Globalization and Female Labor Force Participation in Developing Countries: An Empirical (Re-)Assessment

Abstract

We investigate the impact of foreign direct investment (FDI) and trade, as two measures of globalization, on female labor force participation rate in a sample of 80 developing countries over the time period 1980–2005. Contrary to the mainstream view in the literature, which is mainly based on country-case studies or simple cross-country variation, we find that both, FDI and trade have a generally negative impact on female labor force participation. While the impact is of negligible economic size, it is stronger for younger cohorts, possibly reflecting a higher return to education in open economies. We further find a large degree of cross-regional heterogeneity and that the effect of globalization on female labor force participation depends on the industrial structure, with more positive effects in economies with a higher share of industry in value added. We can thereby explain why country studies find other effects and question the generalization of their results into an overarching globalization tale concerning female labor force participation.

Appendices

Appendix A

Countries Included:

Albania, Armenia, Bangladesh, Belize, Benin, Bolivia, Botswana, Burundi, Cambodia, Cameroon, Central African Republic, Chile, China, Colombia, Congo, Rep., Costa Rica, Cote d’Ivoire, Cuba, Dominican Republic, Ecuador, Egypt, Arab Rep., El Salvador, Fiji, Gambia, Ghana, Guatemala, Guyana, Honduras, India, Indonesia, Jamaica, Jordan, Kazakhstan, Kenya, Kyrgyz Republic, Lao PDR, Lesotho, Liberia, Malawi, Malaysia, Maldives, Mali, Mauritania, Mauritius, Mexico, Mongolia, Morocco, Mozambique, Namibia, Nepal, Nicaragua, Niger, Pakistan, Papua New Guinea, Paraguay, Peru, Philippines, Poland, Rwanda, Senegal, Sierra Leone, Slovak Republic, South Africa, Sri Lanka, Sudan, Swaziland, Syrian Arab Republic, Tajikistan, Tanzania, Thailand, Togo, Tonga, Tunisia, Turkey, Uganda, Ukraine, Vietnam, Yemen, Rep., Zambia, Zimbabwe.

Table 7 Summary statistics based on those observations included in model (5) of Table 6.

Table 8 Multiplicative model

Table 9 Correlation between globalization and female education

Appendix B

Table 10 Overview of related literature

Appendix C: Error Structure of the Model

A concern of our model is the correlation structure of the idiosyncratic error ε in Eq. (4). Despite using a 5-year interval, autocorrelation is one potential issue. Together with potential heteroscedasticity, this can easily be accommodated by using the heteroscedasticity and autocorrelation (HAC) robust approach of Huber (1967) and White (1980) to estimate the variance-covariance (VCV) matrix. However, the hierarchical structure of our model (cf. Wooldridge 2003 and 2010: ch. 20 for an introductory treatment to such models) poses additional problems since, for example, the error ε _ijt is likely to be correlated with the error ε _i,j+1,t+1 because the individuals in cohort j in period t will be in cohort j + 1 in period t + 1. Furthermore, there might be correlation between all errors ε _.jt within country i if there is a systematic measurement error on the country level. All these potential problems with standard inference in linear models point to different forms of error correlation within countries. In line with the conventional panel data literature and given the dimension of our data set, we can assume that N → ∞ and hence the number of countries, which are considered to be the “clusters,” is large while the size of these clusters (i.e. the cohorts by country) is small. As discussed in Wooldridge (2003: 134, see also 2010: 864ff) a robust estimate for the VCV matrix is obtained by clustering the errors on the country level. Assuming that the matrix W _i contains all fixed effects and explanatory variables, classified as X and Z above, for country i and that the corresponding parameter vector δ contains β, θ, μ, and γ, a robust VCV estimator for δ is given by

$$ \widehat{VCV}\left(\widehat{\delta}\right)={\left(\sum\nolimits_{i=1}^N{W}_{i}^{\,\prime} {W}_i\right)}^{-1}\left(\sum\nolimits_{i=1}^N{W}_i^{\,\prime } {\widehat{\varepsilon}}_i{\widehat{\varepsilon}}_i^{\,\,\prime }{W}_i\right)\left(\sum\nolimits_{i=1}^N{W}_i^{\,\prime} {W}_i\right)^{-1}, $$

(A1)

where ${\widehat{\upvarepsilon}}_{\mathrm{i}}$ is the 10 T × 1 vector of residuals for country (i.e. cluster) i.²⁷ Using time-fixed effects is important in this context because it prevents the most likely form of cross-section, i.e. contemporaneous, correlation of the error term. We want to emphasize that clustering the errors at the country level has a tremendous impact on inference, as one would expect (cf. Wooldridge, 2010: 865). If one would (wrongly) cluster on the country-specific cohort level instead, which would be the standard option in most econometric packages, standard errors would be severely underestimated (cf. Table 8 in the appendix to the Working Paper version of this chapter).