The Effectiveness of Affirmative Action Policies in South Africa

Abstract

Affirmative action policies in South Africa are best described as cumulative and becoming increasingly complex over time. The first national affirmative action policy, the Employment Equity Act, was implemented in 1999. This policy was followed by the Broad-Based Black Economic Empowerment Act in 2003, which had a broader focus and stronger compliance incentives. In 2007, the Codes of Good Practice were introduced to support the B-BBEE legislation by prescribing explicit and measurable benchmarks to South African firms and the establishment of institutional structures that monitored its implementation. Since 2009, several industry-specific Sectoral Charters and Codes have also been adopted. A comparison of racial and gender earnings gaps over this period indicates that it was only after the introduction of the Codes of Good Practice that racial and gender earnings gaps started narrowing. This suggests that the details of AA legislation matter: policies that lack clear and quantifiable goals that are poorly monitored, or that offer weak incentives to comply, are unlikely to be effective. Strengthened monitoring and implementation has, in this case, been achieved by clearer General Sector Codes that support the legal framework. This has decentralized implementation to specific economic sectors, moving legal requirements higher up the priorities of business decision-making. Although such legislation was partially successful in reducing between-group earnings gaps, this only occurred at the top of earnings distribution and remained substantial.

Appendix

Data and Methodology

The empirical analysis uses pooled cross-sectional data to investigate the evolution of earnings gaps in the South African labor market. South Africa’s national statistics agency, Statistics South Africa, has collected nationally representative household surveys with an extensive labor module since 1994. These surveys were conducted annually as the October Household Survey (OHS) between 1994 and 1999, semiannually as the Labour Force Survey (LFS) between 2000 and 2008, and quarterly as the Quarterly Labour Force Survey (QLFS) since 2008. We use the Postapartheid Labour Market Series (PALMS) version 3.3 (Kerr et al. 2019) for our analysis.

There are several challenges to comparing the different cross-sectional surveys to obtain time trends. The early surveys experienced several changes in sampling frames and questionnaire design. By the 1997 OHS, many of these issues had been resolved, so we use this as the first survey for our analysis. Although the QLFS contained questions about labor earnings since its inception in 2008, no earnings data were officially released for the surveys between 2008 and 2010. Furthermore, whereas prior to 2008 the unimputed earnings data was made available, since 2010 earnings data included imputations undertaken by Statistics South Africa. As noted in Kerr and Wittenberg (2020: 2), these imputations are of “low quality,” especially after 2012Q2. We therefore ignore all the data for 2013 onward.

Differences in earnings across demographic groups can occur for several reasons, including differences in educational attainment and life-cycle participation patterns. Since this chapter is primarily interested in how affirmative action legislation has affected discrimination that occurs after workers leave school and enter the labor market, we need to control for such differences before estimating earnings gaps. In our empirical analysis, we achieve this in two ways. First, we limit the sample to those with overlapping support for all demographic groups. Effectively this means dropping everyone with fewer than 10 years of schooling from the sample, since this outcome was very rare for the white population group, and increasingly also for young workers from the other population groups.

We also reweight our sample to produce demographic groups with comparable schooling and age distributions across all years. We achieve this by following a method similar to that originally proposed by DiNardo et al. (1996). A logit regression is used to regress each of the eight binary demographic group indicators on a series of education dummy variables and age splines, both interacting with survey dummies. The inverse of the predicted value from this regression is then used to adjust the survey weights, and this adjusted weight is used when calculating the percentiles of the wage distribution in different years for different groups. This approach effectively reweights each of the groups in each of the years so that the education and age distributions more closely resemble that of the entire population for the full period. This implies that our estimated earnings gaps will not be driven by any between-group differences in schooling or age. Furthermore, our estimated time trends will be unaffected by time trends in the schooling levels or age distribution of workers. Confidence intervals that accommodate reweighting are bootstrapped with replicate weights using the procedures outlined by Kolenikov (2010).