Sorting the Gender Earnings Gap: Heterogeneity in the South African Labor Market

Abstract

This study explores inequality among Africans in the South African labor market by investigating heterogeneity in the gender earnings gap. The article adds to the existing literature by applying an innovative sorting method, which provides a complete mapping of heterogeneity in the gender earnings gap, ordering the full distribution of the partial effects from largest to smallest with respect to the underlying characteristics of the population of interest. This makes it possible to identify the extent of the dispersion in earnings inequality as well as those characteristics that are associated with very large, as well as very small (or even reversed), gender gaps in earnings. The study also compares heterogeneity in the hourly and monthly gender earnings gap to assess how adjusting for working hours affects the correlates of earnings inequality.

HIGHLIGHTS

• There is significant heterogeneity in the size of the gender earnings gap in South Africa.

• This is illustrated using a technique that maps the full distribution of earnings differences.

• Marriage, children, and geography are important sources of heterogeneity.

• There is more heterogeneity in the monthly, than hourly, gender earnings gap.

KEYWORDS:

• Gender wage gap
• sorted partial effects
• heterogeneity
• South Africa

JEL Codes:

• J16
• J31

ACKNOWLEDGMENTS

The authors thank Jacqueline Mosomi for all her input with the PALMS data, and the reviewers and the associate editor for the very helpful comments received.

Notes

1 Apartheid was a system of institutionalized racism or segregation that privileged “whites” (or people of European descent) over the other race groups in South Africa. It was formalized in 1948 after the National Party took power and ended in 1994 with the appointment of the first democratic African National Congress government.

2 Four racial categories, which were designated during the Apartheid period, remain commonly identified in all household surveys and the Population Census conducted by the official statistical agency (Statistics South Africa): “Africans,” also referred to as black South Africans, who constitute just over 80 percent of the South African population; “Coloureds” (mixed race), “Indians/Asians,” and “Whites.”

3 This concern was addressed in subsequent household and labor force surveys in South Africa.

4 Note that alternatively we can represent the linear combination of the parameters and the covariates as follows: $\mathrm{\gamma x}=x_i^{\mathrm{\prime }}\gamma ={\sum }_{j=1}^p{\gamma }_j{x}_i$ and $\mathrm{\mu x}=x_i^{\mathrm{\prime }}\mu ={\sum }_{j=1}^p{\mu }_j{x}_i$, where $j$ indicates the column of the covariate matrix, $X$.

5 A negative $\mathrm{\Lambda }\left(x\right)$ is interpreted as a gender earnings penalty experienced by woman $i$ with characteristics $x_i$, while a positive $\mathrm{\Lambda }\left(x\right)$ is interpreted as a gender earnings premium experienced by woman $i$ with characteristics $x_i$.

6 As explained earlier, this is different from existing literature using a decomposition approach (such as Oaxaca-Blinder) which commonly estimates the average partial effect as a counterfactual distribution that allows for decomposing the gap into the effects of coefficients, covariates, and residuals. In a similar vein, literature that uses unconditional quantile regressions analyzes heterogeneity but only at specific points along the outcome distribution (Melly 2005). In contrast to these approaches, the SPE method (i) focuses on the heterogeneity in the explained part among subpopulations defined by covariate values and (ii) uses this heterogeneity as a basis for classifying observational units into most or least affected groups and summarizing their characteristics. The focus on the extremes is arguably better for informing policy targeting.

7 The PALMS version of the QLFS data is created by a team at DataFirst at the University of Cape Town, who collate the labor market data from the official labor force surveys conducted by Statistics South Africa and package them in a comparable and reliable form for public use, along with suitable weights (Kerr, Lam, and Wittenberg 2019).

8 The last wave of the National Income Dynamics Study (NIDS) for South Africa, which was also conducted in 2017, has the advantage over the QLFS of interviewing each adult member of the household, rather than asking only one member of the household to respond on behalf of others, as is the case in the QLFS. However, information on the earnings of the self-employed is largely missing in NIDS and as a longitudinal survey, national representativeness is compromised by attrition across the waves.

9 The Rand-US Dollar exchange rate was R17.62 on February 6, 2023.

10 We use the “spe” command in the R Sorted Effects package (Chen et al. 2019) to produce the estimates in Figure 1. The 90 percent bootstrap uniform confidence bands are shaded in blue and are based on 500 bootstrap replications (this follows algorithm 2.1 in Chernozhukov, Fernandez-Val, and Luo 2018: 1920). The analysis was performed using survey weights as indicated earlier.

11 The APE is marginally larger than the unconditioned average gap, reported in Table 1, signaling that women have characteristics which would favor higher earnings.

12 Although the samples and specifications are not directly comparable, we find considerably more variation in the SPEs than do Chernozhukov et al. (2018) using data from the US.

13 We use the 20 percent cut-off to analyze the tails of the distribution as this provides us with a substantial enough sample size to compare the most and least affected groups, although other papers applying the method have also used 5 percent or 10 percent. As a robustness check, we re-estimated the classification analysis comparing the 10 percent most affected women to the 10 percent least affected women. We found our results to be largely consistent with the 20 percent cut-off analysis, although standard errors are larger because samples are smaller. As would be expected when using more extreme groups, the size of the differences in characteristics between the two groups is often larger than when using the 20 percent samples. To give one example, when the 10 percent sample is used, the difference in the probability of being married between the least and most affected groups of women is −29 percentage points (SE of 0.14) compared to –25 percentage points (SE of 0.11) when the 20 percent sample is used (see Table 2), although in both cases the difference is statistically significant.

14 Although not shown here due to space constraints, we also estimated the analysis on the sample of wage employed only, given concerns that the earnings data for the self-employed might be recorded in the survey with greater error than the earnings for the wage employed. Again, our results suggest substantial heterogeneity in the gender gap, and the classification analysis produces a consistent story (both when comparing the results for monthly and hourly earnings gaps, and when comparing those with and without a full set of controls as done above).

Additional information

Notes on contributors

Dorrit Posel

Dorrit (Dori) Posel is a distinguished professor and holds the Helen Suzman Chair in the School of Economics and Finance at the University of the Witwatersrand. Much of her research over the past thirty years has been within the field of feminist economics, and she has published widely on issues related to marriage and family formation, labor migration, labor force participation and the gender division of labor, the economics and demographics of language use, and measures of well-being.

Dambala Gelo

Dambala Gelo is Associate Professor in the School of Economics and Finance at the University of the Witwatersrand, South Africa. His research concerns applied microeconomics of development and environment with a focus on institutional and organizational questions, human capital, technology adoption, poverty trap, and energy transition.

Daniela Casale

Daniela Casale is Professor in the School of Economics and Finance at the University of the Witwatersrand, South Africa. Daniela is a development economist who has been involved in research in the field of feminist economics for over twenty years. Other broad areas of study include labor, well-being, health, and education. She has published widely across a range of economics, development, and public health journals.

Adeola Oyenubi

Adeola Oyenubi is Associate Professor in the School of Economics and Finance, University of the Witwatersrand, South Africa. Adeola’s research interest is in the areas of health and development economics with a specific focus data driven approaches like Machine Learning.