Education, Income, and Female Labor Force Participation: A Global View, Study Guides, Projects, Research of Literature

Download Education, Income, and Female Labor Force Participation: A Global View and more Study Guides, Projects, Research Literature in PDF only on Docsity! ESSAYS ON GENDER, MIGRATION, AND DEVELOPMENT Dissertation in order to acquire the doctoral degree from the Faculty of Economic Sciences, at the Georg-August-Universität Göttingen Submitted by Manuel Pedro Duarte Santos Silva Born in Porto, Portugal Göttingen, 2018 Contents Acknowledgments i List of Tables v List of Figures vii Introduction 1 1 Gender Inequality as a Barrier to Economic Growth 7 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 1.2 A simple efficiency argument: men and women . . . . . . . . . . . . . . 10 1.3 Unitary households: parents and children . . . . . . . . . . . . . . . . . . 13 1.4 Intra-household bargaining: husbands and wives . . . . . . . . . . . . . 19 1.5 Household formation patterns . . . . . . . . . . . . . . . . . . . . . . . . . 24 1.6 Beyond the household: openness, politics, and corruption . . . . . . . . 27 1.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2 What Drives Female Labor Force Participation? 35 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.2 Data and empirical model . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 2.2.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 2.2.2 Descriptives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 2.2.3 Modeling female labor force participation . . . . . . . . . . . . . . 49 2.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 2.3.1 Selection into marriage and urban areas . . . . . . . . . . . . . . . 59 2.3.2 Selection into education . . . . . . . . . . . . . . . . . . . . . . . . 60 2.4 Decomposition analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 2.4.1 Decomposing changes over time within countries . . . . . . . . . 64 2.4.2 Decomposing differences between countries . . . . . . . . . . . . 67 2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 2.A Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 3 The Roots of Female Emancipation 107 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 3.2 Theoretical discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 3.2.1 Original sources of gender (in)equality . . . . . . . . . . . . . . . 110 3.2.2 Historical household formation patterns . . . . . . . . . . . . . . 112 CONTENTS iv 3.2.3 Implications for gender equality . . . . . . . . . . . . . . . . . . . 114 3.2.4 Origins of household formation patterns . . . . . . . . . . . . . . 117 3.3 Data and descriptives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 3.4 Empirical strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 3.5.1 Ages at first marriage and gender equality . . . . . . . . . . . . . 131 3.5.2 Cool Water breeds late-marriage societies . . . . . . . . . . . . . . 134 3.5.3 Cool Water and historic late marriages . . . . . . . . . . . . . . . . 144 3.5.4 Cool Water and contemporary gender equality . . . . . . . . . . . 146 3.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 3.A Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 4 Can Parental Migration Reduce Petty Corruption in Education? 163 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164 4.2 Related literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 4.3 Moldova and corruption in education . . . . . . . . . . . . . . . . . . . . 167 4.4 Data and descriptives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 4.5 Empirical strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172 4.6 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 4.7 Transmission channels and robustness . . . . . . . . . . . . . . . . . . . . 179 4.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184 4.A Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186 Bibliography 205 List of Tables 2.1 Estimation results: overview . . . . . . . . . . . . . . . . . . . . . . . . . . 53 2.A.1 Data overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 2.A.2 South Africa: sample means . . . . . . . . . . . . . . . . . . . . . . . . . . 87 2.A.3 Brazil: sample means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 2.A.4 Jordan: sample means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 2.A.5 India: sample means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 2.A.6 Bolivia: sample means . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 2.A.7 Indonesia: sample means . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 2.A.8 Vietnam: sample means . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 2.A.9 Tanzania: sample means . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 2.A.10 South Africa: average marginal effects . . . . . . . . . . . . . . . . . . . . 92 2.A.11 Brazil: average marginal effects . . . . . . . . . . . . . . . . . . . . . . . . 93 2.A.12 Jordan: average marginal effects . . . . . . . . . . . . . . . . . . . . . . . 94 2.A.13 India: average marginal effects . . . . . . . . . . . . . . . . . . . . . . . . 95 2.A.14 Bolivia: average marginal effects . . . . . . . . . . . . . . . . . . . . . . . 96 2.A.15 Indonesia: average marginal effects . . . . . . . . . . . . . . . . . . . . . . 97 2.A.16 Vietnam: average marginal effects . . . . . . . . . . . . . . . . . . . . . . 98 2.A.17 Tanzania: average marginal effects . . . . . . . . . . . . . . . . . . . . . . 99 2.A.18 Trends in sample inclusion criteria over time . . . . . . . . . . . . . . . . 100 2.A.19 India and Jordan: decomposition of FLFP . . . . . . . . . . . . . . . . . . 101 2.A.20 Brazil and South Africa: decomposition of FLFP . . . . . . . . . . . . . . 101 2.A.21 Indonesia and Bolivia: decomposition of FLFP . . . . . . . . . . . . . . . 102 2.A.22 Tanzania and Vietnam: decomposition of FLFP . . . . . . . . . . . . . . . 102 2.A.23 South Africa: decomposition of FLFP, 1995–2014 . . . . . . . . . . . . . . 103 2.A.24 Educational attainment and social group: common variables for all coun- tries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 2.A.25 Decomposition of FLFP between countries: first year . . . . . . . . . . . 105 2.A.26 Decomposition of FLFP between countries: last year . . . . . . . . . . . . 106 3.1 Descriptive statistics for selected variables . . . . . . . . . . . . . . . . . . 123 3.2 Determinants of gender gaps: ages at first marriage . . . . . . . . . . . . 132 3.3 Determinants of gender gaps: ages at first marriage; subsample analysis 135 3.4 Determinants of ages at first marriage . . . . . . . . . . . . . . . . . . . . 137 3.5 Determinants of ages at first marriage: subsample analysis . . . . . . . . 138 3.6 Determinants of ages at first marriage: additional controls . . . . . . . . 142 3.7 Assessing unobservable selection: estimates of δ . . . . . . . . . . . . . . 144 3.8 Europe: historical female ages at first marriage . . . . . . . . . . . . . . . 145 Introduction I assume that the reader and I agree on a basic premise: that, above all else, equality of opportunities between men and women is an issue of elementary justice and, therefore, an end in itself. Let us not bother John Rawls or Amartya Sen; let us save our veils of ignorance for other, thornier occasions. In this sense, gender equality is development. As put by Harriet Martineau, in 1837, “[i]f a test of civilization be sought, none can be so sure as the condition of that half of society over which the other half has power,—from the exercise of the right of the strongest” (Martineau, 1837, p. 105). Besides being a topic with a clear normative goal, gender inequality is a fascinating field of economic research. This dissertation focuses, in particular, on the relationship between gender equality and economic development, broadly defined. Since the seminal work of Ester Boserup (1970), gender equality is widely understood as being both an instrument for and a consequence of economic development.1 The first three chapters of this thesis revolve around two fundamental questions. What are the consequences of gender inequality for economic development? (Chapter 1.) And what are its causes—i.e., why do we observe gender inequality in all societies, although to different extents? (Chapters 2 and 3.) To paraphrase Robert Lucas, “once one starts to think about [these questions], it is hard to think about anything else” (Lucas, 1988, p. 5).2 As we shall see, social and cultural norms on gender roles will contribute a great deal to how we think about these issues. In the final chapter (Chapter 4), I investigate an episode of rapid change in social norms, in a context of mass emigration. Even though that essay does not deal with gender equality directly, it suggests that migration may be a powerful vehicle for cultural change in developing countries.3 Since 1980, academic work on gender equality has risen dramatically. Figure 1 plots the combined frequency of the three 2-grams “gender equality”, “gender inequality”, 1For recent surveys of this two-way relationship, see World Bank (2001, 2011); Duflo (2012); Jayachan- dran (2015). 2Lucas’s original remark concerned variation in per capita income across countries. 3The transfer of cultural values from migrants back to their countries of origin is usually known as “social remittances”, a term coined by the sociologist Peggy Levitt (Levitt, 1998, 2001). There is growing research in economics documenting that such cultural transfers exist and are quantitatively important (e.g., Batista and Vicente, 2011; Tuccio and Wahba, 2015; Barsbai et al., 2017; Ivlevs and King, 2017). INTRODUCTION 2 0 .0 00 1 .0 00 2 .0 00 3 2− gr am s fr eq ue nc y (% ) 1950 1960 1970 1980 1990 2000 2010 Year "gender equality", "gender inequality", or "gender gap" "sex equality", "sex inequality", or "sex gap" FIG. 1: Rising scholarship on gender equality Notes: Author’s calculations from Google Books Ngram Viewer, corpus of English books, 2012 version (persistent identifier: googlebooks-eng-all-20120701). See Michel et al. (2011) for details. Data retrieved from http://books.google.com/ngrams [accessed on May 24, 2018]. The full line is the sum of the frequency of the following 2-grams: “gender equality”, “gender inequality”, and “gender gap”. The dashed line is the sum of the frequency of the following 2-grams: “sex equality”, “sex inequality”, and “sex gap”. Searches were case-insensitive. Time period is 1950–2008; 3-year moving averages are shown. and “gender gap” in a large corpus of English books. From virtually no usage in books before 1980, their frequency has been growing fast ever since. Perhaps, this trend simply reflects a change in terminology, if the term “gender” (i.e., the social identity assigned to each biological sex) was replacing the term “sex” (i.e., the biological difference between men and women). But this is not the case: as shown in the figure, the rise of “gender” equality does not coincide with a decline of “sex” equality. While scholarly interest on gender equality was rising, developing countries were making extraordinary progress in reducing gender gaps in several socio-economic di- mensions.4 Since 1980, gender gaps in enrollment rates have been eliminated for primary and secondary schooling in all but a few countries, whereas for tertiary education the gap is now in favor of women (e.g., Ganguli et al., 2014; World Bank, 2011, pp. 106–116). Fertility rates also fell dramatically in this period, with the speed of this decline being much faster than in the past experiences of today’s advanced economies (World Bank, 4During the same period, advanced economies also witnessed progress towards gender equality (e.g., Kleven and Landais, 2017). Claudia Goldin provides a grand narrative for the US (Goldin, 1990, 2006, 2014). 5 positive income effect, we find that the strongest migration-related response in private education expenditure is a substantial decrease in informal payments to public school teachers. Any positive income effect due to migration must hence be overcompensated by some payment-reducing effects. We argue that these effects probably reflect migrants’ changing views on the value and acceptability of corruption, which spillover to their country of origin. 1 Gender Inequality as a Barrier to Economic Growth: a Review of the Theoretical Literature∗ Abstract: In this chapter, we survey the theoretical literature investigating the role of gender inequality in economic development. The vast majority of theories reviewed suggest that gender inequality is a barrier to development, particularly over the long run. Among the many plausible mechanisms through which inequality between men and women affects the aggregate economy, the role of women for fertility decisions and human capital investments is particularly important. Yet, we believe the body of theories could be expanded in several directions. ∗This chapter is co-authored with Stephan Klasen. We gratefully acknowledge funding from the Growth and Economic Opportunities for Women (GrOW) initiative, a multi-funder partnership between the UK’s Department for International Development, the Hewlett Foundation and the International Development Research Centre. All views expressed here and remaining errors are our own. 1 GENDER INEQUALITY AS A BARRIER TO ECONOMIC GROWTH 10 In addition to this descent—from aggregate production factors to households, and then to household members—, the analysis has also expanded horizontally, by consid- ering new arenas in which gender inequality has relevant consequences for economic development. Examples are international trade (Seguino, 2000; Blecker and Seguino, 2002), foreign direct investment (Rees and Riezman, 2012), and politics (Besley et al., 2017). Section 1.6 discusses this literature. The vast majority of theories reviewed suggest that gender inequality is a barrier to economic development, particularly over the long run. In most models, irrespectively of the underlying source of differences between the genders (e.g., biology, socialization, discrimination), the opportunity cost of women’s time is lower than that of men. This gender gap in the value of time affects economic growth through two main mechanisms. First, when women’s time is relatively low, women will be in charge of childrearing and domestic work in the family. A low value of female time means that children are cheap. Fertility will be high, and economic growth will be low, both because population growth has a direct negative impact on long-run economic performance and because human capital accumulates at a slower pace (through the quantity-quality trade-off). Second, if parents expect low returns to female education, due to women specializing in domestic activities, they will invest relatively less in the education of girls. In the words of Harriet Martineau, one of the first to describe this mechanism, “as women have none of the objects in life for which an enlarged education is considered requisite, the education is not given” (Martineau, 1837, p. 107). In the long run, lower human capital investments (on girls) retard economic development. We conclude, in section 1.7, by examining the limitations of the current literature and pointing ways forward. 1.2 A simple efficiency argument: men and women In this section, we review three prominent arguments making the case that gender inequality in productive capabilities generates aggregate inefficiencies. According to this view, more equality between men and women leads to static efficiency gains in the short run. Yet, other authors warn that gender gaps in different dimensions interrelate, and addressing a gap in isolation may have ambiguous short term effects on economic performance. The simplest argument for why gender inequality harms economic growth rests on two premises: (1) men and women are separate inputs in the economy-wide production of goods and services, and (2) each input has positive and diminishing marginal products. 11 An example of this setup is the Solow-type neoclassical growth model of Knowles et al. (2002), where male and female education are imperfect substitutes in production. A gender gap emerges in the level of the education input if men are more educated than women (or vice-versa). A gender gap emerges in the returns to the education input if its output elasticity differs between genders, such that, at any education level, the marginal products of education are also different. In the following, by a reduction of the gender gap, we typically have in mind some sort of re-distribution between the genders to a more egalitarian outcome. Of course, if men are more educated than women, another way of reducing the gender gap would be to increase female education, keeping male education constant. Because average education in society goes up, there is an obvious positive level effect on per capita output. But the interesting question is whether, keeping average education constant, smaller gender gaps in education are more conductive to growth than bigger gaps, i.e., whether there is a distribution effect. In the case where output elasticities are the same for both genders, an economy where men and women contribute equally to aggregate production will maximize real output. Because men and women are imperfect substitutes, gender inequalities in how productive capacities are distributed are inefficient. Simply put, if men contribute more than women, the marginal product to the male input will be lower than the marginal product to the female input. Closing these gender gaps (in education, health, capital access, etc) would boost economic growth. On top of this argument, Knowles et al. (2002) hypothesize that women’s output elasticity of education is larger than men’s. Although they do not explicitly model why this is the case, they justify the hypothesis with positive externalities of female education in reducing fertility and infant mortality, and improving the quantity and quality of children’s education. If the output elasticity of female education is relatively large, a gender gap unfavorable to women reduces per capita output in the long run. In fact, the most efficient outcome would be a gender gap in the opposite direction, i.e., unfavorable to men. The reasoning of Knowles et al. (2002) can easily be extended to other productive capabilities beyond education, such as health and access to capital. Often, female output elasticities are assumed to be larger than male elasticities, due to intergenerational externalities linked to woman’s role as the primary caretaker in the family. A second related argument for why gender inequality leads to aggregate inefficiency concerns the allocation of talent. Assume that talent is randomly distributed in the population. Then, an economy that curbs women’s access to education, market em- 1 GENDER INEQUALITY AS A BARRIER TO ECONOMIC GROWTH 12 ployment, or certain occupations draws talent from a smaller pool than an economy without such restrictions (Klasen, 2002). Gender inequality can thus be viewed as a distortionary tax on talent (Dollar and Gatti, 1999). Indeed, occupational choice models with heterogeneous talent show that exogenous barriers to women’s participation in the labor market and entrepreneurial occupations reduce aggregate productivity and per capita output (Esteve-Volart, 2004; Cuberes and Teignier, 2016, 2018). Thus, if women have lower education, their marginal return to education would be higher than men’s. Similarly, if women are more credit-constrained than men, female returns to capital should be higher than male returns, and so on. The problem with this type of reasoning is that it considers inequalities in separate dimensions as being inde- pendent from each other. In many cases, however, these inequalities are complementary (Duflo, 2012; Bandiera and Does, 2013; Kabeer, 2016). For example, if credit-constrained women face weak property rights, are unable to access certain markets, and have mobil- ity and time constraints, then the marginal return to capital may nevertheless be larger for men. Similarly, the return to male education may well be above the female return if demand for female labor is low or concentrated in sectors with low productivity. In sum, “the fact that women face multiple constraints means that relaxing just one may not improve outcomes” (Duflo, 2012, p. 1076). When applied to a particular productive endowment in isolation, the efficiency argument for gender equality may not hold. A third important economic distortion is discrimination against women in the form of lower wages, holding male and female productivity constant. Cavalcanti and Tavares (2016) estimate the aggregate effects of wage discrimination using a model-based general equilibrium representation of the US economy. In their model, households are unitary and, within the household, women are assumed to be more productive in childrearing than men, so they pay the full time cost of this activity. In the labor market, even though men and women are equally productive, women receive only a fraction of the male wage rate—this is the wage discrimination assumption. Wage discrimination works as a tax on female labor supply. Because women work less than they would without discrimination, there is a negative level effect on per capita output. In addition, there is a second negative effect of wage discrimination operating through endogenous fertility. Since lower wages reduce women’s opportunity costs of childrearing, fertility is relatively high, and output per capita is relatively low. The authors calibrate the model to US steady state parameters and estimate large negative output costs of the gender wage gap. Reducing wage discrimination against women by 50 percent would raise per capita income by 35 percent, in the long run. To sum up, when men and women are imperfect substitutes in production and 15 opportunity costs dominates once again, and fertility declines. The economy moves from a “breadwinner model” to a “dual-earnings model”. Human capital accumulation plays no role in Galor and Weil (1996) and Kimura and Yasui (2010). Each person is exogenously endowed with a unit of brains. The fundamental trade-off in the two models is between the income and substitution effects of rising wages on the demand for children. When Lagerlöf (2003) adds education investments to a gender-based model, an additional trade-off emerges: that between the quantity and the quality of children. Lagerlöf (2003) models gender inequality as a social norm: on average, men have higher human capital than women. Confronted with this fact, parents play a coordination game in which it is optimal for them to reproduce the inequality in the next generation. The reason is that parents expect the future husbands of their daughters to be, on average, relatively more educated than the future wives of their sons. Because, in the model, parents care for the total income of their children’s future households, they respond by investing relatively less on daughters’ human capital. Here, gender inequality does not arise from some intrinsic difference between men and women. It is instead the result of a coordination failure: “[i]f everyone else behaves in a discriminatory manner, it is optimal for the atomistic player to do the same” (Lagerlöf, 2003, p. 404). With lower human capital, women earn lower wages than men and are therefore solely responsible for the time cost of childrearing. But if, exogenously, the social norm becomes more gender egalitarian over time, the gender gap in parental educational investment decreases. As better educated girls grow up and become mothers, their opportunity costs of childrearing are higher. Parents trade-off the quantity of children by their quality; fertility falls and human capital accumulates. However, rising wages have an offsetting positive income effect on fertility because parents pay a (fixed) “goods cost” per child. The goods cost is proportionally more important in poor societies than in richer ones. As a result, in poor economies, growth takes off slowly because the positive income effect offsets a large chunk of the negative substitution effect. As economies grow richer, the positive income effect vanishes (as a share of total income), and fertility declines faster. That is, growth accelerates over time even if gender equality increases only linearly. The natural next step is to model how the social norm on gender roles evolves en- dogenously during the course of development. Hiller (2014) develops such a model by combining two main ingredients: a gender gap in the endowments of brawn (as in Galor and Weil, 1996) generates a social norm, which each parental couple takes as given (as in Lagerlöf, 2003). The social norm evolves endogenously, but slowly; it tracks 1 GENDER INEQUALITY AS A BARRIER TO ECONOMIC GROWTH 16 the gender ratio of labor supply in the market, but with a small elasticity. When the male-female ratio in labor supply decreases, stereotypes adjust and the norm becomes less discriminatory against women. The model generates a U-shaped relationship between economic development and female labor force participation.7 In the preindustrial stage, there is no education and all labor activities are unskilled, i.e., produced with brawn. Because men have a comparative advantage in brawn, they supply more labor to the market than women, who specialize in home production. This gender gap in labor supply creates a social norm that favors boys over girls. Over time, exogenous skill-biased technological progress raises the relative returns to brains, inducing parents to invest in their children’s education. At the beginning, however, because of the social norm, only boys become educated. The economy accumulates human capital and grows, generating a positive income effect that, in isolation, would eventually drive up parental investments in girls’ education.8 But endogenous social norms move in the opposite direction. When only boys receive education, the gender gap in returns to market work increases, and women withdraw to home production. As female relative labor supply in the market drops, the social norm becomes more discriminatory against women. As a result, parents want to invest relatively less in their daughters’ education. In the end, initial conditions determine which of the forces dominates, thereby shaping long-term outcomes. If, initially, the social norm is very discriminatory, its effect is stronger than the income effect; the economy becomes trapped in an equilibrium with high gender inequality and low per capita income. If, on the other hand, social norms are relatively egalitarian to begin with, then the income effect dominates, and the economy converges to an equilibrium with gender equality and high income per capita. In the models reviewed so far, human capital or brain endowments can be understood as combining both education and health. Bloom et al. (2015) explicitly distinguish these two dimensions. Health affects labor market earnings because sick people are out of work more often (participation effect) and are less productive per hour of work (productivity effect). Female health is assumed to be worse than male health, implying that women’s effective wages are lower than men’s. As a result, women are solely responsible for childrearing.9 7The hypothesis that female labor force participation and economic development have a U-shaped relationship—known as the feminization-U hypothesis—goes back to Boserup (1970). See also Goldin (1995). Recently, Gaddis and Klasen (2014) find only limited empirical support for the feminization-U. 8The model does not consider fertility decisions. Parents derive utility from their children’s human capital (social status utility). When household income increases, parents want to “consume” more social status by investing in their children’s education—this is the positive income effect. 9Bloom et al. (2015) build their main model with unitary households, but show that the key conclusions 17 The model produces two growth regimes: a Malthusian trap with high fertility and no educational investments; and a regime of sustained growth, declining fertility, and rising educational investments. Once wages reach a certain threshold, the economy goes through a fertility transition and education expansion, taking off from the Malthusian regime to the sustained growth regime. Female health promotes growth in both regimes, and it affects the timing of the takeoff. The healthier women are, the earlier the economy takes off. The reason is that a healthier woman earns a higher effective wage and, consequently, faces higher opportunity costs of raising children. When female health improves, the rising opportunity costs of children reduce the wage threshold at which educational investments become attractive; the fertility transition and mass education periods occur earlier. In contrast, improved male health slows down economic growth and delays the fertility transition. When men become healthier, there is only a income effect on the demand for children, without the negative substitution effect (because male childrearing time is already zero). The policy conclusion would be to redistribute health from men to women. However, the policy would impose a static utility cost on the household. Because women’s time allocation to market work is constrained by childrearing responsi- bilities (whereas men work full-time), the marginal effect of health on household income is larger for men than for women. From the household’s point of view, reducing the gender gap in health produces a trade-off between short-term income maximization and long-term economic development. In an extension of the model, the authors endogeneize health investments, while keeping the assumption that women pay the full time cost of childrearing. Because women participate less in the labor market (due to childrearing duties), it is optimal for households to invest more in male health. A health gender gap emerges from rational household behavior that takes into account how time-constraints differ by gender; assuming taste-based discrimination against girls or gender-specific preferences is not necessary. Until now, parents invest in their children’s human capital for purely altruistic reasons. This is captured in the models by assuming that parents derive utility directly from the quantity and quality of children. This is the classical representation of children as durable consumption goods (e.g., Becker, 1960). In reality, of course, parents may also have egoistic motivations for investing in child quantity and quality. A typical example is that, when parents get old and retire, they receive support from their children. The quantity and quality of children will affect the size of old-age transfers and parents are robust to a collective representation of the household. 1 GENDER INEQUALITY AS A BARRIER TO ECONOMIC GROWTH 20 quantity) than men do. Prettner and Strulik (2017) build a unified growth theory model with collective households. Men and women have different preferences, but they achieve efficient cooperation based on (reduced-form) bargaining parameters. The authors study the effect of two types of preferences: (i) women are assumed to have a relative preference for child quality, while men have a relative preference for child quantity; and (ii) parents are assumed to have a relative preference for the education of sons over the education of daughters. In addition, it is assumed that the time cost of childcare borne by men cannot be above that borne by women (but it could be the same). When women have a relative preference for child quality, increasing female empower- ment speeds up the economy’s escape from a Malthusian trap of high fertility, low edu- cation, and low income per capita. When female empowerment increases (exogenously), a woman’s relative preference for child quality has a higher impact on household’s decisions. As a consequence, fertility falls, human capital accumulates, and the economy starts growing. The model also predicts that the more preferences for child quality differ between husband and wife, the more effective is female empowerment in raising long-run per capita income, because the sooner the economy escapes the Malthusian trap. This effect is not affected by whether parents have a preference for the education of boys relative to that of girls. If, however, men and women have similar preferences with respect to the quantity and quality of their children, then female empowerment does not affect the timing of the transition to the sustained growth regime. Strulik (2018) goes one step further and endogeneizes why men seem to prefer having more children than women. The reason is a different preference for sexual activity: other things equal, men enjoy having sex more than women.12 When cheap and effective contraception is not available, a higher male desire for sexual activity explains why men also prefer to have more children than women. In a traditional economy, where no contraception is available, fertility is high, while human capital and economic growth are low. When female bargaining power increases, couples reduce their sexual activity, fertility declines, and human capital accumulates faster. Faster human capital accumula- tion increases household income and, as a consequence, the demand for contraception goes up. As contraception use increases, fertility declines further. Eventually, the econ- omy undergoes a fertility transition and moves to a modern regime with low fertility, widespread use of contraception, high human capital, and high economic growth. In the modern regime, because contraception is widely used, men’s desire for sex is de- 12There are lots of empirical studies in line with this assumption, which is rooted in evolutionary psychology. See Strulik (2018) for references. There are several other evolutionary arguments for men wanting more children (including with different women). See, among others, Penn and Smith (2007); Mulder and Rauch (2009); von Rueden and Jaeggi (2016). 21 coupled from fertility. Both sex and children cost time and money. When the two are decoupled, men prefer to have more sex at the expense of the number of children. There is a reversal in the gender gap in desired fertility. When contraceptives are not available, men desire more children than women; once contraceptives are widely used, men desire fewer children than women. If women are more empowered, the transition from the traditional equilibrium to the modern equilibrium occurs faster. Both Prettner and Strulik (2017) and Strulik (2018) rely on gender-specific preferences. In contrast, Doepke and Tertilt (2014) are able to explain gender-specific expenditure patterns without having to assume that men and women have different preferences. They set up a non-cooperative model of household decision making and ask whether more female control of household resources leads to higher child expenditures and, thus, to economic development.13 In their model, household public goods are produced with two inputs: time and goods. Instead of a single home-produced good (as in most models), there is a continuum of household public goods whose production technologies differ. Some public goods are more time-intensive to produce, while others are more goods-intensive. Each specific public good can only be produced by one spouse—i.e., time and good inputs are not separable. Women face wage discrimination in the labor market, so their opportunity cost of time is lower than men’s. As a result, women specialize in the production of the most time-intensive household public goods (e.g., childrearing activities), while men specialize in the production of goods-intensive household public goods (e.g., housing infrastructure). Notice that, because the household is non-cooperative, there is not only a division of labor between husband and wife, but also a division of decision making, since ultimately each spouse decides how much to provide of his or her public goods. When household resources are redistributed from men to women (i.e., from the high- wage spouse to the low-wage spouse), women provide more public goods, in relative terms. It is ambiguous, however, whether the total provision of public goods increases with the re-distributive transfer. In a classic model of gender-specific preferences, a wife increases child expenditures and her own private consumption at the expense of the husband’s private consumption. In Doepke and Tertilt (2014), however, the rise in child expenditures (and time-intensive public goods in general) comes at the expense of male consumption and male-provided public goods. Parents contribute to the welfare of the next generation in two ways: via human capital investments (time-intensive, typically done by the mother) and bequests of physical capital (goods-intensive, typically done by the father). Transferring resources to women 13They do not model fertility decisions. So there is no quantity-quality trade-off. 1 GENDER INEQUALITY AS A BARRIER TO ECONOMIC GROWTH 22 increases human capital, but reduces the stock of physical capital. The effect of such transfers on economic growth depends on whether the aggregate production function is relatively intensive in human capital or in physical capital. If aggregate production is relatively human capital intensive, then transfers to women boost economic growth; if it is relatively intensive in physical capital, then transfers to women may reduce economic growth. There is an interesting paradox here. On the one hand, transfers to women will be growth-enhancing in economies where production is intensive in human capital. These would be more developed, knowledge intensive, service economies. On the other hand, the positive growth effect of transfers to women increases with the size of the gender wage gap, that is, decreases with female empowerment. But the more advanced, human capital intensive economies are also the ones with more female empowerment (i.e., lower gender wage gaps). In other words, in settings where human capital investments are relatively beneficial, the contribution of female empowerment to human capital accumulation is reduced. Overall, Doepke and Tertilt’s (2014) model predicts that female empowerment has at best a limited positive effect and at worst a negative effect on economic growth. Diebolt and Perrin (2013) assume cooperative bargaining between husband and wife, but do not rely on sex-specific preferences or differences in ability. Men and women are only distinguished by different uses of their time endowments, with females in charge of all childrearing activities. In line with this labor division, the authors further assume that only the mother’s human capital is inherited by the child at birth. On top of the inherited maternal endowment, individuals can accumulate human capital during adulthood, through schooling. The higher the initial human capital endowment, the more effective is the accumulation of human capital via schooling. A woman’s bargaining power in marriage determines her share in total household consumption and is a function of the relative female human capital of the previous generation. An increase in the human capital of mothers relative to that of fathers has two effects. First, it raises the incentives for human capital accumulation of the next generation, because inherited maternal human capital makes schooling more effective. Second, it raises the bargaining power of the next generation of women and, because women’s consumption share increases, boosts the returns on women’s education. The second effect is not internalized in women’s time allocation decisions; it is an intergener- ational externality. Thus, an exogenous increase in women’s bargaining power would promote economic growth by speeding up the accumulation of human capital across overlapping generations. 25 extending far back in history. For example, Hajnal (1965, 1982) described a distinct household formation pattern in preindustrial Northwestern Europe (usually referred as the “European Marriage Pattern”) characterized by: (i) late ages at first marriage for women, (ii) most marriages done under individual consent, and (iii) neolocality (i.e., upon marriage, the bride and the groom leave their parental households to form a new household). In contrast, marriage systems in China and India consisted of: (i) very early female ages at first marriage, (ii) arranged marriages, and (iii) patrilocality (i.e., the bride joins the parental household of the groom). Economic historians argue that the “European Marriage Pattern” empowered women, encouraging their participation in market activities and reducing fertility levels. While some view this as one of the deep-rooted factors explaining Northwestern Europe’s earlier takeoff to sustained economic growth (e.g., Hartman, 2004; De Moor and Van Zanden, 2010; Carmichael et al., 2016), others have downplayed the long-run significance of this marriage pattern (e.g., Ruggles, 2009; Dennison and Ogilvie, 2014). Despite this lively debate, the topic has been largely ignored by growth theorists. The few exceptions are Edlund and Lagerlöf (2006) and Tertilt (2005, 2006). Edlund and Lagerlöf (2006) study how rules of consent for marriage influence long- run economic development. In their model, marriages can be formed according to two types of consent rules: individual consent or parental consent. Under individual consent, young people are free to marry whomever they wish, while, under parental consent, their parents are in charge of arranging the marriage. Depending on the prevailing rule, the recipient of the bride-price differs. Under individual consent, a woman receives the bride-price from her husband, whereas, under parental consent, her father receives the bride-price from the father of the groom.15 In both situations, the father of the groom owns the labor income of his son and, therefore, pays the bride-price, either directly, under parental consent, or indirectly, under individual consent. Under individual consent, the father needs to transfer resources to his son to nudge him into marrying. Thus, individual consent implies a transfer of resources from the old to the young and from men to women, relative to the rule of parental consent. Redistributing resources from the old to the young boosts long-run economic growth. Because the young have a longer timespan to extract income from their children’s labor, they invest relatively more in the human capital of the next generation. In addition, under individual consent, the reallocation of resources from men to women can have additional positive effects on growth, by increasing women’s bargaining power (see section 1.4), although this 15The bride-price under individual consent need not be paid explicitly as a lump-sum transfer. It could, instead, be paid to the bride implicitly in the form of higher lifetime consumption. 1 GENDER INEQUALITY AS A BARRIER TO ECONOMIC GROWTH 26 channel is not explicitly modeled in Edlund and Lagerlöf (2006). Tertilt (2005) explores the effects of polygyny on long-run development through its impact on savings and fertility. In her model, parental consent applies to women, while individual consent applies to men. There is a competitive marriage market where fathers sell their daughters and men buy their wives. As each man is allowed (and wants) to marry several wives, a positive bride-price emerges in equilibrium.16 Upon marriage, the reproductive rights of the bride are transferred from her father to husband, who makes all fertility decisions on his own and, in turn, owns the reproductive rights of his daughters. From a father’s perspective, daughters are investments goods; they can be sold in the marriage market, at any time. This feature generates additional demand for daughters, which increases overall fertility, and reduces the incentives to save, which decreases the stock of physical capital. Under monogamy, in contrast, the equilibrium bride-price is negative (i.e., a dowry). The reason is that maintaining unmarried daughters is costly for their fathers, so they are better-off paying a (small enough) dowry to their future husbands. In this setting, the economic returns to daughters are lower and, consequently, so is the demand for children. Fertility decreases and savings increase. Thus, moving from polygny to monogamy lowers population growth and raises the capital stock in the long run, which translates into higher output per capita in the steady state. Instead of enforcing monogamy in a traditionally polygynous setting, an alternative policy is to transfer marriage consent from fathers to daughters. Tertilt (2006) shows that when individual consent is extended to daughters, such that fathers do not receive the bride-price anymore, the consequences are qualitatively similar to a ban on polygyny. If fathers stop receiving the bride-price, they save more physical capital. In the long run, per capita output is higher when consent is transferred to daughters. In summary, the rules regulating marriage and household formation carry relevant theoretical consequences for economic development. While the few studies on this topic have focused on consent rules and polygyny, other features of the marriage market remain largely unexplored. 16In Tertilt (2005), all men are similar (except in age). Widespread polygyny is possible because older men marry younger women and population growth is high. This setup reflects stylized facts for Sub-Saharan Africa. It differs from models that assume male heterogeneity in endowments, where polygyny emerges because a rich male elite owns several wives, while poor men remain single (e.g., Lagerlöf, 2005; Gould et al., 2008). 27 1.6 Beyond the household: openness, politics, and corruption In this section, we review theories that explore how gender inequality affects economic growth in three areas: small open economies, politics, and corruption. Opening to the world All the models reviewed so far considered closed economies. In open economies, however, gender inequality can interact with trade and international capital flows. Seguino (2000) argues that wage discrimination against women promotes economic growth in countries where exports are the main engine of growth and where the ex- port sector is female-intensive. Higher wage discrimination against women—i.e., an increase in the gender wage gap that is unrelated to productivity differences between the genders—increases the country’s export competitiveness. But for an increase in wage discrimination not to result in women leaving to other sectors, a sufficient degree of job segregation is needed, such that women are effectively “trapped” in the export sector. Blecker and Seguino (2002) formalize this argument in a short-run demand-side model. When the price elasticity of export demand is relatively large, a reduction in the gender wage gap will hurt export-led growth. Rees and Riezman (2012) model the effect of globalization on economic growth, through the impact of foreign direct investment on gender equality. Men and women dif- fer in their preferences for child quality, which are relatively higher for women. Women are also assumed to bear the full burden of childrearing. Husbands and wives bargain cooperatively, with the bargaining power of each spouse being a positive function of his or her wage rate. Globalization creates job opportunities in a high productivity sector (factory) for either men or women, who otherwise work in a low productivity sector (farm). If globalization creates job opportunities for women, their bargaining power increases and households trade off child quantity by child quality. Fertility falls, human capital accumulates, and long-run per capita output is high. If, on the other hand, globalization creates jobs for men, their intra-household power increases; fertility increases, human capital decreases, and steady-state income per capita is low. Thus, whether global capital flows generate jobs primarily in female or male intensive sectors matters for long-run growth. Women in politics The extent to which women participate in representative politics can affect economic growth through three different channels: the provision of public goods, role model effects, and politician quality. 1 GENDER INEQUALITY AS A BARRIER TO ECONOMIC GROWTH 30 largest threat in any case. When they resign, they are, on average, replaced by more competent leaders, who then select more competent candidates. Average politician quality goes up.18 The mechanisms of Besley et al. (2017) rest of the assumption of a democratic electoral process with (1) voters deriving utility only from politician competence and gender equality in candidate representation, (2) party leaders facing internal competition from other party candidates (they are all competing for ego rents), (3) more competent candi- dates posing a greater survival threat to a party leader, and (4) a higher share of women candidates posing a greater survival threat to a male leader. On the other hand, in less democratic electoral processes, in places where gender equality is less valued by society, or when party structures are highly centralized and not really open to internal dispute—in all these contexts—, the mechanisms of the model would break down. Women and corruption Women’s underrepresentation in leadership positions, either in politics or business, may also affect economic performance via its effect on corruption. There is suggestive evidence that women engage in less corruption than men (e.g., Dollar et al., 2001; Swamy et al., 2001; Beaman et al., 2009; Brollo and Troiano, 2016).19 Swamy et al. (2001) review several hypotheses explaining this gender difference in corruption with factors that can be expected to persist over time. Women may avoid corruption because they are more risk-averse than men, or because they are more honest—since honesty is a trait they want to pass on to their children (for whose rearing they are mainly responsible)—, or because they put a greater preference on obeying the law—since the law disproportionately benefits the physically weak. But an alternative set of explanations involves factors that result from women’s histor- ical underrepresentation in positions of power. Women may have fewer opportunities to engage in corruption (Goetz, 2007). For example, they may be excluded from cor- ruption networks or have less knowledge on how these operate. These differences are likely to erode, as female representation increases over time, and powerful women become exposed to (and familiar with) corruption practices. Thus, the underlying cause for the gender-differential in corruption will determine whether increasing women’s representation will reduce corruption in the short or in the long run. 18In the model, resignation is voluntary; it comes from the leader trading off ego rents with the utility from the party winning the election. But, in addition, there could be “social pressure” on low quality leaders after the introduction of a quota. Such pressure would reinforce the model’s conclusions. 19There is a broader debate in economics about the effect of corruption on economic growth. The controversy is on whether corruption “greases” or “sands” the wheels of economic growth. See, among others, Shleifer and Vishny (1993); Bardhan (1997); Méon and Sekkat (2005). 31 In summary, if indeed women engage in less corruption than men, it is important to know the underlying cause for this gender difference. If the difference is driven by evolutionary or socialization forces, then more women in leadership positions will likely reduce corruption in the long run. If, on the other hand, the difference stems from a history of underrepresentation, the positive impact of more female leaders is likely to be short-lived. 1.7 Conclusion In this chapter, we surveyed the theoretical literature linking gender inequality to economic development. This literature offers many plausible mechanisms through which inequality between men and women affects the aggregate economy. Yet, we believe the body of theories could be expanded in several directions. We discuss them below and finish by suggesting ways in which the dialogue between theory and empirics on this topic can be improved. The first direction for future research concerns control over fertility. In models where fertility is endogenous, households are always able to achieve their preferred number of children (see Strulik, 2018, for an exception). The implicit assumption is that there is a free and infallible method of fertility control available for all households—a view rejected by most demographers. The gap between desired fertility and achieved fertility can be endogeneized at two levels. First, at the societal level, the diffusion of particular contraceptive methods may be influenced by cultural and religious norms. Second, at the household level, fertility control may be object of non-cooperative bargaining between the spouses, in particular, for contraceptive methods that only women perfectly observe (Ashraf et al., 2014; Doepke and Kindermann, 2016). More generally, the role of asymmetric information within the household is not yet explored. A second direction worth exploring concerns gender inequality in a historical per- spective. In models with multiple equilibria, an economy’s path is often determined by its initial level of gender equality. Therefore, it would be useful to develop theories explaining why initial conditions varied across societies. In particular, there is a large literature on economic and demographic history documenting how systems of marriage and household formation differed substantially across preindustrial societies (e.g., Haj- nal, 1965, 1982; Hartman, 2004; Ruggles, 2009; De Moor and Van Zanden, 2010). In our view, more theoretical work is needed to explain both the origins and the consequences of these historical systems. A third avenue for future research concerns the role of technological change. In 1 GENDER INEQUALITY AS A BARRIER TO ECONOMIC GROWTH 32 several models, technological change is the exogenous force that ultimately erodes gender gaps in education or labor supply (e.g., Galor and Weil, 1996; Doepke and Tertilt, 2009; Bloom et al., 2015). For that to happen, technological progress is assumed to be skill-biased, thus raising the returns to education—or, in other words, favoring brain over brawn. As such, new technologies make male advantage in physical strength ever more irrelevant, while making female time spent on childrearing and housework ever more expensive. Moreover, recent technological progress increased the efficiency of domestic activities, thereby relaxing women’s time constraints (e.g., Greenwood et al., 2005; Cavalcanti and Tavares, 2008). These mechanisms are plausible, but other aspects of technological change need not be equally favorable for women. In many countries, for example, the booming science, technology, and engineering sectors tend to be particularly male-intensive. Even if current technological progress is assumed to weaken gender gaps, historically, technology may have played exactly the opposite role. If technology today is more complementary to brain, in the past it could have been more complementary to brawn. An example is the plow that, relative to alternative technologies for field preparation (e.g., hoe, digging stick), requires upper body strength, on which men have a comparative advantage over women (Boserup, 1970; Alesina et al., 2013). Another, even more striking example, is the invention of agriculture itself—the Neolithic Revolution. The transition from a hunter-gatherer lifestyle to sedentary agriculture involved a relative loss of status for women (Dyble et al., 2015; Hansen et al., 2015). One explanation is that property rights on land were captured by men, who had an advantage on physical strength and, consequently, on physical violence. Thus, in the long view of human history, technological change appears to have shifted from being male-biased towards being female-biased. Endogeneizing technological progress and its interaction with gender inequality is a promising avenue for future research. A final point concerns the role of men in this literature. In most models, gender inequality is not the result of an active male project that seeks the domination of women. Instead, inequality emerges as a rational best response to some underlying gender gap in endowments or constraints. Then, as the underlying gap becomes less relevant—for example, due to skill-biased technological change—, men passively relinquish their power (see Doepke and Tertilt, 2009, for an exception). There is never a male backlash against the short-term power loss that necessarily comes with female empowerment. In reality, it is more likely that men actively oppose losing power and resources towards women (Kabeer, 2016). This possibility has not yet been explored in formal models, even though it could threaten the typical virtuous cycle between gender equality and growth. 2 What Drives Female Labor Force Participation? Comparable Micro-level Evidence from Eight Developing and Emerging Economies∗ Abstract: We investigate the micro-level determinants of labor force participation of urban married women in eight low- and middle-income economies: Bolivia, Brazil, India, Indonesia, Jordan, South Africa, Tanzania, and Vietnam. In order to understand what drives changes and differences in participation rates since the early 2000s, we build a unified empirical framework that allows for comparative analyses across time and space. We find that the coefficients of women’s characteristics differ substantially across countries, and this explains most of the between-country differences in participation rates. In particular, the relationship between a woman’s education and her participation in the labor force varies from being positive and linear (Brazil and South Africa) to being U- or J-shaped (India, Jordan, and Indonesia), or a mixture of both (Bolivia, Vietnam, and Tanzania). Overall, the economic, social, and institutional constraints that shape women’s labor force participation remain largely country-specific. Nonetheless, rising education levels and declining fertility consistently increased participation rates, while rising household incomes contributed negatively in relatively poorer countries, suggesting that a substantial share of women work out of economic necessity. ∗This chapter is co-authored with Stephan Klasen, Janneke Pieters, and Le Thi Ngoc Tu. We are grateful to Esther Heesemann, Lisa Höckel, Bruno Witzel-Souza, and participants at the 26th IAFFE Conference and seminars of the Universities of Goettingen, Hannover, and Stellenbosch for comments and suggestions. For assistance with the Tanzanian data, we thank Novati Buberwa, James Mbongo, and Titus Mwisomba from Tanzania’s National Bureau of Statistics. For assistance with the Indonesian data, we thank Krisztina Kis- Katos, Christoph Kubitza, and Robert Sparrow. Friederike Schilling provided excellent research assistance. We gratefully acknowledge funding from the Growth and Economic Opportunities for Women (GrOW) initiative, a multi-funder partnership between the UK’s Department for International Development, the Hewlett Foundation, and the International Development Research Centre. The views expressed herein do not necessarily represent those of IDRC or its Board of Governors. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 36 2.1 Introduction Worldwide, the current labor force participation rate for women (age 15+) stands at 49 percent, compared to a participation rate of 76 percent for men (ILO, 2017). In the developing world, recent progress in closing this gender gap has been disappointing. In the past two decades, female labor force participation (FLFP, henceforth) rates have increased only modestly, on average, though there is considerable heterogeneity across countries and regions. Female participation rates are lowest in the Middle East and North Africa and in South Asia. South Asia also performed worst in terms of trends, with a declining share of women in the labor force. In contrast, female participation rates increased substantially in Latin America and the Caribbean. The heterogeneity in female participation rates is observed against a background of rising female education, declining fertility, and robust economic growth in almost all developing countries. Women in developing countries have been accumulating skills at an unprecedented pace, while declining fertility reduced the burdens of childrearing and domestic work. Combined with economic growth, one would expect more educated and less time constrained women to enter an expanding labor market. Even if long held gender norms on women working outside the home fail to adjust as quickly, rising opportunity costs in foregone earnings should eventually boost women’s participation rates. But this expectation did not materialize everywhere. In this chapter, we use comparable microdata from eight low and middle-income economies—Bolivia, Brazil, India, Indonesia, Jordan, South Africa, Tanzania, and Vietnam—to analyze how women’s individual and household characteristics are associ- ated with FLFP and what are the key commonalities and differences across countries. The period covered is 2000–2014. We further ask which factors drive FLFP changes over time within countries, and which factors account for differences in FLFP rates between countries. A large literature studies FLFP in the developing world (see Klasen, 2018b, for a review). At the macro level, the feminization-U hypothesis posits that, at low income levels, FLFP declines with economic development but, at some point, as countries get richer, the relationship turns positive (Boserup, 1970; Goldin, 1990, 1995). However, Gad- dis and Klasen (2014) find only weak empirical evidence in support of this hypothesis in a large panel of countries. Instead, countries’ idiosyncratic factors explain most of the worldwide variation in FLFP. Similarly, there is no evidence, at the country level, that closing the gender gap in education reduces the gender gap in labor force participation 37 (Ganguli et al., 2014).1 Studying 101 countries over a long period of time, Aaronson et al. (2017) find large negative effects of fertility on mothers’ labor supply, but only for sufficiently rich countries. At low levels of income, however, the effect of fertility is either small or zero.2 In work closely related to ours, Gasparini and Marchionni (2015) analyze microdata from 18 Latin American countries to investigate changes in FLFP between 1992 and 2012. They conclude that increased education, reduced marriage and fertility, and structural change towards more female-intensive activities contributed significantly to rising female participation throughout this period. However, these factors cannot account for the slowdown in the growth of female labor supply since the 2000s, which the authors link to the decade’s strong economic growth. By improving overall conditions, economic growth “may have reduced the urgency of vulnerable women [rural, low educated, with children and low-earnings spouses] to take low quality jobs” (Gasparini and Marchionni, 2015, p. 13). Several other papers investigate recent trends in FLFP for single countries. Assaad et al. (2014) offer a demand-side explanation for stagnating female participation rates in Jordan since 2000. As public sector hiring tightened since the adjustment policies of the 1980s, so have women’s labor market opportunities; the reason being that women are disproportionately employed in education and health activities.3 In Vietnam, very high FLFP is typically explained by the country’s socialist legacy4, and, to a smaller extent, by excess male mortality during the Vietnam War (Kreibaum and Klasen, 2015). For South Africa, Ntuli and Wittenberg (2013) decompose the increase in the participation rate of black women from 1995 to 2004. They find that changing returns to women’s labor market characteristics account for most of the FLFP increase. Klasen and Pieters (2015) ask why FLFP stagnated in India since the late 1980s and show that rising incomes and 1Ganguli et al. (2014) analyze census data from 40 countries. At the micro level, the authors show that if the education gender gap, the marriage gap (LFP gap between married and single women), and the motherhood gap (LFP gap between mothers and childless women) were to close everywhere, a large unexplained gender gap in participation rates would still remain for most countries. However, Ganguli et al. (2014) assume that education and FLFP are linearly related. As we will show in this chapter, the shape of the education-participation relationship is nonlinear in some countries. 2Aaronson et al. (2017) instrument fertility with twin birth (Rosenzweig and Wolpin, 1980) and sibling sex composition (Angrist and Evans, 1998). Using infertility shocks as a different source of exogenous variation for 26 developing countries, Agüero and Marks (2011) find no effect of fertility on mothers’ labor force participation. Priebe (2010) argues that, in poor settings, child costs push women into the labor market; as fertility declines, this type of distress-driven FLFP falls. The author shows causal evidence of this mechanism operating in Indonesia. 3In the Jordanian context, jobs in public education and health are among the few deemed socially appropriate for married women. 4See Ganguli et al. (2014, p. 184) and Klasen (2018b, p. 15) for further evidence. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 40 In the final part of our analysis, we decompose FLFP differences between countries. We find that differences in covariates cannot explain gaps in participation rates between countries. Instead, the returns to covariates and unobservables account for the bulk of FLFP variation, both around 2000 and 2014. Thus, economic, social, and institutional constraints that shape women’s labor force participation are still largely country-specific. This chapter proceeds as follows. Section 2.2 presents the data, descriptive statistics, and the empirical model. Section 2.3 shows the estimation results. In section 2.4, we decompose labor force participation changes over time and between countries. Section 2.5 concludes. 2.2 Data and empirical model In section 2.2.2, we describe the data sources and present descriptive statistics for the main variables of interest. The empirical model is discussed in section 2.2.3. 2.2.1 Data We select eight non-OECD countries with available good-quality large-scale house- hold surveys allowing us to derive (most of) the variables used in Klasen and Pieters (2015). We purposefully choose a diverse group of countries: two upper middle income countries—Brazil and South Africa—, five lower middle income countries—Bolivia, India, Indonesia, Jordan, and Vietnam—, and one low income country—Tanzania. These countries cover a wide range of geographies, per capita incomes, FLFP rates, economic structures, and urbanization rates (Figure 2.1).8 When compared to the world, India and Jordan have less gender equality and lower FLFP than predicted by their income levels (Figure 2.2). In contrast, Tanzania and Vietnam have more gender equality and higher FLFP than predicted by income alone. For the remaining countries (Bolivia, Brazil, Indonesia, and South Africa), their relative position in the world income distribution predicts well the levels of gender equality and FLFP. The data cover roughly the past one-and-a-half decades, from the early 2000s to the mid-2010s, with the exception of Jordan, whose available time-span is shorter: 2006– 2014. For South Africa, we also include 1995 in some of our analyses, to cover the entire post-apartheid era. 8Figure 2.1 shows data for the first and last year available for each country in our dataset. We obtain similar patterns if we plot data in 2000 and 2014 for all countries. 41 0 2. 5 5 7. 5 10 12 .5 15 G D P p er c ap ita ( in th ou sa nd s P P P − $) Tan za nia In dia Viet na m Boli via In do ne sia Jo rd an Sou th A fri ca Bra zil 20 00 20 14 19 99 20 11 20 02 20 14 20 00 20 14 20 00 20 14 20 06 20 14 19 95 20 14 20 02 20 13 (a) Income per capita 0 20 40 60 80 10 0 IL O m od el ed e st im at es , a ge 2 5− 54 ( % ) Tan za nia In dia Viet na m Boli via In do ne sia Jo rd an Sou th A fri ca Bra zil 20 00 20 14 19 99 20 11 20 02 20 14 20 00 20 14 20 00 20 14 20 06 20 14 19 95 20 14 20 02 20 13 (b) Female labor force participation 0 20 40 60 80 100 Brazil South Africa Jordan Indonesia Bolivia Vietnam India Tanzania 2013 2002 2014 1995 2014 2006 2014 2000 2014 2000 2014 2002 2011 1999 2014 2000 Agriculture (ISIC: 1−5) Manufacturing (ISIC: 15−37) Services (ISIC: 50−99) Other (c) Value added (% of GDP) by sector 0 20 40 60 80 10 0 U rb an p op ul at io n (% ) Tan za nia In dia Viet na m Boli via In do ne sia Jo rd an Sou th A fri ca Bra zil 20 00 20 14 19 99 20 11 20 02 20 14 20 00 20 14 20 00 20 14 20 06 20 14 19 95 20 14 20 02 20 13 (d) Urbanization rate FIG. 2.1: Selected country indicators for the first and last years in our dataset Notes: Sources are ILOSTAT and World Development Indicators. Countries are sorted by income per capita. At the macroeconomic level, 2000–2014 was a period of sustained economic growth. For the eight countries, GDP per capita grew, on average, 3.2 percent per year. India and Vietnam were the best performers, with average annual growth rates of 5.3 and 5.1 percent. South Africa grew the slowest: 1.6 percent per year.9 In general, our survey- years are representative of this macroeconomic period. Of 32 surveys, only two took place during recessions: Brazil, 2009, and Jordan, 2010. We only consider urban households for two reasons.10 First, the analysis requires individual earnings which are difficult to measure in rural areas, given the importance of 9Note that South Africa’s GDP per capita grew much faster between 1995 and 2014, which is the period shown in Figure 2.1. 10For Jordan we consider both urban and rural areas because information on urban status is not available from the 2008 and 2014 surveys. In any case, more than 80 percent of Jordan’s population lives in urban areas, in the period considered, according to data from the World Bank’s World Development Indicators. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 42 Bolivia Brazil India Indonesia Jordan South Africa Tanzania Vietnam .6 .7 .8 .9 1 G en de r D ev el op m en t I nd ex 6 8 10 12 ln(GDP per capita) (a) Gender equality Bolivia Indonesia India Jordan Tanzania Vietnam South Africa Brazil 0 20 40 60 80 10 0 F LF P ( % ) 6 8 10 12 ln(GDP per capita) (b) Female labor force participation FIG. 2.2: Selected gender indicators with respect to per capita income in 2014 Notes: Sources are UNDP’s Human Development Report 2016 [Panel (a)] and ILOSTAT [Panel (b)]. Panel (a)—Linear fit of the two variables shown. 154 countries included. Panel (b)—Quadratic fit of the two variables shown. 174 countries included. FLFP, for ages 25–54, is a ILO modeled estimate. GDP per capita is PPP-adjusted at 2011 international $. smallholder agriculture. Non-marketed agricultural output must be valued in monetary terms, but the necessary detailed price data is often unavailable.11 Moreover, whenever several household members farm the same plots, or agricultural income is aggregated at the household level, it is unavoidable to impute income for each individual. In urban areas, measurement error or missing data are less severe. Second, in settings dominated by agriculture, where many women contribute to household farming, household surveys are more likely to underreport female work. The extent of underreporting likely depends on survey methodology, which varies across countries (and sometimes within countries over time).12 Focusing on urban areas, therefore, improves the comparability of labor force measurements across space and time. The dataset includes nearly 800,000 urban married women of age 25–54. Table 2.A.1 lists the surveys, years, and sample sizes. In Appendix 2.A, we describe each data source in detail, explaining how variables were constructed and harmonized across surveys. 11For example, in its 2000 and 2006 rounds, Tanzania’s Integrated Labor Force Survey only recorded agricultural income in urban areas. Other well-known practical complications are unmeasured product variety and quality. 12For example, in South Africa, the Labor Force Survey (LFS 2001–2007) is better at capturing informal casual employment than the previous October Household Survey (OHS 1995–1999) (Yu, 2007). The number of employment categories in the survey questionnaire increased from three in the OHS to eight in the LFS. 45 0.15 0.08 0.39 0.25 0.14 0.15 0.07 0.36 0.25 0.17 0.14 0.06 0.35 0.29 0.16 0.09 0.05 0.36 0.30 0.20 0.07 0.04 0.34 0.34 0.21 0 .2 .4 .6 .8 1 1995 2001 2003 2010 2014 Less than primary Primary Less than secondary Secondary Tertiary (a) South Africa 0.13 0.33 0.18 0.22 0.14 0.11 0.30 0.18 0.25 0.16 0.08 0.25 0.17 0.31 0.20 0.05 0.20 0.16 0.35 0.24 0 .2 .4 .6 .8 1 2002 2005 2009 2013 Less than primary Elementary (1−4) Elementary (5−8) High school Tertiary (b) Brazil 0.14 0.14 0.22 0.09 0.16 0.24 0.10 0.11 0.22 0.12 0.17 0.28 0.09 0.12 0.19 0.14 0.16 0.30 0.07 0.12 0.20 0.18 0.14 0.29 0 .2 .4 .6 .8 1 2006 2008 2010 2014 Less than primary Primary Preparatory Lower secondary Secondary Tertiary (c) Jordan 0.32 0.08 0.12 0.14 0.21 0.13 0.28 0.07 0.13 0.16 0.22 0.14 0.22 0.07 0.11 0.15 0.27 0.18 0 .2 .4 .6 .8 1 1999 2004 2011 Illiterate Literate Primary Middle school Secondary Tertiary (d) India 0.25 0.19 0.18 0.16 0.22 0.27 0.18 0.17 0.18 0.20 0.20 0.16 0.18 0.23 0.24 0.18 0.15 0.16 0.19 0.31 0.17 0.15 0.16 0.24 0.28 0 .2 .4 .6 .8 1 2000 2005 2008 2011 2014 Less than basic Basic Intermediate Secondary Tertiary (e) Bolivia 0.19 0.31 0.17 0.26 0.07 0.16 0.29 0.20 0.28 0.08 0.13 0.28 0.19 0.29 0.12 0.11 0.24 0.20 0.32 0.14 0 .2 .4 .6 .8 1 2000 2004 2007 2014 Less than primary Primary Junior high school Senior high school Tertiary (f) Indonesia 0.13 0.22 0.25 0.31 0.10 0.12 0.22 0.22 0.35 0.09 0.09 0.22 0.22 0.29 0.19 0.09 0.18 0.22 0.27 0.24 0 .2 .4 .6 .8 1 2002 2006 2010 2014 Less than primary Primary Secondary High school Tertiary (g) Vietnam 0.14 0.14 0.61 0.11 0.00 0.11 0.08 0.67 0.13 0.01 0.07 0.06 0.64 0.17 0.05 0 .2 .4 .6 .8 1 2000 2006 2014 No schooling Less than primary Primary Any secondary Tertiary (h) Tanzania FIG. 2.4: Distribution of educational attainment over time Notes: See Table 2.A.1 for sources. Urban married women, age 25–54; except urban and rural in Jordan. Common Y-axis for all subfigures. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 46 striking differences across countries. The average number of children in a married woman’s household reflects distinct fertility and co-residence patterns across countries (Figure 2.6). Jordan and Tanzania show the highest number of children, both ages 0–4 and 5–14, per household; Brazil and Vietnam have the lowest.14 Overall, most countries experienced a decline in the number of children per household over time. In all countries, working married women are concentrated in a few industries. Most highly educated women work in white-collar services, in particular, public administra- tion, education, and health; the majority of less educated women work in other services, in particular, wholesale and retail trade (Figure 2.A.1). In urban Tanzania, agriculture remains the most important activity for less educated women. Construction and mining employ very few married women in all countries. Based on these descriptive statistics, we can draw several hypotheses. The different patterns we observe in terms of the education-participation relationship imply that rising education levels will have very different impacts on women’s participation rates across countries. In some countries, particularly those with a strong U-shaped relationship, the impact may be limited or even negative. On the other hand, declining fertility is likely to contribute to higher participation rates everywhere, though this depends on the extent to which the presence of children is a barrier to women’s participation in the different countries and how this changed over time. The distribution of female workers across industries suggests that changes in the sectoral structure of employment could have important bearings on women’s likelihood of entering the labor force. While the descriptive patterns are quite similar across countries, the structure of growth may differ and could potentially explain differences in trends in participation rates. Finally, aggregate income growth has two potentially counteracting impacts: rising unearned income and rising earnings. As discussed below, we do not analyze the effect of women’s own expected earnings, which will to some extent be captured by the effects of education. Increases in unearned income are likely to have a negative impact on participation rates in all countries, and here our interest mainly lies in the extent of this force. 14The figure for Jordan is inflated by including rural areas. 47 0 .2 .4 .6 .8 1 LF P Less than primary Primary Less than secondary Secondary Tertiary 1995 2001 2003 2010 2014 (a) South Africa 0 .2 .4 .6 .8 1 LF P Less than primary Elementary (1−4) Elementary (5−8) High school Tertiary 2002 2005 2009 2013 (b) Brazil 0 .2 .4 .6 .8 1 LF P Less than primary Primary Preparatory Lower secondary Secondary Tertiary 2006 2008 2010 2014 (c) Jordan 0 .2 .4 .6 .8 1 LF P Illiterate Literate Primary Middle school Secondary Tertiary 1999 2004 2011 (d) India 0 .2 .4 .6 .8 1 LF P Less than basic Basic Intermediate Secondary Tertiary 2000 2005 2008 2011 2014 (e) Bolivia 0 .2 .4 .6 .8 1 LF P Less than primary Primary Junior high school Senior high school Tertiary 2000 2004 2007 2014 (f) Indonesia 0 .2 .4 .6 .8 1 LF P Less than primary Primary Secondary High school Tertiary 2002 2006 2010 2014 (g) Vietnam 0 .2 .4 .6 .8 1 LF P No schooling Less than primary Primary Any secondary Tertiary 2000 2006 2014 (h) Tanzania FIG. 2.5: FLFP by education level Notes: See Table 2.A.1 for sources. Urban married women, age 25–54; except urban and rural in Jordan. Common Y-axis for all subfigures. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 50 standard errors at the regional level.20 In an alternative specification, we analyze whether FLFP is associated with the sector in which jobs are available locally, as do Klasen and Pieters (2015) for urban India. As a result of norms about the types of work appropriate for women, discriminatory practices, and the extent to which hours and location of work are flexible within a particular occupation, employment opportunities for women may depend especially on employment growth in particular sectors. To capture the structure of local labor demand, we replace the regional fixed effects with the sectoral composition of male employment at the regional level (construction, agriculture, mining, manufacturing, white-collar services, and other services).21 However, we find no clear relationship between these sectoral variables and FLFP. For this reason, we only present results for the specification with regional fixed effects. Our estimates are best interpreted as reduced-form correlations. In this setting, endo- geneity mainly stems from omitted variable bias, due to the individual or household unobservables jointly determining labor force participation, education, fertility, marital matching, and location (urban-rural). We explicitly address some of these concerns in sections 2.3.1 and 2.3.2, where we assess the importance of selection bias related to marriage, settlement in urban areas, and educational attainment. Reverse causality, on the other hand, is less of a concern. We assume that prime-age women completed their education and marriage market histories. Moreover, we assume that each woman takes the labor market status of her spouse as exogenous, since in all countries and years of our sample prime-age married men have nearly universal labor force participation rates. We do not attempt to causally identify structural parameters for two reasons. First, there is no quasi-experimental strategy (such as an instrumental variables approach) applicable to all countries and years similarly.22 Second, the prevailing methods for estimating own-wage effects are notoriously challenging and known to produce unstable results.23 In addition, the quality of existing wage data varies substantially across Africa after 1995, Jordan, and Tanzania). 20For more details on the construction of these variables across countries see Appendix 2.A; for sample means of the variables by country and year see Tables 2.A.2-2.A.9. 21The Indian and Indonesian surveys are representative at the second highest subnational level; this is the level of aggregation used for the regional employment share variables. For the remaining countries, we use the highest subnational administrative level to aggregate the employment shares. 22In principle, one could pursue a country and year-specific IV approach, but the resulting local average treatment effects would be hard to interpret in a unified comparative framework, as the population of compliers would vary across settings and IVs. 23 See Klasen and Pieters (2015, pp. 460–461) for a discussion of the lack of robustness in estimates of own-wage effects in India, as well as a more detailed discussion of the challenges involved in such estimations. 51 surveys. 2.3 Results We first summarize the estimation results for each country (in increasing order of GDP per capita), and then turn to a discussion of the main trends and patterns. Table 2.1 provides an overview of the relationship between key variables (or variable groups) and women’s labor force participation in each country, and their changes over time.24 In Tanzania, FLFP increases linearly with education attainment in 2000 and 2014. The effect of household income is negative but small, and declines in absolute magnitude over time. The number of children aged 0–4 only has a significant (and small) negative effect in 2014. Otherwise, the number of children in the household does not correlate with FLFP. Besides a tiny negative effect of household income, none of the explanatory variables is statistically significant in 2006, which likely reflects the lack of variation in the dependent variable: the participation rate in the estimation sample is 92 percent. India shows a clear U-shaped relationship between own education and FLFP. Relative to the reference group of illiterate women, the average marginal effects are negative and larger in magnitude with each additional level of educational attainment up to completed middle schooling—which is the level associated with the lowest participation rates in all years. The average marginal effect is still negative for complete secondary schooling, but closer to zero. For women with any tertiary education, the positive marginal effect is very large and significant, although declining over time: from 21 percentage points in 1999 to 14 percentage points in 2011. Household head education, household income, and male salaried employment (to proxy security of income) correlate negatively with participation—although the latter effect is no longer significant in 2011. The presence of young children is correlated with lower participation, and this negative effect is becoming stronger over time. For older children, the average marginal effect is actually positive after 1999, but always small. Finally, caste and religion are important correlates of FLFP as well, with lower caste and Hindu women being more active in the labor market than upper caste and Muslim women. The effect of caste is weakening over time; the effect of religion is strengthening. In Vietnam, the relationship between education attainment and FLFP is positive and linear in 2002, but only the effect of tertiary education remains over time. The small negative income effect in 2002 becomes insignificant in the later years. The number 24The average marginal effect estimates for the probit models are reported in Tables 2.A.10-2.A.17. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 52 of young children is negatively associated with FLFP after 2002; the effect is large (in absolute terms) and increases over time: in 2014, one additional young child is associated with a 6 percentage points reduction in women’s participation probability. We do not find clear associations between FLFP and older children, male salaried employment, household head education, or ethnicity. In Bolivia, education is not significantly correlated with FLFP, except for tertiary schooling, which affects participation positively. The estimate fluctuates a bit between 2000 and 2008, after which it increases until 2014. Household income and salaried employment of a male household member reduce FLFP. The effects are substantial, when compared to estimates from other countries. The presence of at least one male salaried employee in the household correlates with a 4 to 10 percentage point decline in the woman’s participation likelihood, depending on the year. Young children have a sizable negative effect. Household head education was negatively related to FLFP in 2008, 2011, and 2014, with the effect getting weaker over time. Native speakers of indigenous languages are more likely to participate in the labor market. In Indonesia, the relationship between own education and FLFP in the first year (2000) resembles the U-shape found for India, with negative effects of primary and junior high school completion (relative to the reference group of women who did not complete primary school), and positive effects of completed secondary schooling and especially tertiary education. Yet, the pattern changes: in 2014, only the positive tertiary education effect remains, and it is somewhat smaller than in 2000. Household income has a sizable negative effect on participation, and this becomes stronger over time. The estimates of male salaried employment are, likewise, negative and increasing (in absolute terms), while the negative effect of household head education decreases. There is a large negative effect of young children and a smaller negative effect of older children. Both are increasing over the years, in absolute terms. In Jordan, tertiary education has a strong positive relationship with FLFP, and the effect is very stable over time. Across lower education levels the relationship with FLFP is flat, except for a small negative effect of lower secondary education, resulting in a J-shaped education-participation relationship. Income has a small but significant negative effect in every year.25 Male salaried employment increases FLFP in the most recent years (2010 and 2014), while the positive effect of a tertiary educated household head disappears after 2010, both suggesting that income security is less relevant. We 25The small size of the income effect should be interpreted with caution. The earnings variable available from the Jordanian surveys is very roughly measured: it is the mid-point of five earning brackets. We thus suspect the average marginal effects of household income to suffer from attenuation bias. 55 Pieters (2015) relate this decline to changes in the selectivity of higher education, an issue we address in section 2.3.2. The patterns suggest that the education-participation relationship moves from weak linear in low-income countries to a U- or J-shape in middle-income countries, before becoming strongly positive in upper-middle income countries. To some extent, this is also the pattern we observe over time within Bolivia and Vietnam. Our results thus illustrate that countries growing from low-income to lower-middle-income status will not necessarily experience an increase in the participation returns to education, and therefore increases in educational attainment levels may have ambiguous effects on FLFP rates. Furthermore, India, Indonesia, and Jordan are not only at the middle of the GDP per capita distribution in this sample of countries, but also form a more or less distinct group in terms of social and religious norms around women’s participation in market activities. It is likely that the U-shape or J-shape at least partly reflects such norms, by which employment outside the home is not deemed appropriate for women at intermediate levels of education. In India and Indonesia, this is further corroborated by a negative relationship between household head education and FLFP, indicating that when the household’s socio-economic status improves, women withdraw from the labor force. Household income is negatively related to women’s participation everywhere, but interestingly the negative effects disappeared in South Africa and Brazil by 2013–14 (Figure 2.8). In these two countries, male salaried employment and household head education have no clear relationship with FLFP either. Hence, in the richest two countries in our sample, income and income uncertainty seem to play no role in ‘pushing’ women to participate in the labor force, whereas their own education is a major factor.26 Thus, women’s own characteristics matter the most for labor force participation; household conditions, except the number of young children, have become irrelevant. The role of children is also noteworthy (Figure 2.9). While women in households with young children are less likely to participate in the labor force in all countries and time periods, older children reduce FLFP only in the relatively high-income countries. In poorer countries we find no evidence for such a relationship, which may reflect income constraints, whereby mothers cannot afford to stay out of the labor force for long in poorer settings.27 26This finding resembles patterns that are taking place in OECD countries in the past decades. In the US, for example, Blau and Kahn (2007) and Heim (2007) show that income elasticities of married women labor supply have plummeted since the 1980s. 27See Priebe (2010) for causal evidence of this mechanism in Indonesia. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 56 − .1 0 .1 .2 .3 .4 .5 A ve ra ge m ar gi na l e ffe ct < Prim Elem (1−4) Elem (5−8) High school Tertiary 2002 95% CI (2002) 2013 95% CI (2013) (a) Brazil − .1 0 .1 .2 .3 .4 .5 A ve ra ge m ar gi na l e ffe ct < Prim Primary < Sec Secondary Tertiary 1995 95% CI (1995) 2014 95% CI (2014) (b) South Africa − .1 0 .1 .2 .3 .4 .5 A ve ra ge m ar gi na l e ffe ct Illiterate Literate Primary Middle school Secondary Tertiary 1999 95% CI (1999) 2011 95% CI (2011) (c) India − .1 0 .1 .2 .3 .4 .5 A ve ra ge m ar gi na l e ffe ct < Prim Primary Junior sec Secondary Tertiary 2000 95% CI (2000) 2014 95% CI (2014) (d) Indonesia − .1 0 .1 .2 .3 .4 .5 A ve ra ge m ar gi na l e ffe ct < Prim Primary Preparatory Lower sec Secondary Tertiary 2006 95% CI (2006) 2014 95% CI (2014) (e) Jordan − .1 0 .1 .2 .3 .4 .5 A ve ra ge m ar gi na l e ffe ct < Basic Basic Intermediate Secondary Tertiary 2000 95% CI (2000) 2014 95% CI (2014) (f) Bolivia − .1 0 .1 .2 .3 .4 .5 A ve ra ge m ar gi na l e ffe ct < Prim Primary Secondary High school Tertiary 2002 95% CI (2002) 2014 95% CI (2014) (g) Vietnam − .1 0 .1 .2 .3 .4 .5 A ve ra ge m ar gi na l e ffe ct No education < Prim Primary Any secondary Tertiary 2000 95% CI (2000) 2014 95% CI (2014) (h) Tanzania FIG. 2.7: Average marginal effects of the woman’s own education Notes: Common Y-axis for all subfigures. Average marginal effects of the full probit model are reported, for each country and year, in Tables 2.A.10-2.A.17. 57 − .0 8 − .0 6 − .0 4 − .0 2 0 .0 2 .0 4 A ve ra ge m ar gi na l e ffe ct 2002 2005 2009 2013 Log income 95% CI (a) Brazil − .0 8 − .0 6 − .0 4 − .0 2 0 .0 2 .0 4 A ve ra ge m ar gi na l e ffe ct 1995 2001 2003 2010 2014 Log income 95% CI (b) South Africa − .0 8 − .0 6 − .0 4 − .0 2 0 .0 2 .0 4 A ve ra ge m ar gi na l e ffe ct 2006 2008 2010 2014 Log income 95% CI (c) Jordan − .0 8 − .0 6 − .0 4 − .0 2 0 .0 2 .0 4 A ve ra ge m ar gi na l e ffe ct 1999 2004 2011 Log income 95% CI (d) India − .0 8 − .0 6 − .0 4 − .0 2 0 .0 2 .0 4 A ve ra ge m ar gi na l e ffe ct 2000 2005 2008 2011 2014 Log income 95% CI (e) Bolivia − .0 8 − .0 6 − .0 4 − .0 2 0 .0 2 .0 4 A ve ra ge m ar gi na l e ffe ct 2000 2004 2007 2014 Log income 95% CI (f) Indonesia − .0 8 − .0 6 − .0 4 − .0 2 0 .0 2 .0 4 A ve ra ge m ar gi na l e ffe ct 2002 2006 2010 2014 Log income 95% CI (g) Vietnam − .0 8 − .0 6 − .0 4 − .0 2 0 .0 2 .0 4 A ve ra ge m ar gi na l e ffe ct 2000 2006 2014 Log income 95% CI (h) Tanzania FIG. 2.8: Average marginal effects of log household per capita earnings (excluding woman’s own earnings) Notes: Common Y-axis for all subfigures. Average marginal effects of the full probit model are reported, for each country and year, in Tables 2.A.10-2.A.17. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 60 find that women migrating more than five years before the survey are 5 percentage points more likely to be in the labor force in 2014; the effect being insignificant in the first two years. For migrants arriving less than five years before the survey, the effects are never significant. In Brazil, migration status (captured by individuals’ place of birth being in a different state or different municipality than their current residence) has no significant effects. In Bolivia, a woman’s migration status (a dummy variable for whether, five years before the survey, she lived outside the municipality of current residence), was associated with lower labor force participation only in the last two survey years (the average marginal effect is around minus 5 percentage points in both years).31 For all three countries, adding the migration controls does not affect the average marginal effects of the remaining explanatory variables.32 In sum, trends in the rates of marriage and urban residence among prime-age women do not influence the determinants of labor force participation. 2.3.2 Selection into education We now consider selection into education, not because of concerns about the robustness of our estimates, but rather out of interest in the forces driving changes in returns to education. Education levels have increased over time in all eight countries. Since, in our sample of prime-age married women, education histories are largely complete, average educational attainment increases because younger, more educated cohorts progressively replace older, less educated ones. This process raises the question of whether the selection of women into education levels varies across cohorts. If it does, trends in the estimated average marginal effects of educational attainment could be driven by changes in the sample’s cohort composition, rather than by changes in the marginal effects of education. We explore this possibility in more detail for India, Indonesia, and South Africa. The three countries experienced rising shares of highly educated women (tertiary level, see Figure 2.10) and, simultaneously, a sizable decrease in the (positive) average marginal effect of being highly educated (see Figure 2.7). We would like to know how much of the decline in the effect of tertiary education could be plausibly explained by decreasing selectivity of women in terms of labor force attachment at the top of the education distribution. Klasen and Pieters (2015) propose 31Full results available upon request. 32Klasen and Pieters (2015) show that, for India, the 1999 results are robust to adding migration variables (both the woman’s and her spouse’s), which are themselves insignificant. There is no migration data available for 2004 and 2011. 61 a thought experiment to estimate an upper bound on the size of the selection effect. Imagine that the initial distribution of women’s educational attainment is a one-to- one match to the distribution of unobserved labor force attachment. If there are K educational levels, there are also K attachment levels; the women achieving the highest level of education being also the ones with the highest level of labor force attachment. As a result, the average marginal effect of education on labor force participation is positively biased. Now, consider a completely supply-driven expansion of education: the government produces and offers cost-free slots of tertiary education. The new slots are filled by women below that educational level in decreasing order of labor force attachment. That is, less attached women are moving up the education ladder. As a consequence, average labor force attachment at the tertiary level is now lower than before, and the estimated effect of education on labor force participation falls. Consider two extreme scenarios of the thought experiment. If all women have the same labor force attachment (or education and labor force attachment are completely unrelated), the education expansion would have no selection effect; over time, any changes in the education estimates result from changes in the effect of education itself. If, on the other hand, the education effect is fully driven by labor force attachment, then the post-expansion education estimates are a weighted sum of the pre-expansion estimates, where the weights are the changes in the attachment composition of each education level. With the last scenario in mind, we can estimate an upper bound of the selection effect. Let us illustrate the procedure for South Africa. In 2014, the share of women with tertiary education was 0.21. Nearly two decades before, in 1995, that share was 0.14. Thus, in 1995, one third of the women in the top 21 percentiles of the education distribution had complete secondary schooling (see Figure 2.10). We can then estimate the average marginal effect of being in the top 21 percentiles of the education distribution in 1995 as two thirds the average marginal effect of tertiary education plus one third the average marginal effect of completed secondary schooling. If this reweighed 1995 estimate comes closer to the average marginal effect of tertiary education in 2014, then the effect of being in the 21 highest education percentiles (relative to the reference group with below primary schooling) did not change over time. What changed instead was the selectivity of women into educational attainment. For India and South Africa, we find that the reweighted estimates closely reproduce the average marginal effects of the latest year. In theory, the selection effect is large enough to explain the declining effect of high education in the two countries (Figure 2.10). For Indonesia, the reweighted estimate is about 30 percent smaller than the average 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 62 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 2014 1995 < Prim Primary < Sec Secondary Tertiary (a) South Africa .1 .2 .3 .4 .5 .6 .7 .8 .9 1 2014 2000 < Prim Primary Junior HS Senior HS Tertiary (b) Indonesia .1 .2 .3 .4 .5 .6 .7 .8 .9 1 2011 1999 Illiterate Literate Primary Middle school Sec school Tertiary (c) India 0.206 0.140 0.140 0.360 0.299 0.212 0.431 0.379 0.373 0 .1 .2 .3 .4 India Indonesia South Africa Estimated: India 1999, Indonesia 2000, SA 1995 Estimated: India 2011, Indonesia 2014, SA 2014 Reweighted: India 2011, Indonesia 2014, SA 2014 (d) Tertiary education: estimated and reweighed average marginal effects FIG. 2.10: Education selectivity Notes: Panels (a), (b), (c) show changes in the education distribution over time. Panel (d) shows the estimated average marginal effects from the probit models and from the reweighting procedure described in the text. marginal effect, implying that the selection effect can account for a stronger decline in returns to higher education than actually observed. This suggests that the participation returns to tertiary education may have in fact increased in Indonesia between 2000 and 2014, even though the estimated effect on labor force participation declined. Yet, since our reweighted estimate reflects the upper bound, it is also possible that the returns to education did not change, or declined. 65 − .1 − .0 5 0 .0 5 .1 .1 5 .2 Tanzania Bolivia India South Africa Vietnam Jordan Brazil Indonesia Labor force: difference Covariates: first year Coefficients & unobservables: first year Covariates: last year Coefficients & unobservables: last year (a) Total covariate contribution − .0 5 − .0 25 0 .0 25 .0 5 .0 75 Tanzania Bolivia India South Africa Vietnam Jordan Brazil Indonesia fir st ye ar las t y ea r fir st ye ar las t y ea r fir st ye ar las t y ea r fir st ye ar las t y ea r fir st ye ar las t y ea r fir st ye ar las t y ea r fir st ye ar las t y ea r fir st ye ar las t y ea r Own education Children Log income Hh head educ Male salaried emp. Age Pop group Region dummies Survey waves (b) Contribution of variable groups FIG. 2.11: Decompositions within countries over time Notes: For point estimates of the decompositions, see Tables 2.A.19-2.A.22. First year and last year refer to the year of the coefficients used to compute the covariate contribution. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 66 between 2006 and 2014. In Brazil, covariates come close to explaining the full FLFP increase between 2002 and 2013 (which was about 5 percentage points). Rising female education and, to a lesser extent, declining fertility are the main forces. In Tanzania, Bolivia, Vietnam, and Indonesia, changes in coefficients and unobserv- ables account for almost all of the change in participation rates.38 The changes in covariates in each of these four countries contributed little or nothing to changes in FLFP rates, whether they declined (in Tanzania and Bolivia) or increased (in Vietnam and, especially, Indonesia). A positive contribution of increasing women’s educational attainment (and declining fertility in Tanzania and Bolivia) was, in all cases, offset mainly by the negative contribution of rising household incomes. Finally, in South Africa, where the participation rate was nearly constant between 2001 and 2014, the positive covariate contribution is offset by a negative contribution of coefficients and unobservables. Similar to Brazil, the covariate contribution is large and is accounted for by rising female education levels and reduced numbers of children.39 Summarizing the main findings, rising educational attainment contributed to higher FLFP in all countries, but most strongly in Brazil and South Africa, reflecting the strong participation-returns to education in these countries. In the other countries, the con- tribution was more limited, but still positive, despite the U- or J-shaped relationship between education and participation in Jordan, India, and Indonesia. This reflects edu- cational attainment increasing predominantly at the highest levels of education, where the participation returns are positive. With the exception of Vietnam, falling fertility also contributed to higher participation rates in all countries. The effect was strongest in Brazil, South Africa, and Jordan. This is mainly because children are more strongly associated with lower participation in Brazil and South Africa; hence a decline in the number of children accounts for a larger increase in the observed participation rate. Rising household incomes contributed to a decline in participation in Tanzania, Bolivia, India, Vietnam, and Indonesia. India is the only country in our sample where rising household head education made a significant 38For Tanzania, results depend on the choice of counterfactual. Using the 2000 coefficients, the covariate effect accounts for 36 percent of the LFP reduction between 2000 and 2014. Increasing household incomes drive the negative covariate effect, being partly offset by the positive effect of rising female education. In 2014, the negative average marginal effect of income shrinks by two thirds relative to 2000. As a result, the total covariate contribution becomes positive when weighted at 2014 coefficients. 39In addition, we decompose the FLFP change in South Africa for the full post-apartheid period: 1995– 2014. Participation rates of urban married women rose substantially from 58.5 percent to 68.1 percent between 1995 and 2001. We find that women’s labor market characteristics account for around 70–74 percent of this increase (Table 2.A.23). Rising education, declining fertility, and a relative increase in the share of black women (in urban areas) were powerful drivers of participation in this period. 67 negative contribution to FLFP rates. Finally, we also find a relatively strong (negative or positive) contribution of changes in the returns to characteristics and unobserved factors in several countries. The direction and relative importance of this component vary widely across countries, and it does not appear to be related to countries’ income level or the observed level or change in FLFP rates. 2.4.2 Decomposing differences between countries We next decompose FLFP differences between countries, using Brazil as the reference country. The decomposition shows the extent to which gaps in participation rates between a particular country and Brazil emanate from differences between women’s observed characteristics versus differences in the returns to those characteristics (or other unobservables). We take Brazil as the counterfactual for two main reasons. First, having the highest per capita income in our sample, it constitutes a natural bench- mark. Second, having the second highest increase in FLFP—entirely accounted for by changes in covariates—it is of particular interest to assess to what extent other countries’ participation rates differ from Brazil’s due to differences in covariates. We run two sets of decompositions: first year, which uses covariates and coefficients from Brazil’s 2002 survey and the other country’s data from the survey year closest to 2002; and last year, which uses covariates and coefficients from Brazil’s 2013 survey and the other country’s data from the survey year closest to 2013. The exercise requires a few data adjustments. First, we recode the educational attain- ment of the woman and household head into four broader categories (less than primary, primary completed, secondary completed, and any tertiary) that are identical for all countries. Similarly, we recode the social group variable—reflecting ethnicity, religion, or nationality—into a dummy variable equal to 1 for the social groups with positive average marginal effects on participation within each country and 0 otherwise.40 To capture regional effects in a comparable way, we compute, for each country and period, the quartiles of the regional average marginal effects on participation. We then create a dummy variable for each quartile.41 Finally, we do not use survey weights for the decompositions. Countries use different sampling strategies in their surveys, and often follow distinct approaches in calculating sample weights. We want to avoid that such methodological differences drive the results. Overall, the unweighted difference in FLFP between each country and Brazil comes very close the weighted difference (Figure 40Table 2.A.24 shows how the education and social group variables are created for each country. 41Excluding the regional dummies altogether does not change the decomposition results in any mean- ingful way. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 70 0 .2 .4 .6 .8 Brazil Jordan India Indonesia Bolivia South Africa Tanzania Vietnam Real FLFP Simulated FLFP at Brazil’s coefficients (a) First year, around 2002 0 .2 .4 .6 .8 Brazil Jordan India Indonesia Bolivia South Africa Tanzania Vietnam Real FLFP Simulated FLFP at Brazil’s coefficients (b) Last year, around 2014 FIG. 2.13: Real FLFP vs. FLFP simulated at Brazil’s coefficients 71 to reality: the mean FLFP in the first year (last year) was 54 (56) percent; ranging from 11 (15) percent in Jordan to 86 (87) percent in Vietnam. The standard deviation was 27 (26) percent.44 In the fictional “Brazilian” market, Jordanian women would have a higher participation rate than women from Tanzania. In reality, in 2014, the participation rate for Jordan was a staggering 67 percentage points lower than in Tanzania. In brief: differences in the observed characteristics of women and their households cannot account for the wide variation in FLFP between countries. Instead, most of the between-country differences result from variation in the returns to those characteristics and other unobservable factors. 2.5 Conclusion Using comparable microdata from eight low and middle-income countries, this chapter sheds light on the impact and relative importance of what are considered key determi- nants of FLFP. We find that the participation-returns to women’s own characteristics and family circumstances—including education, income, and fertility—differ substantially across countries. In fact, heterogeneity in returns to these characteristics explains most of the between-country differences in participation rates, indicating that the economic, social, and institutional constraints that shape women’s labor force participation are still largely country-specific. Nonetheless, some important patterns appear. Overall, rising education levels and declining fertility consistently increase FLFP, although the strength of these two forces differs across countries. At the same time, rising household incomes have a negative effect in all but the three richest countries in our sample (Jordan, South Africa, and Brazil), indicating that, in poorer countries, a substantial share of women work out of economic necessity. In relatively poor countries with high initial participation rates (Vietnam, Tanzania, and, to a lesser extent, Bolivia), improving family circumstances (e.g., higher household incomes, or better educated household heads) have a moderate negative effect on women’s participation. In terms of women’s own characteristics, the positive participation- education gradient is flattening over time, except for relatively high participation returns occurring at the tertiary level. Future gains in female participation rates will depend on the extent to which women achieve educational attainment at the tertiary level. 44The FLFP rates in the paragraph are calculated without survey weights. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 72 In countries with low initial participation rates and strong social barriers to women’s outside-home employment (India, Jordan, and, to a lesser extent, Indonesia), family circumstances have a much stronger grip on women’s participation. Own education has a U- or J-shape relationship with participation, such that rising attainment at intermediate education levels actually depresses FLFP. Once again, expansion of women’s access to tertiary education would be required to raise FLFP further. As shown by the Indonesian experience, however, changing returns to women’s labor supply characteristics can boost participation rates dramatically. In the richest countries (Brazil and South Africa), where social barriers to women’s employment are relatively small, family circumstances other than fertility have become largely irrelevant. With a strong positive education-participation gradient and a strong negative effect of fertility, increases in women’s own education and falling fertility boosted participation in these countries. In the future, higher educational attainment and lower fertility will likely continue to translate into higher FLFP. Finally, we find suggestive evidence of reduced selectivity of tertiary education in India, Indonesia, and South Africa (similar to the results for India in Klasen and Pieters (2015)). This may mitigate the extent to which further educational advancement will translate into higher FLFP in these countries. While this chapter has focused on supply side factors, FLFP might be severely con- strained by demand factors. For example, the unexplained portion of the gender wage gap did not decline substantially in recent decades (Weichselbaumer and Winter-Ebmer, 2005; Oostendorp, 2009), and employment sectors and occupations remain highly seg- regated by gender (Borrowman and Klasen, 2017). Further improvement of women’s labor market characteristics will likely have a limited effect in rising FLFP rates, unless accompanied by the removal of barriers and constraints to female employment both at the household and at the labor market level. 75 employed—those who have worked in the week before the survey interview, or are temporarily not working but will return soon (on holidays, sick leave, parental leave, strike, etc)—and (2) the unemployed—those currently without a job but who are willing to accept one within a week from the survey interview and have actively searched for work during the previous month.46 In the OHS 1995, any positive amount of hours worked in the reference week counts for employment; in the LFS, employment requires a minimum of one hour of work in the reference week. Employment includes all paid employees either in cash or kind, employers and self employed, as well as unpaid workers in family businesses. Unpaid domestic services and begging for money or food do not count as employment. It is clear from the change in survey questionnaires that the LFS is better able to capture informal, casual employment relative to the OHS: there are eight detailed employment categories allowed as answers in the LFS, whereas only three (working full-time, working part-time, with a job but absent from work) are allowed as answers in the OHS. Thus, part of the rise in labor force participation from 1995 to 2001 could reflect improvements in the coverage of casual, low-income employment in the LFS relative to the OHS (Yu, 2007, pp. 17–18). The expanded definition is similar to the strict one but it additionally includes the discouraged job-seekers in the labor force, as part of the unemployed population. Discouraged job-seekers are those individuals who are currently without a job and, although they desire to work, have not actively searched for a job in the month before the survey interview. Unfortunately, with the introduction of the QLFS, the distinction between the strict and expanded concepts becomes less clear (Yu, 2009; Kerr and Wittenberg, 2016). Yu (2009) shows that comparability between the last LFS (in 2007) and the first QLFS (in 2008) is much better for the strict definition than for the expanded definition. Accordingly, we consider strict labor force status in our analyses. The household labor income variable is created from the consistent real earnings variable in the PALMS dataset for wage employment and self-employment.47 However, a substantial fraction (20–30 percent) of employed individuals in the OHS 1995, LFS 2001 and 2003 did not report their earnings as a point value, but rather used the earnings brackets available in the questionnaire. Individuals who are self-employed or in high skilled occupations disproportionately used the bracket option. As a result, creating an household income variable without accounting for bracket responses will underestimate 46Unfortunately, the OHS1995 metadata does not include a detailed explanation on how the employment status is derived from the survey questions (Yu, 2009, pp. 17 & 49–58). However, the derivation of employment is very similar for the later OHS (1996–1999), so we refer to it in the text. 47See Burger and Yu (2006) for more details on constructing a consistent earnings series from the OHS and LFS surveys. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 76 the household incomes of such individuals. As a simple solution, whenever necessary, we impute earnings with the bracket midpoint and, for the top bracket, which is open- ended, we set the midpoint to be 10 percent higher than the lower bound. Von Fintel (2007) shows that, for the LFS, this “simplistic” midpoint imputation performs as well (for purposes of statistical inference) as more complex distributional assumptions (e.g. interval regressions), given that skewness is not too extreme and the share of right- censored observations is not too high.48 For the years 1995, 2001 and 2003, we calculate the midpoint of the corresponding bracket variable from PALMSv3.1, convert it into monthly earnings and deflate it to 2000 Rands.49 For the year 2010, Stats SA already imputes refusals and categorical responses; for 2014, only categorical responses are imputed by Stats SA. Unfortunately, the imputation methods used by Stats SA are not described in the surveys’ metadata (Kerr and Wittenberg, 2016). After the imputations described, around 7 percent of employed individuals in the pooled sample have missing earnings information. For the household-level income variable, we sum up the earnings of all individuals in the household, with the exception of those households were at least one employed individual has missing earnings, for whom we assign missing household income. Finally, we convert the values to international dollars using the World Bank’s 2011 PPP exchange rate for private consumption. A limitation of the PALMS is the absence of information on the relationships between household members. As a result, there is no information on who is the household head. Given these limitations, we cannot capture the household’s lifetime income using the household head’s or the husband’s education as a proxy. In practice, we use the maximum education level of any adult (18+) married male household member, coding as an explicit missing category those households where no adult married men are listed. We aggregate employment at the province level by industry using the classification of Klasen and Pieters (2015). For South Africa, this means recoding the PALMS variable jobindcode. When in doubt about whether a particular industry should be classified as blue or white collar, the education distribution of urban married female employees was used as an auxiliary tool: thus, the seemingly ambiguous “Services” category in the raw data was included in the white-collar services category due to the much higher 48The number of right-censored observations for each wave of the LFS 2001 and 2003 is low, with a minimum of 38 observations in the first wave (March) of LFS 2001 and a maximum of 83 observations in the second wave (September) of LFS 2003. 49For the year 1995, the lowest earning bracket for wage employment is too wide (R1 - R999). As a result, all of the observations reporting daily earnings and 97.23 percent of the observations reporting weekly earnings fall in this category (c.f. with 26 and 0.4 percent of the observations for monthly and yearly wage earnings). Therefore we do not impute a midpoint value for the 531 observations with daily or weekly reference periods; they are set to missing. We proceed similarly for self-employment earnings. 77 prevalence of highly educated employees (see, e.g., Figure 2.A.1a). For the estimation of population means, average marginal effects of regression covari- ates, and decomposition analyses, we use the individual cross-entropy weights available in the PALMSv3.1.50 Brazil For Brazil, we use four yearly household surveys called Pesquisa Nacional por Amostra de Domicı́lios (PNAD) from the Brazilian Institute of Geography and Statistics (IBGE): PNAD 2002, 2005, 2009 and 2013. The surveys are harmonized using the Stata code created by Data Zoom at the PUC-Rio.51 A particular feature of the PNAD surveys is the distinction between different family units within a given household (see Alves, 2005). For example, multigenerational households are usually classified as different families living in a single household. We code as currently married family heads and their spouses, including couples who are officially married or living together as husband and wife. Otherwise, for consistency with surveys from the other countries, we construct all household-level variables using the household identifiers, disregarding their sub-classification into families. There were several education reforms in the past three decades. As a result, some levels of education attainment changed names and duration. We reclassify these different levels into five broader groups: less than primary; elementary (levels 1–4); elementary (levels 5–8); high school completed; any tertiary. We proxy the household’s lifetime earnings potential with the education level of the household head, creating an additional missing category for the cases when the married woman is the household head, as her education level is already captured in the own education variable. There are two reference periods available to define labor force participation: previous week or previous year. For consistency with most of the other countries, we use labor force participation in the week of reference (last week of September) in the empirical analyses. In the PNAD, employment status covers all individuals of age 10 or above that work in: (1) paid activities; (2) unpaid activities in support of a self-employed or employer household member in the production of primary goods; (3) unpaid activities in support of a religious institution or cooperative; (4) food production and/or construction work 50Variable ceweight2. This is an update by Takwanisa Machemedze at DataFirst, University of Cape Town of the original cross-entropy weights created by Nicola Branson. See Branson and Wittenberg (2014) for details on the cross-entropy approach. 51Available at http://www.econ.puc-rio.br/datazoom/english/index.html. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 80 Bolivia For Bolivia, we use the household survey Encuesta de Hogares collected by the Bolivian National Institute of Statistics. We use the years 2000, 2005, 2008, 2011 and 2014.53 The variable currently married includes both married couples as well cohabiting couples that are living as husband and wife. We create a consistent educational attain- ment over time with five categories: less than basic; basic; intermediate; secondary; any tertiary. We proxy the household’s lifetime earnings potential with the education level of the household head, creating an additional missing category for the cases when the married woman is the household head, as her education level is already captured in the own education variable. Unfortunately, it is difficult to construct a consistent ethnicity variable across surveys due to changes in coding and lack of appropriate codebook for the year 2014. We opt for an approximation based on the self-reported languages spoken: we create a dummy dividing the surveyed individuals into those only speaking Spanish, and those speaking Spanish and an Indigenous language. Very few individuals speak no Spanish at all, so we add them to the latter category. A labor force participant is anyone above age seven who was either employed for at least one hour in the week before the survey interview, or unemployed (i.e., no current job, but actively searching). Employment encompasses the production of goods and services for the market or the production of goods for own consumption. Those who have a job but were temporarily absent during the reference week (e.g., on holidays, sick leave, maternal leave, striking) are also part of the employed population. Unpaid domestic services, voluntary service work or other unpaid work for a salaried family members do not count as employment. Those without jobs and not actively searching are coded as inactive. The income variable is constructed from individual-level monthly earnings from the main job, as done for the other countries. We recode the missing earnings of unpaid family workers or unpaid apprentices with zero. We then inflate the earnings to 2010 prices. Finally, we sum up the individual real earnings excluding the woman’s own earnings to construct the household income variable and convert it to 2011 PPP-$ using the World Bank PPP exchange rate for private consumption. The coding of employment industries follows Klasen and Pieters (2015) and employ- ment shares are aggregated at the department level. Although information on rural-urban migration is unavailable, we measure (overall) 53In 2000, the survey was known as Encuesta Continua de Hogares. 81 migration with a dummy variable for whether the woman was living somewhere other than the municipality of current residence, five years before the survey. There is a change in the wording of the question used to derive this dummy variable. In 2000 and 2005, the question asks: Between [5 years before survey year] and [survey year], did you live somewhere else? After 2008, the question asks: Where did you live five years ago?. Vietnam For Vietnam, we use four years of a national representative general purpose household survey: the Vietnam Household Living Standards Survey (VHLSS) 2002, 2006, 2010, and 2014. For the education variable, we classify the highest completed school grade into five broader education levels, based on the Vietnam education system, namely: less than primary education, completed primary, completed lower secondary, completed high school, and any tertiary education. We proxy the household’s lifetime earnings potential with the education level of the household head, creating an additional missing category for the cases when the married woman is the household head, as her education level is already captured in the own education variable. Vietnam has approximately 56 ethnic groups, but around 88 percent of household heads are Kinh. In our analysis, we create a dummy with two categories: Kinh and non-Kinh. All individuals aged 10 or older were asked to state whether they had a job in the 12 months before the survey interview. Having a job is defined as working as a wage earner, or being self-employed in agriculture or non-farm activities. Unfortunately, after 2002, we cannot distinguish between those not working but actively searching for a job. That is, we cannot distinguish between the unemployed and the inactive. We thus define labor force participants as those having a job in the previous 12 months. For 2002, however, there is a job search question, so we can construct a labor force participation variable that classifies the unemployed as active. For this year, the difference of participation rate between the two definitions of labor force is very small (around 2 percent). Notice that the reference period for the job question (12 months) is much longer than the reference period in most other countries (usually, one week). Thus, on the one hand, excluding the unemployed from the active population will, in general, underestimate the rate of labor force participation. But, on the other hand, unemployment will be much lower with such a long reference period. In the end, at least for the year 2002, the two effects seem to cancel out. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 82 Whenever employed individuals have missing earnings, we impute them using a simple hotdeck procedure for each year separately based on age (5-years groups, from 16–20 to 61–65, and 65+), gender, educational attainment, and rural/urban. Finally, we sum up the individual real earnings excluding the woman’s own earnings to construct the household income variable and convert it to 2011 PPP-$ using the World Bank PPP exchange rate for private consumption. The coding of employment industries follows Klasen and Pieters (2015) and employ- ment shares are aggregated at the province level. Tanzania For Tanzania, we use the Integrated Labour Force Surveys 2000–01, 2005–06 and 2014 collected by the National Bureau of Statistics. These are quarterly surveys and we use all four waves for each year. The coding of the education variables is slightly different from the one used in other countries. Given the lower education attainment levels compared with the remaining countries, we do not distinguish between completed and not completed secondary schooling. We create five attainment levels: never attended school; primary not com- pleted; primary completed; any secondary schooling; any tertiary. We proxy the house- hold’s lifetime earnings potential with the education level of the household head, cre- ating an additional missing category for the cases when the married woman is the household head, as her education level is already captured in the own education vari- able. We did not identify any variable capturing a meaningful social identity (or discrimi- natory) marker such as the ethnicity or religion markers used for other countries. A labor force participant is anyone above age 10, in 2000 and 2006, and 15, in 2014, who was either employed for at least one hour in the week before the survey interview, or unemployed (i.e., no current job, but actively searching). Employment is defined as working for cash or in-kind pay, employers and self employed, unpaid family workers in family enterprises, production of primary products for own consumption, or production of other fixed assets (including housing) for own use. Those who have a job but were temporarily absent during the reference week (e.g., on holidays, sick leave, maternal leave, striking) are also part of the employed population. Unpaid domestic services are excluded from this definition. Those without jobs and not actively searching are coded as inactive. Notice that, according to the employment definition above, individuals engaged in subsistence agriculture are part of the labor force. Even in urban areas, 85 employment shares at the regency level. We use the borders as of 1998, to obtain units that are consistent over time. Additional Tables and Figures TABLE 2.A.1: Data overview Country Year Survey N† South Africa: 1995 October Household Survey 8,262 2001 Labor Force Survey 12,862 2003 Labor Force Survey 12,050 2010 Quarterly Labor Force Survey 21,438 2014 Quarterly Labor Force Survey 20,744 Brazil: 2002 Pesquisa Nacional por Amostra de Domicı́lios 46,562 2005 Pesquisa Nacional por Amostra de Domicı́lios 48,637 2009 Pesquisa Nacional por Amostra de Domicı́lios 49,360 2013 Pesquisa Nacional por Amostra de Domicı́lios 45,423 Jordan‡: 2006 Employment and Unemployment Survey 26,140 2008 Employment and Unemployment Survey 33,629 2010 Employment and Unemployment Survey 32,993 2014 Employment and Unemployment Survey 30,593 India: 1999 NSS Employment and Unemployment Survey 33,507 2004 NSS Employment and Unemployment Survey 30,489 2011 NSS Employment and Unemployment Survey 28,252 Bolivia: 2000 Encuesta Continua de Hogares (MECOVI) 1,563 2005 Encuesta de Hogares 1,283 2008 Encuesta de Hogares 1,183 2011 Encuesta de Hogares 3,113 2014 Encuesta de Hogares 3,863 Indonesia: 2000 Susenas 51,363 2004 Susenas 73,447 2007 Susenas 75,713 2014 Susenas 87,462 Vietnam: 2002 Living Standard Survey 5,281 2006 Household Living Standard Survey 1,704 2010 Household Living Standard Survey 1,970 2014 Household Living Standard Survey 2,043 Tanzania: 2000 Integrated Labour Force Survey 2,051 2006 Integrated Labour Force Survey 2,899 2014 Integrated Labour Force Survey 4,325 Notes: †Number of observations of urban married women age 25–54. Estimation samples are smaller due to missing covariate data. ‡For Jordan, sample sizes refer to both urban and rural areas. 2 WHAT DRIVES FEMALE LABOR FORCE PARTICIPATION? 86 0 20 40 60 Agr icu ltu re M ini ng M an uf ac tu rin g Utili tie s Con str uc tio n Tra de Tra ns po rt Fina nc e Ser vic es Dom es tic se rv ice s Low education, 1995 High education, 1995 Low education, 2014 High education, 2014 (a) South Africa 0 10 20 30 40 Agr icu ltu re Oth er in du str ial M an uf ac tu rin g Con str uc tio n Tra de & re pa ir Acc om m od at ion & fo od Tra ns po rt, st or ag e & co m m un ica ti Pub lic a dm ini str at ion Edu ca tio n, h ea lth & so cia l s er v. Dom es tic se rv ice s Oth er se rv ice s Oth er a cti vit ies Low education, 2002 High education, 2002 Low education, 2013 High education, 2013 (b) Brazil 0 20 40 60 80 Agr icu ltu re M ini ng M an uf ac tu rin g Utili tie s Con str uc tio n W ho les ale & re ta il t ra de Tra ns po rt & st or ag e Acc om m od at ion & fo od In fo rm at ion & co m m un ica tio n Fina nc e & in su ra nc e Rea l e sta te & b us ine ss Pub lic a dm ini str at ion Edu ca tio n Hea lth & so cia l w or k Oth er a cti vit ies Low education, 2006 High education, 2006 Low education, 2014 High education, 2014 (c) Jordan 0 20 40 60 80 Agr icu ltu re M ini ng M an uf ac tu rin g Utili tie s Con str uc tio n Tra de , a cc om m od at ion , f oo d Tra ns po rt & co m m un ica tio n Fina nc e & b us ine ss se rv ice s Pub lic a dm in, h ea lth , e du ca tio n Oth er se rv ice s Low education, 1999 High education, 1999 Low education, 2011 High education, 2011 (d) India 0 10 20 30 40 Agr icu ltu re M ini ng M an uf ac tu rin g Utili tie s Con str uc tio n Tra de & re pa ir Acc om m od at ion & fo od Tra ns po rt, st or ag e, co m m un ica tio Fina nc ial se rv ice s Rea l e sta te & b us ine ss se rv ice s Pub lic a dm in & so cia l s ec ur ity Edu ca tio n Hea lth & so cia l s er vic es Dom es tic se rv ice s Oth er se rv ice s Low education, 2000 High education, 2000 Low education, 2014 High education, 2014 (e) Bolivia 0 10 20 30 40 50 Agr icu ltu re M ini ng M an uf ac tu rin g Utili tie s Con str uc tio n Tra de , r et ail & re pa ir Acc om m od at ion & fo od Tra ns po rts & co m m un ica tio n Fina nc e, in su ra nc e & re al es ta te Pub lic a dm in, d ef en se & se cu rit y Edu ca tio n & h ea lth Oth er se rv ice s Low education, 2000 High education, 2000 Low education, 2014 High education, 2014 (f) Indonesia 0 10 20 30 40 Agr icu ltu re M ini ng M an uf ac tu rin g W at er su pp ly Con str uc tio n W ho les ale & re ta il t ra de Tra ns po rta tio n & st or ag e Acc om m od at ion & fo od In fo rm at ion & co m m un ica tio n Fina nc e, re al es ta te , in su ra nc e Pro fe ss ion al & te ch nic al se rv ice Adm ini str at ive & su pp or t s er vic e Pub lic a dm in, e du ca tio n, h ea lth Oth er se rv ice Dom es tic se rv ice s Low education, 2002 High education, 2002 Low education, 2014 High education, 2014 (g) Vietnam 0 10 20 30 40 50 Agr icu ltu re M ini ng M an uf ac tu rin g Utili tie s Con str uc tio n W ho les ale & re ta il t ra de Acc om m od at ion & fo od Tra ns po rt, re pa ir, co m m un ica tio n Fina nc e, in su ra nc e, b us ine ss Pub lic a dm ini str at ion Edu ca tio n Oth er se rv ice s Priv at e ho us eh old s Low education, 2000 High education, 2000 Low education, 2014 High education, 2014 (h) Tanzania FIG. 2.A.1: Distribution of female workforce across industries, by education Notes: See Table 2.A.1 for sources. Urban married women (25–54), except urban and rural in Jordan; employed only (including self-employed). Low education is below secondary schooling; high education is completed secondary or higher (any secondary or higher for Tanzania). 87 TABLE 2.A.2: South Africa: sample means 1995 2001 2003 2010 2014 Labor force 0.58 0.68 0.67 0.67 0.68 Own education: Less than primary 0.15 0.16 0.15 0.09 0.07 Primary 0.08 0.07 0.06 0.05 0.04 Secondary not completed 0.39 0.37 0.37 0.36 0.35 Secondary completed 0.25 0.24 0.28 0.30 0.33 Tertiary 0.14 0.16 0.15 0.20 0.21 Log income 4.46 3.75 3.50 4.35 4.16 Male salaried emp. 0.77 0.65 0.63 0.66 0.65 Max adult married male education: Less than primary 0.14 0.15 0.14 0.09 0.08 Primary 0.05 0.06 0.06 0.04 0.03 Secondary not completed 0.36 0.32 0.32 0.31 0.30 Secondary completed 0.24 0.23 0.25 0.28 0.29 Tertiary 0.16 0.16 0.15 0.19 0.21 Missing: no adult married male 0.05 0.07 0.08 0.08 0.08 Ethnicity: Black 0.50 0.56 0.58 0.60 0.62 Coloured 0.15 0.14 0.15 0.15 0.15 Indian/Asian 0.06 0.06 0.06 0.05 0.05 White 0.29 0.23 0.20 0.20 0.18 Age 37.37 37.67 37.81 38.77 38.79 Children 0-4 0.56 0.52 0.50 0.44 0.41 Children 5-14 1.07 0.92 0.90 0.85 0.80 N 7,601 11,361 10,658 20,713 17,890 TABLE 2.A.3: Brazil: sample means 2002 2005 2009 2013 Labor force 0.62 0.66 0.69 0.67 Own education: Less than primary 0.12 0.10 0.07 0.05 Elementary (1-4) 0.33 0.30 0.24 0.19 Elementary (5-8) 0.18 0.18 0.17 0.16 High school 0.23 0.26 0.31 0.35 Tertiary 0.14 0.16 0.20 0.24 Log income 4.90 4.91 5.10 5.27 Male salaried emp. 0.63 0.64 0.67 0.68 Household head education: Less than primary 0.13 0.11 0.08 0.05 Elementary (1-4) 0.32 0.30 0.23 0.18 Elementary (5-8) 0.16 0.16 0.14 0.13 High school 0.19 0.21 0.23 0.25 Tertiary 0.13 0.13 0.14 0.16 Missing: woman is hh head 0.06 0.08 0.17 0.23 Ethnicity: White 0.61 0.57 0.54 0.51 Black 0.05 0.06 0.07 0.08 Mixed 0.33 0.36 0.38 0.40 Asian 0.01 0.01 0.01 0.01 Indigenous 0.00 0.00 0.00 0.00 Age 37.80 38.14 38.38 38.63 Children 0-4 0.35 0.32 0.29 0.28 Children 5-14 0.92 0.87 0.78 0.71 N 39,193 42,189 42,855 38,596