Sources of the Persistent Gender Wage Gap along the Unconditional Earnings Distribution: Findings from Kenya

Abstract

Past studies on gender wage inequality in Africa typically attribute the gender pay gap either to gender differences in characteristics or in the return to characteristics. The authors suggest, however, that this understanding of the two sources may be far too general and possibly overlook the underlying covariates that drive the gender wage gap. Moreover, past studies focus on the gender wage gap exclusively at the conditional mean. The authors go further to evaluate the partial contribution of each wage-determining covariate to the magnitude of the gender pay gap along the unconditional earnings distribution. The authors' data are from Kenya, and their empirical technique mirrors re-centered influence function regressions. The authors' results are novel and suggest that while gender differences in characteristics and the return to characteristics widen the gender pay gap at the lower end of the wage distributions, gender differences in characteristics widen the gender wage gap at the upper end of the wage distributions. Importantly, the authors find that the underlying covariates driving gender differences in characteristics and the return to characteristics are the industry, occupation, higher education and region covariates. In the middle of the distributions, however, the authors find that gender differences in the return to characteristics, fueled by education and experience covariates, exert the strongest influence on the magnitude of the gender pay gap.

Notes

¹ Ntuli (2009) uses quantile regressions which have also found widespread application in both developed and developing countries. For example, Nielsen & Rosholm (2001) and Mueller (1998) use quantile regressions to examine public–private sector wages in Zambia and Canada, respectively. Albrecht et al. (2007) use quantile regressions to examine urban–rural inequality in Vietnam. Quantile regressions have also been used to examine wage inequality in China (e.g. Knight & Song, 2003) and to examine gender pay gaps in Spain (e.g. Garcia et al., 2001), Chile (e.g. Montenegro, 2001), and the Philippines (Sakellarious, 2004).

² It is important to mention that reverse causality is also plausible here. In other words, male-dominated occupations could be seen as market oriented, while female-dominated occupations could be seen as domestic oriented. The point here is to avoid the arbitrary assignment of the labels “market” and “domestic.”

³ Experience rather than age is typically used in Mincerian wage equations. However, measures of actual experience are unavailable in our data. For this reason, and consistent with Mincer (1974), we use potential experience defined as age-S-6, where S represents years of schooling. And since potential experience is a linear function of age and years of schooling, age and experience can be used interchangeably in the earnings equations. We choose experience.

⁴ In the Kenyan education system, primary education consists of classes Standards 1–8 and secondary of Forms 1–6.

⁵ Let v be a distributional statistic of interest such as a quantile. The influence function (IF) of v at a point y in robust statistics and econometrics is defined as:where is a slight perturbation of F by point mass at y. The re-centered influence function (RIF) is obtained by adding the original quantile back to its . For example, RIF for a quantile q _τ is given bywhere f _y is the marginal density function of Y, and I(·) is an indicator function. In practice, the RIF may be estimated by replacing unknown quantities by the estimators, that is , where is the τth sample quantile and is the kernel density estimator.

⁶ Indeed, in our sub-samples, the female sub-sample (2687) is relatively smaller than the male sub-sample (4834). Furthermore, our results from the propensity score estimates in Table 3 suggest no a priori reason that would impact our results in Tables 4 and 5.

⁷ The impact of either composition effects or wage structure effects and the influence of various categories and covariates on composition and wage structure effects at other quantiles will be determined and interpreted in the same way as at the 20th quantile. In this light, composition effects explain the gender pay gap at the 10th quantile and would follow the same interpretation as at the 20th quantile.