Empirical approach

We want to estimate the impact of a firm’s ownership status on the GWG in order to establish how transnational corporation (TNC) activity affects wage-based gender inequality in South Africa. To do this, we follow a standard empirical approach used in the literature on trade and gender equality.⁴ That is, we estimate a Mincerian (Mincer, 1974) wage equation at the individual level that controls for an individual’s gender, whether he or she works for a firm with foreign ownership and an interaction between these two variables. Foreign ownership is determined by the company

4 See, for instance, Boler et al. (2018) and Bezuidenhout et al. (2019).

report in tax forms on whether their ultimate holding company⁵ is a resident of another country. We also control for the individual characteristics that are observed in the data, namely age and age squared, and whether a firm trades, given that the large majority of foreign-owned firms export and import which could cause us to assign a trade effect to the TNC variable. As the data do not provide sufficient individual or firm-level characteristics to exclude the possibility that our estimates suffer from omitted variable bias, we add sequentially several multi-dimensional fixed effects to account for this issue. To obtain a baseline estimate for the GWG and the role of foreign-owned firms that is only conditional on the observables included in the regression, we start by estimating the following equation without fixed effects:

(1) where wijst is the monthly income of worker i employed by firm j in industry s at time t, femi is a dummy variable equal to one for women, foreignjt is a dummy variable equal to one if the firm is foreign owned and tradejt is a dummy variable equal to one if the firm imports, exports or does both.

Coefficient β2 in equation (1) indicates how and if foreign ownership of a firm affects the GWG. This coefficient can be biased if a sectoral selection bias is present.

The GWG between foreign-owned and domestically-owned firms can thus appear larger or smaller than it is, if it is driven by variables related to the type of industry firms are in, rather than their ownership status. If, for example, women are mainly employed in high-paying industries that also have a higher share of foreign-owned firms, then the differential GWG of foreign-owned firms will be underestimated because of this gender-based clustering of workers in certain industries. We account for this potential source of bias by adding industry-year fixed effects () to equation (1). Accordingly, equation (2) is then specified as follows (for conciseness, we summarize age and age squared in the vector X):

(2) Equation (2) controls for sectoral selection bias over time. Yet, this might not be sufficient for an unbiased estimation of β2 as another type of selection bias can occur within industries if certain types of workers select into certain types of firms.

In this case, the GWG can be driven by individual characteristics that are not captured by the control variables in equations (1) and (2). For example, if male or

5 A holding company is defined as a company that controls enough voting stock in a subsidiary company to elect the board of directors and thereby control management of the subsidiary (Legwaila, 2010).

ln 𝑤𝑤_{𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖}= 𝛽𝛽₁𝑓𝑓𝑓𝑓𝑓𝑓_𝑖𝑖+ 𝛽𝛽₂𝑓𝑓𝑓𝑓𝑓𝑓_𝑖𝑖∗ 𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓_{𝑖𝑖𝑖𝑖}+ 𝛽𝛽₃𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓_{𝑖𝑖𝑖𝑖} + 𝛽𝛽₄𝑓𝑓𝑓𝑓𝑓𝑓_𝑖𝑖∗ 𝑡𝑡𝑓𝑓𝑡𝑡𝑡𝑡𝑓𝑓_{𝑖𝑖𝑖𝑖}+ 𝛽𝛽₅𝑡𝑡𝑓𝑓𝑡𝑡𝑡𝑡𝑓𝑓_{𝑖𝑖𝑖𝑖} + 𝛽𝛽6𝑡𝑡𝑓𝑓𝑓𝑓𝑖𝑖𝑖𝑖+ 𝛽𝛽7𝑡𝑡𝑓𝑓𝑓𝑓𝑖𝑖𝑖𝑖2+ 𝜀𝜀𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖

ln 𝑤𝑤𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖= 𝛽𝛽1𝑓𝑓𝑓𝑓𝑓𝑓𝑖𝑖+ 𝛽𝛽2𝑓𝑓𝑓𝑓𝑓𝑓𝑖𝑖∗ 𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑖𝑖𝑖𝑖+ 𝛽𝛽3𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑖𝑖𝑖𝑖 + 𝛽𝛽4𝑓𝑓𝑓𝑓𝑓𝑓𝑖𝑖∗ 𝑡𝑡𝑓𝑓𝑡𝑡𝑡𝑡𝑓𝑓𝑖𝑖𝑖𝑖+ 𝛽𝛽5𝑡𝑡𝑓𝑓𝑡𝑡𝑡𝑡𝑓𝑓𝑖𝑖𝑖𝑖 + 𝛽𝛽′𝑋𝑋_{𝑖𝑖𝑖𝑖}+ 𝛼𝛼_{𝑖𝑖𝑖𝑖}+ 𝜀𝜀_{𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖}.

female workers with higher education and better skills are more prone to work for foreign-owned firms, the GWG will be driven by the worker’s level of education or skills rather than the firm’s ownership status. To account for this, we add individual, or employee, fixed effects (𝛼i) to equation (2) in order to arrive at equation (3), which now controls for unobservable individual characteristics:⁶

(3) In addition to controlling for unobserved industry-specific (equation (2)) and individual-specific (equation (3)) characteristics that can bias the GWG between firms that are foreign owned and those that are not, it is also necessary to control for firm-specific characteristics that can influence both the GWG and a firm’s ownership status. The literature on wage differences shows, for instance, that there is a strong relation between a firm’s wages and its productivity and size.⁷ As an example, if larger firms are more likely to be foreign owned, have higher wages and employ either more women or men, the coefficient of interest β2 will be biased. To control for this, we add employer-employee, or job, fixed effects. By using a job fixed effect, the coefficient is identified only from firms that switch their ownership status while holding the workforce composition in these firms constant. This restriction provides us with a very conservative estimate for the effect of foreign ownership on the GWG and, thus, serves as our preferred specification:

(4)

We consider that the fixed effects, in particular in our preferred specification given by equation (4), effectively control for omitted variables bias and thus allow us to identify the causal effect of foreign ownership on the GWG.⁸ This, in turn, allows us to discuss the role of foreign ownership (i.e. FDI) in gender inequality and broader inequality in a developing-country context.

6 The new fixed effect absorbs the female dummy such that from equation (4) on, we cannot identify the degree of the GWG anymore.

7 For a review, see Bhorat et al. (2017).

8 We do not consider it likely that foreign firms invest on the basis of the targeted firm’s GWG. Hence, reverse causality is not a probable source of bias.

ln 𝑤𝑤𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖= 𝛽𝛽1𝑓𝑓𝑓𝑓𝑓𝑓𝑖𝑖+ 𝛽𝛽2𝑓𝑓𝑓𝑓𝑓𝑓𝑖𝑖∗ 𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑖𝑖𝑖𝑖+ 𝛽𝛽3𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑖𝑖𝑖𝑖 + 𝛽𝛽4𝑓𝑓𝑓𝑓𝑓𝑓𝑖𝑖∗ 𝑡𝑡𝑓𝑓𝑡𝑡𝑡𝑡𝑓𝑓𝑖𝑖𝑖𝑖+ 𝛽𝛽5𝑡𝑡𝑓𝑓𝑡𝑡𝑡𝑡𝑓𝑓𝑖𝑖𝑖𝑖 + 𝛽𝛽′𝑋𝑋_{𝑖𝑖𝑖𝑖}+ 𝛼𝛼_{𝑖𝑖𝑖𝑖}+ 𝛼𝛼_𝑖𝑖+ 𝜀𝜀_{𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖}.

ln 𝑤𝑤_{𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖}= 𝛽𝛽₁𝑓𝑓𝑓𝑓𝑓𝑓_𝑖𝑖+ 𝛽𝛽₂𝑓𝑓𝑓𝑓𝑓𝑓_𝑖𝑖∗ 𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓_{𝑖𝑖𝑖𝑖}+ 𝛽𝛽₃𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓_{𝑖𝑖𝑖𝑖} + 𝛽𝛽₄𝑓𝑓𝑓𝑓𝑓𝑓_𝑖𝑖∗ 𝑡𝑡𝑓𝑓𝑡𝑡𝑡𝑡𝑓𝑓_{𝑖𝑖𝑖𝑖}+ 𝛽𝛽₅𝑡𝑡𝑓𝑓𝑡𝑡𝑡𝑡𝑓𝑓_{𝑖𝑖𝑖𝑖} + 𝛽𝛽′𝑋𝑋_{𝑖𝑖𝑖𝑖}+ 𝛼𝛼_{𝑖𝑖𝑖𝑖}+ 𝛼𝛼_{𝑖𝑖𝑖𝑖}+ 𝜀𝜀_{𝑖𝑖𝑖𝑖𝑖𝑖𝑖𝑖}.