As a result, low-wage nonunion workers are often thought to be the most likely to desire unionization. In response, a cost minimizing firm might tilt any wage increases toward these low-wage workers in order to reduce the probability of becoming unionized in the cheapest possible fashion.

Such reasoning is laid out in a formal model by Kahn and Curme |14~ but exhibits a serious limitation. The limitation stems from the insight that the probability of retaining a union job may vary positively with the level of nonunion earnings. That is, while low paid nonunion workers would be most likely to want a union job, they are also the least likely to be given one. This fact has been confined by union queue and selection equations and makes good sense |1~. Those workers whose current wage is well below the union wage are those whom the unionized employer will not select because of the costly gap between their marginal productivity and the union wage. Thus, nearly every characteristic which correlates positively with being in the union queue, correlates negatively with being selected from it. The consequence is that it is sensible to think about two probabilities, the probability of a worker voting for unionization if continued employment is certain, and the probability of retaining employment once unionized. If unemployment has positive costs, the rational nonunion employee will ultimately base support for the union on the product of these two probabilities, a probability of support not conditioned on continued employment.

Thought of this way, the firm might tilt nonunion wage supplements toward high wage workers if those workers have a higher unconditional probability of supporting the union. Such a higher probability could result if high wage workers knew their productivity advantage made them more likely to be retained in the face of a common union wage. Thus, the relative size of the optimal nonunion wage supplements cannot be predicted when there are productivity differences which influence the probability of worker retention after unionization. This point is made explicitly in a theoretical model that modifies that of Kahn and Curme and it leaves the issue of the relative sizes of supplements to be settled by empirical estimation, the task turned to now.(4)

The logic just discussed would seem to suggest a straightforward strategy: divide a sample of nonunion workers into low- and high-wage groups and compare the union effect. This rather obvious strategy has not been applied. Instead, the focus has been on the influence of unions on measures of nonunion wage dispersion. Kahn and Curme computed the nonunion log earnings variance for each occupation/industry cell and used this as the dependent variable in a regression which included percent of the industry unionized as an independent variable.

This approach, while closely related to the issue of which workers receive larger supplements suffers from several problems. First, a strong union presence may reduce wage variance within industries without a general pattern of wages being tilted. For example, a slight increase in the earnings of all high wage workers and a large increase in the earnings of the single lowest paid worker might generate a reduced variance and so a negative correlation. The point is not that such an outcome is necessarily likely, but rather that a variance measure is very sensitive to outliers because large deviations are squared.

Second, any examination of wage dispersion is subject to difficulties of endogeneity. The dependent variable of dispersion has typically been explained by controls which include the wage level as well as the percent of the industry organized. Yet, dispersion might well explain the percent unionized rather than the other way around. Both Hirsch |10~ and Farber and Saks |4~ suggest that workers demand unions in response to earnings inequality. Alternatively, as suggested by Hirsch |10~ more nearly equal earnings may be the result of worker homogeneity which could also be positively associated with the demand for unionization. Further, the wage level, an explanatory variable, is obviously influenced itself by the presence of unionization. The point remains that such simultaneity makes the influence of percent organized on dispersion difficult to distill.(5)

Table I. Selected Descriptive Statistics

 Upper = 1 Upper = 0

LnWage 2.195 1.514
 (0.426) (0.396)

Education 13.40 12.15
 (2.658) (2.469)

Experience 17.79 15.48
 (12.71) (14.37)

Craft .1159 .0695
 (0.320) (0.254)

Service .0926 .1761
 (0.290) (0.381)

Clerical .1470 .2529
 (0.354) (0.427)

Sales .1503 .1419
 (0.357) (0.349)

Professional .1551 .0537
 (0.362) (0.225)

Manager .1805 .0525
 (0.385) (0.223)

Technical .0480 .0293
 (0.213) (0.168)

Union Coverage 15.23 15.22
 (14.17) (14.16)

Note: The numbers presented are the means and those in parentheses are the
standard deviations for nonunion workers from the 1983 CPS.

Third, generating measures of dispersion by industry or occupation changes the fundamental weighting of the data. Aggregation can bring with it difficulties of interpretation that would seem unwarranted given that the fundamental hypothesis is not about dispersion but the size of supplements.

In light of these difficulties we execute the more immediate test of examining supplements for high and low wage workers. A sample of nonunion, private employees is taken from the May 1983 Current Population Survey. This sample is then bifurcated using the median wage of each industry. Each worker is assigned to either the upper or lower half of his or her nonunion industry wage distribution. Table I presents selected descriptive statistics for the two halves of the distribution. Nearly every variable traditionally associated with higher wages takes on a larger value in the upper half. Of particular importance for the ultimate conclusion of this paper is the high concentration of managers and professionals in the upper half.

Several variants on traditional log earnings equations are estimated and following convention, the coefficient on the percent of the industry unionized indicates the size of the threat supplement. The percent unionized comes from a several year moving average as computed by Curme, Hirsch and Macpherson |3~. At issue is whether the coefficient for the percentage measure takes different sizes for workers in the separate halves of the wage distribution. The first series of tests includes the full sample and uses as regressors human capital and personal characteristics known to determine wages. A second set of tests, reported in the subsequent section, excludes those workers least likely to have unionization as a realistic alternative.

The research reported here and in related work, is subject to several important caveats. First, actual wage supplements are not observed but inferred from the correlation between the percent of an industry unionized and the pattern of wages. To the extent that this correlation reflects factors other than supplements, the results from this paper are subject to alternative interpretations. Second, at no point is the behavior of the union ever detailed. The objectives of the union and resulting pattern of compensation could influence both the pattern of supplements and support for unionization. Instead, it is simply assumed that unions follow the general pattern of wage compression. Despite these concerns, the testing of previous theory is improved by recognizing that the association between low earnings and support for unions is ambiguous and by providing a more direct examination of whether low wage nonunion workers receive larger/supplements than their high wage counterparts.

III. The First Series of Tests

The results of the first series of tests indicate that workers in the lower half of their industry's distribution receive larger supplements. Initially it was assumed that workers in each half of the distribution have identical returns to each characteristic. The only differences are returns to the dummy "Upper" and to a variable which interacts the percent union with this dummy. The controls include a constant, education, potential experience, potential experience squared, regional dummies, occupational dummies, marital status, residency in an SMSA, years of tenure, gender, race, plant size and the percent of each worker's industry which is unionized. The coefficients on these controls exhibit very typical signs and sizes. As column one of Table II shows, the dummy "Upper" has an enormous positive sign and is highly significant as one would expect given that it is a specific bifurcation of the underlying dependent variable. Of more immediate interest are the coefficients on the coverage variables. In the lower part of the distribution the coefficient is that on the coverage variable alone, .0061, while for the upper part of the distribution it is the sum of the two coefficients, .0047. Thus, a ten point increase in union coverage yields a predicted supplement of 4.81 percent for high wage workers and a supplement of 6.29 percent for low wage workers.(6)

The first modification of the testing equation separates the estimation regimes for those above and those below their industry medians. In order to retain standard errors for testing, a full interaction model was chosen in which each of the independent variables is interacted with the "upper" variable. The highlights of the estimations are shown in the second column of Table II. Many of the new interactions are significant and an F-test rejects the restricted form in which coefficients are assumed to be identical in the two halves of the distribution. Despite the expanded estimation, those nonunion members in the upper half still receive significantly smaller induced wage supplements.

A second modification is motivated by concern that industry effects not captured in the existing variables might generate biased measures of statistical significance. A random industry effect model in which the error is thought of as the sum of an individual and an industry component is reported in column 3 of Table II. Mundlak's |18~ GLS estimation provides consistent estimation of the complete interaction model correcting for the possibility of biased statistical tests. The Lagrange Multiplier test rejects the hypothesis of no random industry effects and suggests that the results in column 3 are superior to those in 2. Despite the existence of industry effects, the basic results remain. The standard errors seem somewhat larger but the point estimates for the unionization coefficients change only modestly.

Table II. Examining the Size of Wage Supplements

 1 2 3 4 5 6

Coverage .0061 .0061 .0076 .0054
 (17.44) (17.45) (13.87) (8.558)

Upper .5143 -.0716 .0741 -.0719
 (49.61) (1.372) (1.455) (1.381)

Upper x -.0014 -.0017 -.0013 -.0016
Coverage (2.969) (3.306) (2.518) (3.277)

Lambda .0832
 (1.992)

Coverage .0069
(.25) (14.42)

Coverage .0093
(.50) (12.31)

Coverage -.0029
x Upper(.25) (4.368)

Coverage -.0066
x Upper(.50) (6.397)

R-squared .6198 .6354 .5966 .6521 .7207

Chi-squared 9733.2

n 9645 9645 9645 9645 5198 2329

Notes: t-statistics are placed in parentheses and all variables mentioned in
the text were included in the estimation but have been suppressed to save
space. The full results are available from the author.

We next recognize that union status is not randomly allocated. Workers who ultimately support unionization and are, indeed, hired by a union employer may differ systematically from others. Moreover, the excluded (and potentially very hard to measure) characteristics which result in union status may be correlated with the independent variables in the wage regression. To correct for the resulting bias the standard selection criteria variable, the inverse Mills ratio, is generated from a union probit equation on the full underlying sample, union and nonunion (n = 12050). The estimation used the full set of variables without the interactions and without the inclusion of experience squared but with a slightly finer occupational breakdown.(7) The full interaction specification of the nonunion wage regression is then estimated including the selection variable, lambda. This estimation is summarized in column 4 and despite the evidence of selection, the results are unchanged. The significant difference in the coverage effect remains.

Dividing the range into those above and below the median puts a large number of workers near the median into the separate halves based on differences of only a few pennies in earnings. A more dramatic test might be provided by excluding those workers who have very similar wages and are located near the median. In the first of such tests, the full interaction formulation is estimated on a reduced sample which excluded any worker within twenty-five cents of their industry's median. In this way, those with earnings at least twenty-five cents above their industry's median are compared to those with earnings at least twenty-five cents below their industry's median. A second test replicates this methodology but compares those with earnings more than fifty cents above their industry's median to those with earnings more than fifty cents below their industry's median.

The results presented in Columns 5 and 6 highlight that excluding those near the median results in somewhat larger coefficients for the coverage variables. While retaining only a fraction of the original sample, the fundamental results emerge in even sharper relief.(8) Those in the lowest portion of the distribution receive supplements larger than any indicated so far while the negative interaction for those in the upper portion is also larger than any of the others presented. This pattern is replicated in analogous GLS estimations available from the author. In short, it appears to this point that union induced supplements are far more strongly directed toward lower paid nonunion workers.

IV. The Second Series of Tests: The Sample Composition

It might be argued that using a representative sample of nonunion workers is inappropriate because many of the workers are professionals and managers who are unlikely to be unionized regardless of their industry and hence should receive no supplements. In particular, if these workers are disproportionately in the upper half of their industry's distribution, the results so far might be spurious. Instead of reflecting the strategic use of wage supplements in response to a threat of unionization, the results may simply reflect that those workers with the highest earnings are unlikely to organize given current industrial relations law and practice.

In an attempt to examine this possibility, a series of subsamples were examined. The first subsample excluded managers and professionals and the second subsample examined only blue collar workers. Within each sample the median nonunion wage was again computed and workers were again assigned to either the upper or lower half of the distribution. In the smaller subsamples the relevant occupational dummies and interactions were obviously excluded but otherwise the estimations were identical.

The entire tenor of the results switches dramatically, as column 1 of Table III highlights. Excluding managers and professionals causes the coefficient on the coverage interaction to lose half its size and statistical significance. No evidence remains that workers in the upper half receive smaller supplements. This result persists in the GLS estimation and the sample selection estimate. It also persists if the original median wage is used to divide the sample. The coverage variable itself, however, retains both size and significance.

As shown in column 2, the coefficient on the interaction actually takes a positive sign but continues to lack statistical significance when examining the highly unionized blue collar workforce. Among this group, where the threat of unionization could be expected to be the greatest, there is no evidence that wage supplements are tilted toward the lower half of the nonunion distribution.

Table III. Results for Subsamples

 1983 1983 1988 1988 1988
 W/O Managers Blue Full W/O Managers Blue
 or Prof. Collar Sample or Prof. Collar

Coverage .0077 .0065 .0070 .0068 .0063
 (11.97) (7.155) (19.52) (18.87) (12.65)

Coverage -.0006 .0007 -.0007 -.0001 -.0001
x Upper (1.136) (1.029) (5.311) (0.036) (0.335)

r-squared .6000 .5866 .6469 .5935 .6107

N 7519 2884 112745 82078 30510

Notes: t-statistics are placed in parentheses and all variables mentioned in
the text were included in the estimation but have been suppressed to save
space. The full results are available from the author.

It might be argued that these results are to be expected because the sample size shrinks as increasingly narrow subsamples are examined. This contention is without merit. The other variables, including coverage retain significance despite the smaller sample size and an examination of the outgoing rotation 1988 CPS reveals a similar pattern. Using this alternative data required sacrificing some variables, specifically tenure and plant size, but provides a nonunion sample of over one hundred thousand observations. Thus, even when examining subsamples, small sample size cannot explain the results.

Columns 3-5 replicate the original interaction specification on the full sample and each of the two subsamples just examined using the alternative data set. Other than those variables not available in the new survey, the estimation is identical. New median wage figures were computed in order to assign workers and new union density figures were taken from Curme, Hirsch and Kahn |3~. The full sample equation shows a large and highly significant coefficient on coverage and a negative and highly significant coefficient on the interaction variable. This larger sample mirrors the results shown earlier. When managers and professionals are excluded the interaction coefficient loses significance despite the very large number of observations. This lack of significance persists in the blue collar sample as well.

Several notes are in order. First, removing the managers alone is sufficient for the pattern of tilting to lose statistical significance although the change in the size of the coefficients is larger removing professionals as well. Second, once the two occupations are removed, attempts to recover evidence of tilting were unsuccessful. In addition to those already mentioned, a squared unionization measure was entered to try to capture nonlinearities and narrower ranges of union coverage were examined.

It is of interest that Kahn and Curme |4~ examined a fully representative nonunion sample that retained the range of white collar occupations. Their findings together with the earlier full sample results indicate that the finding that supplements are tilted toward lower paid workers depends crucially on the composition of the sample. In those samples which focus more closely on workers for whom unionization is a more genuine possibility, there is no such evidence. Instead, the evidence in favor of such tilting would seem to be somewhat of a statistical artifact. Those workers least likely to be unionized for institutional reasons (managers and professionals) are disproportionately in the upper half of the distribution and receive little or no supplements. Thus, one should be more circumspect in claiming a general conclusion that tilting of wage supplements occurs.

It may be the case that managers do deserve to be in the sample and are subject to threat effects. The argument could be that occupational differentials within establishments are reasonably stable so when wages of nonmanagerial workers rise, so do those of managers. If so, managers could receive higher wages when the nonmanagerial workers are unionized or when they receive nonunion supplements. Whether or not this is true, remains an empirical question. To the extent it is true, the full sample results may be more meaningful than previously indicated. Yet, even then tilting would have more to do with the nature of internal labor markets than with threat induced supplements.

V. Conclusions

This work began by questioning the claim that lower paid nonunion workers would receive larger union driven wage supplements. Once it is recognized that those workers with lower nonunion pay are less productive and less likely to be retained by a newly unionized employer, no clear theoretical prediction emerges. The empirical section of this paper devised a more direct test of the relative size of nonunion supplements. The initial results indicated that workers in the lower half of the distribution seemed to receive larger supplements. This finding was robust in the face of changes in specification but, was highly sensitive to the sample examined.

When managers and professionals were eliminated from the sample, or when blue collar workers were examined, the initial results could not be recovered under any specification. When the sample eliminated those workers least likely for institutional reasons to be unionized, there was no evidence that those in the lower half of the distribution received any greater union induced supplements. The supplements continued, however, to be large and statistically significant in all subsamples and across all specifications.

The absence of any tilting in the supplements fits with the ambiguous theoretical prediction. To the extent that wages contain information about productivity not captured elsewhere in the explanatory variables, they provide information on the probability of retaining a job following unionization. As a consequence, it is unclear whether low wage workers should receive a disproportionate share of nonunion wage supplements. The evidence is that they do not.

1. For a description and analysis of union avoidance strategies see Lawler and West |15~.

2. Earlier studies that used industry rather than individual data tended to be somewhat more ambiguous. See Kahn |13~ and Holzer |12~.

3. Lewis |16~ is quick to point out that correlations between wages and percent organized could be spurious and be driven by industry specific effects which are, in turn, correlated with union organization. Neumark and Wachter |19~ argue previous tests are incomplete and that a full test of die threat effect must examine the influence of increasing union wages on both nonunion wages and nonunion employment.

4. The theoretical model mentioned is in Heywood |7~ and will be provided to the interested reader.

5. It remains likely that wage levels and percent organized are similarly simultaneous in the estimations in this paper. Yet, examining dispersion does not eliminate this problem as wage levels have been included as a control by both Kahn and Curme |14~ and Belman and Heywood |2~. Moreover, this issue will be partially addressed in a sample selection estimation to be presented.

6. These percentage increases, g, are generated from the net coefficients, b : g = |e.sup.(b x 10)~ - 1.

7. There is no particularly strong justification for this set of exclusions but the resulting wage regression did not prove particularly sensitive to modest variations.

8. It is worth emphasizing that the median measure is within each worker's industry. If the measure were inappropriately taken to be the overall median, fewer workers would be excluded by the fifty cent band.

References

1. Abowd, John and Henry S. Farber, "Job Queues and the Union Status of Workers." Industrial and Labor Relations Review, April 1982, 354-67.

2. Belman, Dale and John S. Heywood, "Union Membership, Union Organization and the Dispersion of Wages." Review of Economics and Statistics, February 1990, 148-53.

3. Curme, Michael, Barn Hirsch, and David Macpherson, "Union Membership and Contract Coverage in the United States, 1983-1988." Industrial and Labor Relations Review, October 1990, 5-33.

4. Farber, Henry S. and Daniel H. Saks, "Why Workers Want Unions: The Role of Relative Wages and Job Characteristics." Journal of Political Economy, April 1950, 349-69.

5. Freeman, Richard B., "Unionism and the Dispersion of Wages." Industrial and Labor Relations Review, October 1980, 3-23.

6. -----, "Union Wage Practices and Wage Dispersion Within Establishments." Industrial and Labor Relations Review, October 1982, 3-21.

7. Heywood, John S. "Unionization and the Pattern of Nonunion Wage Supplements." Working Paper, Department of Economics, University of Wisconsin-Milwaukee, 1992.

8. -----, "Who Queues for a Union Job." Industrial Relations, Winter 1990, 119-27.

9. Hills, Stephen M., "The Attitudes of Union and Nonunion Male Workers Toward Union Representation." Industrial and Labor Relations Review, January 1985, 179-94.

10. Hirsch, Batty, "The Interindustry Structure of Unionism, Earnings and Earnings Dispersion." Review of Economics and Statistics, October 1982, 22-39.

11. ----- and Jon Neufeld, "Nominal and Real Union Wage Differentials and the Effects of Industry and SMSA Density." Journal of Human Resources, Winter 1987, 138-48.

12. Holzer, Harry J. "Unions and the Labor Market Status of White and Minority Youth." Industrial and Labor Relations Review, October 1982, 22-39.

13. Kahn, Lawrence M., "The Effects of Unions on the Earnings of Nonunion Workers," Industrial and Labor Relations Review, January 1978, 205-16.

14. ----- and Michael Curme, "Unions and Nonunion Wage Dispersion." Review of Economics and Statistics, November 1987, 600-607.

15. Lawler, John and Robin West, "Impact of Union-avoidance Strategy in Representation Elections." Industrial Relations, Fall 1985, 412-33.

16. Lewis, H. Gregg. Union Relative Wage Effects. Chicago: University of Chicago Press, 1986.

17. Moore, William, Robert Newman, and James Cunningham, "The Effect of the Extent of Unionism on Union and Nonunion Wages." Journal of Labor Research, Winter 1985, 21-44.

18. Mundlak, Yair, "On Pooling of Time Series and Cross Section Data." Econometrica, 1978, 69-85.

19. Neumark, David and Michael Wachter. "Union Threat Effects and Nonunion Industry Wage Differentials." National Bureau of Economic Research, Working Paper No. 4046, April 1992.

20. Podgursky, Michael, "Unions, Establishment Size, and Intra-Industry Threat Effects." Industrial and Labor Relations Review, January 1986, 277-84.

21. Rosen, Sherwin, "Trade Union Power, Threat Effects and the Extent of Organization." Review of Economic Studies, April 1969, 185-96.