77469 Asymmetries in the Union Wage Premium in Ghana Niels-Hugo Blunch and Dorte Verner There is little evidence on the size of the union wage premium in developing economies. The article uses a matched employer-employee data set for Ghana and adopts a quantile regression approach that allows the effects of unionization to vary across the conditional wage distribution. It is shown that if there are intrafirm differences in unionization, there does appear to be a premium among poorer paid workers in the formal sector. Although this cannot be given a causal interpretation, it suggests import- ant issues about how unions may affect one part of the labor market. I did not begin my research on Ghanaian trade unions with the idea that organized labor could rescue this small, developing West African state from its economic difficulties. But neither did I have the pre-conceived notion, today popular in some quarters, that unions are irrelevant to the struggles of Third World peoples to maximize their own freedom. —Paul S. Gray (1981) Experience in developed economies suggests that unions may be a mechanism for providing a positive work environment by reducing labor turnover and wage nego- tiation costs and promoting worker training, increased benefits, and higher produc- tivity (Standing 1992). Unions have also been found to reduce wage inequality and wage discrimination (Chaykowski and Slotsve 2002; Freeman 1980; Panagides and Patrinos 1994; Standing 1992). Much less is known about unions in developing economies, particularly in Sub-Saharan African countries.1 Such analysis seems especially warranted for these countries as a basis for policy proposals. Formal sector jobs are scarce and wages are generally low, leading both to poverty and—because of low tax revenues—low levels of goods and services from the public sector. To shed some light on these issues in developing economies, this article analyzes wage determinants in Ghanaian manufacturing industries, focusing on Niels-Hugo Blunch is a consultant in the Social Protection Unit of the Human Development Network at the World Bank and a Ph.D. candidate at The George Washington University; his e-mail address is nblunch@worldbank.org. Dorte Verner is a senior country economist in the Social Development Family of the Latin America and Caribbean Region at the World Bank; her e-mail address is dverner@world- bank.org. The authors thank Franc ¸ ois Bourguignon, Sudharshan Canagarajah, Donald Parsons, and participants at a seminar at the Centre for Labour Market and Social Research, Aarhus, Denmark, for helpful comments and suggestions. Comments and suggestions from three anonymous referees and a journal review board member helped greatly improve this article. 1. See, however, Kristensen and Verner (1999), Rama (2000), and Schultz and Mwabu (1998). THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 2, Ó The International Bank for Reconstruction and Development / THE WORLD BANK 2004; all rights reserved. doi:10.1093/wber/lhh040 18:237–252 237 238 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 2 the association of union membership and wages and the possible asymmetries in this association across the conditional wage distribution. Ghana is an ideal candidate for such analysis because of its long history of active labor unions.2 I. HYPOTHESES A union premium could arise through three possible channels. First is the direct effect through individual union membership. This is the standard union premium, well known from the empirical literature on the union relative wage effect begin- ning with Lewis (1963).3 Second, because this does not take into account possible spillovers to nonunion workers, the analysis here allows unionism to potentially affect all workers through an industry-level union density variable. Third, there might be an additional union effect arising from training (Booth and Chatterji 1998; Booth and others 1999), because unions may promote training more than management does. For example, unions can have a longer time horizon than management, which may focus on maximizing profits and stock values in the short term, or unions may counterbalance the firm’s ex post monopsonistic power in wage determination. Thus where unions are active there may be increased recognition of the value of training, so that trained workers receive a wage premium.4 (Booth and Chatterji 1998 and Booth and others 1999 develop theoretical models in which a key prediction is that union workers receive more training and higher returns to training than do nonunion workers.) The hypothesis here is that union effects are more likely at the lower part of the wage distribution, because unions are generally seen as (and generally perceive themselves as) proponents of workers’ rights and wages for the poorer or less skilled segment of the workforce. This view is also in line with previous research, which generally finds that unions reduce wage inequality and wage discrimination. Chamberlain (1994) finds distinct asymmetries in the union premia in U.S. manu- facturing, which is 28 percent at the bottom decile but declines monotonically to 0.3 percent at the top decile. This contrasts with an ordinary least squares (OLS) estimate of the mean union premium of 15.8 percent, driven primarily by the bottom decile of the conditional distribution (for Ghana, the result is still more 2. Unionism is low or even absent in many countries of Sub-Saharan Africa because of the economic dominance of smallholder agriculture and the historically prohibition against independent trade unions in many countries. Ghana, however, has a long tradition of labor unions, originating with the many guilds and artisan associations in the early nineteenth century (Gray 1981). The potential synergies between labor organizations and government were anticipated early, even before Ghana’s independence from British rule in 1957: ‘‘As early as 1930, Lord Passfield (Sidney Webb) had noted in a dispatch that regu- lation of wage laborer organizations was of importance, and that colonial governments should act to facil- itate the passage of unions into constitutional channels’’ (Gray 1981, p. 14). 3. To avoid a causal interpretation, we use the term union wage premium rather than union relative wage effect, which has previously been widely used in the literature but unfortunately might suggest a causal relationship. 4. Alternatively, unions may enable trained workers to extract the rents created by workers. Blunch and Verner 239 striking, with the mean union premium disappearing altogether). More recently, Card (1996) finds similar results for the United States and Chaykowski and Slotsve (2002) for Canada. Thus, although there is some evidence of asymmetry in the union wage premium for developed countries, such asymmetry may be particularly relevant for developing economies, where the high costs of monitoring may prevent effective monitoring of adherence to minimum wage legislation. Unions in developed areas may bargain more on behalf of workers from the middle part of the wage distribution because minimum wage legislation is generally adhered to. Thus for developing and developed economies alike, the upper part of the wage distribution would not be expected to exhibit a positive and statistically significant association between wages and individual union membership. Analyzing the impact of unions across the entire wage distribution is thus important for at least two reasons (Chaykowski and Slotsve 2002). First, it may shed light on whether unions have an impact on particular socioeconomic groups in the labor market, enriching the social and economic view of unions, including whether unionism should be reinforced or retrenched. Second, such analyses provide insight into which workers benefit most from unions and, therefore, help identify which groups unions may most successfully appeal. Before getting on with the analysis, some words of caution about interpreting the results are in order. First, because matched employer-employee panel data sets for Sub-Saharan Africa are rare, the empirical analysis is limited to cross- section data—as is the case for most wage equation studies for Sub-Saharan Africa. That means that individual effects, which might otherwise overcome the possible selection of workers into unions (on unobservables such as type or ability; see Card 1996 for a study for the United States using panel data), cannot be isolated. Second, to account for selection into unions using cross-section data requires an exclusion restriction (or instrument). This is also not feasible for the present analysis because there are no readily available variables that affect selection into unions but do not affect wages directly. Controlling for firm size should somewhat mitigate the biases arising from the nonrandom place- ment of unions (at the aggregate level) into sectors or industries,5 but the possibility of selection of workers into unions at the individual level based on unobservables (ability, motivation) remains. The results should therefore be seen as suggestive rather than as explicitly causal.6 5. Previous studies of the association of union membership and wages have not always been able to control for firm size. Schultz and Mwabu (1998), for example, analyzing the association of individual union membership and wages and unemployment in South Africa using a quantile regression framework, note the importance of this variable with respect to the nonrandom placement of unions but are unable to include the variable in their analyses because it is not part of their data set. 6. It may be that unions are simply more active in higher-wage sectors. Again, because high-wage sectors are also typically found to be those with larger firms, controlling for firm size should at least mitigate some of this bias. 240 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 2 II. DATA AND DESCRIPTIVE ANALYSIS The data are from the Regional Program on Enterprise Development survey for Ghana, organized by the World Bank and funded by the British Overseas Development Administration and conducted by the Centre for the Study of African Economies at the University of Oxford and the University of Ghana at Legon in 1994.7 Although the survey yields information on firms and workers in 180 manufacturing firms, missing observations for one or more variables resulted in an effective estimation sample of 683 workers in 108 firms. The main variables applied in this study include a ‘‘core’’ of (log) monthly wages (including allowances); the standard human capital variables of age and age squared (to capture potential general experience), tenure in the firm and tenure squared (to capture potential specific experience), highest level of education completed, training variables (whether a worker has received training within the firm), and occupational control variables and firm-level control variables, most notably firm size by number of employees (the appendix lists the variables and their definitions).8 The possible existence of a union wage premium may be determined in various ways. The first thing to consider is the channel through which unions may affect wages: Is there an individual, direct effect through union membership, or is there an indirect effect through the degree of unionization at an establishment or in an industry as a reflection of the bargaining power of a union within a firm or industry, thus allowing for spillovers to nonunion workers? This is examined as an empirical question, through a dummy variable for individual union membership and an industry-level union density variable. The aggregate union wage premium (spillover mechanism) is modeled using the degree of unionization at the industry rather than at the firm level largely because collective bargaining in Ghana occurs mostly at the more aggregated industry level (Gray 1981). This approach further ensures that collinearity is not likely to be a serious problem, because these two union variables are only weakly correlated (a simple correlation of 0.08), whereas the individual union membership and firm union density variables have a simple correlation of 0.76, so including both of these in the same regression could result in substantial collinearity problems.9 7. Surveys were conducted in 1992, 1993, and 1994. Although the surveys have a panel structure for collecting firm-level data, the data on workers were collected as independent cross-sections. The analysis therefore includes only the most recent of the three surveys, treating the data as cross-sectional for both firms and workers. 8. Larger establishments tend to pay higher wages than smaller establishments, controlling for other factors (Schaffner 1998; Velenchik 1997). Not controlling for this, therefore, could lead to substantial omitted-variable bias. 9. Maloney and Ribeiro (1999) include the union density variable at the firm level, arguing that this is a proxy for the bargaining power of the union over the firm’s rents. However, the high correlation between this variable and the dummy variable for individual union membership in the Ghana sample means that inclusion of both would likely yield problems with multicollinearity, and so it does not appear valid to simultaneously allow the two different channels of a union relative effect (individual and aggre- gate spillover) proposed here. Blunch and Verner 241 There may still be an endogeneity issue related to the use of the variable of individual union membership, however. What is needed is a good instrument that may help explain individual union membership without also explaining wages. Because the data set lacks such a variable, the analysis follows Schultz and Mwabu (1998) in arguing that it is beyond the scope of the data to endogenize union membership, including explaining who gets a union job and who does not, and to explain the extent to which unions enhance the productivity of workers with the same observable characteristics. As a result, any estimated union wage premium may overstate or understate the true union wage premium. Because the estimated coefficient picks up the effect from the omitted variable as well, the omission of any variable influencing wage determination that is positively correlated with the union variable will cause the estimated coefficient on the union variable to be upwardly biased, whereas the omission of a nega- tively correlated variable will cause the estimated coefficient to be downwardly biased. For example, it may be conjectured that nonunion jobs are typically found in smaller establishments or industries because it takes a certain size for an establishment or industry to become interesting for a union in terms of potential members. Hence, failure to control for firm size may cause bias in the union premium. However, the present data set permits incorporating firm size as an explanatory variable, which likely decreases such bias considerably. Furthermore, it may be conjectured that firms that have become unionized respond to unionization by carefully vetting prospective new workers so as to employ even higher quality workers than before because of increased wage demands. Because ‘‘quality’’ or ‘‘ability’’ is unmeasurable, omitted variable bias is again possible in the union wage premium estimates. The same applies to the labor turnover argument: If unions reduce labor turnover, and this effect cannot be directly observed and included as an explanatory variable, the result is once again omitted variable bias in the union premium estimate. With a panel data set one solution is to extract the individual fixed effect of the data, thus mitigating the potential bias (see, for example, Card 1996). But with a cross-section data set, as in this case, unionism has to be modeled based on individual union membership (and possibly a variable of union density at the firm or industry level, constructed from this). Even though the data do not allow for fully endogenizing union membership, it is possible to shed some light on whether the endogeneity of individual union membership will pose severe problems for the subsequent analyses by doing some descriptive work. Data on individual and firm characteristics across union membership reveal that union workers earn higher wages than nonunion work- ers, not controlling for other characteristics (table 1). However, union workers (and their firms) are different from nonunion workers (and their firms) in several ways (although not always statistically significantly so). They are better educated, older, and more experienced (within the firm). They are also more likely to have a permanent contract. At 127 employees, the average firm size of unionized workers is much larger than that of nonunionized workers, at 242 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 2 T A B L E 1 . Worker and Firm Characteristics across Individual Union Membership Nonunion worker Union member Variable Mean SD Mean SD Difference t-Statistic Log(wages) 10.522 0.077 10.752 0.054 0.230*** 2.65 Age 34.1 0.870 36.1 0.949 2.0 1.58 Gender 0.204 0.032 0.153 0.031 À0.050 1.21 None 0.094 0.021 0.094 0.030 0.001 0.01 Primary 0.037 0.012 0.015 0.008 À0.023 1.53 Middle 0.509 0.032 0.446 0.045 À0.064 1.15 Secondary 0.123 0.024 0.124 0.030 0.001 0.03 Vocational 0.102 0.017 0.183 0.032 0.081** 2.29 Polytechnic 0.073 0.016 0.124 0.033 0.051 1.44 Professional 0.042 0.012 0.015 0.008 À0.027* 1.98 University 0.021 0.007 0.000 0.000 À0.021*** 2.99 Production 0.443 0.040 0.376 0.039 À0.067 1.22 Administration 0.104 0.024 0.193 0.030 0.089** 2.44 Commercial 0.062 0.014 0.079 0.020 0.017 0.70 Professional 0.046 0.012 0.025 0.015 À0.021 1.06 Support staff 0.098 0.030 0.178 0.036 0.081* 1.86 Manager 0.127 0.017 0.099 0.026 À0.028 0.93 Experience (years) 5.9 0.562 8.3 0.926 2.4** 2.27 Permanent contract 0.969 0.012 1.000 0.000 0.031** 2.50 Received training in firm 0.320 0.045 0.282 0.049 À0.038 0.59 Accra 0.611 0.056 0.629 0.088 0.017 0.19 Number of employees in firm 64.8 12.072 126.7 27.601 61.8** 2.21 *Significant at the 10 percent level. **Significant at the 5 percent level. ***Significant at the 1 percent level. Note: Number of observations in the full sample = 683. The clustering of workers within firms is taken into account in the calculation of SDs and t-statistics for H0: Difference = 0. Source: Regional Program on Enterprise Development for Ghana (Wave III 1994). 65 employees. Firms of unionized workers are also more likely to be located in Accra, the capital. That firm size is an important correlate of individual union membership is reassuring, because including this variable in subsequent analyses is likely to take care of many of the concerns with nonrandom placement of unions, at least at the aggregate level. However, selection is still possible and even probable at the individual level, where workers may select into unions based on preferences or latent personal characteristics or be chosen by employees based on latent personal characteristics, all of which are unobserved by the researcher. Again, although the data prevent a detailed analysis of individual determinants of union membership, the finding from the previous descriptive analysis that union workers (and their firms) are quite different from nonunion workers (and their firms) on observable characteristics suggests the likelihood of differences Blunch and Verner 243 on unobservable characteristics such as ability and motivation as well. Yet additional calculations reveal that union membership is fairly evenly distributed across occupations and industries. For example, a production worker is only slightly more likely to be member of a union than a manager (26.3 and 24.7 percent).10 Because 98 percent of the sample has a permanent contract (table A-1), including that as an explanatory variable may appear fruitless because of the low variation. However, there is the possibility of an indirect effect from training. Employers would be more inclined to invest in training for employ- ees who are likely to stay with the firm for some time, so the quantity or quality of training may be different for workers with a permanent contract. The variable for contract status is thus interacted with training. Because only about 30 percent of the sample both has a permanent contract and received training, this is likely to have some explanatory power because of its higher variation. In sum, because the data are unsuitable for making explicit causal inferences about the effect of unions on wages, the analysis explores only the association between union membership and wages, keeping in mind the differing character- istics of union workers (and firms) and nonunion workers (and firms) and the possible nonrandom placement of unions into high-wage sectors. III. METHODOLOGY The theoretical framework for the analysis is standard human capital theory. An individual builds up knowledge and skills through education and experience (Becker 1964; Mincer 1974). Formally, the economic model is derived from the theory of individual demand for schooling, which views education as an invest- ment in human capital (Becker 1964). In the traditional human capital litera- ture, wages are determined by education and other individual characteristics. Because the Ghanaian data set allows inclusion of union- and firm-level vari- ables, the standard Mincerian wage function is augmented with union- and firm-level characteristics ð1Þ Wi ¼ W ðIi ; Fi ; Ui Þ; where W, the wages of individual i, is the dependent variable; I is a vector of individual characteristics (such as age and age squared proxying general experi- ence; tenure in the firm, capturing firm-specific experience; the level of educa- tion; and gender); F is a vector of characteristics for the firm of individual i, including the size of the firm (measured by the number of employees) and geographical location; and U is a vector of variables capturing the possible 10. The industry-level union densities are 19.8 for wood, 25.7 for food, 32.1 for textiles, and 33.1 for metal. 244 THE WORLD BANK ECONOMIC REVIEW, VOL. 18, NO. 2 union wage premium for individual i, as measured by individual union member- ship or the union density of the industry or firm (as discussed in the previous section). Lewis (1963) defines the union–nonunion wage differential (referred to here as the union wage premium) as11 ð2Þ ri ¼ ðWiu À Win Þ=Win : The possible existence and magnitude of the wage differential depend on the extent to which the union affects the wages of union workers relative to the wages of nonunion workers. At one extreme the union may, through its bar- gaining power, merely extract and subsequently share the existing rents of the firm (in the form of profits) with its members. At the other extreme, the union may generate rents through its potential adverse effects on labor turnover and the costs of wage negotiations between management and workers. In reality, a combination of these two effects would be expected. The theory is silent on the precise empirical implementation of this notion of a union wage premium, which is left to individual researchers. This study uses quantile regression analysis. It enables simultaneously estimating the marginal effects for different quantiles (where the quantile of interest may be chosen arbitrarily) of the dependent variable, thus exploring the entire conditional distribution. The main advantage here is the semi-parametric nature of the approach, which relaxes the restrictions on the parameters being constant across the entire conditional distribution of the dependent variable. A priori, the union wage premium might be expected to differ across the conditional wage distribution. For example, unions may be bargaining mainly on behalf of workers at the lower end of the wage distribution. As noted, this is especially likely in devel- oping economies, where minimum wage legislation may not always be strictly adhered to because of the high costs of monitoring. Furthermore, it seems likely that the returns to education and tenure, for example, would differ across the wage distribution. For example, education might be thought to be a more important determinant at higher quantiles than at lower quantiles.12 The method has other virtues, as well. Allowing the parameter estimates for the marginal effects of the explanatory variables to differ across quantiles of the dependent variable achieves greater robustness to potential heteroscedasticity. This contrasts with OLS regression analysis, which requires homoscedasticity 11. The union–nonunion wage differential or union wage premium—which is the focus here—is only one aspect of the economics of unions. For an excellent recent survey of the economics of unions in a broader context, see Booth (1995). See also Lewis (1963, 1986), Freeman and Medoff (1984), and Pencavel (1995). 12. This turns out to be the case. An extended version of this article with the full set of results is available online at www.niels-hugo.dk. Blunch and Verner 245 (indeed, in most of the empirical literature, the presence of homoscedasticity is merely a maintained hypothesis). Additionally, when the error terms are non- normal, for instance, quantile regression estimators may be more efficient than least squares estimators. Furthermore, because the quantile regression objective function is a weighted sum of absolute deviations, it yields a robust measure of location. As a consequence, the estimated coefficient vector is not as sensitive to extreme observations of the dependent variable. The method, developed by Koenker and Basset (1978), can be formulated as ð3Þ Yi ¼ Xi by þ uyi ¼ Quanty ðYi jXi Þ ¼ Xi by where Quanty(YijXi) denotes the yth conditional quantile of Y given X for individual i. In general, the yth sample quantile (0 < y < 1) of Y solves 8 9 1< X X = ð4Þ min ¼ yjYi À X0i bj þ ð1 À yÞjYi À X0i bj b n :i:Y !X0 b i:Y