Labour unions and wage inequality among African men in South Africa

Abstract

One Achilles' heel of post-Apartheid South Africa is the growing intra-racial income inequality, particularly among Africans. This paper examines the role of labour unions in explaining this phenomenon among African men given that labour markets are at the core of income inequality in South Africa. Using cross-sectional data drawn from Labour Force Surveys for 2001–10, we find a monotonically declining union wage premium. Further, our results indicate that unions have both compressionary and disequalising effects on wages. The disequalising effect dominates the compressionary effect, suggesting that unions have a net effect of increasing wage inequality among African men in South Africa. This finding implies that there is scope for unions to reduce inequality through initiatives that promote wage compression.

Keywords:

• income inequality
• wage distribution
• labour unions
• inequality decomposition
• South Africa

JEL codes:

• D31
• J31
• J51

1. Introduction

That an unequal dispersion of labour market earnings lies at the heart of an extraordinarily large and increasing problem of income inequality in post-Apartheid South Africa cannot be overemphasised. More recently, this has manifested through growing intra-racial income inequality, where the increase is relatively higher for Africans than other races. Hence conclusions abound that increasing intra-African earnings inequality dictates the evolution of aggregate income inequality in the country (Leibbrandt et al., 2012).

Following South African history and international literature, some studies attribute the increase in earnings inequality to the distribution of and returns to education among Africans (Mwabu & Schultz, 1996, 2000; Branson et al., 2012). While human capital partly explains the de facto wage distribution, the segmentation of the labour market into union and non-union sectors cannot be ignored as unions play an important part in wage determination. Following the post-Apartheid Labour Relations Act 1995, which promotes collective bargaining, unionisation increased significantly and fortified unions. Their strength also stems from an inconspicuous line between their efforts and those of political parties in resisting the Apartheid regime (Schultz & Mwabu, 1998; Butcher & Rouse, 2001). As such, labour unions in South Africa are labelled as being far too powerful for the country's level of income (Banerjee et al., 2008).

The deracialisation and strengthening of labour unions in post-Apartheid South Africa motivated research on unions' inter-racial and intra-racial effects on wages (see, for instance, Moll, 1993; Azam & Rospabé, 2007). A common finding from the literature is that union members receive high wage premiums. Based on the presence of a union premium, there are perceptions that unions are instrumental to wage inequality in South Africa, yet there is little empirical evidence on this notion (Hofmeyr & Lucas, 2001). International literature shows that the social effect of unions is two-pronged; that is, unions have a compressionary (within-group) effect on their members' wages and a disequalising (between-group) effect on wages of union and non-union members via the wage premium (cf. Card et al., 2004). Thus, focusing on the premium leaves a dearth of knowledge on the net effect of unions on wages.

In light of this, the contributions to this paper are two-fold. First, since unions disequalise wages through the premium, we examine the evolution of the wage premium to assess whether the distortionary effect was growing over time. In addition, this will update available studies that used data for 1985–2004 (i.e. Hofmeyr & Lucas, 2001; Banerjee et al., 2008). Results from this will also enable us to verify whether labour unions have successfully preserved the benefits of their members over the 2001–10 decade. This is topical given contemporary worker protests that are linked to loss in confidence towards unions inter alia. Second, the paper will assess the net effect of unions on wage inequality among African men over the same period. If the compressionary (disequalising) effect dominates the disequalising (compressionary) effect, this implies that unions in South Africa reduce (increase) wage inequality. The findings could signal whether labour unions are among entities driving income inequality in South Africa. The study will use individual data drawn from South African Labour Force Surveys.

The rest of the paper is organised as follows. In Section 2 we present a brief background on inequality and labour unions in South Africa. We review relevant literature on inequality and labour unions in Section 3. The methodology and data are discussed in Section 4. Section 5 discusses the results and Section 6 concludes the paper.

2. Background on inequality and labour unions among Africans

Income inequality in South Africa has its roots in the system of political Apartheid, which was institutionalised in 1948. Apartheid was geared to advance the economic welfare of white people at the expense of Africans, coloureds and Indians (van der Berg & Bhorat, 1999; Leibbrandt et al., 2001). This saw the creation of separate economies for the different racial groups, with whites being more advanced and serviced by other racial groups. In the process, whites had relatively higher access to quality education, superior jobs, wages and working conditions, and income-generating opportunities than non-whites – in all these Indians fared better than coloureds, who in turn fared better than Africans (van der Berg & Bhorat, 1999; Leibbrandt et al., 2001). Furthermore, poor whites had the privilege of state grants while these were rationed for non-whites. This undoubtedly skewed the distribution of wages, income and access to opportunities in the economy.

More importantly, the disadvantaged position of African workers was institutionalised in labour relations, the Bantu Labour Act 1953 and the Industrial Conciliation Act 1956 prohibited Africans to form or join registered labour unions (van der Berg & Bhorat, 1999:7; Bendix, 2001:67). This strategically left white unions with power to set wages for all workers in the various sectors of the economy. This arrangement enabled white workers to ensure that their wages remained higher than others. African workers resisted this oppression, which culminated in legislation of their labour unions in 1980 leading to the expansion of union membership (Bendix, 2001:77). Most of the new unions represented the interests of unskilled and semi-skilled workers (mainly non-whites) who were militant in the face of entrenched unfair labour practices (Bendix, 2001:78). Despite having a voice in the industrial relations systems, African workers were still politically oppressed. The oppression led them into forming alliances with political parties in fighting for democracy (Finnemore & van der Merwe, 1996:33; Bendix, 2001:81).

Since the attainment of democracy in 1994, the post-Apartheid government has made major changes to the South African constitution and effected new policies aimed at reducing income inequality. As mentioned earlier, the wage determination process in the new dispensation is underpinned by the Labour Relations Act 1995, which gives workers freedom to join labour unions and provides a platform for collective bargaining. The collective bargaining process occurs at two echelons; centralised and plant levels. Centralised bargaining occurs when one or more registered labour unions bargain with one or more registered employer organisations for wages and working conditions in a particular industry or sector (Bendix, 2001:271; Butcher & Rouse, 2001; Bhorat et al., 2012). This happens when labour unions and employer organisations are representative of the majority of their members, and they have been granted ministerial approval (Bendix, 2001:271; Butcher & Rouse, 2001). The bargaining council agreements can be extended to non-union workers in covered sectors. Non-members can, however, seek ministerial exemption from the conditions (Bhorat et al., 2012).

The second level of collective bargaining occurs at individual plants: unions for workers at a plant can bargain with their employers for plant-specific wage adjustments and conditions of employment. This level of bargaining can supplement agreements from bargaining councils – union workers covered by bargaining councils receive the same union premiums as those outside the bargaining council (Butcher & Rouse, 2001). While wages in the union sector are set as discussed above, those in the competitive sector are set by employers presumably on the basis of their profit-maximising decisions. In conclusion, the wage-setting process for the union sector has been well documented by earlier studies and hence here we only presented a summary. See, for instance, Moll (1993), Butcher & Rouse (2001) and Bhorat et al. (2012) for more details.

3. Literature review

This section presents a brief review of literature on income inequality in South Africa, as well as extant literature on the impact of labour unions on wages.

There are numerous studies in South Africa that document the extent and evolution of income inequality at household level and in the labour market. According to Hoogeveen & Ozler (2005), the overall Gini coefficient of household per-capita income increased from 0.57 in 1995 to 0.58 in 2000. Concomitantly, the contribution of between population group inequality to overall inequality decreased from 38% to 33%. This implies that an increase in intra-racial inequality accounted for the overall increase in inequality between 1995 and 2000. The increase was mostly attributed to a rise in intra-African inequality as the associated Gini coefficient increased from 0.47 to 0.50 over the same period. This result is also supported by Leibbrandt et al. (2012), who further distinguished the labour market as the main source of aggregate inequality in South Africa. Labour market income contributed between 85% and 88% to the overall Gini coefficient. This finding concurs well with Leite et al. (2006) and Tregenna &Tsela (2012). These authors also found that there has been a slight decrease in earnings inequality from 2000 to 2007, when all races are considered. It is noteworthy that although the literature has similar qualitative conclusions on inequality trends, it lacks consensus on the magnitude of inequality due to use of different datasets, different ways of constructing incomes from interval data, and different samples.

Given that labour markets are at the core of earnings inequality in South Africa, concerted efforts have been invested towards understanding the role of labour market institutions, particularly unions, in shaping income inequality. Interest in unions was partly roused by the strengthening of African labour unions post their legislation in 1980, a need to verify whether labour unions also compress the wage distribution as posited by international literature, and a need to understand whether unions lead to wage inflexibility in an era of high unemployment (Moll, 1993; Schultz & Mwabu, 1998; Butcher & Rouse, 2001; Hofmeyr & Lucas, 2001; Card et al., 2004; Azam & Rospabé, 2007; Banerjee et al., 2008).

A general result from these studies is that union members receive a wage premium relative to non-union members. Although these studies provide irrefutable evidence for the existence of a union wage premium, they do not present a consistent measure of the premium. Some report values as low as 6% (Bhorat et al., 2012), others around 24% (Moll, 1993; Barnajee et al., 2008) and some in excess of 100% (Michaud & Vencatachellum, 2001; Azam & Rospabé, 2007). This large variation in the premium is largely driven by use of different estimation samples, methodologies, and datasets.

An additional finding pertains to the evolution of the union wage premium. For instance, Hofmeyr & Lucas (2001) explored the trend controlling for selection into employment and unions. The results showed that the premium increased from 1985 to 1993, although the magnitudes were not robust across models. In a related study, Barnajee et al. (2008) found a premium that increased from 14% in 1995 to 27% in 2000 and then declined to 23% in 2004. The decrease in recent years raises curiosity – was it fortuitous or did it mark the beginning of a declining trend in the premium. This raises a question of whether the impact of unions on wage inequality has weakened concomitantly.

Other studies provide evidence suggesting that African labour unions compress the wage distribution of their members (Schultz & Mwabu, 1998; Butcher & Rouse, 2001). Using quantile regression, Schultz & Mwabu (1998) found a union wage premium of 145% at the bottom of the wage distribution and 19% at the top. This huge variation suggests that unions reduce wage inequality among their members.

In synopsis, this discussion shows that unions have two relative wage effects. First they reduce wage dispersion in the union sector. Second, they disequalise wages between union and non-union sectors. While this information is important, it does not allow us to deduce the overall wage effect of unions, yet this is critical for understanding whether unions increase or reduce wage inequality in South Africa.

4. Methodology and data analysis

4.1 Methodology

The standard treatment effects model is widely used to estimate the union/non-union wage differential. In this model, a union membership dummy is included in the wage function to capture the union premium or penalty. One drawback of this approach is that it assumes that the coefficients of personal characteristics such as education and age are the same across sectors (Lee, 1978; Heitmueller, 2006; Azam & Rospabé, 2007). This assumption implies that the wage-setting mechanisms are the same in union and non-union sectors, which might not be the case. In light of this inherent limitation of the standard treatment effect model, we estimate the union wage premium using an endogenous switching regression model applied by Lee (1978). This model allows us to estimate separate wage functions for union and non-union members, thereby yielding richer results with regards to structural differences in wage determination between the two sectors.

In the endogenous switching regression model, a switching equation sorts individuals over two states (Lokshin & Sajaia, 2004) – in this case, union and non-union sectors. The state in which an individual is selected into determines the wage regime faced by that individual. Formally, the model is specified as follows:

where I_i is a latent variable that determines the sector in which individual i is working – if , i is selected into regime U (union sector), otherwise regime N (non-union sector) is the outcome; w_ji is the wage of individual i in state j = U, N;X_i is a vector of individual and job characteristics that influence wages (i.e. age, education, occupation, economic sector and location); Z_i is a vector of observed characteristics that influence the sector into which individual i is selected – in this case it consists of age, education, economic sector, location and a dummy variable indicating whether an individual lives with other union members (an inverse mills ratio to control for selection into labour-force participation and subsequent employment is also included following the correction procedure suggested by Bhorat et al., 2001); δ, γ, β are vectors of parameters to be estimated; and u_ji, ϵ_i are disturbance terms.

Equation (3) is the switching function that determines the sector in which an individual is employed. Equations (1) and (2) are the wage functions for a specific sector. It is assumed that the error terms have a trivariate normal distribution, with mean vector zero and the following covariance matrix:

Following standard practice, the variance of the error term of the reduced form switching equation is set to one in order to identify the parameters δ and γ (Harthog & Oosterbeek, 1991). Given the assumptions placed on the distribution of the error terms, the logarithmic function for the wage equations and switching function is (cf. Lee, 1978; Lokshin & Sajaia, 2004):

where ; F is the cumulative normal distribution and f is a normal density distribution function; is an optimal weight for observation i and is the correlation between μ_ji and μ_i. Since both w_U and w_N cannot be observed simultaneously for individual i, does not appear in the likelihood function as this parameter is unidentified.

Conditional on union status, the union wage equation is given by:

Conditional on non-union status, the union wage equation is given by:

The estimated wage equations allow comparison of wage differences between union and non-union sectors. To measure the average percentage increment of the wage rate for union membership, we predict wages using the estimates:

Following Lee (1978), the union-membership premium is computed as follows:

where n is the sample size.

Using the estimates from the wage functions, we then examine the impact of unions on earnings inequality (see Appendix A for formulas). First, inequality measures are computed using predicted earnings from Equations (5) and (6). Second, while maintaining the predicted earnings for non-union members using Equation (6), we replace predicted earnings of union-members by the following counterfactual earnings distribution:

Equation (10) provides the earnings of union members under a non-union regime. We then re-compute inequality measures and assess the extent to which they differ. This exercise will enable us to assess the extent to which the premium contributes to inequality.

The second approach used in assessing the impact of unions on earnings inequality is adopted from Card et al. (2004). Using a two-sector model, let be the log wage of individual i when employed in the non-union sector and be the log wage of the same individual i when employed in the union sector. Assume that:

where , are random error terms with conditional means of zero. Assuming that, in the absence of unionisation, current union workers will receive the same average wage as non-union members, then the union/non-union wage gap is given by:

Let and denote the variance of potential wage outcomes of workers in the union and non-union sectors respectively; then the variance gap is given by:

Let α be the unionisation rate, then overall variance of log wages is given by (cf. Freeman, 1980; Card et al., 2004):

The effect of unions on the variance of wages relative to what would prevail if all workers were paid according to the non-union wage regime is:

Equation (16) shows that the presence of unions exerts two effects on the dispersion of wages relative to the counterfactual V^N. First, the within-sector effect that is associated with the fact that wage dispersion is different in union and non-union sectors is captured by . Second, the between-sector effect reflecting the wage disequalising effect caused by unions is captured by the term . Although this analysis does not incorporate differences in the rate of unionisation across different types of skill groups, this simple framework provides basic insights into the impact of union on earnings inequality.

4.2 Data

The data used for the study were obtained from Statistics South Africa's September Labour Force Surveys for 2001, 2004, 2007 and the Third Quarter of the 2010 Quarterly Labour Force Survey. The surveys are nationally representative, containing socio-economic information for individuals living in 30 000 households across the country. The Labour Force Surveys were reweighted to establish comparability with the Quarterly Labour Force Survey on the basis of the 2001 population census. We restrict our sample to African men aged 15 to 64 years provided they had information on our key variables defined in Table A1 in Appendix B. The full samples for 2001, 2004, 2007 and 2010 had 50 485, 51 384, 50 352 and 82 998 individuals, respectively. Table 1 presents the economic positions of these individuals; that is, strict and broad labour-force participation rates, employment and union status. The broad labour force includes employed people, active job seekers and discouraged workers, while the strict labour force excludes discouraged workers. In the initial sample, the broad labour-force participation rates range from 59 to 68%, while the strict labour-force participation rates are relatively lower ranging from 51 to 54%. Among the participants, approximately 55% were employed in 2001 and 2004; this increased to 60% in 2007 and 2010.

Table 1: Economic status by period

	2001	2004	2007	2010
Panel A: labour-force participation status
Broad labour-force participation	0.667 (0.471)	0.672 (0.470)	0.676 (0.468)	0.587 (0.492)
Strict labour-force participation	0.539 (0.498)	0.511 (0.500)	0.537 (0.499)	0.510 (0.500)
Number of observations	50 485	51 384	50 352	82 998
Panel B: employment status
Employed	0.559 (0.497)	0.558 (0.497)	0.604 (0.489)	0.608 (0.488)
Strict unemployment	0.250 (0.433)	0.203 (0.402)	0.191 (0.393)	0.260 (0.439)
Broad unemployment	0.441 (0.497)	0.442 (0.497)	0.396 (0.489)	0.392 (0.488)
Number of observations	32 890	32 373	31 643	45 832
Panel C: union status
Union member	0.537 (0.499)	0.417 (0.493)	0.406 (0.491)	0.412 (0.492)
Number of observations	4 511	4 332	5 340	8 822

Note: Standard deviations in parentheses.

The samples in panels A and B of Table 1 are used for estimating sample selection correction terms (for labour-force participation and subsequent employment) to be included in our union membership functions. To derive the final sample used in wage functions, we restrict the sample of employed workers to full-time employees in the formal non-agricultural sector. The excluded workers fall outside the domain of South African labour unions. This restriction leaves a final sample composed of 4511, 4332, 5340 and 8822 African men in 2001, 2004, 2007 and 2010, respectively. The corresponding union membership rates for these sub-samples are 54%, 42%, 41% and 41%. Our statistics suggest that there was some de-unionisation among African men between 2001 and 2004, but this stabilised in later years.

Based on the final sample depicted in panel C of Table 1, Table 2 presents some basic descriptive statistics of individual characteristics by union status and period. We find that log hourly wages increased over the period, and they are on average higher for union than non-union workers. The higher union wages unveiled here are consistent with previous studies in South Africa and other countries. As indicated in Table 2, there are notable differences across union and non-union workers – union members have higher marriage rates and are generally older than non-union workers. In addition, the proportion of individuals with education above matric is higher among union-members relative to non-union members. For instance, in 2001 the proportion of workers with education above matric is 11% in non-union sectors compared with 19% in union sectors. This pattern is consistent across all years and is more evident in 2007, where the proportion of workers with education above matric is 10% in the non-union sector compared with 25% in the union sector. In contrast, workers with education levels below matric are somewhat concentrated in the non-union sector.

Table 2: Descriptive statistics for union and non-union sub-samples by period

	2001		2004		2007		2010
	Non-union	Union	Non-union	Union	Non-union	Union	Non-union	Union
Log hourly wage	2.259 (0.655)	2.510 (0.610)	2.445 (0.706)	2.932 (0.670)	2.751 (0.784)	3.229 (0.762)	3.114 (0.720)	3.723 (0.801)
Demographics
Age	36.002 (9.935)	39.190 (9.014)	35.190 (9.667)	39.761 (9.799)	34.322 (10.081)	39.280 (9.999)	34.468 (9.622)	39.641 (9.954)
Married	0.595 (0.491)	0.741 (0.438)	0.593 (0.491)	0.729 (0.445)	0.493 (0.500)	0.681 (0.466)	0.472 (0.499)	0.635 (0.482)
Education
No schooling	0.052 (0.222)	0.052 (0.221)	0.038 (0.191)	0.050 (0.218)	0.031 (0.173)	0.030 (0.170)	0.022 (0.146)	0.018 (0.131)
Primary	0.210 (0.407)	0.268 (0.443)	0.193 (0.395)	0.215 (0.411)	0.174 (0.379)	0.164 (0.370)	0.135 (0.342)	0.113 (0.317)
Incompletesecondary	0.354 (0.478)	0.279 (0.449)	0.365 (0.482)	0.299 (0.458)	0.404 (0.491)	0.320 (0.466)	0.415 (0.493)	0.266 (0.442)
Matric	0.271 (0.445)	0.217 (0.412)	0.295 (0.456)	0.251 (0.433)	0.273 (0.446)	0.283 (0.450)	0.329 (0.470)	0.350 (0.477)
Certificate	0.017 (0.130)	0.020 (0.140)	0.018 (0.133)	0.016 (0.127)	0.015 (0.121)	0.022 (0.147)	0.011 (0.103)	0.022 (0.146)
Diploma	0.052 (0.223)	0.117 (0.322)	0.056 (0.230)	0.111 (0.314)	0.067 (0.249)	0.130 (0.336)	0.068 (0.252)	0.166 (0.373)
Degree	0.044 (0.204)	0.047 (0.212)	0.034 (0.182)	0.057 (0.233)	0.036 (0.187)	0.052 (0.222)	0.020 (0.141)	0.066 (0.248)
Employment sector / industry
Public sector	0.161 (0.367)	0.392 (0.488)	0.119 (0.323)	0.360 (0.480)	0.090 (0.286)	0.335 (0.472)	0.056 (0.229)	0.386 (0.487)
Mining andquarrying	0.044 (0.205)	0.234 (0.424)	0.029 (0.168)	0.185 (0.388)	0.027 (0.161)	0.148 (0.355)	0.025 (0.156)	0.120 (0.325)
Manufacturing	0.247 (0.431)	0.197 (0.398)	0.202 (0.402)	0.206 (0.405)	0.203 (0.402)	0.202 (0.402)	0.191 (0.393)	0.205 (0.404)
Electricity, gasand watersupply	0.018 (0.133)	0.018 (0.132)	0.009 (0.093)	0.020 (0.139)	0.007 (0.084)	0.013 (0.112)	0.005 (0.072)	0.020 (0.140)
Construction	0.108 (0.310)	0.038 (0.192)	0.146 (0.353)	0.035 (0.184)	0.179 (0.383)	0.035 (0.183)	0.176 (0.381)	0.046 (0.209)
Wholesale andretail trade	0.216 (0.412)	0.091 (0.288)	0.230 (0.421)	0.092 (0.289)	0.250 (0.433)	0.133 (0.340)	0.266 (0.442)	0.078 (0.268)
Transport,storage andcommunication	0.099 (0.299)	0.069 (0.254)	0.077 (0.267)	0.069 (0.254)	0.085 (0.279)	0.062 (0.241)	0.070 (0.256)	0.080 (0.271)
Financialservices	0.131 (0.337)	0.060 (0.238)	0.176 (0.381)	0.085 (0.279)	0.148 (0.355)	0.104 (0.305)	0.180 (0.384)	0.105 (0.307)
Community, socialand personal services	0.137 (0.344)	0.292 (0.455)	0.130 (0.337)	0.308 (0.462)	0.102 (0.302)	0.304 (0.460)	0.086 (0.281)	0.347 (0.476)
Occupation andlocation dummies	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Number ofobservations	1 819	2 692	2 266	2 066	2 887	2 453	4 856	3 966

Note: Standard deviations in parentheses.

We also find that most individuals are employed in the private sector, 60 to 67% of union workers and 80 to 94% of non-union workers, compared with the public sector. A relatively large proportion of non-union workers were employed in trade industries and financial services, while those in union jobs were concentrated in mining, manufacturing and services industries. For the manufacturing sector, the workers seem to be equally concentrated across union and non-union sectors. Most workers are employed as sales workers, artisans, operators and in elementary occupations regardless of union status. Also about 80% of non-union workers were in urban areas, whereas the same applies to 68 to 78% of union workers. In general, the majority of workers were employed in Gauteng and KwaZulu Natal provinces.

5. Results

Given that our model estimates a choice model for union membership and wage functions, our discussion of results is structured as follows: first, we discuss the heckprobit models of employment and labour-force participation, and then the correlates of union membership. This is followed by a discussion of wage functions. Thereafter we discuss the link between unions and wage inequality and robustness checks.

5.1 Heckprobit models of employment and labour-force participation

The findings for the Heckprobit models of labour-force participation and employment are presented in panels A and B of Table 3 respectively. They show that there is an inverted U-shaped age–labour-force participation profile for African men, and similarly for employment. The probabilities of participating and employment are also shown to be higher for married men relative to their non-married counterparts. This suggests that family responsibilities could be a financial drive for African men to not only participate but to be actually employed. In line with human capital theory, the prospects of participating, and finding a job are positively correlated to education relative to no schooling (i.e. having matric education and above for employment). The results for the exclusion restrictions for the labour-force participation model (i.e. living with a pension-eligible woman aged above 59 years or man aged above 64 years in a household) – proxies for non-labour income – show that having an elderly woman in a household has non-labour income effects on participation, while the presence of an elderly man is statistically insignificant. The finding for African women is in line with Bertrand et al. (2003).

Table 3: Labour-force participation and employment probit models (coefficients shown)

	2001	2004	2007	2010
Panel A: labour-force participation
Age	0.349***	0.338***	0.716***	0.347***
Age-squared	–0.004***	–0.004***	–0.004***	–0.004***
Married	0.496***	0.385***	0.412***	0.526***
Primary	0.299***	0.286***	0.282***	0.191***
Incomplete secondary	0.124***	0.119**	0.117***	0.281***
Matric	0.720***	0.906***	0.831***	0.767***
Certificate	0.367***	0.390***	0.301***	0.652***
Diploma	0.609***	1.139***	0.888***	1.114***
Degree	0.658***	0.829***	0.962***	1.039***
Presence of elderly men in household (age > 64)	–0.033	–0.073	–0.048	–0.080**
Presence of elderly women in household (age > 59)	–0.254***	–0.250***	–0.273***	–0.241***
Constant	–5.391***	–5.249***	–5.341***	–5.812***
Panel B: employment
Age	0.152***	0.156***	0.164***	0.155***
Age-squared	–0.002***	–0.002***	–0.002***	–0.002***
Married	0.911***	0.937***	0.936***	0.864***
Primary	–0.002	0.038	0.095**	–0.187**
Incomplete secondary	0.032	–0.075	0.054	–0.226***
Matric	0.113*	0.026	0.152**	–0.055
Certificate	0.258*	0.287**	0.293**	0.181**
Diploma	0.357***	0.488***	0.555***	0.205*
Degree	0.518***	0.797***	0.590***	0.538***
Proportion of other adults in a householdwho are employed	5.165***	5.366***	5.199***	4.932***
Constant	–4.826***	–4.993***	–4.877***	–4.726***
athrho	0.472***	0.468***	0.451***	0.367***

Notes: Location dummies are also included in the labour-force participation and employment equations. Reference group: no schooling. Significance level: ***1%, **5%, *10%.

Panel B also shows that the prospects that an individual will be employed are positively correlated with living in a household comprised of a large proportion of other adults who are employed. In line with South African literature on social networks and job status, this result suggests that other employed household members act as a proxy for the respondent's social networks on employment opportunities (e.g. Schoer & Leibbrandt, 2006). Nonetheless, this finding should be cautiously interpreted as it is difficult for us to establish the timelines of the employment events for the respondent and other household members given that we are using cross-sectional data. It is also noteworthy that this variable could be overestimated, because Burns et al. (2010) maintain that it does not fully capture all network-related effects.

The findings for athrho are positive and statistically significant, suggesting that the error terms for the labour-force participation and employment equations are positively correlated; hence it is suitable to estimate the employment equation controlling for sample selection bias. Furthermore, we maintain that the exclusion restrictions used herein are plausible given the national situation of high unemployment and poverty – that is, they do not have significant influence on the final outcome of the study (i.e. wages). As such, some South Africans are willing to work for market wages that are below their reservation wages – especially in the formal sector, which pays higher than the informal sector; involuntary unemployment (Kingdon & Knight, 2004). This also supports the findings by Walker (2003), who concludes that a person's position in the labour market is not as a result of his/her reservation wage. Hence the standard theoretical prediction that we only observe market wages for individuals whose market wages exceed their reservation wages does not generally apply in South Africa. This suggests that there is no strong (if any) correlation between non-labour income and market wages in South Africa.

5.2 Union membership models

Table 4 shows marginal effects for union membership probit models for 2001, 2004, 2007 and 2010 in reduced and structural forms. The results for reduced form models show an inverted U-shaped age–union membership profile. This suggests that unions attract middle-aged workers relative to younger and older counterparts. This result also applies to job tenure. Residing with other union members is also positively correlated with the prospects of joining a union. According to Moll (1993), this could be reflecting household tastes for unionism or firms' recruitment strategies. Apart from the 2004 sample, the findings for education show that having some form of education increases the probability of joining a union relative to those without schooling. The prospects were, in many cases, highest for those with certificates and diplomas, and the effects increased over time. These findings suggest that workers of all education levels could be seeking union protection against possible employer exploitation given high unemployment rates in the country.

Table 4: Marginal effects for union membership probit models, by period

	2001		2004		2007		2010
	Reduced form	Structural form	Reduced form	Structural form	Reduced form	Structural form	Reduced form	Structural form
Age	0.030***	0.030***	0.017***	0.019***	0.030***	0.026***	0.031***	0.033***
Age-squared	0.000***	0.000***	0.000***	0.000***	0.000***	0.000***	0.000***	0.000***
Married	0.015***	0.024***	–0.038***	–0.035***	0.053***	0.069***	–0.024***	–0.024***
Primary	0.055***	0.073***	–0.051***	–0.037***	0.102***	0.076***	0.088***	0.077***
Incomplete secondary	0.037***	0.057***	–0.004**	0.012***	0.174***	0.142***	0.132***	0.126***
Matric	0.044***	0.054***	0.049***	0.059***	0.221***	0.148***	0.223***	0.221***
Certificate	0.161***	0.191***	–0.039***	0.022***	0.261***	0.184***	0.320***	0.311***
Diploma	0.153***	0.182***	0.098***	0.115***	0.239***	0.220***	0.300***	0.288***
Degree	0.043***	0.087***	0.051***	0.123***	0.129***	0.216***	0.213***	0.229***
Mining and quarrying	0.266***	0.266***	0.248***	0.236***	0.385***	0.376***	0.364***	0.370***
Electricity, gas and water supply	–0.139***	–0.151***	–0.082***	–0.070***	0.023***	0.084***	0.119***	0.150***
Construction	–0.243***	–0.261***	–0.217***	–0.192***	–0.244***	–0.219***	–0.215***	–0.209***
Wholesale and retail trade	–0.119***	–0.126***	–0.147***	–0.109***	–0.092***	–0.026***	–0.189***	–0.182***
Transport, storage and comm.	–0.085***	–0.093***	–0.076***	–0.088***	–0.063***	–0.068***	–0.018***	–0.021***
Financial services	–0.081***	–0.087***	–0.118***	–0.083***	–0.011***	0.015***	–0.088***	–0.071***
Community, social and pers.services	–0.043***	–0.054***	–0.070***	–0.076***	0.042***	0.056***	–0.049***	–0.058***
Large firms (50/more employees)	0.144***	0.043***	0.157***	0.042***	0.147***	0.037***	0.209***	0.039***
Job tenure	0.041***	–0.001***	0.043***	–0.001***	0.041***	–0.001***	0.038***	–0.001***
Job tenure squared	–0.001***	0.248***	–0.001***	0.317***	–0.001***	0.336***	–0.001***	0.180***
Selection term	–0.019***	–0.019***	–0.012***	–0.011***	–0.014***	–0.013***	–0.034***	–0.034***
Other union members	0.248***	0.148***	0.314***	0.150***	0.334***	0.150***	0.179***	0.208***
Wage differential	0.117***	0.153***	0.234***	0.072***

Notes: Other variables included in all the models are a constant term, urban, provincial and public-sector dummies. Reference groups: no schooling, manufacturing and Gauteng province. Results for age-squared are negative but appear as positive due to rounding off. Significance level: ***1%, **5%, *10%.

Working in large firms increases the chances of joining a union relative to small firms because it is easier for unions to organise workers in large firms. The findings for 2004–10 also show that being in an urban area increases opportunities to join a union relative to a rural area; the converse applies to 2001. Furthermore, working in the mining sector avails higher opportunities to join a union relative to the base category (manufacturing); the reverse is the case for other industries, except electricity where the effect is not robust. This is expected since mining and manufacturing industries are traditionally union hubs. Also, being in the public sector increases chances of unionism relative to the private sector. Provincial variations in the probability of unionism were also detected. This could be due to regional idiosyncratic factors that affect attitudes towards unionism (Moll, 1993). Sample selection correction terms are statistically significant, implying that unobservable characteristics which influence the employment decision also affect the probability of joining a union once employed. These were computed from the findings of heckprobit models of employment as union and non-union wages are only observed for employed people who are a non-random sample of the labour force. We therefore controlled for selection into employment following Bhorat & Leibbrandt (2001). Findings for the structural models are qualitatively similar to those of the reduced form models. They also show that the union wage differential is positively associated with the chances of joining a union.

5.3 Wage models

Some of the findings for age and tenure exhibit an inverted U-shaped age/tenure–earnings profile in both the union and non-union sectors as indicated in Table 5. This suggests that, regardless of union status, middle-aged workers receive more wage benefits than other workers. The results for education demonstrate that earnings increase with education levels in both sectors. While this outcome accords well with the theory of human capital, it could also be reflecting the skills-biasedness of the South African labour market. Overall, union wages are higher than non-union wages in urban than rural areas and this urban wage effect has increased over time. The sample selection terms were also negative and statistically significant, which implies that the samples of union and non-union workers selected justifies our correction for sector choice.

Table 5: Wage equations by union status

	2001		2004		2007		2010
	Union	Non-union	Union	Non-union	Union	Non-union	Union	Non-union
Age	0.015*	0.019*	0.015*	0.025***	0.010	–0.008	–0.007	0.008
Age-squared	0.000	0.000	0.000	0.000**	0.000	0.000	0.000	0.000
Married	0.023	0.096***	0.039	0.056**	0.012	0.084***	–0.003	0.002
Primary	–0.023	0.136**	0.010	0.100*	0.110	0.003	0.185**	–0.073
Incomplete secondary	0.116***	0.290***	0.088*	0.195***	0.223***	0.088	0.271***	0.067
Matric	0.338***	0.420***	0.332***	0.381***	0.535***	0.215***	0.412***	0.226***
Certificate	0.376***	0.666***	0.476***	0.789***	0.721***	0.361***	0.386***	0.049
Diploma	0.493***	0.743***	0.749***	0.696***	0.890***	0.695***	0.889***	0.427***
Degree	0.807***	0.973***	0.854***	1.194***	1.214***	1.347***	0.996***	1.042***
Urban	0.093***	–0.007	0.112***	0.065*	0.142***	0.141***	0.180***	0.192***
Mining and quarrying	0.266***	0.266***	0.089**	0.006	0.068	0.024	0.117**	0.027
Electricity, gas and water supply	–0.139***	–0.151***	–0.062	0.008	0.006	0.281**	–0.273***	0.122
Construction	–0.243***	–0.261***	–0.160**	0.060	–0.052	0.077**	–0.129*	–0.031
Wholesale and retail trade	–0.119***	–0.126***	–0.319***	0.012	–0.311***	–0.004	–0.260***	–0.063
Transport, storage and comm.	–0.085***	–0.093***	0.044	–0.062	–0.087	–0.115**	–0.027	–0.124***
Financial services	–0.081***	–0.087***	–0.282***	0.058	–0.215***	–0.062	–0.413***	–0.044
Community, social and pers. services	–0.043***	–0.054***	0.002	–0.006	0.014	0.056	0.024	–0.099*
Large firms (50/more employees)	0.012	0.056*	0.077***	0.034	0.073***	0.086***	0.010	–0.029
Job tenure	0.022***	0.014**	0.012**	0.002	0.023***	0.007	0.020***	0.015**
Job tenure squared	0.000*	0.000*	0.000	0.000	0.000**	0.000	0.000***	–0.001***
Selection term	–0.398***	–0.189**	–0.128**	–0.387***	–0.094*	–0.330***	–0.103	–0.218**
Chow test (F-statistic)	(3.660)		(7.240)		(6.170)		(6.820)
P-value	(0.000)		(0.000)		(0.000)		(0.000)

Notes: Other variables included in all the models are a constant term, urban, provincial, public sector and occupational dummies. Reference groups: no schooling, manufacturing, Gauteng province and machine and plant operators. Significance level: ***1%, **5%, *10%.

Using Equation (9) we computed union wage premia for 2001–10 and the findings are presented in Table 6. The union premium exhibits a declining trend over the period 2001–10. In 2001 union workers earned 57% more than non-union workers. The premium decreased to 54%, 48% and 42% in 2004, 2007 and 2010 correspondingly. Differences in these premia are statistically significant at the 5% level. This shows that the wage gain of union workers decreased by 15% points over the period 2001–10. This trend harmonises with that for Banerjee et al. (2008) for the period 2000–04. It suggests a weakening in the ability of unions to preserve the benefits of their members.

Table 6: Union wage premium

Year	Union premium	Difference	95% confidence interval
2001	0.567 (0.003)	—	[0.572; 0.562]
2004	0.541 (0.004)	0.025	[0.549; 0.534]
2007	0.476 (0.004)	0.066	[0.483; 0.469]
2010	0.424 (0.002)	0.052	[0.428; 0.420]

Notes: Standard errors in parentheses.

5.4 Unions and wage inequality

We further examine the impact of unions on wage inequality using two procedures. The first procedure uses predicted wages for union and non-union workers, computed using Equations (5) and (6), to compute measures of inequality. These results are presented in Table 7, panel A labelled ‘with premium’. We then re-compute inequality measures using a counterfactual distribution for union members (computed using Equation (10)) – wages that they would receive under the non-union wage regime, maintaining predicted wages for non-union members. The inequality measures derived from this distribution are labelled ‘without premium’ in Table 7, panel A. The second procedure is based on Card et al.'s (2004) model discussed earlier and the variance decomposition (based on Equation (16)) results are presented in Table 7, panel B.

Table 7: Impact of unions on wage inequality

	2001	2004	2007	2010
Panel A: Inequality measures
With premium
Gini	0.297	0.349	0.367	0.370
Theil index	0.147	0.210	0.243	0.249
Atkinson (e = 0.5)	0.069	0.096	0.108	0.110
Atkinson(e = 1)	0.129	0.175	0.191	0.196
Atkinson(e = 2)	0.225	0.289	0.304	0.311
Without premium
Gini	0.267	0.286	0.323	0.320
Theil index	0.137	0.163	0.237	0.215
Atkinson (e = 0.5)	0.061	0.071	0.098	0.091
Atkinson(e = 1)	0.110	0.126	0.163	0.156
Atkinson(e = 2)	0.179	0.200	0.240	0.240
Panel B: Effect of unions on wage structure
Mean predicted log wages
Union workers	2.725	2.833	3.099	3.639
Non-union workers	2.159	2.299	2.622	3.208
Union gap	0.566	0.534	0.477	0.431
Variance decomposition
Overall variance	0.296	0.379	0.410	0.516
Variance in non-union sector	0.226	0.306	0.355	0.477
Within-sector effect	–0.014	–0.013	–0.017	–0.019
Between-sector effect	0.084	0.086	0.072	0.058
Total effect	0.070	0.073	0.056	0.039

In accordance with the evolution of South African income inequality, the Gini coefficient with union wage effects increased over the period 2001–10. In 2001, the Gini coefficient was 0.3 and increased by 5 points in 2004. Although small, a further increase in inequality is registered in 2007 and 2010. When considering the distribution ‘without premium’, where all workers are paid under the non-union wage regime, we find that the Gini coefficient slightly decreased for all years. In 2001 the Gini coefficient decreases from 0.3 (with premium) to 0.27 (without premium), while in 2004 it decreases from 0.35 to 0.29. This change in the Gini coefficient suggests that unions have a modest effect on wage inequality. This finding is consistent across the years, and is supported by the Theil index and Atkinson's measures of inequality.

It is noteworthy that the magnitude of the Gini coefficient reported here is much smaller than in previous studies, which report Gini coefficients in excess of 0.5 (see for instance Tregenna & Tsela, 2012). This difference stems largely from the sample used in this study, which is restricted to African men employed in formal-sector non-agriculture occupations, while other studies included all workers irrespective of gender and race. This restriction is, however, relaxed in the robustness check, which is based on an all-inclusive sample.

The above findings on inequality trends are corroborated by the overall variance of earnings. We find that the variance increased from 0.3 in 2001 to 0.52 in 2010. With regards to the decomposition of variance, the negative sign of the within-sector effect indicates that unions have a compressionary effect on wages within the union sector. This result is consistent throughout the 2001–10 period. The between-sector effect is positive, suggesting that unions also have a disequalising effect on wages between union and non-union sectors. This disequalising outcome is consistent with the common finding of a union wage premium. Overall, unions in South Africa seem to increase wage dispersion since the between-sector effect outweighs the within-sector effect over the period. Relative to the overall variance the compressionary effect ranges from 4 to 5%, while the between-sector effect ranges from 12 to 29%. From this result it is clear that unions have a weak compressionary effect relative to the disequalising effect.

5.5. Robustness checks

To ascertain that the results presented in this paper are robust, this section presents a series of robustness checks. The first check is concerned with the sensitivity of the results to the selected sample, while the second set of checks examines the sensitivity of the results to compositional changes of union and non-union sectors overtime.

5.5.1 Selected sample

The results discussed earlier are based on a sample that excludes individuals working in the informal sector. To check whether the results are not sensitive to this, we re-estimate the model on a ‘broader’ sample that includes both informal-sector and formal-sector workers. The results from this exercise are presented in Table 8 along with the baseline results. The findings indicate that when using a ‘broader’ sample, we find a larger union wage premium relative to the baseline result (which excludes informal-sector workers). This result is expected given that informal-sector workers often engage in precarious activities that offer relatively low wages (see Perry et al., 2007). Consistent with the baseline results, we also find a declining union wage premium from 86% in 2001 to 50% in 2010. With regards to variance decomposition, the results suggest that the net effect of unions is to increase wage dispersion since the disequalising effect dominates the compressionary effect. The qualitative similarity of the results is reassuring, suggesting that our findings are not too sensitive to the selected sample.

Table 8: Robustness checks

	Baseline: excluding informal sector workers				Including informal sector workers				Workers with age > 25 years				Workers with education < matric
	2001	2004	2007	2010	2001	2004	2007	2010	2001	2004	2007	2010	2001	2004	2007	2010
Mean predicted log wages
Union workers	2.725	2.833	3.099	3.639	2.442	2.740	2.968	3.551	2.719	2.826	3.185	3.607	2.590	2.634	2.903	3.373
Non-union workers	2.159	2.299	2.622	3.208	1.580	2.050	2.434	3.055	2.127	2.247	2.640	3.108	2.009	2.079	2.419	2.971
Union gap	0.566	0.534	0.477	0.431	0.863	0.690	0.534	0.496	0.592	0.579	0.544	0.499	0.581	0.556	0.483	0.402
Variance decomposition
Overall variance	0.296	0.379	0.410	0.516	0.646	0.520	0.531	0.553	0.320	0.426	0.421	0.503	0.298	0.309	0.336	0.470
Variance in non-union sector	0.226	0.306	0.355	0.477	0.475	0.423	0.474	0.504	0.239	0.337	0.358	0.454	0.219	0.241	0.283	0.436
Within-sector effect	–0.014	–0.013	–0.017	–0.019	–0.040	–0.025	–0.016	–0.020	–0.010	–0.016	–0.027	–0.027	–0.010	–0.015	–0.008	–0.018
Between-sector effect	0.084	0.086	0.072	0.058	0.211	0.123	0.077	0.069	0.091	0.105	0.090	0.076	0.089	0.083	0.061	0.053
Total effect	0.070	0.073	0.056	0.039	0.171	0.097	0.060	0.049	0.081	0.089	0.063	0.049	0.079	0.068	0.054	0.035

5.5.2 Sector composition

One potential pitfall of the inter-temporal analysis in this study is that it might be simply reflecting changes in the composition of union and non-union workers. Indeed, there are a number of trends during the 2001–10 period that are likely to affect the composition of workers in union and non-union sectors. For instance, the financial crisis might have affected the employment probabilities of low-skilled workers, which might influence the results. In addition, the implementation of the employment equity policy in this decade is likely to have influenced the wages of skilled African men as employers competed to employ skilled African men. Given that skilled workers are less likely to join unions, the returns in the non-union sector might affect the estimated average union wage premium. To address these potential pitfalls; first, we re-estimate the model on a sub-sample that excludes young workers (aged 15 to 25 years). The exclusion of young workers enables us to obtain results based on a sample purged of workers who were likely to become unemployed during the financial crisis. Second, we re-estimate the model on a sub-sample composed of low-skilled workers (i.e. individuals with an education level less than or equal to matric) to examine the sensitivity of the results to the implementation of the employment equity policy. The results presented in Table 8 indicate that when excluding workers aged 15 to 25 years and low-skilled workers, we find a union wage premium that is above the baseline result. The higher premium is in line with the notion that unions avail higher premiums at the bottom of the wage distribution (Schultz & Mwabu, 1998; Casale & Posel, 2012). We also find a declining union wage premium despite the possibility of sector compositional changes. This suggests that our results are fairly robust. Similar to the baseline results, we also find that the disequalising effect dominates the compression effect, indicating that the net effect of unions is to increase wage inequality for both sub-samples.

6. Discussion and conclusion

Understanding sources of intra-African wage inequality in South Africa is important for initiatives to reduce rising income inequality in the country. While there are various drivers of wage inequality such as human capital, historical factors and labour market institutions, this paper focuses on the latter by considering labour unions.

The paper has two interesting findings. First, we find that the union wage premium among full-time, wage-employed African men has been decreasing over the 2001–10 decade. This finding is contrary to the results for the 1990s, which show an upward trend for African men (Hofmeyr & Lucas, 2001). Although the latter studied a different sample (i.e. African, urban men) from the one herein, it still provides insights on the evolution of union wage premium. The initial increase in the wage premium as uncovered in Hofmeyr & Lucas (2001) was expected given that South Africa was emerging from a highly discriminatory system characterised by low wages for Africans. Consequently, Africans' unions made concerted efforts to secure sizeable premiums for their members. In recent years, the declining trend uncovered here suggests that unions have been weakening. Possible reasons for this could be the skills-biasedness (towards highly skilled workers) of the labour market, which is juxtaposed to high unemployment. In this context, it might be difficult for unions to bargain for higher premiums as most of their members are lowly educated and vulnerable to unemployment.

Second, we find that unions in South Africa have both compressionary and disequalising effects on wages among full-time, wage-employed African men. The compressionary effect is, however, dominated by the disequalising effect for this particular sample, suggesting that the net effect of unions is to increase wage inequality. On the basis of the above findings, it is clear that unions partly contribute to wage inequality among African men.

When placed in a global context, our findings differ from what has been found elsewhere. For instance, in the United Kingdom, the United States and Canada, Card et al. (2004) found that unions reduced overall wage inequality since the compressionary effect outweighed the disequalising effect. This could be due to the differences in human capital, historical factors, labour market institutions and sample of analysis. To some extent this implies that South African unions can contribute to the overarching goal of inequality reduction through more compression of the wage distribution.

We conclude with a discussion of the limitations of this study. First, Card et al.'s (2004) model used the assumption that non-union members do not benefit from provisions of bargaining council agreements. In South Africa, some non-union workers benefit from bargaining council wage agreements; therefore it is important to verify the extent of the overlap. Unfortunately, this is impossible to check using our data. Nonetheless, Bhorat et al. (2012) analysed the bargaining council wage premia and found that non-union workers covered by bargaining councils receive wage premiums of 9 to 10%. This premium is small in the South African context where premiums can be in excess of 100%, suggesting that the spillover effect can be deemed as trivial. Also, Butcher & Rouse (2001) reported that only 16% of non-union workers were covered by bargaining councils, which further indicates that the spillover effect is not substantial. The second limitation of this study is that it does not account for unobservable characteristics that might affect sector choice and wages. Future studies can address this using panel data. Third, although we have corrected for sample selection bias in our endogenous switching regression model, our results could potentially be biased. Casale & Posel (2012) point out that union wage equations are sensitive when controlling for selection through different techniques, and there is no consensus on a superior technique.

Appendix A. Inequality measures

This appendix presents details of the inequality measures used in Section 5 of the study.

Gini coefficient (G)

The Gini coefficient is the most commonly used measure of inequality. It can be computed as follows (cf. Gini, 1912):

where y_i is the income of individual i, is mean income and n is the sample size. When G = 1 it represents perfect inequality, while G = 0 represents zero inequality.

General entropy (GE) measure

The general entropy measure is another class of inequality measures defined as:

Measures of the GE are sensitive to changes at the lower end of the distribution for α close to zero and sensitive to changes across the distribution for α = 1. It is sensitive to changes at the upper end of the distribution for higher values of α. When α = 1, this yields the Theil (1967) index. The values of GE ranges from zero to ∞, with zero representing an equal distribution and higher values representing higher levels of inequality (Litchfield, 1999).

Atkinson measures of inequality (A_ϵ)

This measure has a weighting parameter ϵ that measures aversion to inequality. The higher the value of ϵ, the more society is concerned about inequality (Atkinson, 1970). The Atkinson class of measures can be computed as follows:

where denotes mean income. Similar to the Gini coefficient, the Atkinson class of measures range from zero to one, with zero representing no inequality and one representing complete inequality.

Appendix B. Definitions of variables used in the study

The log hourly earnings in Table A1 were constructed from earnings information that was reported in either points or intervals. Both categories were reflected in weekly, monthly or annual payment intervals. We obtained monthly earnings from point data through multiplying gross weekly wages by 4.3, dividing yearly wages by 12 and retaining the gross monthly earnings. To those who reported intervals, we assigned midpoints of the bands as converted to monthly terms. This gave us a series of monthly earnings for 2001, 2004, 2007 and 2010. These were deflated to 2001 values using deflators from Statistics South Africa. In turn, the real monthly earnings were converted to hourly wages as we divided them by hours usually worked per month (i.e. hours usually worked per week in main job multiplied by 4.3).

Table A1: Variable definitions

Variable	Description
Union (dependent variable)	Dummy variable: 1 if individual is a union member, 0 otherwise
Wage (dependent variable)	An individual's log hourly wage deflated to 2001 values
Age	An individual's age in completed years
Married	Dummy variable: 1 if individual is married civilly or traditionally, 0 otherwise
Education
No education^a	Dummy variable: 1 if individual has 0 years of schooling or Grade R, 0 otherwise
Primary	Dummy variable: 1 if individual's schooling is in the range Grade 1 to 7, 0 otherwise
Incomplete secondary	Dummy variable: 1 if individual's schooling is in the range Grade 8 to 11, 0 otherwise
Secondary	Dummy variable: 1 if individual's schooling is Grade 12, 0 otherwise
Certificate	Dummy variable: 1 if individual's schooling is vocational or technical training, 0 otherwise
Diploma	Dummy variable: 1 if individual's schooling is diploma, 0 otherwise
Degree	Dummy variable: 1 if individual's schooling is bachelor's degree and above, 0 otherwise
Provinces
Gauteng^a	Dummy variable: 1 if individual resides in Gauteng province, 0 otherwise
Eastern Cape	Dummy variable: 1 if individual resides in Eastern Cape province, 0 otherwise
Free State	Dummy variable: 1 if individual resides in Free State province, 0 otherwise
KwaZulu Natal	Dummy variable: 1 if individual resides in KwaZulu Natal province, 0 otherwise
Mpumalanga	Dummy variable: 1 if individual resides in Mpumalanga province, 0 otherwise
North West	Dummy variable: 1 if individual resides in North West province, 0 otherwise
Northern Cape	Dummy variable: 1 if individual resides in Northern Cape province, 0 otherwise
Northern Province	Dummy variable: 1 if individual resides in Northern province, 0 otherwise
Western Cape	Dummy variable: 1 if individual resides in Western Cape province, 0 otherwise
Urban	Dummy variable: 1 if individual resides in an urban area, 0 otherwise
Industries
Mining	Dummy variable: 1 if an individual works in mining sector, 0 otherwise
Manufacturing^a	Dummy variable: 1 if an individual works in manufacturing sector , 0 otherwise
Electricity	Dummy variable: 1 if an individual works in electricity sector, 0 otherwise
Construction	Dummy variable: 1 if an individual works in construction sector, 0 otherwise
Trade	Dummy variable: 1 if an individual works in trade sector, 0 otherwise
Transport	Dummy variable: 1 if an individual works in transport sector, 0 otherwise
Finance	Dummy variable: 1 if an individual works in finance sector, 0 otherwise
Services	Dummy variable: 1 if an individual works in services sector, 0 otherwise
Firm-size	Dummy variable for firms with more than 50 employees
Job tenure	number of years that an individual has spent with his/her main employer
Public sector	Dummy variable: 1 if an individual is employed in the public sector, 0 otherwise
Occupations
Manager	Dummy variable: 1 if an individual is a manager, 0 otherwise
Professional	Dummy variable: 1 if an individual is a professional, 0 otherwise
Technician	Dummy variable: 1 if an individual is a technician, 0 otherwise
Clerk	Dummy variable: 1 if an individual is a clerk, 0 otherwise
Service	Dummy variable: 1 if an individual is a services worker, 0 otherwise
Skilled agriculture/fishery	Dummy variable: 1 if an individual is a skilled worker, 0 otherwise
Artisan	Dummy variable: 1 if an individual is a craftsman, 0 otherwise
Operator^a	Dummy variable: 1 if an individual is an operator, 0 otherwise
Elementary	Dummy variable: 1 if an individual does elementary jobs, 0 otherwise
Labour-force participationbroad	Dummy variable: 1 if employed, active job seeker or discouraged worker
Employed	Dummy variable: 1 if employed full time in formal non-agriculture sector
Other household unionmembers^b	Dummy variable for co-residence with other union members
Old men^c	Dummy variable for co-residence with men aged 65 and above
Old women^c	Dummy variable for co-residence with women aged 60 and above
Proportion of other adultsemployed^d	Proportion of other working age people in a respondent's household who are employed

Notes: ^aBase category. ^bExclusion restriction in the union membership probits. ^cExclusion restrictions in the broad labour-force participation model. ^dExclusion restrictions in the employment model.